Add Batch d2b94079-fd1c-47a6-9d8d-e8e376f69527 data

.gitattributes

@@ -8962,3 +8962,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/SelfCodeAlign_[[:space:]]Self-Alignment[[:space:]]for[[:space:]]Code[[:space:]]Generation/3a798739-2faf-43f6-a5ff-0673cd78e6ae_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/SemCoder_[[:space:]]Training[[:space:]]Code[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Comprehensive[[:space:]]Semantics[[:space:]]Reasoning/dd072104-6bac-4f30-b120-ddb21fd79226_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/SemFlow_[[:space:]]Binding[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]and[[:space:]]Image[[:space:]]Synthesis[[:space:]]via[[:space:]]Rectified[[:space:]]Flow/86147f1c-010a-4580-882b-88bd6f71555e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semantic[[:space:]]Density_[[:space:]]Uncertainty[[:space:]]Quantification[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Confidence[[:space:]]Measurement[[:space:]]in[[:space:]]Semantic[[:space:]]Space/285668d0-7ad8-4306-a6f6-31a0ec7d1b03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semantic[[:space:]]Feature[[:space:]]Learning[[:space:]]for[[:space:]]Universal[[:space:]]Unsupervised[[:space:]]Cross-Domain[[:space:]]Retrieval/fbb932f8-ccea-4027-b256-f6079f904c10_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semantic[[:space:]]Routing[[:space:]]via[[:space:]]Autoregressive[[:space:]]Modeling/406f4ee0-0b4e-46fd-9f4c-1723f1bf8a42_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semantics[[:space:]]and[[:space:]]Spatiality[[:space:]]of[[:space:]]Emergent[[:space:]]Communication/50cbb227-8a2f-45c8-b39a-d2448bff013d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semi-Open[[:space:]]3D[[:space:]]Object[[:space:]]Retrieval[[:space:]]via[[:space:]]Hierarchical[[:space:]]Equilibrium[[:space:]]on[[:space:]]Hypergraph/d14549f3-c4ca-4b7a-939e-03f66bcf2477_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semi-Random[[:space:]]Matrix[[:space:]]Completion[[:space:]]via[[:space:]]Flow-Based[[:space:]]Adaptive[[:space:]]Reweighting/2dda777c-38bc-43d0-9234-0a6bbebc35cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semi-Supervised[[:space:]]Sparse[[:space:]]Gaussian[[:space:]]Classification_[[:space:]]Provable[[:space:]]Benefits[[:space:]]of[[:space:]]Unlabeled[[:space:]]Data/cd0e33bc-dac7-45d8-aa8d-258dec283cd3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semi-Truths_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Dataset[[:space:]]of[[:space:]]AI-Augmented[[:space:]]Images[[:space:]]for[[:space:]]Evaluating[[:space:]]Robustness[[:space:]]of[[:space:]]AI-Generated[[:space:]]Image[[:space:]]detectors/0f6b1de4-2630-4a2e-b926-9b729ed99dc3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semi-supervised[[:space:]]Knowledge[[:space:]]Transfer[[:space:]]Across[[:space:]]Multi-omic[[:space:]]Single-cell[[:space:]]Data/d73c5942-d075-4e9e-9a0b-bf43bf67424b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semi-supervised[[:space:]]Multi-label[[:space:]]Learning[[:space:]]with[[:space:]]Balanced[[:space:]]Binary[[:space:]]Angular[[:space:]]Margin[[:space:]]Loss/c9c10c64-54c0-4589-950a-9ce6882360b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semidefinite[[:space:]]Relaxations[[:space:]]of[[:space:]]the[[:space:]]Gromov-Wasserstein[[:space:]]Distance/cb546905-0eba-48be-9a15-72c43b35827f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Separate[[:space:]]and[[:space:]]Reconstruct_[[:space:]]Asymmetric[[:space:]]Encoder-Decoder[[:space:]]for[[:space:]]Speech[[:space:]]Separation/2beba807-9753-4d02-8bc5-d552f944824e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Separation[[:space:]]and[[:space:]]Bias[[:space:]]of[[:space:]]Deep[[:space:]]Equilibrium[[:space:]]Models[[:space:]]on[[:space:]]Expressivity[[:space:]]and[[:space:]]Learning[[:space:]]Dynamics/48e3d9d2-cedc-45fe-b488-7893122a9d72_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Separations[[:space:]]in[[:space:]]the[[:space:]]Representational[[:space:]]Capabilities[[:space:]]of[[:space:]]Transformers[[:space:]]and[[:space:]]Recurrent[[:space:]]Architectures/f4ad2de7-10d5-4b38-bd9b-d34d31841c15_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sequence-Augmented[[:space:]]SE(3)-Flow[[:space:]]Matching[[:space:]]For[[:space:]]Conditional[[:space:]]Protein[[:space:]]Generation/9db84361-f33b-49dc-a314-bb3a4f715d4f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sequential[[:space:]]Decision[[:space:]]Making[[:space:]]with[[:space:]]Expert[[:space:]]Demonstrations[[:space:]]under[[:space:]]Unobserved[[:space:]]Heterogeneity/95500556-3089-47d7-889b-7ca4b586c018_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sequential[[:space:]]Harmful[[:space:]]Shift[[:space:]]Detection[[:space:]]Without[[:space:]]Labels/c24ea6ad-ba4c-431a-8fd5-d6d5d531cc84_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sequential[[:space:]]Probability[[:space:]]Assignment[[:space:]]with[[:space:]]Contexts_[[:space:]]Minimax[[:space:]]Regret,[[:space:]]Contextual[[:space:]]Shtarkov[[:space:]]Sums,[[:space:]]and[[:space:]]Contextual[[:space:]]Normalized[[:space:]]Maximum[[:space:]]Likelihood/8d51dd41-131f-4dea-aac1-0b232eb7f04a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sequential[[:space:]]Signal[[:space:]]Mixing[[:space:]]Aggregation[[:space:]]for[[:space:]]Message[[:space:]]Passing[[:space:]]Graph[[:space:]]Neural[[:space:]]Networks/6d1aa47f-a57b-40ee-858c-a6f5c52b4d93_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SequentialAttention++[[:space:]]for[[:space:]]Block[[:space:]]Sparsification_[[:space:]]Differentiable[[:space:]]Pruning[[:space:]]Meets[[:space:]]Combinatorial[[:space:]]Optimization/772515d2-be6f-45fa-bb17-ffcc79990bff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sequoia_[[:space:]]Scalable[[:space:]]and[[:space:]]Robust[[:space:]]Speculative[[:space:]]Decoding/33d65f1d-f7c2-49b6-a0e6-df37dcb0aae0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Set-based[[:space:]]Neural[[:space:]]Network[[:space:]]Encoding[[:space:]]Without[[:space:]]Weight[[:space:]]Tying/1e355695-fd67-431f-aaa9-789dd48eab46_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SfPUEL_[[:space:]]Shape[[:space:]]from[[:space:]]Polarization[[:space:]]under[[:space:]]Unknown[[:space:]]Environment[[:space:]]Light/d066b6e4-0b36-4f97-b4ec-033c48b19675_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shadowcast_[[:space:]]Stealthy[[:space:]]Data[[:space:]]Poisoning[[:space:]]Attacks[[:space:]]Against[[:space:]]Vision-Language[[:space:]]Models/2831dd59-32e7-4229-a880-2b11412019f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shadowheart[[:space:]]SGD_[[:space:]]Distributed[[:space:]]Asynchronous[[:space:]]SGD[[:space:]]with[[:space:]]Optimal[[:space:]]Time[[:space:]]Complexity[[:space:]]Under[[:space:]]Arbitrary[[:space:]]Computation[[:space:]]and[[:space:]]Communication[[:space:]]Heterogeneity/a57aa501-06de-4c39-936b-70a09b37d0bf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shape[[:space:]]analysis[[:space:]]for[[:space:]]time[[:space:]]series/59b8bd3a-b442-4e7b-ab4e-ab1667df827f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shaping[[:space:]]the[[:space:]]distribution[[:space:]]of[[:space:]]neural[[:space:]]responses[[:space:]]with[[:space:]]interneurons[[:space:]]in[[:space:]]a[[:space:]]recurrent[[:space:]]circuit[[:space:]]model/9f13e97e-ac6e-4523-b3bd-27b5f0e142c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ShareGPT4Video_[[:space:]]Improving[[:space:]]Video[[:space:]]Understanding[[:space:]]and[[:space:]]Generation[[:space:]]with[[:space:]]Better[[:space:]]Captions/a61fb035-43c0-4f84-835b-d0b22e4d9da8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shared[[:space:]]Autonomy[[:space:]]with[[:space:]]IDA_[[:space:]]Interventional[[:space:]]Diffusion[[:space:]]Assistance/c93da49d-6b48-4371-8a73-bb656b8f56ae_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sharing[[:space:]]Key[[:space:]]Semantics[[:space:]]in[[:space:]]Transformer[[:space:]]Makes[[:space:]]Efficient[[:space:]]Image[[:space:]]Restoration/15161f2d-9d47-4503-bf2d-2bf77a1847b0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sharpness-Aware[[:space:]]Minimization[[:space:]]Activates[[:space:]]the[[:space:]]Interactive[[:space:]]Teaching's[[:space:]]Understanding[[:space:]]and[[:space:]]Optimization/16579085-644f-4756-846b-f78041c1c922_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sharpness-diversity[[:space:]]tradeoff_[[:space:]]improving[[:space:]]flat[[:space:]]ensembles[[:space:]]with[[:space:]]SharpBalance/f6e05aba-19e5-45d4-a05e-77bde9b6ee3f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shaving[[:space:]]Weights[[:space:]]with[[:space:]]Occam's[[:space:]]Razor_[[:space:]]Bayesian[[:space:]]Sparsification[[:space:]]for[[:space:]]Neural[[:space:]]Networks[[:space:]]using[[:space:]]the[[:space:]]Marginal[[:space:]]Likelihood/b0b343b2-88c1-43c8-a2f0-5d872962b3a5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ShiftAddLLM_[[:space:]]Accelerating[[:space:]]Pretrained[[:space:]]LLMs[[:space:]]via[[:space:]]Post-Training[[:space:]]Multiplication-Less[[:space:]]Reparameterization/c674459a-e78a-4e26-9d70-c7e760140210_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shopping[[:space:]]MMLU_[[:space:]]A[[:space:]]Massive[[:space:]]Multi-Task[[:space:]]Online[[:space:]]Shopping[[:space:]]Benchmark[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/152a9919-d469-4761-9c88-2b95d0329129_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Should[[:space:]]We[[:space:]]Really[[:space:]]Edit[[:space:]]Language[[:space:]]Models_[[:space:]]On[[:space:]]the[[:space:]]Evaluation[[:space:]]of[[:space:]]Edited[[:space:]]Language[[:space:]]Models/42acc268-d21e-46ab-a8ce-48caf54c12bb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ShowMaker_[[:space:]]Creating[[:space:]]High-Fidelity[[:space:]]2D[[:space:]]Human[[:space:]]Video[[:space:]]via[[:space:]]Fine-Grained[[:space:]]Diffusion[[:space:]]Modeling/bd3deab5-5dae-41cc-8760-1e2b282ca85c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shuffling[[:space:]]Gradient-Based[[:space:]]Methods[[:space:]]for[[:space:]]Nonconvex-Concave[[:space:]]Minimax[[:space:]]Optimization/0be4d69c-3cd7-4254-a3f6-72c78db52b19_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sigmoid[[:space:]]Gating[[:space:]]is[[:space:]]More[[:space:]]Sample[[:space:]]Efficient[[:space:]]than[[:space:]]Softmax[[:space:]]Gating[[:space:]]in[[:space:]]Mixture[[:space:]]of[[:space:]]Experts/d4950be0-e982-45b7-a1da-c2dc88f67e0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sim2Real-Fire_[[:space:]]A[[:space:]]Multi-modal[[:space:]]Simulation[[:space:]]Dataset[[:space:]]for[[:space:]]Forecast[[:space:]]and[[:space:]]Backtracking[[:space:]]of[[:space:]]Real-world[[:space:]]Forest[[:space:]]Fire/005837ad-5876-48ca-be86-665669fc1194_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SimGen_[[:space:]]Simulator-conditioned[[:space:]]Driving[[:space:]]Scene[[:space:]]Generation/40e346aa-b31d-4a6f-9fb9-77f5a298b395_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SimPO_[[:space:]]Simple[[:space:]]Preference[[:space:]]Optimization[[:space:]]with[[:space:]]a[[:space:]]Reference-Free[[:space:]]Reward/e2a7ad92-2b86-4fa2-a756-07537424404d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SimVG_[[:space:]]A[[:space:]]Simple[[:space:]]Framework[[:space:]]for[[:space:]]Visual[[:space:]]Grounding[[:space:]]with[[:space:]]Decoupled[[:space:]]Multi-modal[[:space:]]Fusion/92c24147-f04d-40b4-a459-03f365181c2d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Similarity-Navigated[[:space:]]Conformal[[:space:]]Prediction[[:space:]]for[[:space:]]Graph[[:space:]]Neural[[:space:]]Networks/82eb227c-4b02-456b-89b3-10eb04834db1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Simple[[:space:]]and[[:space:]]Effective[[:space:]]Masked[[:space:]]Diffusion[[:space:]]Language[[:space:]]Models/2179133b-8569-47ad-a065-870759392169_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Simple[[:space:]]and[[:space:]]Fast[[:space:]]Distillation[[:space:]]of[[:space:]]Diffusion[[:space:]]Models/083050b1-21e7-44c0-bc89-159abe53cc6f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Simplified[[:space:]]and[[:space:]]Generalized[[:space:]]Masked[[:space:]]Diffusion[[:space:]]for[[:space:]]Discrete[[:space:]]Data/45fd68c4-a8cd-43f8-8407-23e419510bf8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Simplifying[[:space:]]Constraint[[:space:]]Inference[[:space:]]with[[:space:]]Inverse[[:space:]]Reinforcement[[:space:]]Learning/b06ef132-0302-4341-b47b-535b8f9f17b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Simplifying[[:space:]]Latent[[:space:]]Dynamics[[:space:]]with[[:space:]]Softly[[:space:]]State-Invariant[[:space:]]World[[:space:]]Models/78c63b10-6170-4d1a-af70-82992d9e01f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Simulation-Free[[:space:]]Training[[:space:]]of[[:space:]]Neural[[:space:]]ODEs[[:space:]]on[[:space:]]Paired[[:space:]]Data/deecd092-ad55-4ed8-bdf0-c6b22afdb290_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Single[[:space:]]Image[[:space:]]Reflection[[:space:]]Separation[[:space:]]via[[:space:]]Dual-Stream[[:space:]]Interactive[[:space:]]Transformers/f810c14d-7dc4-4a0a-ab3f-ffa564686604_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Single[[:space:]]Image[[:space:]]Unlearning_[[:space:]]Efficient[[:space:]]Machine[[:space:]]Unlearning[[:space:]]in[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/72e5519e-0eaf-450f-bd85-60aa2f07a67d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Single-Loop[[:space:]]Stochastic[[:space:]]Algorithms[[:space:]]for[[:space:]]Difference[[:space:]]of[[:space:]]Max-Structured[[:space:]]Weakly[[:space:]]Convex[[:space:]]Functions/465d1d0f-0553-4da0-8b18-42c9e1bdad8b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sketched[[:space:]]Lanczos[[:space:]]uncertainty[[:space:]]score_[[:space:]]a[[:space:]]low-memory[[:space:]]summary[[:space:]]of[[:space:]]the[[:space:]]Fisher[[:space:]]information/51eb2d0b-fb20-4bd6-b5aa-bbbbe6f9e937_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sketching[[:space:]]for[[:space:]]Distributed[[:space:]]Deep[[:space:]]Learning_[[:space:]]A[[:space:]]Sharper[[:space:]]Analysis/ce013905-a73e-46c1-87fb-f2aa6bfba026_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sketchy[[:space:]]Moment[[:space:]]Matching_[[:space:]]Toward[[:space:]]Fast[[:space:]]and[[:space:]]Provable[[:space:]]Data[[:space:]]Selection[[:space:]]for[[:space:]]Finetuning/7dedb859-3756-4e84-8ede-900254ef9017_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SkiLD_[[:space:]]Unsupervised[[:space:]]Skill[[:space:]]Discovery[[:space:]]Guided[[:space:]]by[[:space:]]Factor[[:space:]]Interactions/f6ed1722-e07e-4e09-b799-53a59492e0eb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Skill-aware[[:space:]]Mutual[[:space:]]Information[[:space:]]Optimisation[[:space:]]for[[:space:]]Zero-shot[[:space:]]Generalisation[[:space:]]in[[:space:]]Reinforcement[[:space:]]Learning/0454a2b9-e4d9-41a6-9e71-8603357742d6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Skinned[[:space:]]Motion[[:space:]]Retargeting[[:space:]]with[[:space:]]Dense[[:space:]]Geometric[[:space:]]Interaction[[:space:]]Perception/40e3b94c-ecce-4175-a106-f4b707dca1f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SkipPredict_[[:space:]]When[[:space:]]to[[:space:]]Invest[[:space:]]in[[:space:]]Predictions[[:space:]]for[[:space:]]Scheduling/720a587c-0edf-4b13-a1d2-6967b51050f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Slack-Free[[:space:]]Spiking[[:space:]]Neural[[:space:]]Network[[:space:]]Formulation[[:space:]]for[[:space:]]Hypergraph[[:space:]]Minimum[[:space:]]Vertex[[:space:]]Cover/5500609d-a1be-4915-b77e-cc04d2f3d898_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SleeperNets_[[:space:]]Universal[[:space:]]Backdoor[[:space:]]Poisoning[[:space:]]Attacks[[:space:]]Against[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]Agents/5ff83d54-5ee4-4e1b-8936-4fbf4c8d0f38_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Slice-100K_[[:space:]]A[[:space:]]Multimodal[[:space:]]Dataset[[:space:]]for[[:space:]]Extrusion-based[[:space:]]3D[[:space:]]Printing/14479786-4232-41b4-b07a-feb7f21c9445_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Slicing[[:space:]]Vision[[:space:]]Transformer[[:space:]]for[[:space:]]Flexible[[:space:]]Inference/ec4aa152-568c-42ac-b748-00908670ba44_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/Semantic Density_ Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space/285668d0-7ad8-4306-a6f6-31a0ec7d1b03_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33a3d36c1bf1f722f52dc47a62b66b8d87588f62c119384c920300e792670251
+size 1126743

2024/Semantic Density_ Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2dd1dec6dd1956095cd59d715163a131b205f160313b6cf03173cb2b0d651b9f
+size 1039568

2024/Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval/fbb932f8-ccea-4027-b256-f6079f904c10_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a36b21e0436b5b8c92f3b772b13f330801249a5b1c8cdc0657a3a09a0bea1e7
+size 4249898

2024/Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25b6a9ede982a55a732bd03be0af2629ec3066a6101b5c768e164612070f90fa
+size 1044522

2024/Semantic Routing via Autoregressive Modeling/406f4ee0-0b4e-46fd-9f4c-1723f1bf8a42_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35c554967b57c08f9538c6515585d4104aa714fa6db4468af8fc3787765b0d90
+size 8354247

2024/Semantic Routing via Autoregressive Modeling/full.md

@@ -0,0 +1,603 @@
+# Semantic Routing via Autoregressive Modeling
+
+Eric Zhao
+
+UC Berkeley and Google Research  
+eric.zh@berkeley.edu
+
+Pranjal Awasthi
+
+Google Researchlawasthi@gmail.com
+
+Zhengdao Chen
+
+Google Research zhengdaoc@google.com
+
+Sreenivas Gollapudi
+
+Google Research  
+sgollapu@google.com
+
+Daniel Delling
+
+Google
+delling@google.com
+
+# Abstract
+
+We study learning-based approaches to semantic route planning, which concerns producing routes in response to rich queries that specify various criteria and preferences. Semantic routing is already widely found in industry applications, especially navigational services like Google Maps; however, existing implementations only support limited route criteria and narrow query sets as they rely on repurposing classical route optimization algorithms. We argue for a learning-based approach to semantic routing as a more scalable and general alternative. To foster interest in this important application of graph learning, we are releasing a large-scale publicly-licensed benchmark for semantic routing consisting of real-world multi-objective navigation problems—expressed via natural language queries—on the richly annotated road networks of US cities. In addition to being intractable with existing approaches to semantic routing, our benchmark poses a significant scaling challenge for graph learning methods. As a proof-of-concept, we show that—at scale—even a standard transformer network is a powerful semantic routing system and achieves non-trivial performance on our benchmark. In the process, we demonstrate a simple solution to the challenge of scaling up graph learning: an autoregressive approach that decomposes semantic routing into smaller "next-edge" prediction problems.
+
+# 1 Introduction
+
+The route planning problem arises in diverse applications including navigating transportation networks and supply chain logistics (Dantzig and Ramser, 1959). Route planning is classically formulated in the language of graph theory as searching for paths between a source and a destination that optimize a single static objective (Dijkstra, 1959), usually travel time. In practice, route planning systems are not as simple. For example, navigation services like Google Maps do not only optimize for travel time—they allow drivers to specify multiple modes of transportation, add intermediate stops, and convey preferences about avoiding roads with bike lanes (Ratliff et al., 2009). Modern routing systems are better described as performing semantic routing: a contextual and multi-objective generalization of route planning that involves processing a rich user query that specifies not only a source and destination but also various criteria—such as bike lanes preferences—and responding with a satisfactory route (Delling et al., 2017).
+
+Current routing systems implement semantic routing capabilities in an ad hoc way, customizing classical route optimization algorithms to support specific and limited types of route criteria. For example, the technique of "cost modifiers" rescales edge weights according to whether the edge type is preferred/dispreferred, such as doubling travel time along small roads when it is indicated that small roads should be avoided (Ratliff et al., 2009), and then runs a standard shortest-paths
+
+![](images/82e6374a7258a2f6a0479188c2014b2a3b64f8c62338edfe72ea15dd4b7434b9.jpg)  
+Figure 1: Examples of semantic routing tasks from our benchmark that involve navigating between two locations in the Glenview neighborhood of Chicago. The user query is specified in pink, while we outline the path of the optimal route (according to our automated evaluation scheme) in red.
+
+![](images/95272bfdc9dd4d37aa8c811672d707585f849ac68d026c94a73c99eeaf58a22c.jpg)
+
+algorithm. As industry routing services transition to offering increasingly rich experiences, it has become clear that the current paradigm is unscalable. First, this ad hoc approach can only support the limited types of route criteria that can be efficiently achieved through classical graph algorithms; for example, cost modifiers only support "linear" criteria (Ratliff et al., 2009). Second, modern routing systems are complex ensembles of combinatorial algorithms and expensive to maintain; supporting new features requires adding new customized algorithms into an already messy ensemble and even then different route criteria can often not be efficiently combined. Third, current systems cannot take full advantage of rich graph metadata as they require explicit instructions on what and how to use each data field (Cormen et al., 2022); this is a major opportunity cost as, for example, modern road network metadata include live road conditions, the businesses along each road (OpenStreetMap contributors, 2017), and even the locations and types of fruit trees (Welty and Phillips, 2013). One can directly observe how these limitations manifest in industry services. For example, Google Maps allows for specifying an intermediate stop if an explicit address is given but cannot automatically plan a short stop to purchase dental floss; it can provide routes that use public transportation but not when an intermediate stop is specified. In this paper, we study the open question of building a semantic routing system that is truly scalable and offers general open-ended capabilities.
+
+We argue for a learning-based approach to semantic routing, where systems are trained on historical, and potentially simulated, data consisting of tuples of queries, rich graphs, and satisfactory routes. A learning-based approach avoids all three key limitations of the current paradigm. It transfers the complexity of semantic routing from handcrafting ensembles of combinatorial graph algorithms to instead creating a rich dataset that demonstrates desired capabilities. Moreover, as we will demonstrate, learned models can approximately fulfill route criteria that may not admit efficient classical solutions and take advantage of available rich graph metadata that may be impractical to manually engineer support for. Semantic routing also exhibits the usual hallmarks of a practical machine learning application: though undesirable, having latency and errors in a semantic routing system is not typically catastrophic or dangerous, unlike higher-stakes applications such as autonomous driving. With the cost of deploying large neural networks plummeting and the toolkit of graph learning methods exploding in recent years, we believe that learned semantic routing systems are now a viable short-term goal. Our contributions are two-fold and aimed at fostering interest in this important application of graph learning.
+
+1. First, we release a large-scale publicly-licensed benchmark for semantic routing, which can be found at github.com/google-research/google-research/tree/master/semantic Routing. This benchmark, and the remainder of our paper, focuses on real-world multi-objective navigation problems that involve providing route suggestions to drivers. These problems include providing efficient routes for errand running, planning recreational day trips, avoiding highway driving, and providing waypoint routes that make stops by, e.g., grocery stores. Each problem in this benchmark is a tuple consisting of (1) a subgraph of a real-world road network with rich metadata including local businesses, sourced from the OpenStreetMap project (OpenStreetMap contributors, 2017), and (2) an unstructured natural language user query. Our benchmark also includes an automated evaluation mechanism that, given a problem tuple, scores candidate routes. Consisting of  $1,000,000+$  user queries and  $300,000+$  miles of real-world road networks, our benchmark provides a comprehensive platform for evaluating semantic routing systems and tests capabilities that are intractable with existing methods.  
+2. Second, we demonstrate the viability of learning-based semantic routing systems with a proof-of-concept. Specifically, we show that—when done at sufficient scale—training even a standard
+
+<table><tr><td>Find a suitable place for dog walking and a post box along the way.
+{need: (dog park or park) and (post box)}</td></tr><tr><td>Help me plan a kid-friendly adventure through town, with all places no more than a 20-minute drive away.
+{visit: (game store or video game store or toy store or arcade)}</td></tr><tr><td>Please plan a route for me that includes a stop at a bakery and a car repair shop, avoiding highways whenever possible.
+{visit: (bakery) and (car repair shop), dislikes: highways}</td></tr></table>
+
+Figure 2: Examples of queries drawn from our benchmark, demonstrating an errand route planning task (top), a trip planning task (middle), and the combination of a waypoint routing task and highway avoidance task (bottom). Annotated labels are underneath queries in gray and are hidden to models.
+
+transformer network on our benchmark's dataset creates a powerful semantic routing system, with a  $60 - 70\%$  success rate on benchmark tasks even on previously unseen user queries and in a previously unseen American city. As existing semantic routing methods entirely fail on our benchmark—even when implemented with unfair advantages—our proof-of-concept also establishes the first non-trivial baseline of our benchmark. A major challenge we encountered was scaling up graph learning methods to large road graphs (e.g., 10,000 vertices and 25,000 edges); we identified an autoregressive approach to be critical: performing repeated next-edge predictions to unroll a route instead of predicting an entire route at once, allowing us to only provide the model a small local subsection of the road network at a time. This approach is similar to that of a language model, except instead of unrolling a sentence by predicting tokens, we unroll a route by predicting edges.
+
+# 1.1 Related Work
+
+Route Planning in Navigation. Traditional routing tasks are approached as finding shortest paths within road networks and can be solved with dynamic programming algorithms such as Dijkstra's method (Dijkstra, 2022) and its generalizations (Shimbel, 1954; Bellman, 1958; Ford, 1956). Their efficiency has been improved via priority-queue-based implementations (Goldberg, 2001; Meyer, 2001), bidirectional search (Sint and de Champeaux, 1977),  $\mathbf{A}^*$  search (Hart et al., 1968), and contraction hierarchies (Bast et al., 2016; Delling et al., 2009; Bauer et al., 2010) that specifically make use of the near planar structure of the road networks. However, incorporating custom contextual information into these algorithms is challenging, requiring solving a (often NP-complete) combinatorial problem per context. For a very narrow set of contextual routing tasks where the "cost" of a route can be decomposed into sum of its component edges, it is possible to define context-dependent cost modifiers on edges such that the running shortest-paths on the new costs results in satisfactory solutions (Ratliff et al., 2009). However, this approach is not possible for most routing tasks of interest and—even when available—still requires one to solve the challenging problem of learning how to design good cost modifiers from complex natural language queries. A more generally available alternative is to run unguided search algorithms such as  $\mathbf{A}^*$  search or spectral methods based on electrical flows (Sinop et al., 2021) to compute a large set of candidate routes from source to destination, and then extract the best candidate. However, this approach needs one to have a good way of scoring candidate routes and may require iterating through intractably many candidates to find a solution.
+
+Learning on Graphs. In recent years, many data-driven approaches have been explored for combinatorial problems on graphs. The works of Xu et al. (2019, 2020) show that graph neural networks (GNNs) can learn dynamic-programming-based algorithms; Velicković et al. (2019); Ibarz et al. (2022) create GNN based neural learners to simulate various combinatorial algorithms by "hinting" them with intermediate computations of the algorithms. Other works have explored the custom models, such as pointer networks and attention-based GNNs, for solving graph problems such as the traveling salesman (Kool et al., 2018; Vinyals et al., 2015; Khalil et al., 2017; Nowak et al., 2017; Deudon et al., 2018). However, these approaches are all empirically tested only at small scales: graphs around 100 nodes large. In comparison, the semantic routing applications we explore in this paper concern road networks that are—at smallest—10,000 vertices large; naively scaling graph learning algorithms to this setting would require an impractically expensive distributed GNN system. One exception to the small scale studies in previous works is Graves et al. (2016), who conducted routing experiments on the London underground system using a highly specialized and complex variant of a
+
+GPS, please guide me to charging station, public restroom, and minimize time spent on small roads.
+
+![](images/dc841bf8bf93975dd2085d43f75a88ad5079d6c30f8db4bd369ec16f526a24a9.jpg)  
+Figure 3: Example of semantic routing tasks from our main benchmark (left) and our grid-world (right), visualized in false-color. The optimal route is illustrated in orange/pink and colored dots denote points-of-interest.
+
+![](images/89cfcf86b85aeb32985261bddcf8116b8361263ae101179991f9485afaa97049.jpg)
+
+differential neural computer. In contrast, we present in Section 3 a simple approach to scaling graph learning methods to large-scale semantic routing problems: train on an autoregressive objective so that the semantic routing problem decomposes into smaller next-edge prediction problems where it suffices to only consider a local neighborhood of the graph.
+
+Autoregressive Models. Autoregressive approaches based on the transformer model (Vaswani et al., 2017) have achieved remarkable progress for generative tasks from language modeling to image and video generation. The ability of such models to generate sequences via the next-token-prediction paradigm and to incorporate multi-modal information makes them a natural fit to tackle semantic routing. Moreover, when trained at scale and augmented with prompting techniques or external tools and operations, autoregressive models appear to demonstrate capabilities of planning (e.g., Wei et al. (2022); Huang et al. (2022); Schick et al. (2023); Yao et al. (2023); Valmeekam et al. (2023)). We take inspiration from autoregressive methods to approach semantic routing, re-discovering strong parallels between the domains of routing and language.
+
+# 2 A Benchmark for Semantic Routing
+
+We are releasing a public dataset of semantic routing problems concerning complex navigation tasks on American roads. To facilitate discussion, we will formalize a semantic routing problem as consisting of a graph  $G = (V, E)$  with vertices  $V$  and edges  $E$ , a source and destination pair  $u, v \in E$ , and a query  $s \in \Sigma^*$  where  $\Sigma^*$  denotes all strings of alphabet  $\Sigma$ . We can denote the metadata associated with each edge in  $G$  using the map  $\theta : E \to \Xi$ , where  $\Xi$  is the space of possible metadata. Note that a single query  $s$  may communicate multiple subtasks and objectives. The goal of a semantic routing problem is to find a path  $p \in E^*$  from source to destination that "satisfies" the query  $s$ . We can formalize this goal as finding a path  $p$  that starts at the source  $u$ , ends at the destination  $v$ , and maximizes  $f(s, \{\theta(e)\}_{e \in p})$  where  $f : \Sigma^* \times \Xi^* \to \mathbb{R}$  is a scoring function and  $\{\theta(e)\}_{e \in p}$  is the metadata associated with each edge in the path. In our benchmark, each query  $s$  is a natural language user query, so  $\Sigma$  can be understood as the English alphabet. Our graph  $G$  is a road network where each edge corresponds to a road segment, and edge metadata  $\theta$  contains road information such as colocated businesses and the speed limit.
+
+# 2.1 Benchmark Overview
+
+This benchmark evaluates semantic routing algorithms on whether they can successfully return a satisfactory route when provided a natural language semantic routing query and a road graph on which to navigate. To this end, we collected and labeled a large dataset of natural language semantic routing queries and curated open-source data from OpenStreetMap contributors (2017) on the road
+
+<table><tr><td>Method Name</td><td>Waypoint Routing Success Rate</td></tr><tr><td>Cost modifiers</td><td>0 % ± 0%</td></tr><tr><td>Electrical flows (4096 routes)</td><td>1.3% ± 1.3%</td></tr></table>
+
+Table 1: This table depicts the success rate of classical semantic routing methods (named in the "Baseline" column) on waypoint routing task sampled from our benchmark. We define a method as being successful if it returns at least one route that visits points-of-interest that together fulfill each part of the user's query. The "cost modifiers" method returns every route that can be elicited by a cost modifying rule (Delling et al., 2009); the "Electrical flows" method returns 4,096 routes sampled with the electrical flows algorithm (Sinop et al., 2021).  $\pm$  denotes standard error ( $n = 100$ ).
+
+networks and points-of-interest (facilities, businesses, restaurants, etc.) of major urban centers in the United States. The benchmark is released under an Apache 2.0 license; see supplemental materials. Structured similarly to a reinforcement learning benchmark, it is accessible as a software package with three main functions: (1) generates semantic routing tasks by randomly drawing a user query from a data bank and choosing the road graph of an American city, (2) given a semantic routing task, featurizes the task by generating feature arrays that describe the road graph in the form of a Tensorflow record (Abadi et al., 2015), and (3) given a semantic routing task and a candidate solution, returns a numerical score that describes estimated route quality. Examples of these functionality are provided in Python notebooks available in the previous link.
+
+The benchmark consists of over 1 million natural language queries and navigation tasks drawn from an over 11 million vertex graph representing 375,000 miles of real-world road networks and 1.3 million points-of-interest in 23 major US cities. Figure 1 illustrates samples of semantic routing tasks from our benchmark and their solutions.
+
+Benchmarked Tasks. Our benchmark measures semantic routing capability with important real-world problems that involve navigating drivers on roads. The majority of tasks in our benchmark involve a variant of waypoint routing: routing tasks that require planning out (usually multiple) intermediate stops on one's route. For example, a user might ask to get coffee and drop off a package en route to work. Two types of waypoint routing tasks appear in our benchmark: where the user specifies a specific location category to visit (e.g., a grocery store), or where the user specifies a need or errand that they need to complete en route (e.g., buy a water bottle). The latter allows for more flexibility as, for example, a water bottle can be purchased at many types of locations including both grocery stores and gas stations. Fulfilling a waypoint query involves up to four expensive steps:
+
+1. Identifying points-of-interest that fulfill the request (e.g., nearby coffee shops and mail boxes).  
+2. Choosing which points-of-interest to stop by (e.g., picking a coffee shop and mailbox to visit).  
+3. Picking an order to make the stops (e.g., determining that getting coffee first would be faster).  
+4. Efficiently navigating through stops by minimizing travel time or fulfilling additional objectives.
+
+One fifth of the tasks in our benchmark require also taking into account a driver's personal comfort preferences, i.e., whether a driver has an aversion to highways or a preference against driving on small local roads. Another subset of tasks in our benchmark involve planning day trips, i.e. itinerary planning. In these tasks, the user query specifies a time budget and the type of day trip that the user wants to go on; the goal is to return a route that takes the user to as many relevant points-of-interest as possible within the time budget. Whereas waypoint routing is a variant of the traveling salesman and set cover problems, trip planning is a variant of the knapsack problem.
+
+We provide examples of each of these tasks in Table 2. A challenging aspect of our benchmark is that a single problem typically involves multiple tasks; for example, a single query might describe multiple errands that need to be run en route and also an aversion to highways. On average, a single query conveys 1.8 distinct tasks or objectives, with a maximum of up to 5 tasks per query.
+
+Baselines. In this section, we adapt classical semantic routing methods to our benchmark and evaluate them on their success rate on waypoint routing task. Here, we define a route as being successful if it visits points-of-interest that together fulfill every request in the user's query; we will ignore further considerations like travel time for simplicity. As noted in Section 1, we expect the
+
+inherent limitations of classical methods to result in poor performance and will therefore generously adapt these methods to afford them unfair advantages.
+
+First, we consider the method of cost modifiers (Delling et al., 2009) where—depending on the user query—we rescale travel times along each edge according to the edge's metadata, and then apply a classical shortest-paths algorithm to find a route. To upper bound the performance of systems that use cost modifiers on our benchmark, for every semantic routing problem, we perform a brute-force search over the routes returned by each possible cost modifier scheme (i.e., edge rescaling rule) and check if any of the routes is successful; if so, we mark down the overall method as being successful. Second, we consider a guess-and-check method that uses electrical flows (Sinop et al., 2021) to generate a large and diverse set of candidate routes and check if any of the routes is successful; once again, we mark down the method as being successful if such a route is found.
+
+The performance of these methods, as shown in Table 1, is unsurprisingly poor. Whereas waypoint routing problems are provably unsolvable with cost modifier methods, the electrical flows method is similarly doomed due to the intractably large search space of possible routes from source to destination. For a baseline level of performance on our benchmark, we refer readers to Section 3.
+
+# 2.2 Technical Details
+
+Road network data. We curated and sourced road network data and points-of-interest data from the OpenStreetMap project for the American cities of New York City, Santa Barbara, Miami, Cambridge (MA), Providence, San Francisco, Orlando, Boulder, Albuquerque, Portland, Madison, New Orleans, Chicago, Palo Alto, Philadelphia, Austin, San Jose, Memphis, Denver, Washington DC, Seattle, Sacramento, Santa Cruz, and Berkeley. Our curated road networks embed rich metadata in every road segment, including each segment's road type (e.g., interstate, service road), number of lanes, speed limit, length, estimated average transit time, geometry (start and end coordinates), and a list of points-of-interest that can found along the road segment. Every point-of-interest in our road networks exists in the real world and is annotated with its real-world name and assigned to one of over two hundred point-of-interest categories, such as "gas station" or "郵信箱".
+
+Our benchmark produces semantic routing problems by randomly selecting pairs of source and destination edges within the road network of Berkeley, California; for tasks involving trip planning, the destination is the same as the source. We perform rejection sampling to omit source-destination pairs if the optimal route between the two requires traversing more than 600 road segments. Due to its diverse composition, we chose the city of Berkeley as our benchmark's default hold-out city, meaning that the training datasets that accompany our benchmark do not contain any examples of semantic routing in Berkeley. This means that our benchmark evaluates semantic routing systems on a road network they have never encountered and with points-of-interest and roads they have never seen in training. To simulate live traffic conditions on our road networks, we randomly rescale each road's default transit time in every new semantic routing problem instance.
+
+Query dataset. We collected one million natural language user queries through a combination of human labeling and language model augmentation. The first step in our process produced  $1,000+$  distinct types of semantic routing queries by manually taxonomizing common road situations (e.g. ran out of gas, passenger spilled drink in backseat), and writing representative queries for each situation in the taxonomy. An industry language model (OpenAI, 2023) was then applied to augment these queries into  $10,000+$  natural language variations. Multi-objective queries were then formed by randomly merging together combinations of up to 5 queries, where we again applied a language model to refine the writing of combined queries. The resulting queries were spot-checked for correctness and are depicted in Figure 2 and Appendix D. Tour planning queries were not merged with other query types.
+
+To define an evaluation rule for our benchmark, i.e. the route scoring function  $f$ , we label every query in our dataset with a structured interpretation of its content. For example, queries that specify an aversion to driving on highways have a special flag in their label indicating such. For every waypoint routing task communicated in a query, the query's label contains a list of point-of-interest categories that would satisfy the task; for example, a "get waterbottle" task might be labeled with the list ("grocery store", "gas station", etc.). Similarly, for every trip planning task communicated in a query, the query's label contains a list of point-of-interest categories that the driver might want to visit on the requested trip. Queries are human labeled, though when two queries are rewordings
+
+![](images/48eedad46cebb7bb79ac4dcd15f00c61975e4515026735e47877dc338910d300.jpg)  
+Figure 4: Architecture of proof-of-concept semantic routing model: a transformer network trained on an autoregressive objective. Arrows illustrate the flow of data from input to next-edge prediction.
+
+of one another, we may only label one query and propagate its label to the other. Figure 2 depicts random examples of queries and their labels.
+
+Automated route evaluation. Although it is possible to use human raters to evaluate routing systems on our benchmark's collection of semantic routing problems, we also provide an automated system for evaluating candidate routes against problems in our benchmark. Our system—which simulates a generic user's preferences with standard heuristics—returns various metrics about the quality of the route, including an overall numerical score and task specific performance metrics like the waypoint routing success metric of Table 1. Our evaluation system's overall score metric should be understood as providing a ground-truth ordering over routes in terms of goodness; similarly, the ground-truth optimal route is the route maximizing this score. We define the overall score as follows. First, we refer to the problem's query label and identify all waypoint requests and points-of-interest that would satisfy the requests. If the route does not visit points-of-interest that satisfy all waypoint requests, an infinitely large penalty is subtracted from the score. The total seconds of travel time is also deducted from the overall score; although if the user specifies an aversion/preference for driving on a particular road type, travel time on such roads is increased/decreased by a factor  $5/0.5$ . In trip planning tasks, a reward of 1,000 is added to the score for every relevant and unique type of point-of-interest visited. When a time budget is communicated in the query, an infinitely large penalty is subtracted from the score if the time budget is exceeded by more than a fourth.
+
+Although this evaluation scheme can be succinctly stated, its optimization is non-trivial and generalizes various combinations of traveling salesman, shortest path, set cover, and knapsack problems. Moreover, this reward scheme leads to empirically high-quality results as demonstrated in Figure 1.
+
+Training dataset. We are also releasing a training dataset consisting of 3 million pairs of semantic routing problems generated from our benchmark and their ground-truth optimal routes. We brute-force solved for these routes by peeking at the query labels and applying an ensemble of combinatorial algorithms, mimicking a very inefficient classical semantic routing system. Beyond query labeling, this requires 0.3 CPU hours per route.
+
+Grid-world test bed. To provide a more flexible test bed for semantic routing algorithms, we are also releasing a purely synthetic semantic routing benchmark. This benchmark simulates a large grid graph with highway roads and populates it with artificial points of interest, as illustrated in Figure 3. Each grid world generated is of size  $30 \times 30$  and consists of 3560 road segments; these graphs can be understood as randomly sampled subgraphs of an infinitely large grid world, taken around a navigation task's source vertex. The query dataset for this grid world consists of simple template-generated natural language queries and represents only waypoint routing tasks. Some experiments in Section 3 are performed on this grid-world benchmark.
+
+# 3 A Proof-of-Concept Autoregressive Model
+
+One challenge that has limited the adoption of graph learning methods in routing applications is that routing is a naturally large-scale decision-making process. Road networks form highly connected
+
+<table><tr><td>Success Rate (Waypoint Routing, Errands)</td><td>92.0% ± 3.7%</td></tr><tr><td>Success Rate (Waypoint Routing, Locations)</td><td>38.0% ± 6.6%</td></tr><tr><td>Success Rate (Waypoint Routing, Mixed)</td><td>34.0% ± 6.9%</td></tr><tr><td>Success Rate (Trip Planning)</td><td>68% ± 4%</td></tr><tr><td>Median Excess Penalty (Waypoint Routing, Mixed)</td><td>0.8 seconds</td></tr><tr><td>Median Total Travel Time (Waypoint Routing, Mixed)</td><td>509 seconds</td></tr></table>
+
+Table 2: The task-level performance metrics of the proof-of-concept autoregressive transformer model on our benchmark. The evaluated problems represent various semantic routing tasks with queries drawn from a holdout test set and conducted on a road network from a previously unseen city (Berkeley, CA). "Waypoint Routing Success Rate" is the fraction of waypoint routing problems that are successfully completed (see Table 1). "Errands" are waypointing needs that can be fulfilled by multiple point-of-interest categories (e.g., buy water bottle), "locations" are those that correspond to a specific point-of-interest category (e.g., gas station), and "mixed" refers to waypoint routing problems consisting of both errand and location subtasks. "Trip Planning Success Rate" is the fraction of trip planning tasks that are completed under the time budget and where the predicted route provides at least half the experiences of an optimal itinerary. "Excess Penalty" is the median gap between the achieved overall route score and best possible score, conditioned on successfully completed waypoint routing problems; this penalty can be understood as total travel time (in seconds) rescaled by  $5x$  when transiting along an unfavorable road type. Standard error is denoted by  $\pm$ .
+
+graphs, and even a 10 kilometer stretch of road involves, on average, 600 different road segments (OpenStreetMap contributors, 2017). Meanwhile, popular deep learning methods for graph-structured data such as GNNs are often difficult to scale, leading to an impression that graph learning methods are an impractical solution to semantic routing.
+
+In this section we demonstrate a simple graph learning method that serves as a strong baseline for semantic routing tasks. Our key observation is that we can train autoregressive networks on a natural next-vertex prediction objective to unroll routes. We show that, by decomposing semantic routing into next-vertex subproblems, it is possible to construct high-quality routes while only needing to engage a small neighborhood of the road network at each point in time. This approach allows standard transformer networks to achieve strong results on our semantic routing benchmark.
+
+# 3.1 Autoregressive Modeling
+
+Autoregressive learning is an approach to learning generative models where one trains a network on the task of predicting the next element of a string (Brown et al., 2020), which can be a sentence in language generation or, in semantic routing, an edge. These networks can be used to generate strings by unrolling a solution path through repeated next-element predictions or, in the case of semantic routing, through repeated next-edge predictions. From a probabilistic standpoint, this approach circumvents the need to estimate the distribution of the entire solution path by decomposing it into a sequence of conditional distributions via the chain rule.
+
+Despite their greedy nature, autoregressive predictors have demonstrated remarkable success in language generation and showcased their ability to effectively navigate complex decision spaces (Huang et al., 2022; Bubeck et al., 2023). Moreover, they are proven theoretically to be universal learners if trained on adequate data (Malach, 2023). These results suggest autoregressive models to be a viable graph learning solution to semantic routing: one can avoid needing to scale graph learning algorithms to large road networks if they can decompose routing into next-edge prediction problems where only a small neighborhood of the graph is relevant at any single time.
+
+# 3.2 A Proof of Concept
+
+We trained a standard transformer network on our benchmark's training dataset. By truncating the example routes in the dataset at random positions, we optimized our model on an autoregressive objective of predicting subsequent edges. As inputting entire road networks into our model was impractical due to its  $O(|V|^2)$  memory footprint, we instead input a small neighborhood of the road network around the head of each truncated route. The resulting model fits on a single GPU.
+
+![](images/f8546b7bdd0a08be6683704340f36fa56cadcb2f00490b4d6d4466c13e7de57c.jpg)  
+Figure 5: Visualization of a receptive field consisting of 512 road segments; a receptive field is a neighborhood that is the taken around the head of a partial route. As discussed in Section 3.3, our semantic routing model observes a 512-edge receptive field instead of the entire road network.
+
+We report the performance metrics of this model in Table 2. The model successfully interprets and fulfills roughly 4 in 5 waypoint routing tasks; recall that this evaluation is performed on previously unencountered queries and a previously unseen city's road network. The model incurs a median excess travel time of less than  $0.2\%$ ; this is including a 5x travel-time penalization for unnecessary travel on roads (e.g., highways) that the user query specifies an aversion against. Our model also provides a satisfactory itinerary for 2 in 3 trip planning queries. Although it is trivial to compare these results to those of classical methods, we refer readers to Appendix D for random examples of our model's predictions for an intuitive sense of these performance metrics. We believe these figures to be a strong demonstration of the viability of learning-based semantic routing; these success rates on challenging combinatorial problems expressed in natural language is non-trivial in any application domain. Moreover, we are confident that our proof-of-concept is far from an upper bound of what is achievable with learned semantic routing models.
+
+# 3.3 Technical Details
+
+Model description. Our proof of concept implements a simple decoder-only transformer network (Vaswani et al., 2017; Brown et al., 2020) which we train using a standard cross entropy loss on the task of predicting the subsequent edge to a truncated route. Our network—as illustrated in Figure 4—takes as input a sequence of “tokens” that contain the user’s query, the destination, the partial route that we are trying to predict the subsequent edge for, a set of possible subsequent edges, and the small neighborhood (the “receptive field”) around the head of the partial route. To featurize the tokens in this sequence, we create a feature vector for each edge from their metadata, and process it with a multi-layer perceptron. This is in contrast to language models which learn an embedding table for their alphabet, which in our setting would correspond to learning an embedding vector for every segment in our road networks; we deviate from this embedding approach so we can take advantage of road segment metadata and encourage the model to generalize to edges not seen during training. Because of this, our proof-of-concept can provide high-quality routes in a city it has never encountered during training.
+
+Receptive field. Our autoregressive approach allows us to treat each next-edge inference as a local problem: instead of passing in the entire road graph into our model, we only pass in a small neighborhood around the head of the partial route. We refer to this small neighborhood as the receptive field of our model. Our proof-of-concept uses a 500-edge receptive field, which we illustrate the size of in Figure 5. In comparison, our navigation tasks are defined on 10,000-node (roughly 25,000 edge) subgraphs of urban road networks, meaning that our use of a receptive field reduced the memory footprint of our model by a factor of over 2,000.
+
+Inference strategy. Given a next-vertex prediction model, there are several approaches to obtaining a candidate path for a given routing task. The first approach is greedy decoding, a popular and computationally efficient baseline approach in language modeling (Chowdhery et al., 2022). To see
+
+![](images/5f5256e1875344b9e9ab9cf2b688729e2d9b3bbf891e4bf245dd51b57f551c54.jpg)  
+Test-Set Beam Search Performance using Various Scorers
+
+![](images/a98a409736f189238f468935264321b4937940f22a3dae03a96fb472ebf843dc.jpg)  
+Figure 6: The semantic routing performance metrics of an autoregressive transformer under a 10-width beam-search algorithm. The route prediction is extracted from the final beam either using the model's logits (red), a secondary model (purple), or omnipotently choosing the best route (blue).
+
+this, note that the time complexity of our next-vertex prediction framework is independent of the size of the road network. Rather, since each step requires quadratic computation in the length of the route so far, the full time complexity is cubic in the path length  $O(|\mathrm{Route}|^3)$  for greedy decoding, with improved complexity possible through more efficient attention mechanisms (e.g. Wang et al. 2020; Zaheer et al. 2020; Dao et al. 2022).
+
+Inspired by classical AI and natural language generation, we also apply beam search (Reddy, 1977) to our next-vertex prediction model during inference. Beam search returns a set of candidate paths (as many as the beam width) generated by the model and allows one to cleanly trade-off accuracy for inference complexity, with a run time of  $O(\mathrm{BeamWidth}^2 \cdot |\mathrm{Route}|^3)$ . We train a secondary scoring model to choose a final path from this candidate set. The scoring model is trained in a contrastive fashion, where ground-truth paths are taken from our training set as positive examples and random paths are sampled from the outputs of our model as negative examples. After embedding these routes with our next-vertex model, we train the scoring model as a 2-layer MLP with a logistic loss. At inference time, we select from candidate paths by choosing the one receiving the largest logits from the scoring model. In Figure 6, we ablate this scoring model and notice a significant drop in semantic routing performance. The main results presented in this paper, i.e. Table 2, use beam-search with a secondary scoring model.
+
+# 3.4 Ablation Studies and Benchmark Demonstrations
+
+Additional experiments, including ablation studies, can be found in Appendix A and illustrate how increasing receptive field sizes only marginally impacts model performance and evaluate how semantic route planning performance scales with model size and aspects of the road graph's topology. Appendix D lists randomly sampled examples of semantic route planning tasks from our benchmark along with the optimal route according to our benchmark's labeling system.
+
+# 4 Discussion
+
+We study a learning-based approach to semantic routing. To facilitate further research on this important graph learning application, we release a large-scale benchmark of real-world multi-objective navigation problems. We also introduce a proof-of-concept model to establish a competitive baseline for the benchmark and to demonstrate the viability of learning-based approaches. To this end, we formulated an autoregressive architecture that reduces semantic routing to the scalable problem of next-edge prediction. We refer readers to Appendix B for a continued discussion of open questions.
+
+# References
+
+M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, and X. Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015.  
+H. Bast, D. Delling, A. Goldberg, M. Müller-Hannemann, T. Pajor, P. Sanders, D. Wagner, and R. F. Werneck. Route planning in transportation networks. Algorithm engineering: Selected results and surveys, pages 19-80, 2016.  
+R. Bauer, D. Delling, P. Sanders, D. Schieferdecker, D. Schultes, and D. Wagner. Combining hierarchical and goal-directed speed-up techniques for dijkstra's algorithm. Journal of Experimental Algorithmics (JEA), 15:2-1, 2010.  
+R. Bellman. On a routing problem. Quarterly of Applied Mathematics, 16(1):87-90, 1958.  
+T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020.  
+S. Bubeck, V. Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y. T. Lee, Y. Li, S. Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. 2023.  
+A. Chowdhery, S. Narang, J. Devlin, M. Bosma, G. Mishra, A. Roberts, P. Barham, H. W. Chung, C. Sutton, S. Gehrmann, et al. Palm: Scaling language modeling with pathways. 2022.  
+T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to algorithms. MIT press, 2022.  
+G. B. Dantzig and J. H. Ramser. The truck dispatching problem. Management Science, 6(1):80-91, 1959.  
+T. Dao, D. Fu, S. Ermon, A. Rudra, and C. Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022.  
+D. Delling, P. Sanders, D. Schultes, and D. Wagner. Engineering route planning algorithms. In *Algorithms of large and complex networks: design, analysis, and simulation*, pages 117-139. Springer, 2009.  
+D. Delling, A. V. Goldberg, T. Pajor, and R. F. Werneck. Customizable route planning in road networks. Transp. Sci., 51(2):566-591, 2017.  
+M. Deudon, P. Cournut, A. Lacoste, Y. Adulyasak, and L.-M. Rousseau. Learning heuristics for the tsp by policy gradient. In *Integration of Constraint Programming*, Artificial Intelligence, and Operations Research: 15th International Conference, CPAIOR 2018, Delft, The Netherlands, June 26–29, 2018, Proceedings 15, pages 170–181. Springer, 2018.  
+E. W. Dijkstra. A note on two problems in connexion with graphs. Numer. Math., 1(1):269-271, 1959.  
+E. W. Dijkstra. A note on two problems in connexion with graphs. In *Edsger Wybe Dijkstra: His Life, Work, and Legacy*, pages 287-290. 2022.  
+V. P. Dwivedi and X. Bresson. A generalization of transformer networks to graphs. 2020.  
+L.R.Ford.Network flow theory.1956.  
+A. V. Goldberg. A simple shortest path algorithm with linear average time. In European Symposium on Algorithms, pages 230-241. Springer, 2001.
+
+A. Graves, G. Wayne, M. Reynolds, T. Harley, I. Danihelka, A. Grabska-Barwińska, S. G. Colmenarejo, E. Grefenstette, T. Ramalho, J. Agapiou, et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626):471-476, 2016.  
+P. E. Hart, N. J. Nilsson, and B. Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100-107, 1968.  
+W. Huang, P. Abbeel, D. Pathak, and I. Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022.  
+B. Ibarz, V. Kurin, G. Papamakarios, K. Nikiforou, M. Bennani, R. Csordás, A. J. Dudzik, M. Bošnjak, A. Vitvitskyi, Y. Rubanova, et al. A generalist neural algorithmic learner. In Learning on Graphs Conference, pages 2-1. PMLR, 2022.  
+E. Khalil, H. Dai, Y. Zhang, B. Dilkina, and L. Song. Learning combinatorial optimization algorithms over graphs. Advances in neural information processing systems, 30, 2017.  
+W. Kool, H. Van Hoof, and M. Welling. Attention, learn to solve routing problems! 2018.  
+E. Malach. Auto-regressive next-token predictors are universal learners. 2023.  
+U. Meyer. Single-source shortest-paths on arbitrary directed graphs in linear average-case time. In Proceedings of the twelfth annual ACM-SIAM symposium on Discrete algorithms, pages 797-806, 2001.  
+A. Nowak, S. Villar, A. S. Bandeira, and J. Bruna. A note on learning algorithms for quadratic assignment with graph neural networks. 1050:22, 2017.  
+OpenAI. Gpt-4 technical report, 2023.  
+OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org, 2017.  
+N. D. Ratliff, D. Silver, and J. A. Bagnell. Learning to search: Functional gradient techniques for imitation learning. Autonomous Robots, 27:25-53, 2009.  
+R. Reddy. Multi-system approach to speech understanding. NADC, Proceedings: Voice Technology for Interactive Real-Time Command (Control Systems Application, 1977.  
+T. Schick, J. Dwivedi-Yu, R. Dessi, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom. Toolformer: Language models can teach themselves to use tools. 2023.  
+A. Shimbel. Structure in communication nets. In Proceedings of the symposium on information networks, pages 119-203. Polytechnic Institute of Brooklyn, 1954.  
+A. K. Sinop, L. Fawcett, S. Gollapudi, and K. Kollias. Robust routing using electrical flows. In Proceedings of the 29th International Conference on Advances in Geographic Information Systems, pages 282-292, 2021.  
+L. Sint and D. de Champeaux. An improved bidirectional heuristic search algorithm. Journal of the ACM (JACM), 24(2):177-191, 1977.  
+K. Valmeekam, M. Marquez, S. Sreedharan, and S. Kambhampati. On the planning abilities of large language models - a critical investigation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023.  
+A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.  
+P. Velicković, R. Ying, M. Padovano, R. Hadsell, and C. Blundell. Neural execution of graph algorithms. 2019.  
+O. Vinyals, M. Fortunato, and N. Jaitly. Pointer networks. Advances in neural information processing systems, 28, 2015.
+
+S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma. Linformer: Self-attention with linear complexity. 2020.  
+J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022.  
+E. Welty and C. Phillips. Falling Fruit: A Global Collaborative Foraging Map, 2013.  
+K. Xu, J. Li, M. Zhang, S. S. Du, K.-i. Kawarabayashi, and S. Jegelka. What can neural networks reason about? 2019.  
+K. Xu, M. Zhang, J. Li, S. S. Du, K.-i. Kawarabayashi, and S. Jegelka. How neural networks extrapolate: From feedforward to graph neural networks. 2020.  
+S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y. Cao, and K. R. Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023.  
+S. Yun, M. Jeong, R. Kim, J. Kang, and H. J. Kim. Graph transformer networks. Advances in neural information processing systems, 32, 2019.  
+M. Zaheer, G. Guruganesh, K. A. Dubey, J. Ainslie, C. Alberti, S. Ontanon, P. Pham, A. Ravula, Q. Wang, L. Yang, et al. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297, 2020.
+
+# A Additional Experiments
+
+In this section, we describe some additional brief experiments into the behavior of autoregressive transformers that are trained on semantic routing tasks.
+
+![](images/671a10502b28cf5c1627825e1e36c54642eb3a946cc03dfce29a2a4874fa9abd.jpg)  
+Figure 7: The grid world semantic routing performance of transformer networks of varying receptive field sizes. The largest possible receptive field size is the entire road network, which is 3516 edges.
+
+Receptive field experiment. In this experiment, we measure how model performance scales with receptive field size. As demonstrated by the strong performance of the proof-of-concept model, the greedy approach using receptive fields to make next-edge predictions with only local information is surprisingly effective. To validate that the use of a receptive field does not significantly degrade model performance, we consider train autoregressive transformer models on our grid-world benchmark, varying the size of each model's receptive fields. As depicted in Figure 7, there are no statistically significant gains in performance from passing in the entire road network (3560 edges) or even a large subset of it (1024 edges); rather, the gains from using a larger receptive field almost entirely disappear even when only a small fraction of the graph is included (256 edges).
+
+We attribute this to two factors. First, real-world road networks are intentionally designed to be amenable to human drivers, which, like our model, are often learning agents making greedy navigation decisions with only local information; similarly, businesses and facilities can usually be found in predictable places. Second, autoregressive methods have demonstrated significant utility even in applications where greedy approaches appear to be a poor fit. For example, recent findings increasingly indicate that language models—at scale—are capable of long-term planning despite their greedy nature (Huang et al., 2022; Valmeekam et al., 2023).
+
+![](images/0ddadb935999d5ca284c4a7e3ee38dc5114b6ff7d8394ffafc8c57d900505f86.jpg)  
+Figure 8: The grid world performance of an autoregressive transformer under beam-search with varying width.
+
+Beam search width tuning. This experiment studies how beam search width affects the performance of autoregressive semantic routing models. To this end, we implement and train our autoregressive Transformer architecture on our grid-world semantic routing dataset, and evaluate the performance of the beam search inference strategy for varying beam widths. The results, depicted in Figure 8, demonstrate monotone but diminishing returns from increasing beam width.
+
+![](images/04d4eb73a5f21dad34985d663982009c0481daa007e7443655cbcad38e0b3619.jpg)  
+Next-Vertex Accuracy vs Points-of-Interest Density
+
+![](images/5ea2e2a4bf5212b39a81a82ee274a6ffa2a77c4ce060a4888d0061d44aadfc0c.jpg)  
+Figure 9: The next-edge prediction accuracy (left) and waypoint routing success (right) of autoregressive transformers on tasks in our grid-world benchmark.
+
+Point-of-interest density experiment. In this experiment, we attempt to understand how the difficulty of fulfilling a query affects the difficulty of semantic routing and, in particular, waypoint routing tasks. One should expect that, when points-of-interest are extremely plentiful, waypoint routing reduces to finding the shortest path since the shortest path likely encounters the desired points-of-interest by chance. In this experiment, we train autoregressive transformers models on multiple variants of the grid-world benchmark where we vary the density of points-of-interest on our road network. The results plotted in Figure 9, indicate that while—as expected—a higher point-of-interest density results in higher waypoint routing success rates. However, we surprisingly also observe that a higher point-of-interest density makes next-edge prediction more difficult. Our interpretation of this apparent contradiction is that point-of-interest density indeed makes semantic routing easier, but also increases the number of reasonable routes that may exist; the resulting higher entropy distribution over solution routes results in a lower next-edge prediction accuracy. This highlights a trend we frequently observed while training autoregressive semantic routing models: relative differences in next-edge accuracy often significantly disagree with relative differences in system performance, such as waypoint routing success.
+
+![](images/aa72efa2a04952b1b377ae08eb3c38edcc36ed4087d3384f5fd07e28de29bb52.jpg)  
+Figure 10: The semantic routing performance (left) and next-edge prediction accuracy (middle, right) of autoregressive transformers on our grid-world benchmark. On the left, both embedding dimension and network depth are rescaled. In the middle, only depth is rescaled. On the right, only embedding dimension is rescaled.
+
+![](images/df619a7dc7ccc4eb677520a4401c88b4eb12eb095fa4f6cba2c02f794e4b686c.jpg)
+
+![](images/9b17116760990fa005e2b5881422cee6433b216feb45a302969dfb6d62c3ffac.jpg)
+
+A scaling study. This experiment studies how network scale interacts with the performance of autoregressive semantic routing models. To this end, we implement our autoregressive Transformer architecture and train multiple copies of it on our grid-world semantic routing dataset, where each copy is configured with a different model capacity. We rescale our model in two directions. We first vary the depth of our network from its default depth of 6 transformer blocks to either 3 or 10 blocks deep. We separately vary the width of our network from its default embedding dimension of 256 to
+
+either a dimension of 64 and 1024; we follow the canonical scaling strategy for transformer networks and maintain intermediate dimension at twice that of embedding dimension. The impact of these rescalings on next-vertex prediction accuracy is depicted in the two right-most plots of Figure 10. As expected, scale leads to gains in prediction power but these gains quickly saturate. If we rescale our model simultaneously in both directions, we again observe quickly diminishing gains from scale; this result is depicted at the left of Figure 10 and is stated in terms of semantic routing success rate rather than next-edge prediction accuracy. This saturation in gains from scaling up model capacity is one reason to believe that our choice of a simple transformer architecture may not be optimal.
+
+# B Open Questions
+
+Longer Trips. Most timesteps in real-world routing tasks are uninteresting and, for example, correspond to driving straight on a highway for hundreds of road segments. Typically, this is addressed with hierarchical methods that add "shortcut" edges to summarize common sequences of road segments (Delling et al., 2009). To maintain the simplicity of our proof-of-concept, we only implemented a simple alternative of contracting vertices with edges of the same road type, which only reduced effective road segment length by a factor of 2-3. This posed an issue as, in longer routes, next-vertex pairs contain sparser signals as more of the route is trivial, making training more expensive. We accordingly configured our benchmark to restrict navigation tasks to those can be solved in 600 road segments. The benchmark can be made more challenging by increasing this cap, which will require integrating techniques like strategically adjusting the coarseness of road segments in contraction hierarchies (Delling et al., 2009).
+
+Alternatives to Next-Vertex Prediction. Although we implemented an autoregressive approach that unrolls routes from source to destination, autoregressive strategies need not unroll routes in this order. For example, one can unroll routes from the destination, or in a hierarchical middle-out fashion. As we found that unrolling source to destination resulted in sufficient good performance by real-world measures, we defer consideration of alternative approaches to future work.
+
+Receptive Field Representation. Our model encodes its receptive field inputs in a simple fashion: concatenate all edges into a sequence and pass the sequence through a transformer network so different edges can attend to one another. As our proof-of-concept was not aimed at demonstrating that any particular network architecture is superior, we opted for a basic transformer for simplicity and scalability. We expect that performance improvements can be gained by both enlarging the size of the receptive field and including richer graph metadata (e.g., allowing the model to retrieve the reviews of relevant local businesses during route finding), for which graph representation learning methods such as GNNs and graph transformers (Yun et al., 2019; Dwivedi and Bresson, 2020) are potentially useful thanks to their ability to combine the information from graph connectivity and metadata through learning.
+
+Efficiency and Scaling. The significant attention received by large language models have spurred a number of advancements in implementing autoregressive models, especially transformer networks, to be more efficient and scalable Wang et al. (2020); Zaheer et al. (2020); Dao et al. (2022). As we have shown that autoregressive transformers are indeed viable solutions to semantic routing, these efficiency gains can similarly be realized by semantic routing applications. In particular, the inference efficiency of our model can be greatly improved by adopting architectural modifications like more efficient attention mechanisms (Dao et al., 2022).
+
+# C Experiment Details
+
+In this section, we review the hyperparameters for each experiment in this paper. In terms of grid world experiments, Table 6a describes the hyperparameters for the point-of-interest density experiment, the beam width experiment and the secondary scoring model ablation experiment; Table 3 describes the hyperparameters for the receptive field size experiment; and Table 5 and Table 4 describe the hyperparameters for the scaling study experiment. The hyperparameters for the experiment on our benchmark are described in Table 6b. We use "Embedding Dim.", "Intermediate Dim.", and "Attention Heads" to denote the embedding dimension, intermediate dimension, and attention head count of the transformer blocks in the network. Similarly, we use "MLP Layers (Road)" and "Intermediate Dim.
+
+(Road)” to denote the depth and intermediate dimensions of the multi-layer perceptron blocks used to encode road segment features. The “Receptive Field Size” denotes the number of edges provided in the receptive field input. In the “Learning Rate Schedule” key, “Cosine” denotes that a cosine one-cycle learning rate schedule is used. The “Seeds” key refers to the number of random seeds the experiment is performed on. Experiments performed in this paper were conducted using computer resources on Google Cloud. Approximately 10,000 CPU hours were used to generate training data from the benchmark software. The experiments described in Appendix A were performed on a GPU cluster for approximately 1,000 GPU hours. The experiments described in Section 3 were performed on a GPU cluster for approximately 20,000 GPU hours.
+
+<table><tr><td>Hyperparameter</td><td colspan="6">Values</td></tr><tr><td>Datapoints (# Routes)</td><td>1 million</td><td>1 million</td><td>1 million</td><td>1 million</td><td>1 million</td><td>1 million</td></tr><tr><td>Learning Rate</td><td>0.0001</td><td>0.0001</td><td>0.0001</td><td>0.0001</td><td>0.0001</td><td>0.0001</td></tr><tr><td>L2 Weight Decay</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>Dropout</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Batch Size</td><td>1024</td><td>1024</td><td>1024</td><td>1024</td><td>128</td><td>128</td></tr><tr><td>Iterations</td><td>200,000</td><td>200,000</td><td>200,000</td><td>200,000</td><td>400,000</td><td>400,000</td></tr><tr><td>Seeds</td><td>5</td><td>5</td><td>5</td><td>5</td><td>5</td><td>5</td></tr><tr><td>Learning Rate Schedule</td><td>Cosine</td><td>Cosine</td><td>Cosine</td><td>Cosine</td><td>Cosine</td><td>Cosine</td></tr><tr><td>Receptive Field Size</td><td>8</td><td>64</td><td>128</td><td>256</td><td>1024</td><td>3516</td></tr><tr><td>Attention Heads</td><td>4</td><td>4</td><td>4</td><td>4</td><td>4</td><td>4</td></tr><tr><td>Embedding Dim.</td><td>256</td><td>256</td><td>256</td><td>256</td><td>256</td><td>256</td></tr><tr><td>Intermediate Dim.</td><td>512</td><td>512</td><td>512</td><td>512</td><td>512</td><td>512</td></tr><tr><td>Transformer Blocks</td><td>6</td><td>6</td><td>6</td><td>6</td><td>6</td><td>6</td></tr><tr><td>Intermediate Dim. (Road)</td><td>512</td><td>512</td><td>512</td><td>512</td><td>512</td><td>512</td></tr><tr><td>MLP Layers (Road)</td><td>2</td><td>2</td><td>2</td><td>2</td><td>2</td><td>2</td></tr></table>
+
+Table 3: Hyperparameters for the receptive field size experiment on the grid-world dataset.  
+
+<table><tr><td>Hyperparameter</td><td>Values</td><td></td><td></td><td></td></tr><tr><td>Datapoints (# Routes)</td><td>4 million</td><td>4 million</td><td>4 million</td><td>4 million</td></tr><tr><td>Learning Rate</td><td>0.0001</td><td>0.0001</td><td>0.0001</td><td>0.0001</td></tr><tr><td>L2 Weight Decay</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>Dropout</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Batch Size</td><td>1024</td><td>1024</td><td>1024</td><td>1024</td></tr><tr><td>Iterations</td><td>400,000</td><td>400,000</td><td>400,000</td><td>400,000</td></tr><tr><td>Seeds</td><td>5</td><td>5</td><td>5</td><td>5</td></tr><tr><td>Learning Rate Schedule</td><td>Cosine</td><td>Cosine</td><td>Cosine</td><td>Cosine</td></tr><tr><td>Receptive Field Size</td><td>256</td><td>256</td><td>256</td><td>256</td></tr><tr><td>Attention Heads</td><td>4</td><td>4</td><td>4</td><td>4</td></tr><tr><td>Embedding Dim.</td><td>256</td><td>64</td><td>1024</td><td>256</td></tr><tr><td>Intermediate Dim.</td><td>512</td><td>128</td><td>2048</td><td>512</td></tr><tr><td>Transformer Blocks</td><td>6</td><td>6</td><td>6</td><td>10</td></tr><tr><td>Intermediate Dim. (Road)</td><td>512</td><td>128</td><td>2048</td><td>512</td></tr><tr><td>MLP Layers (Road)</td><td>2</td><td>2</td><td>2</td><td>2</td></tr></table>
+
+Table 4: Hyperparameters for the scaling study experiment on the grid-world dataset.
+
+<table><tr><td colspan="4">Hyperparameter</td></tr><tr><td>Datapoints (# Routes)</td><td>4 million</td><td>4 million</td><td>4 million</td></tr><tr><td>Learning Rate</td><td>0.0001</td><td>0.0001</td><td>0.0001</td></tr><tr><td>L2 Weight Decay</td><td>0</td><td>0</td><td>0</td></tr><tr><td>Dropout</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Batch Size</td><td>1024</td><td>1024</td><td>1024</td></tr><tr><td>Iterations</td><td>400,000</td><td>400,000</td><td>400,000</td></tr><tr><td>Seeds</td><td>5</td><td>5</td><td>5</td></tr><tr><td>Learning Rate Schedule</td><td>Cosine</td><td>Cosine</td><td>Cosine</td></tr><tr><td>Receptive Field Size</td><td>256</td><td>256</td><td>256</td></tr><tr><td>Attention Heads</td><td>4</td><td>4</td><td>8</td></tr><tr><td>Embedding Dim.</td><td>256</td><td>64</td><td>1024</td></tr><tr><td>Intermediate Dim.</td><td>512</td><td>128</td><td>2048</td></tr><tr><td>Transformer Blocks</td><td>6</td><td>3</td><td>10</td></tr><tr><td>Intermediate Dim. (Road)</td><td>512</td><td>128</td><td>2048</td></tr><tr><td>MLP Layers (Road)</td><td>2</td><td>2</td><td>2</td></tr></table>
+
+Table 5: Hyperparameters for the scaling study experiment on the grid-world dataset (continued).  
+
+<table><tr><td>Hyperparameter</td><td>Values</td></tr><tr><td>Datapoints (# Routes)</td><td>1 million</td></tr><tr><td>Learning Rate</td><td>0.0001</td></tr><tr><td>L2 Weight Decay</td><td>0</td></tr><tr><td>Dropout</td><td>-</td></tr><tr><td>Batch Size</td><td>1024</td></tr><tr><td>Iterations</td><td>200,000</td></tr><tr><td>Seeds</td><td>5</td></tr><tr><td>Learning Rate Schedule</td><td>Cosine</td></tr><tr><td>Receptive Field Size</td><td>256</td></tr><tr><td>Attention Heads</td><td>4</td></tr><tr><td>Embedding Dim.</td><td>256</td></tr><tr><td>Intermediate Dim.</td><td>512</td></tr><tr><td>Transformer Blocks</td><td>6</td></tr><tr><td>Intermediate Dim. (Road)</td><td>512</td></tr><tr><td>MLP Layers (Road)</td><td>2</td></tr></table>
+
+(a) Hyperparameters for grid-world experiments varying point-of-interest density, beam width, and the scoring model.
+
+<table><tr><td>Hyperparameter</td><td>Values</td></tr><tr><td>Datapoints (# Routes)</td><td>2 million</td></tr><tr><td>Learning Rate</td><td>0.00003</td></tr><tr><td>L2 Weight Decay</td><td>0</td></tr><tr><td>Dropout</td><td>-</td></tr><tr><td>Batch Size</td><td>1024</td></tr><tr><td>Iterations</td><td>600,000</td></tr><tr><td>Seeds</td><td>3</td></tr><tr><td>Learning Rate Schedule</td><td>Cosine</td></tr><tr><td>Receptive Field Size</td><td>512</td></tr><tr><td>Attention Heads</td><td>8</td></tr><tr><td>Embedding Dim.</td><td>1024</td></tr><tr><td>Intermediate Dim.</td><td>2048</td></tr><tr><td>Transformer Blocks</td><td>8</td></tr><tr><td>Intermediate Dim. (Road)</td><td>2048</td></tr><tr><td>MLP Layers (Road)</td><td>2</td></tr></table>
+
+(b) Hyperparameters for the proof-of-concept experiment on our main benchmark, i.e., the results in Table 2.
+
+<table><tr><td>Hyperparameter</td><td>Values</td></tr><tr><td>Datapoints</td><td>5,000</td></tr><tr><td>Learning Rate</td><td>0.0005</td></tr><tr><td>L2 Weight Decay</td><td>0.01</td></tr><tr><td>Batch Size</td><td>256</td></tr><tr><td>Iterations</td><td>20,000</td></tr><tr><td>Intermediate Dim.</td><td>64</td></tr><tr><td>Hidden Layers</td><td>1</td></tr><tr><td>Activation</td><td>GELU</td></tr></table>
+
+(c) Hyperparameters for the secondary scoring model in all experiments.
+
+Table 6: Additional hyperparameters.
+
+# D Example of Predicted Routes
+
+This section provides eight randomly sampled examples of the semantic routing tasks in our benchmark. These are depicted in Figure 11, Figure 12, Figure 13 and Figure 14.
+
+Find me the nearest supermarket to get ingredients for dinner and also suggest a place for a quick lunch. Avoid local roads when possible.
+
+![](images/6126e1ff28f673633919626e9aca4650ddba9aa968220f18b8b261efbcc80fd7.jpg)
+
+Can you guide me to nearby grocery stores to get dinner ingredients, a gas station to refuel my car, and stores selling energy drinks?
+
+![](images/f11ffe31ec59a4b1fca24288fd70bec2996518e5e22abd782ab0269a29361b82.jpg)  
+Figure 11: Randomly sampled semantic routing tasks from our benchmark, and their optimal routes (orange).
+
+![](images/0250e8feb65dda346ebbda808f2e2df599d0c94884c7a54a6c0532470bafb83f.jpg)  
+Mind finding a nearby pharmacy?
+
+![](images/c3b11393ec1ece4f55dee4504fe6b5a6fe1989cda56375026c8a2eea10dedd1c.jpg)  
+Figure 12: Randomly sampled semantic routing tasks from our benchmark, and their optimal routes (orange).
+
+Find the fastest way to a snack store and a fast food place, avoiding highways.
+
+![](images/b0addc28f10765e873cf30364ecdb9d2c019a5f3b80472195dc9c9a9c774c943.jpg)
+
+Can you guide me to nearby grocery stores to get dinner ingredients, a gas station to refuel my car, and stores selling energy drinks?
+
+![](images/1b898c75c11609e21f5a31749b0c516e6e85f555c9155dff48ad396260758d0d.jpg)  
+Figure 13: Randomly sampled semantic routing tasks from our benchmark, and their optimal routes (orange).
+
+![](images/0f571758c29d516ddb152c0081d5bd1bcbafb44a138a17c32ad6a675432690e8.jpg)  
+Games Video games Amusement arcade Toys Start End
+
+Figure 14: Randomly sampled semantic routing tasks from our benchmark, and their optimal routes (orange).  
+![](images/c4f7eeec6f623355864cf749596ed98f644704d69fd1c65e56465ce96c9bd693.jpg)  
+I need to stop at shops and buy dinner ingredients, car snacks, and also get gas.  
+Deli Vending machine Supermarket Butcher Fuel Convenience Greengrocer Start End
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: Yes, the abstract and introduction clearly state the main two contributions of the paper: (1) a benchmark for semantic routing (Section 2) and (2) a proof-of-concept demonstration that autoregression can overcome scaling challenges (Section 3).
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: The primary limitation of our benchmark is that, as an automated benchmark, it relies on an automated system for scoring solutions rather than human raters; this is discussed in Section 2.2 along with the note that the benchmark's scoring system nonetheless corresponds to reasonable high-quality solutions. The limitations of our proof-of-concept model are described extensively in our discussion of open questions and future directions in Section B.
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory Assumptions and Proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [NA]
+
+Justification: The paper does not include theoretical results.
+
+Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental Result Reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+Answer: [Yes]
+
+Justification: Experiment details including hyperparameters are provided in detail in Section 3.3 and Appendix C.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.
+
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example
+
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [Yes]
+
+Justification: All data and code described in this paper, for both our benchmark and proof-of-concept, are available open-source under an Apache 2.0 license.
+
+Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental Setting/Details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes]
+
+Justification: All training and test details are described in the tables of Appendix C.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment Statistical Significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+Answer: [Yes]
+
+Justification: Standard error is provided in the main results of Table 2.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.  
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).
+
+- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.  
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments Compute Resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: Computer resource usage, though estimated and not known exactly, is described in Section C.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code Of Ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: The Code of Ethics is conformed to.
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader Impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [NA]
+
+Justification: This paper focuses on semantic routing (e.g., navigation services like Google Maps), an application that does not lend itself to malicious use via disinformation, surveillance, discrimination, privacy violations, or security risks.
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.
+
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.  
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [Yes]
+
+Justification: The paper poses no such risks.
+
+# Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: The original creators of both data and software used in the paper are cited.
+
+# Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URI.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.
+
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New Assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [Yes]
+
+Justification: The released software is accompanied with documentation and example Python notebooks to facilitate usage.
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and Research with Human Subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification: The paper does not involve crowdsourcing nor research with human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: The paper does not involve crowdsourcing nor research with human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

2024/Semantic Routing via Autoregressive Modeling/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0599b2d8a49ca30a1f87166dbb4e7518bcbe5c5f8259af5409301768d44daf60
+size 1178824

2024/Semantics and Spatiality of Emergent Communication/50cbb227-8a2f-45c8-b39a-d2448bff013d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ef10e758d8313327f21dc5d9cb9005c59c76d099488799d80f17becf88fc874
+size 1094398

2024/Semantics and Spatiality of Emergent Communication/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bba099d344877b6729d811e8cdc12a82760ff6de06250443f2c0cc090ec7c919
+size 1840161

2024/Semi-Open 3D Object Retrieval via Hierarchical Equilibrium on Hypergraph/d14549f3-c4ca-4b7a-939e-03f66bcf2477_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aae4252e21aecdac399b68318a7406976f0990fa0c85bbb042d3531a8b77f99f
+size 5373310

2024/Semi-Open 3D Object Retrieval via Hierarchical Equilibrium on Hypergraph/full.md

@@ -0,0 +1,556 @@
+# Semi-Open 3D Object Retrieval via Hierarchical Equilibrium on Hypergraph
+
+Yang Xu $^{1}$ , Yifan Feng $^{1}$ , Jun Zhang $^{2}$ , Jun-Hai Yong $^{1}$ , and Yue Gao $^{1*}$ $^{1}$ BNRist, THUIBCS, KLISS, BLBCI, School of Software, Tsinghua University, China  
+ $^{2}$ Tencent AI Lab  
+{xuyang9610, evanfeng97}@gmail.com, junejzhang@tencent.com, {yongjh, gaoyue}@tsinghua.edu.cn
+
+# Abstract
+
+Existing open-set learning methods consider only the single-layer labels of objects and strictly assume no overlap between the training and testing sets, leading to contradictory optimization for superposed categories. In this paper, we introduce a more practical Semi-Open Environment setting for open-set 3D object retrieval with hierarchical labels, in which the training and testing set share a partial label space for coarse categories but are completely disjoint from fine categories. We propose the Hypergraph-Based Hierarchical Equilibrium Representation (HERT) framework for this task. Specifically, we propose the Hierarchical Retrace Embedding (HRE) module to overcome the global disequilibrium of unseen categories by fully leveraging the multi-level category information. Besides, tackling the feature overlap and class confusion problem, we perform the Structured Equilibrium Tuning (SET) module to utilize more equilibrial correlations among objects and generalize to unseen categories, by constructing a superposed hypergraph based on the local coherent and global entangled correlations. Furthermore, we generate four semi-open 3DOR datasets with multi-level labels for benchmarking. Results demonstrate that the proposed method can effectively generate the hierarchical embeddings of 3D objects and generalize them towards semi-open environments.
+
+# 1 Introduction
+
+3D objects are of paramount significance, finding extensive applications from computer graphics [1] to security [31] and autonomous robotics [2]. As the fundamental task of data acquisition, 3D object retrieval (3DOR) [22, 10] plays a pivotal role in the computer vision community [3]. 3DOR methods learn to represent 3D objects from the training set and then extract features from query objects to effectively align similar samples. According to the category overlap between training and testing sets, existing 3DOR algorithms can be divided into closed-set and open-set types. The former conducts retrieval for objects whose categories have been seen in the training set [32, 36], while the latter handles objects of unseen categories [7].
+
+Existing open-set 3D learning methods are based on the assumption that the labels of object categories are at a single level [46, 23]. In practical scenarios, objects are typically described by multiple hierarchical labels. This leads to categories in training and testing sets showing varying degrees of overlap at different levels. As illustrated in Figure 1(a), we term this scenario, where the training and testing set share a partial label space of coarse categories but are completely disjoint from fine categories, as Semi-Open Environment. Besides, existing methods typically extract basic features using pre-trained models, followed by further open-set learning and optimization. However, for
+
+![](images/9d45f4f37ebf37daa3bc091ddc82849b4d0bdf08965d836606fa0e65253f4a5f.jpg)  
+(a) Hierarchical labels in semi-open environment.
+
+![](images/345cca5ca092df573bed810070a353652d8e0a5ac411b09f9b080c8cbd7eb308.jpg)  
+(b) Optimization contradiction of hierarchical labels.  
+Figure 1: Illustration of motivation from open-set to semi-open 3DOR. Objects may be more accurately described by hierarchical labels than single-level labels in real-world scenarios. In this semi-open setting, the training and testing set share a partial label space for coarse categories but are completely disjoint from fine categories.
+
+categories with the same coarse label but different fine labels, the training of the basic model and the open-set learning will result in contradictory optimizations as shown in Figure 1(b).
+
+Specifically for 3D object retrieval techniques, only a few methods [7, 32] have been explored for open-set retrieval. These methods consider only the single-layer labels of objects and strictly assume no overlap between the training and testing data distributions. The existing open-set 3DOR method [7] treats objects from seen and unseen categories as isomorphic vertices and constructs a graph model based on global correlations. On one hand, this isomorphic model overlooks hierarchical correlations inherent in multi-level categories, particularly the shared coarse labels. Consequently, the embedding distribution from structure-aware learning tends to be unbalanced towards the feature space of seen categories in the training set. On the other hand, this structure focuses on global correlations between seen and unseen categories, neglecting intricate local correlations within unseen categories themselves. This lack of local attention may result in issues like feature overlap and class confusion within these unseen categories.
+
+Focusing on this practical semi-open environment for retrieval, where the training and testing set share a partial label space for coarse categories but are completely disjoint from fine categories, we introduce the semi-open 3D object retrieval task and construct four datasets with multi-level labels to expand the application of 3DOR. We propose the Hypergraph-Based Hierarchical Equilibrium Representation (HERT) framework for semi-open 3DOR. To overcome the global disequilibrium of unseen categories, we propose the Hierarchical Retrace Embedding module (HRE) to achieve balanced representation across multi-level categories. This module generates multi-level retrace embeddings for capturing the hierarchical semantics of objects. To tackle the feature overlap and class confusion problem, we propose the Structured Equilibrium Tuning (SET) module. This module utilizes high-order correlations among objects for unseen category generalization, by constructing a superposed Hypergraph based on local coherent and global entangled correlations. In summary, our main contributions are fourfold:
+
+- We introduce the semi-open 3D object retrieval task to refine the setting of the 3D object retrieval task in real-world scenario applications, and we construct four datasets for benchmarking downstream tasks.  
+- We propose the Hypergraph-Based Hierarchical Equilibrium Representation (HERT) framework for the semi-open 3DOR task, including the Hierarchical Retrace Embedding (HRE) and the Structured Equilibrium Tuning (SET) modules, which are designed to overcome the distribution disequilibrium and confusion of unseen categories.  
+- We propose a superposed Hypergraph structure to capture high-order correlations among objects, under the guidance of local coherent correlations and global entangled correlations from hierarchical category information.  
+- Experimental results on the four datasets demonstrate that our method can outperform state-of-the-art retrieval methods towards the semi-open environment.
+
+# 2 Related Work
+
+3D Object Retrieval. 3D object retrieval (3DOR) aims to find the relevant objects from the target set for the query objects, most 3DOR methods construct feature alignment models for 3D objects through metric learning. MMJN [20] proposes a discrimination loss to minimize the distance between objects belonging to the same categories, and learn discriminative embeddings for retrieval. MIFN [17] constructs fusion networks by weighted concatenation for modality-specific features. PVNet [42] proposes a joint network for the fusion of multi-view and point cloud features. PVRNet [43] proposes an attention mechanism to generate the unified embedding of different modalities. CMCL [14] designs a cross-modal center loss to compress features of different modalities to a modal-invariant space. However, most existing methods mainly focus on close-set retrieval. Currently, only a few methods [7, 32, 39, 38, 37] have explored open-set retrieval, but they typically assume no overlap between training and testing data distributions, which is at odds with the semi-open setting.
+
+Open-Set Learning. Open-set learning (OSL) methods can be roughly separated into two categories [45], i.e., discriminative methods and generative models. In discriminative models, traditional methods achieve the classification of unknown and known categories based on the Support Vector Machine (SVM) [27, 28] or the Extreme Value Theory (EVT) [12, 26]. Recently, deep learning methods for OSL have made remarkable rapid progress. OpenMax [4] is the first algorithm proposed to replace the SoftMax layer and calibrate the output probability with the Weibull distribution. Then, [29] and [41] utilize the one-vs-rest units and reconstructed latent representation for unknown detection. Generative OSL models are designed to anticipate the distribution of novel classes through training. PROSER [45] allocates placeholders for both data and classifier to detect the unknown classes, C2AE [21] propose a two-step framework to tackle open-set recognition problem with close-set training and open-set training, respectively. Both types of open-set learning methods [9, 19] are based on the assumption that there is no overlap between known and unknown data distribution, thus facilitating the design of classifiers for unseen object detection, but are challenging to generalize to practical scenarios that involve hierarchical superposed categories.
+
+Multi-Label Learning. Multi-label learning aims to create a model that can assign multiple labels for each instance simultaneously. Existing multi-label learning methods are mostly designed for recognition tasks, and based on one-vs-all classifiers [40, 24] and embeddings [44, 18]. As for the retrieval task, DMSSPH [33] proposes a multi-level preserving hashing network, and AMD-GCN [16] designs a GCN-based network for multi-label pattern image retrieval. TranGCN [15] proposes a cross-modal attention mechanism at each layer for multi-label embeddings. Although these methods construct effective models for image embedding, they overlook the hierarchical relationship and dependencies between different categories. In this paper, we aim to model the correlations between hierarchical label information and generalize to unseen categories.
+
+# 3 Problem Setup
+
+# 3.1 3D Object Retrieval
+
+The goal of the 3D object retrieval (3DOR) is to develop a method using the training set  $S_{trn} = \{(o_i, y_i^c, y_i^f)\}_{i=1}^L$ , which is then employed to identify similar objects in the retrieval (testing) set  $S_{ret} = \{(o_i, \hat{y}_i^c, \hat{y}_i^f)\}_{i=1}^R$ , which is comprised of the query set  $S_q$  and the target set  $S_t$ . Here,  $L$  and  $R$  represent the number of samples in the training and testing (retrieval) sets, respectively. The expressions  $y_i^f \in \mathcal{Y}^f = \{c_j^f\}_{j=1}^{Y^f}$  and  $y_i^c \in \mathcal{Y}^c = \{c_k^c\}_{k=1}^{Y^c}$  denote the fine and coarse category labels of 3D object  $o_i$ , respectively. Typically,  $y_i^c$  provides the category label from a more general (coarse) level, such as the basic geometric shapes or other common attributes between  $S_{trn}$  and  $S_{ret}$ , whereas  $y_i$  provides a semantic-specific (fine) category label. Generally, the number of coarse categories  $Y^c$  is much smaller than the number of fine categories, which is  $Y^c \ll Y^f$ . For 3D objects, The  $o_i = \{m_r\}_{r=1}^M$  denotes the representation by  $M$  modalities, i.e., multi-view, point cloud and voxel.
+
+# 3.2 Semi-Open 3D Object Retrieval
+
+In traditional open-set 3DOR, the category spaces of the training set and the retrieval set are not the same fine indicating  $y_{i}^{f}\in \mathcal{V}^{f} = \{c_{j}^{f}\}_{j = 1}^{Y^{f}},\hat{y}_{i}^{f}\in \hat{\mathcal{V}}^{f} = \{\hat{c}_{j}^{f}\}_{j = 1}^{\hat{Y}^{f}}$ , and  $\mathcal{V}^f\neq \hat{\mathcal{V}}^f$ .  $Y^{f}$  and  $\hat{Y}^f$  denote
+
+![](images/dcd7e1e1ceee76bbaae7403f0286b2d531c45bb78311132fe3c8ec3db6ab1088.jpg)  
+Figure 2: An overview of the proposed Hypergraph-Based Hierarchical Equilibrium Representation framework (HERT) framework for semi-open 3D object retrieval. Our framework is composed of the Hierarchical Retrace Embedding (HRE) and the Structured Equilibrium Tuning (SET) modules, which are designed for multi-level semantic embedding and hierarchical structure-aware tuning.
+
+the numbers of fine categories in the training set and the retrieval set, respectively. In the practice of semi-open 3DOR, the training set and the testing set share the same space of coarse category, which  $y_{i}^{c} \in \mathcal{Y}^{c} = \{c_{k}^{c}\}_{k=1}^{Y^{c}}$ ,  $\hat{y}_{i}^{c} \in \hat{\mathcal{Y}}^{c} = \{\hat{c}_{k}^{c}\}_{k=1}^{\hat{Y}^{c}}$ , and  $\mathcal{Y}^{c} = \hat{\mathcal{Y}}^{c}$ ,  $Y^{c}$  and  $\hat{Y}^{c}$  denote the numbers of coarse categories in the training set and the testing set, respectively. During the retrieval phase, each query object is provided with a coarse category label. For better representation, the semi-open 3DOR defines the retrieval model  $r := (o_{i}|y_{i}^{c}) \mapsto z_{i}$  that maps the 3D object  $o_{i}$  into a semantic embedding  $z_{i} \in \mathbb{R}^{d}$  under coarse category condition  $y_{i}^{c}$ . The task seeks to minimize the expected risk under the conditional constraints of coarse labels:
+
+$$
+\mathbb {E} _ {\left(S _ {i}, S _ {j}\right) \sim \left(\mathcal {S} _ {q}, \mathcal {S} _ {t}\right)} \left[ \mathbb {I} \left(\hat {y} _ {i} ^ {f} \neq \hat {y} _ {j} ^ {f}\right) e ^ {- \mathbb {D} \left(r \left(o _ {i} \mid y _ {i} ^ {c}\right), r \left(o _ {j} \mid y _ {j} ^ {c}\right)\right)} + \mathbb {I} \left(\hat {y} _ {i} ^ {f} = \hat {y} _ {j} ^ {f}\right) \left(1 - e ^ {- \mathbb {D} \left(r \left(o _ {i} \mid y _ {i} ^ {c}\right), r \left(o _ {j} \mid y _ {j} ^ {c}\right)\right)}\right) \right], (1)
+$$
+
+where  $S_{i} = (o_{i},y_{i}^{c},\hat{y}_{i}^{f})$  and  $S_{j} = (o_{j},y_{j}^{c},\hat{y}_{j}^{f})$  are object instances sampled from the query set  $S_{q}$  and target set  $S_{t},\hat{y}_{i}^{f}$  and  $\hat{y}_j^f$  denote the predicted fine labels,  $y_{i}^{c}$  and  $y_{j}^{c}$  are the conditional constraint, which are coarse labels.  $\mathbb{I}(\cdot)$  is the indicator function, which returns 1 if the expression holds and 0 otherwise.  $\mathcal{H}$  is the hypothesis space of map  $r(\cdot |\cdot)$ .  $\mathbb{D}(z_i,z_j)$  is the distance metric function between different embeddings.
+
+# 4 Methodology
+
+# 4.1 Overall Framework
+
+The overall framework of the Hypergraph-Based Hierarchical Equilibrium Representation (HERT) framework is illustrated in Figure 2. HERT is composed of the Hierarchical Retrace Embedding (HRE) and the Structured Equilibrium Tuning (SET) modules. Given 3D objects represented by multiple modalities, common-used backbones are used to extract the basic features for each modality. Then, the HRE module is introduced to generate the multi-level retrace embeddings of hierarchical semantic information. Next, in the SET module, the superposed hypergraph is constructed based on the local coherent and global entangled correlations. Finally, the hypergraph convolution and memory bank under the superposed structure are used to smooth and distill for feature generalization between seen and unseen categories.
+
+# 4.2 Hierarchical Retrace Embedding
+
+To obtain fully multi-level embeddings of objects based on hierarchical categories, the HRE module is designed here. Specifically, the HRE utilizes two hierarchical auto-encoders in series as shown in Figure 2. The multi-modal auto-encoder  $\mathcal{A}_m$  encodes the multi-modal basic features of 3D objects
+
+to get the unified embeddings. The retrace auto-encoder  $\mathcal{A}_r$  encodes the coarse label aligning with the unified embeddings to get retrace embeddings.
+
+Given  $N$  3D objects  $\{o_i\}_{i = 1}^N$  and feature extractors  $\{\mathcal{F}^k\}_{k = 1}^M$ , the basic feature of  $M$  modalities  $\{\mathbf{F}^k\}_{k = 1}^M = \mathcal{F}^k\big(\{o_i\}_{i = 1}^N\big)$  can be generated, where  $\mathbf{F}^k\in \mathbb{R}^{N\times d_f}$ . As shown in Fig. 2, we first compress the input basic features from  $M$  modalities into the latent modal-invariant space and generate the unified embeddings from multi-modal auto-encoder  $\mathcal{A}_m$ . Specifically, for the basic features  $f_{i}^{k}$  of object  $o_i$ , the unified embedding of it can be denoted as  $c_{i} = \mathcal{T}\left(\mathcal{A}_{m}(\{f_{i}^{k}\}_{k = 1}^{M})\right)$  where  $\mathcal{X}$  denotes the aggregation function among modalities and  $c_{i}\in \mathbb{R}^{d_{c}}$ .
+
+After getting the unified embedding  $c_{i}$ , the retrace auto-encoder  $\mathcal{A}_r$  takes the coarse category label  $y_i^c$  into retrace encoding  $e_i$ . As shown in Fig. 2, then  $\mathcal{A}_r$  compresses the unified embedding  $c_{i}$  aligned with  $e_i$  into the retract space  $\mathbb{S}_r$  and does the reverse reconstruction to the mixed space  $\mathbb{S}_x$ , which can be defined as follows:  $\Psi \coloneqq \mathbb{S}_x\to \mathbb{S}_f$  and  $\Phi \coloneqq \mathbb{S}_f\to \mathbb{S}_x$ , where  $\Psi (\cdot)$  is the encoder that maps the unified embedding  $c_{i}$  aligned with  $e_i$  into the retract space  $\mathbb{S}_r$ , the retrace embedding  $r_i$  can be generated by as  $r_i = \Psi (c_i + e_i)$ ,  $r_i\in \mathbb{R}^{d_r}$ .  $\Phi (\cdot)$  is the decoder that maps the retrace embedding to reconstruction feature  $\hat{m}_i = \Phi (r_i)$ ,  $\hat{m}_i\in \mathbb{R}^{d_c}$ .
+
+Through the HRE stage, we got the unified embedding  $c_{i}$ , retrace embedding  $r_{i}$ , and mixed feature  $\hat{m}_{i}$  for each 3D object.
+
+# 4.3 Structured Equilibrium Tuning
+
+To endow the unified and retrace embeddings of 3D objects with the ability to generalize to unseen categories, we introduce the SET module as shown in Figure 2. Specifically, the Superposed Hypergraph structure is employed to capture the local coherent and global entangled correlations under the constraint of hierarchical category information. Then, hypergraph convolution is employed to utilize the collaborative high-order correlation under the guidance of this superposed Hypergraph. Finally, we use the hypergraph convolution and memory bank to generalize the structure-aware knowledge to generate unbiased features for unseen categories.
+
+# 4.3.1 Superposed Hypergraph
+
+Despite the lack of fine labels in objects from unseen categories, the implicit information within these categories can enhance the generalization capabilities of 3D object embeddings, especially when prompted by coarse labels. To establish the high-order correlations between 3D objects from both seen and unseen categories under common general space, we design a superposed hypergraph structure. The hypergraph in this paper can be represented as  $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$ , where  $\mathcal{V}$  and  $\mathcal{E}$  are the vertex set and the hyperedge set, respectively.
+
+We construct the vertices in our superposed hypergraph by the multi-level retrace embeddings. Specifically, we combine the unified embedding  $c_{i}$  and retrace embedding  $r_i$  to generate the heterogeneous retrace vertices  $v_{i}\in \mathbb{R}^{N\times d_{c}}$ , which can be defined as follows:
+
+$$
+v _ {i} = \lambda c _ {i} + (1 - \lambda) r _ {i} \tag {2}
+$$
+
+where  $\lambda$  is the hyper-parameter to trade-off between embeddings of different levels.
+
+Distinct from the traditional hypergraph [8], we construct a superposed hypergraph with two types of conditional hyperedges: coherent hyperedges and entangled hyperedges.
+
+**Coherent Hyperedge.** The coherent hyperedges capture the local coherent correlations in the coarse category space. We define the coherent hyperedge as  $e^{c} \in \mathcal{E}_{c}$  under the condition of coarse labels:
+
+$$
+\mathcal {E} _ {c} = \left\{\mathcal {S} _ {v} \left(y ^ {c}\right) \mid y ^ {c} \in \mathcal {Y} ^ {c} \right\}, \tag {3}
+$$
+
+where  $S_{v}(y^{c})$  denotes the vertex subset that shares the same coarse label  $y^{c}$  and  $\mathcal{Y}^c$  denotes the coarse label space of 3D objects.
+
+Entangled Hyperedge. The entangled hyperedges model the global collaborative correlations among objects between seen and unseen categories. For each vertex, we define its entangled hyperedge as the set of nearest  $K - 1$  neighboring vertices. Specifically,
+
+$$
+\mathcal {E} _ {t} = \left\{\mathcal {D} _ {\mathrm {K N N} _ {k}} (v) \mid v \in \mathcal {V} \right\} \tag {4}
+$$
+
+where  $\mathcal{D}_{\mathrm{KNN}_k}(v)$  denotes the k-nearest neighbors in the feature space of vertex  $v$ .
+
+In this way, we construct  $Y^{c}$  coherent hyperedges and  $N$  entangled hyperedges, where  $Y^{c}$  is the number of coarse categories and  $N$  is the number of vertex in the superposed hypergraph. Finally, we combine the two sets of hyperedges to obtain a complete superposed hypergraph for structure-aware tuning.
+
+$$
+\mathcal {E} = \mathcal {E} _ {c} \cup \mathcal {E} _ {t} \tag {5}
+$$
+
+where  $\mathcal{E}_c$  and  $\mathcal{E}_t$  are the set of coherent and entangled hyperedges, respectively.
+
+# 4.3.2 Structure-Aware Tuning
+
+To get better generalization on unseen categories, hypergraph convolution and memory bank are adopted here for feature smoothing and distillation.
+
+For the convenience of computation, the hypergraph can be represented by the incidence matrix  $\mathbf{H} \in \{0,1\}^{|\mathcal{V}| \times |\mathcal{E}|}$ , where the  $i$ -th hyperedge is the  $i$ -th column of  $\mathbf{H}$ , and  $\mathbf{H}(v,e) = 1$  if the hyperedge  $e$  contains the vertex  $v$ .  $\mathbf{W} \in \mathbb{R}^{|\mathcal{E}| \times |\mathcal{E}|}$  is a diagonal matrix, where  $\mathbf{W}_{i,i}$  denotes the weight of the  $i$ -th hyperedge.
+
+To learn the conditional embeddings  $\tilde{\mathbf{V}}\in \mathbb{R}^{N\times d_c}$  from multi-level embedding  $\mathbf{V}\in \mathbb{R}^{N\times d_c}$  under the guidance of hypergraph, the hypergraph convolution (HGNNCov [8]) can be represented as:
+
+$$
+\tilde {\mathbf {V}} = \sigma \left(\mathbf {D} _ {v} ^ {- \frac {1}{2}} \mathbf {H} \mathbf {W} \mathbf {D} _ {e} ^ {- 1} \mathbf {H} ^ {\top} \mathbf {D} _ {v} ^ {- \frac {1}{2}} \mathbf {V} \boldsymbol {\Theta}\right), \tag {6}
+$$
+
+where  $\mathbf{D}_v$  and  $\mathbf{D}_e$  are the diagonal degree matrices for vertex and hyperedge, respectively.  $\Theta \in \mathbb{R}^{d_u\times d_u}$  is the trainable parameter for the HGNNConv layer [8].
+
+To increase the generalization ability of the SET module, we construct a memory bank  $\mathcal{M}$  that contains  $L$  invariant memory anchors for conditional embedding  $\tilde{v}_i$  of the 3D object  $o_i$ , we compute the activation score for each memory anchor in the memory bank by  $t_{ij} = \mathcal{D}_m(\tilde{v}_i,a_j)$ , where  $a_j$  denotes the anchor and  $D_{m}(\cdot ,\cdot)$  denotes the distance metric function. We rebuild the aligned multi-level embedding of each object by  $z_{i} = \sum_{j = 1}^{L}t_{i,j}^{n}a_{j},z_{i}\in \mathbb{R}^{d_{c}}$ , where  $t_{ij}^{n}$  denotes the normalized values of activation score.
+
+# 4.4 Training Objective
+
+Loss Function for the HRE. To get a better representation of multi-level label information, we first use the Homology Loss  $\mathcal{L}_{homo}$  and Bi-reconstruction Loss  $\mathcal{L}_{br}$  followed [7] for  $\mathcal{A}_m$  to leverage the collaborative information across modalities, then we adopt the Retrace Cross-Entropy loss  $\mathcal{L}_{ce}$  to guide the retrace embedding of coarse category information, which can be defined as follows:
+
+$$
+\mathcal {L} _ {c e} = - \sum_ {k = 1} ^ {Y ^ {c}} y _ {i, k} ^ {c} \log (p _ {i, k}), \tag {7}
+$$
+
+where  $p_{i,k} = \frac{e^{\hat{m}_{i,k}}}{\sum_{k=1}^{Y^c} e^{\hat{m}_{i,k}}}$  is the predicted probability score of the 3D object  $o_i$  in  $k$ -th coarse category for the reconstructed mixed feature  $\hat{m}_i$ .  $y_{i,k}^c$  is the  $k$ -th value of the one-hot encoded ground truth coarse label of  $o_i$ , and  $Y^c$  is the number of the coarse categories.
+
+In the hierarchical retrace embedding stage, the overall loss function is given:
+
+$$
+\mathcal {L} _ {H R E} = \mu \left(\mathcal {L} _ {h o m o} + \mathcal {L} _ {b r}\right) + (1 - \mu) \mathcal {L} _ {c e} \tag {8}
+$$
+
+where  $\mu$  is the hyper-parameter to trade-off between the loss of multi-modal and multi-level representations.
+
+Loss Function for the SET. To train the hypergraph convolution and learnable memory anchors, we adopt Memory Reconstruction Loss  $\mathcal{L}_{mr}$  and the Cross-entropy Loss  $\mathcal{L}_{ce}$ :
+
+$$
+\mathcal {L} _ {m r} = \left\| \tilde {u} _ {i} - z _ {i} \right\| _ {2}, \tag {9}
+$$
+
+$$
+\mathcal {L} _ {c e} = - \sum_ {k = 1} ^ {Y} \left(y _ {i, k} \log \left(\tilde {p} _ {i, k}\right) + y _ {i, k} \log \left(p _ {i, k}\right)\right), \tag {10}
+$$
+
+Table 1: Comparisons of retrieval performance on SO-ESB and SO-NTU dataset.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">SO-ESB</td><td colspan="4">SO-NTU</td></tr><tr><td>mAP↑</td><td>Recall↑</td><td>NDCG↑</td><td>ANMRR↓</td><td>mAP↑</td><td>Recall↑</td><td>NDCG↑</td><td>ANMRR↓</td></tr><tr><td>SDML</td><td>0.4947</td><td>0.8027</td><td>0.1858</td><td>0.5430</td><td>0.4384</td><td>0.7009</td><td>0.1937</td><td>0.5764</td></tr><tr><td>CMCL</td><td>0.4990</td><td>0.8154</td><td>0.1880</td><td>0.5457</td><td>0.4440</td><td>0.7053</td><td>0.1946</td><td>0.5721</td></tr><tr><td>MMSAE</td><td>0.5036</td><td>0.8503</td><td>0.1931</td><td>0.5523</td><td>0.4454</td><td>0.7046</td><td>0.1935</td><td>0.5745</td></tr><tr><td>TranGCN</td><td>0.5063</td><td>0.9011</td><td>0.1968</td><td>0.5408</td><td>0.4548</td><td>0.7121</td><td>0.1961</td><td>0.5624</td></tr><tr><td>C2AE</td><td>0.4809</td><td>0.7863</td><td>0.1824</td><td>0.5501</td><td>0.4303</td><td>0.6987</td><td>0.1915</td><td>0.5828</td></tr><tr><td>HGM2R</td><td>0.5049</td><td>0.8831</td><td>0.1939</td><td>0.5551</td><td>0.4821</td><td>0.7364</td><td>0.2026</td><td>0.5438</td></tr><tr><td>Ours</td><td>0.5756</td><td>0.9346</td><td>0.2045</td><td>0.4874</td><td>0.5678</td><td>0.8116</td><td>0.2251</td><td>0.4677</td></tr></table>
+
+![](images/c815e0c32dfc3a9a64d206f838fec60cc20877249d8b2a742ebb86843132a24f.jpg)  
+(a) PR-C on SO-ESB.
+
+![](images/e3938532e36adec7e92c885359240de0cf87d2de9cc4c301adfc9adf36189c4d.jpg)  
+(b) PR-C on SO-NTU.
+
+![](images/562ea95283481928192ac89f6860852a52c6870ddd9e64cdc6a4b55854357073.jpg)  
+Figure 3: The Precision-Recall Curves (PR-C) of the proposed method and compared methods on the four datasets, respectively.
+
+![](images/d65a795bec51b5c8960f9c4e0d57ffa4e918feba128d2d2f5217240c395961f4.jpg)  
+(c) PR-C on SO-MN40.  
+(d) PR-C on SO-ABO.
+
+where  $\tilde{p}_{i,k} = \frac{e^{\tilde{u}_{i,k}}}{\sum_{m=1}^{Y} e^{\tilde{u}_{i,m}}}$  and  $p_{i,k} = \frac{e^{z_{i,k}}}{\sum_{m=1}^{Y} e^{z_{i,m}}}$  is the predicted probability score of the 3D object  $o_i$  in  $k$ -th fine category for the multi-level embedding  $\tilde{v}_i$  and memory reconstruction embedding  $z_i$ .  $y_{i,k}$  is the  $k$ -th value of the one-hot encoded ground truth fine label of  $o_i$ , and  $Y$  is the number of the coarse categories.
+
+In the structured equilibrium tuning stage, the overall loss function is given by:
+
+$$
+\mathcal {L} _ {S E T} = \eta \mathcal {L} _ {m r} + (1 - \eta) \mathcal {L} _ {c e} \tag {11}
+$$
+
+where  $\eta$  is the hyper-parameter for trade-off between them.
+
+# 5 Experiments
+
+# 5.1 Datasets and Evaluation Metrics
+
+Datasets. We generate four semi-open 3DOR datasets, including SO-ESB, SO-NTU, SO-MN40, and SO-ABO, based on the public datasets ESB [13], NTU [5], ModelNet40 [35], and ABO [6], respectively. We add coarse category labels for each object based on the basic geometric shapes such as solid of revolution, rectangular-cubic, etc. Also, we remove some objects that are difficult to categorize based on their shapes. Then, we split the fine categories into seen categories for training and unseen categories for testing, the training and testing sets share the same coarse label space according to the semi-open environment setting. Each object has three modalities including multi-view, voxel, and point cloud. Specifically, the detailed descriptions of dataset generation and setting are shown in Appendix B.
+
+Evaluation Protocols. As for the evaluation criteria, We employ commonly used retrieval metrics for comparison, including Mean Average Precision (mAP), Recall, Normalized Discounted Cumulative Gain (NDCG), Average Normalized Modified Retrieval Rank (ANMRR), and the Precision-Recall Curve (PR-C). For the mAP, Recall, and NDCG metrics, the higher scores is better. For the ANMRR metric, the lower score is better.
+
+Table 2: Comparisons of retrieval performance on SO-MN40 and SO-ABO dataset.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">SO-MN40</td><td colspan="4">SO-ABO</td></tr><tr><td>mAP↑</td><td>Recall↑</td><td>NDCG↑</td><td>ANMRR↓</td><td>mAP↑</td><td>Recall↑</td><td>NDCG↑</td><td>ANMRR↓</td></tr><tr><td>SDML</td><td>0.5018</td><td>0.3241</td><td>0.6082</td><td>0.5106</td><td>0.4380</td><td>0.3425</td><td>0.4726</td><td>0.5564</td></tr><tr><td>CMCL</td><td>0.5086</td><td>0.3281</td><td>0.6128</td><td>0.5060</td><td>0.4520</td><td>0.3657</td><td>0.4816</td><td>0.5458</td></tr><tr><td>MMSAE</td><td>0.5189</td><td>0.3335</td><td>0.6226</td><td>0.4938</td><td>0.4783</td><td>0.3863</td><td>0.4929</td><td>0.5264</td></tr><tr><td>TranGCN</td><td>0.5188</td><td>0.3358</td><td>0.6131</td><td>0.4957</td><td>0.5175</td><td>0.3956</td><td>0.5127</td><td>0.4801</td></tr><tr><td>C2AE</td><td>0.4865</td><td>0.3152</td><td>0.5977</td><td>0.5231</td><td>0.4669</td><td>0.3674</td><td>0.4794</td><td>0.5313</td></tr><tr><td>HGM2R</td><td>0.5779</td><td>0.3698</td><td>0.6482</td><td>0.4407</td><td>0.6069</td><td>0.4675</td><td>0.5463</td><td>0.4154</td></tr><tr><td>Ours</td><td>0.6336</td><td>0.3993</td><td>0.6874</td><td>0.3972</td><td>0.6339</td><td>0.4793</td><td>0.5622</td><td>0.3836</td></tr></table>
+
+Table 3: Ablation studies on SO-ESB and SO-NTU dataset.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">SO-ESB</td><td colspan="4">SO-NTU</td></tr><tr><td>mAP↑</td><td>Recall↑</td><td>NDCG↑</td><td>ANMRR↓</td><td>mAP↑</td><td>Recall↑</td><td>NDCG↑</td><td>ANMRR↓</td></tr><tr><td>HRE w/o ReEnz</td><td>0.5159</td><td>0.9086</td><td>0.1953</td><td>0.5431</td><td>0.4913</td><td>0.7534</td><td>0.2053</td><td>0.5355</td></tr><tr><td>HRE w/o Lce</td><td>0.5133</td><td>0.8738</td><td>0.1934</td><td>0.5365</td><td>0.5161</td><td>0.7902</td><td>0.2162</td><td>0.5162</td></tr><tr><td>SET w/o Ec</td><td>0.5358</td><td>0.8957</td><td>0.1975</td><td>0.5184</td><td>0.5285</td><td>0.7898</td><td>0.2184</td><td>0.4986</td></tr><tr><td>GCN-based SET</td><td>0.5405</td><td>0.8999</td><td>0.2003</td><td>0.5192</td><td>0.5144</td><td>0.7703</td><td>0.2140</td><td>0.5138</td></tr><tr><td>MLP-based SET</td><td>0.5014</td><td>0.8483</td><td>0.1930</td><td>0.5476</td><td>0.4689</td><td>0.7304</td><td>0.2023</td><td>0.5561</td></tr><tr><td>HRE+SET</td><td>0.5756</td><td>0.9346</td><td>0.2045</td><td>0.4874</td><td>0.5678</td><td>0.8116</td><td>0.2251</td><td>0.4677</td></tr></table>
+
+# 5.2 Experimental Settings
+
+Implemental Details. In our experiments, we choose three modes of multi-view (12 views), point cloud (1024 points), and voxel (32 dimensions) as the representation of 3D objects. The basic features for framework input are extracted by MVCNN [30], PointNet [25], and 3D ShapeNets [35], respectively. The HRE and SET modules are trained separately with 40 and 120 epochs. The SGD optimizers are used for both two modules with learning rates of 0.1 and 0.001, respectively. As for the hyper-parameters in HERT, we set  $\lambda = 0.5$ ,  $\mu = 0.8$ , and  $\eta = 0.9$ . Detailed implemental settings for our framework are provided in Appendix C.
+
+Compared Methods. Under this semi-open setting, since there is no 3D object retrieval method designed specifically for this multi-level settings, we choose the current state-of-the-art methods of close-set 3D retrieval (SDML [11], CMCL [14], MMSAE [34]), close-set multi-label retrieval(Trang-GCN [15]), and open-set 3D learning (C2AE [21], HGM $^2$ R [7]). For each method, we add a multi-label learning mechanism [15] on their basis for comparison. We provide more implemented details of compared methods in Appendix D.
+
+# 5.3 Comparison with the State-of-the-Arts
+
+To evaluate the effectiveness of the proposed HERT framework, we conduct experiments on SO-ESB, SO-NTU, SO-MN40, and SO-ABO datasets. The comparison of quantitative results is presented in Table 1 and Table 2, respectively. Results demonstrate that the proposed HERT outperforms the state-of-the-art methods on all four datasets. Especially on the SO-NTU and SO-MN40 datasets, our method achieves mAP of  $0.5678 / 0.6336$  with about  $17.7\% /9.6\%$  improvements compared with the second-best method, and achieves Recall of  $0.8116 / 0.3993$  with about  $10.2\% /7.9\%$  improvements compared with the second-best method. Besides, we also provide the qualitative results through precision-recall curves as shown in Figure 3, in which the larger area below the curve indicates better performance.
+
+The better results indicate that the proposed HERT framework has the capability to understand and generalize unseen fine categories under the guidance of coarse labels. The proposed HRE and SET modules can fully leverage the hierarchical category information into multi-level retrace embedding and generalize them to unseen categories. This approach better captures the hierarchical semantic correlations in the wild and provides a practical framework for the representation learning of multi-label tasks in semi-open environments. We provide more visualized results in Appendix E.
+
+Table 4: Ablation studies on SO-MN40 and SO-ABO dataset.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">SO-MN40</td><td colspan="4">SO-ABO</td></tr><tr><td>mAP↑</td><td>Recall↑</td><td>NDCG↑</td><td>ANMRR↓</td><td>mAP↑</td><td>Recall↑</td><td>NDCG↑</td><td>ANMRR↓</td></tr><tr><td>HRE w/o ReEnz</td><td>0.5791</td><td>0.3710</td><td>0.6479</td><td>0.4410</td><td>0.6055</td><td>0.4523</td><td>0.5535</td><td>0.4062</td></tr><tr><td>HRE w/o Lce</td><td>0.5967</td><td>0.3783</td><td>0.6756</td><td>0.4309</td><td>0.5885</td><td>0.4269</td><td>0.5413</td><td>0.4230</td></tr><tr><td>SET w/o Ec</td><td>0.5913</td><td>0.3757</td><td>0.6669</td><td>0.4347</td><td>0.6006</td><td>0.4263</td><td>0.5494</td><td>0.4132</td></tr><tr><td>GCN-based SET</td><td>0.5602</td><td>0.3573</td><td>0.6410</td><td>0.4628</td><td>0.5686</td><td>0.4253</td><td>0.5314</td><td>0.4415</td></tr><tr><td>MLP-based SET</td><td>0.5088</td><td>0.3290</td><td>0.6149</td><td>0.5073</td><td>0.4880</td><td>0.3787</td><td>0.5023</td><td>0.5159</td></tr><tr><td>HRE+SET</td><td>0.6336</td><td>0.3993</td><td>0.6874</td><td>0.3972</td><td>0.6339</td><td>0.4793</td><td>0.5622</td><td>0.3836</td></tr></table>
+
+![](images/b5a04a551fcc4175dce3722e6113e07eb65e949f10de634460408d93022693c2.jpg)  
+(a) PR-C on SO-ESB.
+
+![](images/f146757de903e4d6a8b80aa30e7829640a5d89372c946c44b0fa8c195b89a728.jpg)  
+(b) PR-C on SO-NTU.
+
+![](images/2c9508c04330ddad135b3c8f98dbe40b4a3babbd1dd01bd7983af42a22c452f7.jpg)  
+Figure 4: The Precision-Recall Curves (PR-C) of the ablation studies on four datasets, respectively.
+
+![](images/3bc02295c7e5216710db6ba64a7969b168da3b64a5b35e78fe74afbc37a92509.jpg)  
+(c) PR-C on SO-MN40.  
+(d) PR-C on SO-ABO.
+
+# 5.4 Ablation Study
+
+We conduct ablation experiments on each module of HERT to demonstrate their effectiveness. First, we remove the retrace encoding  $e_i$  (HRE w/o ReEnz) and cross-entropy loss in the HRE module (HRE w/o  $\mathcal{L}_{ce}$ ), where "w/o" denotes "without". This is equivalent to using a naive hierarchical embedding approach for coarse labels. As shown in Table 3, 4, and Figure 4, the proposed retrace embedding approach in the HRE module achieves an mAP improvement of  $9.4\% / 6.2\%$  in SO-MN40 dataset and  $7.8\% / 4.7\%$  in SO-ABO dataset. These results effectively demonstrate the effectiveness of the HRE module for hierarchical categories.
+
+As for the SET module, we remove the proposed coherent hyperedges (SET w/o  $\mathcal{E}_c$ ) for comparison, also we replace the hypergraph convolution with Graph Convolutional Networks (GCN-based SET) and Multilayer Perceptron (MLP-based SET). Ablation results in Table 3, 4, and Figure 4 show that the proposed SET outperforms all other structure learning methods, and the combination of the HRE and the SET yields the best performance. These results demonstrate the proposed SET can effectively utilize the semi-supervised correlations among objects.
+
+# 6 Conclusion
+
+In this paper, we introduce a more practical Semi-Open Environment setting for open-set 3D object retrieval with hierarchical labels, in which the training and testing set share a partial label space for coarse categories but are completely disjoint from fine categories. We propose the Hypergraph-Based Hierarchical Equilibrium Representation (HERT) framework for semi-open 3D object retrieval. Specifically, to overcome the global disequilibrium of unseen categories, we propose the Hierarchical Retrace Embedding (HRE) module to fully leverage the multi-level category information. Besides, we perform the Structured Equilibrium Tuning (SET) module to tackle the feature overlap and class confusion problem. This module utilizes more equilibrial correlations among objects and generalizes to unseen categories, by constructing a superposed hypergraph based on the local coherent and global entangled correlations. Furthermore, we construct four 3D object datasets with multi-level category labels for semi-open 3DOR tasks, i.e., SO-ESB, SO-NTU, SO-MN40, and SO-ABO. Results demonstrate that the proposed method can effectively generate and generalize the hierarchical embeddings of 3D objects in semi-open environments. However, due to dataset limitations, we are currently unable to verify the balanced representation effect on more than three levels of labels, which is one of our future research directions. We believe this paper can provide new insights for future research in more practical scenarios of open-set learning.
+
+# 7 Acknowledgement
+
+This work was supported by Beijing Natural Science Foundation (No. L242167), CCF-Tencent Open Research Fund, and Jiangxi Provincial Natural Science Foundation (20224ACB218002).
+
+# References
+
+[1] Ferran Argelaguet and Carlos Andujar. A survey of 3d object selection techniques for virtual environments. Computers & Graphics, 37(3):121-136, 2013.  
+[2] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 20(10):3782-3795, 2019.  
+[3] Song Bai, Peng Tang, Philip HS Torr, and Longin Jan Latecki. Re-ranking via metric fusion for object retrieval and person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 740–749, 2019.  
+[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1563-1572, 2016.  
+[5] Ding-Yun Chen, Xiao-Pei Tian, Yu-Te Shen, and Ming Ouhyoung. On visual similarity based 3d model retrieval. Computer graphics forum, 22(3):223-232, 2003.  
+[6] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022.  
+[7] Yifan Feng, Shuyi Ji, Yu-Shen Liu, Shaoyi Du, Qionghai Dai, and Yue Gao. Hypergraph-based multimodal representation for open-set 3d object retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2206-2223, 2023.  
+[8] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3181-3199, 2022.  
+[9] ZongYuan Ge, Sergey Demyanov, Zetao Chen, and Rahul Garnavi. Generative openmax for multi-class open set classification. arXiv preprint arXiv:1707.07418, 2017.  
+[10] Alexander Grabner, Peter M Roth, and Vincent Lepetit. 3d pose estimation and 3d model retrieval for objects in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3022-3031, 2018.  
+[11] Peng Hu, Liangli Zhen, Dezhong Peng, and Pei Liu. Scalable deep multimodal learning for cross-modal retrieval. In Annual Conference of the Association for Computing Machinery Special Interest Group in Information Retrieval, pages 635-644, 2019.  
+[12] Lalit P Jain, Walter J Scheirer, and Terrance E Boult. Multi-class open set recognition using probability of inclusion. In European Conference on Computer Vision, pages 393-409. Springer, 2014.  
+[13] Subramaniam Jayanti, Yagnanarayanan Kalyanaraman, Natraj Iyer, and Karthik Ramani. Developing an engineering shape benchmark for cad models. Computer-Aided Design, 38(9):939-953, 2006.  
+[14] Longlong Jing, Elahe Vahdani, Jiaxing Tan, and Yingli Tian. Cross-modal center loss for 3d cross-modal retrieval. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3142-3151, 2021.  
+[15] Ying Li, Chunming Guan, Rui Cai, Ye Erwan, Ding Yuxiang, and Jiaquan Gao. Tran-gen: Multi-label pattern image retrieval via transformer driven graph convolutional network. In ACM International Conference on Multimedia, pages 6301-6310, 2023.  
+[16] Ying Li, Hongwei Zhou, Yeyu Yin, and Jiaquan Gao. Multi-label pattern image retrieval via attention mechanism driven graph convolutional network. In ACM International Conference on Multimedia, pages 300-308, 2021.  
+[17] Qi Liang, Mengmeng Xiao, and Dan Song. 3d shape recognition based on multi-modal information fusion. Multimedia Tools and Applications, 80:16173-16184, 2021.
+
+[18] Weiwei Liu, Donna Xu, Ivor W Tsang, and Wenjie Zhang. Metric learning for multi-output tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):408-422, 2018.  
+[19] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In European Conference on Computer Vision, pages 613–628, 2018.  
+[20] Weizhi Nie, Qi Liang, An-An Liu, Zhendong Mao, and Yangyang Li. Mmjn: Multi-modal joint networks for 3d shape recognition. In ACM International Conference on Multimedia, pages 908–916, 2019.  
+[21] Poojan Oza and Vishal M Patel. C2ae: Class conditioned auto-encoder for open-set recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2307–2316, 2019.  
+[22] Panagiotis Papadakis, Ioannis Pratikakis, Theoharis Theoharis, and Stavros Perantonis. Panorama: A 3d shape descriptor based on panoramic views for unsupervised 3d object retrieval. International Journal of Computer Vision, 89:177-192, 2010.  
+[23] Jitendra Parmar, Satyendra Chouhan, Vaskar Raychoudhury, and Santosh Rathore. Open-world machine learning: applications, challenges, and opportunities. ACM Computing Surveys, 55(10):1-37, 2023.  
+[24] Yashoteja Prabhu, Anil Kag, Shrutendra Harsola, Rahul Agrawal, and Manik Varma. Parabel: Partitioned label trees for extreme classification with application to dynamic search advertising. In The International Conference of World Wide Web, pages 993-1002, 2018.  
+[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017.  
+[26] Ethan M Rudd, Lalit P Jain, Walter J Scheirer, and Terrance E Boult. The extreme value machine. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(3):762-768, 2017.  
+[27] Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(7):1757-1772, 2012.  
+[28] Walter J Scheirer, Lalit P Jain, and Terrance E Boult. Probability models for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11):2317-2324, 2014.  
+[29] Lei Shu, Hu Xu, and Bing Liu. Doc: Deep open classification of text documents. arXiv preprint arXiv:1709.08716, 2017.  
+[30] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 945-953, 2015.  
+[31] Jian Wang, Fan Li, Xuchong Zhang, and Hongbin Sun. Adversarial obstacle generation against lidar-based 3d object detection. IEEE Transactions on Multimedia, 2023.  
+[32] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Open-set fine-grained retrieval via prompting vision-language evaluator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19381-19391, 2023.  
+[33] Dayan Wu, Zheng Lin, Bo Li, Mingzhen Ye, and Weiping Wang. Deep supervised hashing for multi-label and large-scale image retrieval. In International Conference on Multimedia Retrieval, pages 150-158, 2017.  
+[34] Yiling Wu, Shuhui Wang, and Qingming Huang. Multi-modal semantic autoencoder for cross-modal retrieval. Neurocomputing, 331:165-175, 2019.  
+[35] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1912-1920, 2015.  
+[36] Zhi Xiong, Dayan Wu, Wen Gu, Haisu Zhang, Bo Li, and Weiping Wang. Deep discrete attention guided hashing for face image retrieval. In International Conference on Multimedia Retrieval, pages 136-144, 2020.  
+[37] Yang Xu, Yifan Feng, and Lin Bie. Triadic elastic structure representation for open-set incremental 3d object retrieval. In ACM International Conference on Multimedia Retrieval, pages 20-28, 2024.
+
+![](images/05c97db5b14a78df31468edd874dd9dc39e8a3c59b2080f074a9c6ce9fec20d1.jpg)  
+Figure 5: Visualizations of the hierarchical labels and multi-modal representations of 3D objects in the SO-MN40 datasets.
+
+[38] Yang Xu, Yifan Feng, and Yue Gao. Negative prompt driven complementary parallel representation for open-world 3d object retrieval. In International Joint Conference on Artificial Intelligence, pages 1498-1506, 2024.  
+[39] Yang Xu, Yifan Feng, and Yu Jiang. Structure-aware residual-center representation for self-supervised open-set 3d cross-modal retrieval. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2024.  
+[40] Ian EH Yen, Xiangru Huang, Wei Dai, Pradeep Ravikumar, Inderjit Dhillon, and Eric Xing. Ppdsparse: A parallel primal-dual sparse method for extreme classification. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 545-553, 2017.  
+[41] Ryota Yoshihashi, Wen Shao, Rei Kawakami, Shaodi You, Makoto Iida, and Takeshi Naemura. Classification-reconstruction learning for open-set recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4016-4025, 2019.  
+[42] Haoxuan You, Yifan Feng, Rongrong Ji, and Yue Gao. Pynet: A joint convolutional network of point cloud and multi-view for 3d shape recognition. In ACM International Conference on Multimedia, pages 1310-1318, 2018.  
+[43] Haoxuan You, Yifan Feng, Xibin Zhao, Changqing Zou, Rongrong Ji, and Yue Gao. Pvrnet: Point-view relation neural network for 3d shape recognition. In AAAI Conference on Artificial Intelligence, pages 9119-9126, 2019.  
+[44] Hsiang-Fu Yu, Prateek Jain, Purushottam Kar, and Inderjit Dhillon. Large-scale multi-label learning with missing labels. In International Conference on Machine Learning, pages 593-601. PMLR, 2014.  
+[45] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Learning placeholders for open-set recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2021.  
+[46] Zhi-Hua Zhou. Open-environment machine learning. National Science Review, 9(8):nwac123, 2022.
+
+# A Semi-Open Environment
+
+In practical scenarios, the categories of 3D objects can be labeled from various perspectives such as appearance and functionality, or from different levels of granularity, which means hierarchical labels. For example, consider a 3D object, which can simultaneously be labeled as "sphere" and "blue" at the basic geometric and color level. Moving up the semantic hierarchy, the same object may be labeled as "transportation", "car", and "red" to indicate
+
+Table 5: The statistics of the semi-open 3DOR datasets.  
+
+<table><tr><td></td><td></td><td>SO-ESB</td><td>SO-NTU</td><td>SO-MN40</td><td>SO-ABO</td></tr><tr><td rowspan="4">Categories</td><td>Coarse</td><td>3</td><td>3</td><td>3</td><td>3</td></tr><tr><td>Fine</td><td>41</td><td>67</td><td>40</td><td>21</td></tr><tr><td>Seen</td><td>17</td><td>13</td><td>8</td><td>4</td></tr><tr><td>Unseen</td><td>24</td><td>54</td><td>32</td><td>17</td></tr><tr><td rowspan="4">Number</td><td>Training</td><td>98</td><td>378</td><td>2821</td><td>1082</td></tr><tr><td>Retrieval</td><td>457</td><td>1232</td><td>7591</td><td>4432</td></tr><tr><td>Query</td><td>96</td><td>216</td><td>128</td><td>68</td></tr><tr><td>Target</td><td>361</td><td>1016</td><td>7463</td><td>4364</td></tr></table>
+
+its coarse semantic, fine semantic, and color. Further up the hierarchy, labels could include descriptors related to the object's specific function or detailed features. For instance, a 3D architectural model might be labeled as "residential", "multi-story building", and "with a garage"to convey its functional and structural attributes. These hierarchical labels not only enhance the descriptive richness of 3D object representations but also play a crucial role in applications spanning computer graphics, virtual reality, machine learning, and game development. They enable a more comprehensive understanding of an object's characteristics and context, ultimately enhancing its usability and interpretability across diverse domains.
+
+Most open-set learning methods are based on the assumption that the labels of object categories are at a single level. This setting implies that existing methods consider category visibility only between the training and testing sets at a single layer, ignoring the distribution overlap among categories at other layers. Besides, the existing strict setting of the single-level open-set also brings more deviation bias to the feature space of unseen categories. In this paper, we do not follow the strict assumption of the existing open-set approach regarding the disjoint distribution between training and testing sets, instead leveraging the different overlaps across multi-level categories to enhance representation and retrieval performance. We term this scenario, where the training and testing set share a partial label space of coarse categories but are completely disjoint from fine categories, as Semi-Open Environment.
+
+# B Semi-Open Dataset Generation
+
+Considering that there is no existing dataset for semi-open 3DOR task, we generate four semi-open 3DOR datasets, including SO-ESB, SO-NTU, SO-MN40, and SO-ABO, based on the public datasets ESB [13], NTU [5], ModelNet40 [35], and ABO [6], respectively. We add coarse category labels for each object according to the basic geometric shape and remove some objects that are difficult to categorize based on their shapes. Thus, the number of 3D objects in the released datasets is smaller than that of the original datasets. Specifically, the coarse labels for the four datasets are:
+
+- SO-ESB: Flat-thin wall components, Rectangular-cubic prism, Solids of revolution.  
+- SO-NTU: Rectangular-cubic prism, Solids of revolution, Miscellaneous shape.  
+- SO-MN40: Rectangular-cubic prism, Solids of revolution, Miscellaneous shape.  
+- SO-ABO: Rectangular-cubic prism, Solids of revolution, Miscellaneous shape.
+
+Then, we split the fine categories into seen categories for training and unseen categories for testing, the training and testing sets share the same coarse label space according to the semi-open environment setting. The statics of the four semi-open 3DOR datasets are shown in Table 5. Each object has three modalities including multi-view, voxel, and point cloud. The multi-view data are rendered by Blender  $3.0^{2}$  to get 12 images for each object. The point number for point cloud is 1024, and the dimension for voxel data is  $32 \times 32 \times 32$ . Point cloud and voxel data are sampled by Open3D  $0.13.0^{3}$ . The illustration of hierarchical labels and multiple modalities of the semi-open 3DOR datasets are shown in Figure 5.
+
+# C Method Details
+
+The overall framework of HERT is composed of the Hierarchical Retrace Embedding (HRE) and the Conditional Structure Learning (SET) modules. Given 3D objects represented by multiple modalities, common-used
+
+Algorithm 1 Training the HRE module
+
+Input: Basic features  $\{r^k\}_{k = 1}^M$  and their coarse labels  $\{y_i^c\}_{i = 1}^N$  of  $N$  instances  $\{o_i\}_{i = 1}^N$
+
+Parameter:  $\mu = 0.5$
+
+Output 1: Unified embeddings  $\{c_i\}_{i = 1}^N$
+
+Output 2: Retrace embeddings  $\{f_i\}_{i=1}^N$ .
+
+1: Let epoch  $= 0$  
+2: Initialize multi-modal auto-encoder  $\mathcal{A}_m$ ;  
+3: Initialize aggregation function  $\mathcal{T}$  
+4: Initialize retrace auto-encoder  $\mathcal{A}_r = \{\Psi, \Phi\}$ ;  
+5: Encoding coarse labels as retrace encoding  $\{e_i\}_{i = 1}^N$  
+6: while epoch  $\leq 40$  do  
+7: Get unified embedding for each modality  $\{c_i^1, c_i^2, c_i^3\}_{i=1}^N = \mathcal{A}_m(\{f_i^1, f_i^2, f_i^3\}_{i=1}^N)$ .  
+8: Get unified embedding for each object  $\{c_i\}_{i = 1}^N = \mathcal{T}(\{c_i^1,c_i^2,c_i^3\}_{i = 1}^N)$  
+9: Calculate the Homology Loss  $\mathcal{L}_{homo} = \mathcal{L}_{homo}(\{c_i\}_{i=1}^N)$ .  
+10: Calculate the Bi-reconstruction Loss  $\mathcal{L}_{br} = \mathcal{L}_{br}(\{c_i\}_{i=1}^N)$ .  
+11: Get retrace embedding of each object  $\{r_i\}_{i = 1}^N = \Psi (\{c_i + e_i\}_{i = 1}^N)$  
+12: Get mixed feature of each object  $\{\hat{m}_i\}_{i = 1}^N = \Phi (\{f_i\}_{i = 1}^N)$  
+13: Calculate the Retrace Cross-Entropy Loss  $\mathcal{L}_{ce} = \mathcal{L}_{ce}(\{\hat{m}_i\}_{i=1}^N)$ .  
+14: Calculate loss for the HRE module  $\mathcal{L}_{HRE} = \mu (\mathcal{L}_{homo} + \mathcal{L}_{br}) + (1 - \mu)\mathcal{L}_{ce}$  
+15: if  $\mathcal{L}_{HRE}$  does not converge then  
+16: Update parameters of  $\mathcal{A}_m$ ,  $\mathcal{T}$ , and  $\mathcal{A}_r$  by  $\mathcal{L}_{HRE}$ .  
+17: epoch  $+ = 1$  
+18: else  
+19: Break.  
+20: end if  
+21: end while  
+22: return Unified embeddings  $\{c_i\}_{i = 1}^N$  , and retrace embeddings  $\{r_i\}_{i = 1}^N$
+
+backbones are used to extract the basic features for each modality. Specifically, the basic features of multi-view, point cloud, and voxel are extracted by the pre-trained MVCNN [30], PointNet [25], and 3D ShapeNets [35], respectively.
+
+The implemental details of the HRE module are provided in Algorithm 1, the HRE utilizes two hierarchical auto-encoders in series. The multi-modal auto-encoder  $\mathcal{A}_m$  encodes the multi-modal basic feature of 3D objects to get the unified embeddings. The retrace auto-encoder  $\mathcal{A}_a$  encodes the coarse label aligning with the unified embeddings to get retrace embeddings.
+
+As shown in Algorithm 2, we provide the implemental details of the SET module, the superposed hypergraph structure  $\mathcal{G}$  is employed to capture the local coherent and global entangled correlations under the constraint of hierarchical category information. Then, hypergraph convolution is employed to utilize the collaborative high-order correlation under the guidance of this superposed hypergraph. Finally, we use the hypergraph convolution (HGNNConv) and memory bank  $\mathcal{M}$  to generalize the structure-aware knowledge to generate unbiased features for unseen categories.
+
+Table 6: The hyper-parameters of the HERT framework.  
+
+<table><tr><td></td><td>HRE</td><td>SET</td></tr><tr><td>Optimizer</td><td>SGD</td><td>SGD</td></tr><tr><td>Learning Rate</td><td>0.1</td><td>0.001</td></tr><tr><td>Momentum</td><td>0.9</td><td>0.9</td></tr><tr><td>Weight Decay</td><td>0.1</td><td>0</td></tr><tr><td>LR Scheduler</td><td>Cosine Annealing</td><td>Cosine Annealing</td></tr><tr><td>Tmax</td><td>40</td><td>60</td></tr><tr><td>etamin</td><td>0.00001</td><td>0.00001</td></tr><tr><td>Max Epoches</td><td>40</td><td>120</td></tr></table>
+
+Our experiments were conducted on a Tesla V100-32G GPU and an Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz. Besides, We provide more hyper-parameters of HERT in Table 6. The hyper-parameters "k" in the SET module are set to 12, 10, 50, and 50 for SO-ESB, SO-NTU, SO-MN40, and SO-ABO, respectively.
+
+Algorithm 2 Training the SET module  
+Input: Unified embeddings  $\{c_i\}_{i = 1}^N$  ,retrace embeddings  $\{f_i\}_{i = 1}^N$  , and coarse labels  $\{y_i^c\}_{i = 1}^N$  of  $N$    
+instances  $\{o_i\}_{i = 1}^N$    
+Parameter:  $\lambda = 0.8$ $\eta = 0.9$    
+Output: Final embeddings  $\{\hat{v}_i\}_{i = 1}^N$    
+1: Let epoch  $= 0$  .   
+2: Initialize superposed hypergraph  $\mathcal{G} = \{\mathcal{V},\mathbf{W},\mathcal{E}\}$  .   
+3: Construct vertices  $\mathcal{V} = \{v_{ij}\}_{i = 1}^{N} = \{\lambda c_{i} + (1 - \lambda)f_{i}\}_{i = 1}^{N};$    
+4: Construct coherent hyperedge  $\mathcal{E}_c = \{\mathcal{S}_v(y^c)\mid y^c\in \mathcal{V}^c\}$  .   
+5: Construct entangled hyperedge  $\mathcal{E}_t = \{\mathcal{D}_{\mathrm{KNN}_k}(v)\mid v\in \mathcal{V}\}$  .   
+6: Construct hyperedges  $\mathcal{E} = \mathcal{E}_c\cup \mathcal{E}_t$  .   
+7: Initialize weight diagonal matrix W of G;   
+8: Calculate diagonal degree matrices  $\mathbf{D}_v$  and  $\mathbf{D}_e$  .   
+9: Calculate incidence matrix H of G;   
+10: Initialize HGNNConv parameters  $\Theta$  of G;   
+11: Construct memory bank M;   
+12: while epoch  $\leq 120$  do   
+13: Get  $\{\tilde{v}_i\}_{i = 1}^N = \sigma (\mathbf{D}_v^{-\frac{1}{2}}\mathbf{H}\mathbf{W}\mathbf{D}_e^{-1}\mathbf{H}^\top \mathbf{D}_v^{-\frac{1}{2}}\{v_i\}_{i = 1}^N\Theta)$  .   
+14: Calculate the Cross-Entropy Loss  $\mathcal{L}_{ce} = \mathcal{L}_{ce}(\{\tilde{v}_i\}_{i = 1}^N)$  .   
+15: Rebuild  $\{\tilde{v}_i\}_{i = 1}^N$  to get  $\{z_i\}_{i = 1}^N$  by M;   
+16: Calculate the Momery Reconstruction Loss  $\mathcal{L}_{mr} = \mathcal{L}_{mr}(\{\tilde{v}_i\}_{i = 1}^N,\{z_i\}_{i = 1}^N)$  .   
+17: Calculate loss for the SET module  $\mathcal{L}_{SET} = \eta \mathcal{L}_{mr} + (1 - \eta)\mathcal{L}_{ce}$    
+18: if LSET does not converge then   
+19: Update parameters of  $\Theta$  and M by LSET.   
+20: epoch  $+ = 1$    
+21: else   
+22: Break.   
+23: end if   
+24: end while   
+25: return Final embeddings  $\{\tilde{v}_i\}_{i = 1}^N$    
+26: Retrieval by  $\{\tilde{v}_i\}_{i = 1}^N$
+
+# D Comparison Implement
+
+Under this semi-open setting, since there is no 3D object retrieval method designed specifically for this multi-level settings, we choose the current state-of-the-art methods of close-set 3D retrieval (SDML [11], CMCL [14], MMSAE [34]), close-set multi-label retrieval(Tran-GCN [15]), and open-set 3D learning (C2AE [21], HGM $^2$ R [7]). For each method, we add a multi-label learning mechanism [15] on their basis for comparison. The coarse embeddings are supervised by the coarse labels by cross-entropy loss.
+
+- SDML [11]: SDML is a metric learning based method for multi-modal retrieval, which learns projection functions for different modalities independently. We implement the multi-label embedding by adding an auto-encoder after the encoder of the DSAE module.  
+- CMCL [14]: CMCL is a typical 3DOR network based on cross-modal center loss, which is designed to eliminate cross-modal discrepancy. We construct an auto-encoder and take the center representation of each object to generate the coarse embeddings.  
+- MMSAE [34]: MMSAE is an auto-encoder based method that compresses features from different modalities into a unified space. We take the semantic code vector of the MMSAE network as the input and generate coarse embeddings by another auto-encoder.  
+- TranGCN [15]: TranGCN is a multi-label retrieval method based on a graph convolutional network. We treat the fine and coarse labels as the two labels of supervision.  
+- C2AE [21]: C2AE is an open-set 3D object recognition method with a two-step structure, which is designed to identify objects of the out-of-the-distribution categories. We add an auto-encoder to generate the coarse embeddings after the first step.  
+- HGM²R [7]: HGM²R is a hypergraph-based 3DOR method specifically designed for the open-set environment, which assumes no distribution overlap between training and testing set. We construct an
+
+![](images/2f4dbcf6dd487546a271c20c2e2a2991fe6ab9ae4d4643295284b200108fb919.jpg)  
+(a) Retrace embeddings.
+
+![](images/5576d83a44e84793e6410432646296497fc3d4b5f0d3f54812f86e593b46dc05.jpg)  
+(b) Final embeddings.  
+Figure 6: The t-SNE visualization of the embeddings from unseen categories in the OS-MN40 dataset.
+
+auto-encoder for coarse embedding after the MM3DOE module, and we construct only the knn-based hyperedges in the SAIKL module.
+
+# E Performance Result
+
+The HRE module is designed to obtain multi-level embeddings of objects based on hierarchical categories. Therefore, we provide two visualized results of the retrace embeddings and final conditional embeddings of unseen categories in Figure 6(a) and Figure 6(b), respectively. These two t-SNE visualizations show that the proposed HRE and SET modules can fully leverage the hierarchical category information into multi-level retrace embedding and generalize them to unseen categories. This approach better captures the hierarchical semantic correlations in the wild and provides a practical framework for the representation learning of multi-label tasks towards a semi-open environment.
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes],
+
+Justification: We provide comprehensive motivations and detailed introductions of all the contributions in the abstract and introduction.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: We discussed the current limitations of this work in the conclusion.
+
+# 3. Theory Assumptions and Proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [Yes]
+
+Justification: We provided a formalized, step-by-step explanation of each procedure in the methodology.
+
+# 4. Experimental Result Reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+Answer: [Yes]
+
+Justification: We provide detailed experiment and reproduction settings in the experiments and appendix.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [No]
+
+Justification: We are in the preparation of the codes, models, and datasets, and they will be released at the earliest opportunity.
+
+# 6. Experimental Setting/Details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes]
+
+Justification: We provide all experimental settings and details settings in the experiments and appendix.
+
+# 7. Experiment Statistical Significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+Answer: [Yes]
+
+Justification: We explain the evaluation protocols of the experiments.
+
+# 8. Experiments Compute Resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: We provide the sufficient information on the computer resources in appendix.
+
+# 9. Code Of Ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: We have carefully read and adhered to the NeurIPS Code of Ethics.
+
+# 10. Broader Impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [Yes]
+
+Justification: We discussed the broader impacts of this work in the conclusion.
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA]
+
+Justification: This paper poses no such risks.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: We have obtained and adhered to the permissions for the datasets used and have cited them in the paper.
+
+# 13. New Assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [Yes]
+
+Justification: We will release the dataset and provide the corresponding documentation immediately upon the paper's acceptance.
+
+# 14. Crowdsourcing and Research with Human Subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification: This paper does not involve crowdsourcing nor research with human subjects.
+
+# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: the paper does not involve crowdsourcing nor research with human subjects.

2024/Semi-Open 3D Object Retrieval via Hierarchical Equilibrium on Hypergraph/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c248db845e22e15d2fae34450a129ee94a363436a3b4588b37a9f9ed64aa7c97
+size 770769

2024/Semi-Random Matrix Completion via Flow-Based Adaptive Reweighting/2dda777c-38bc-43d0-9234-0a6bbebc35cf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c04abc68d481144329f27bf295d62a0fc3d5fbc5a1472399bab8ca23447723f
+size 808400

2024/Semi-Random Matrix Completion via Flow-Based Adaptive Reweighting/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8845968a3a1cb831df2a4bc499182ddd0c24ce343df5c76a9295cdfde8ac46db
+size 1396804

2024/Semi-Supervised Sparse Gaussian Classification_ Provable Benefits of Unlabeled Data/cd0e33bc-dac7-45d8-aa8d-258dec283cd3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7aa3b857959dabe8194f966ecd9e6610c3a120905cfcf02ed001f601bd8208b4
+size 1243591

2024/Semi-Supervised Sparse Gaussian Classification_ Provable Benefits of Unlabeled Data/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:559142df7fce9b9ec64f57774dde95e8419d0bdb5d8ff2771f42b687d3dc6ea7
+size 1710212

2024/Semi-Truths_ A Large-Scale Dataset of AI-Augmented Images for Evaluating Robustness of AI-Generated Image detectors/0f6b1de4-2630-4a2e-b926-9b729ed99dc3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8de7240f31d7cceb0000bcceeb8e7e43362a7e2ee6cb2894bbce6beebc61c3de
+size 19589457

2024/Semi-Truths_ A Large-Scale Dataset of AI-Augmented Images for Evaluating Robustness of AI-Generated Image detectors/full.md

@@ -0,0 +1,744 @@
+# Semi-Truths: A Large-Scale Dataset of AI-Augmented Images for Evaluating Robustness of AI-Generated Image detectors
+
+Anisha Pal $^{1*}$  Julia Kruk $^{1*}$  Mansi Phute $^{1}$  Manognya Bhattaram $^{1}$  Diyi Yang $^{2}$  Duen Horng Chau $^{1}$  Judy Hoffman $^{1}$
+
+<sup>1</sup>Georgia Institute of Technology, <sup>2</sup>Stanford University {apal72, jkruk3, mphute6, polo, judy}@gatech.edu {manognya.work}@gmail.com, {diyiy}@stanford.edu https://huggingface.co/datasets/semi-truths/Semi-Truths
+
+![](images/26b6a9c8710a3309d1b9c44656fcc1121229032aea277bab6c08a456eb34b47d.jpg)  
+Visualizing Varying Magnitudes of Change by Area Ratio (AR)  
+Figure 1: SEMI-TRUTHS image augmentations that are measured by the size of the augmented region (Area Ratio) and the semantic change achieved (Semantic Magnitude), categorized into 3 levels - small (col1), medium (col2), and large (col3).
+
+# Abstract
+
+Text-to-image diffusion models have impactful applications in art, design, and entertainment, yet these technologies also pose significant risks by enabling the creation and dissemination of misinformation. Although recent advancements have produced AI-generated image detectors that claim robustness against various augmentations, their true effectiveness remains uncertain. Do these detectors reliably identify images with different levels of augmentation? Are they biased toward specific scenes or data distributions? To investigate, we introduce SEMI-TRUTHS, featuring 27,600 real images, 223,400 masks, and 1,329,155 AI-augmented images that feature targeted and localized perturbations produced using diverse augmentation techniques, diffusion models, and data distributions. Each augmented image is accompanied by metadata for standardized and targeted evaluation of detector robustness. Our findings suggest that state-of-the-art detectors exhibit varying sensitivities to the types and degrees of perturbations, data distributions, and augmentation methods used, offering new insights into their performance and limitations. The code for the augmentation and evaluation pipeline is available at https://github.com/J-Kruk/SemiTruths.
+
+*Equal contribution.
+
+# 1 Introduction
+
+The rise of text-to-image generative models has democratized automated image creation for machine learning practitioners and the general public alike. While existing architectures like Variational Autoencoders [84, 32] and GANs [4, 100, 30, 35, 38] have produced realistic images for several years, diffusion models [15, 69, 13] have enhanced image quality, diversity, and usability, driving their rapid adoption. However, this technology presents a double-edged sword: despite its applications in fields like art, design, marketing, and entertainment [34, 96], its growing ubiquity brings a heightened risk of misuse for spreading misinformation [94, 55]. Recent incidents reveal an alarming rise in AI-modified images being used to perpetrate harmful acts such as misinformation campaigns [19, 91, 95], fraud, defamation, and identity theft [26, 79, 21, 9]. One concerning capability of these models is their ability to alter small attributes of an original image. Rather than creating images from scratch, individuals can alter only specific parts or attributes of an image to drastically change the narrative while decreasing the likelihood of detection. An example of one such "semi-truth" is the "Sleepy Joe" [72] video circulated on Twitter in 2020, where President Joe Biden's face was edited to appear as if he fell asleep during an interview. The implications of such subtle perturbations and their potential to spread misinformation [19, 91, 95] underscore the critical need for automated detection of such attacks.
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Magnitude of Change</td><td rowspan="2">Targeted Perturb.</td><td rowspan="2">Saliency Check</td><td rowspan="2">Data Collection</td><td colspan="3">Generation</td><td colspan="2">Data Dist.</td><td colspan="2">Scale</td></tr><tr><td>GANs</td><td>Diffusion</td><td>#Methods</td><td>Scene</td><td>#Real Bench.</td><td>Real</td><td>Fake</td></tr><tr><td>1 DFDC [16]</td><td>X</td><td>X</td><td>X</td><td>Generated</td><td>✓</td><td>X</td><td>8</td><td>Face</td><td>1</td><td>488.4k</td><td>~1.7M</td></tr><tr><td>2 FaceForensics++ [71]</td><td>X</td><td>X</td><td>X</td><td>Generated</td><td>✓</td><td>X</td><td>4</td><td>Face</td><td>1</td><td>509.9k</td><td>~1.8M</td></tr><tr><td>3 Celeb-DF [47]</td><td>X</td><td>X</td><td>✓</td><td>Generated</td><td>✓</td><td>X</td><td>1</td><td>Face</td><td>1</td><td>225.4k</td><td>~2.1M</td></tr><tr><td>4 DeepFakeFace [76]</td><td>X</td><td>X</td><td>X</td><td>Generated</td><td>X</td><td>✓</td><td>3</td><td>Face</td><td>1</td><td>30k</td><td>90k</td></tr><tr><td>5 CIFAKE [5]</td><td>X</td><td>X</td><td>X</td><td>Generated</td><td>X</td><td>✓</td><td>1</td><td>General</td><td>1</td><td>60k</td><td>60k</td></tr><tr><td>6 DiffusionDB [90]</td><td>X</td><td>X</td><td>✓</td><td>Sourced</td><td>X</td><td>✓</td><td>1</td><td>General</td><td>0</td><td>0</td><td>14M</td></tr><tr><td>7 MidJourney prompts [83]</td><td>X</td><td>X</td><td>X</td><td>Sourced</td><td>X</td><td>✓</td><td>1</td><td>General</td><td>0</td><td>0</td><td>248k</td></tr><tr><td>8 TWIGMA [10]</td><td>X</td><td>X</td><td>X</td><td>Sourced</td><td>X</td><td>✓</td><td>unknown</td><td>General</td><td>0</td><td>0</td><td>800k</td></tr><tr><td>9 GenImage [102]</td><td>X</td><td>X</td><td>X</td><td>Generated</td><td>✓</td><td>✓</td><td>8</td><td>General</td><td>1</td><td>1.33M</td><td>1.35M</td></tr><tr><td>10 SEMI-TRUTHS</td><td>✓</td><td>✓</td><td>✓</td><td>Generated</td><td>X</td><td>✓</td><td>8</td><td>General</td><td>6</td><td>27,635</td><td>~1.33M</td></tr></table>
+
+Table 1: SEMI-TRUTHS vs other AI-generated image datasets. We compare SEMI-TRUTHS with other AI-generated image datasets across multiple categories: (1) Magnitude of Change: provides metadata on the magnitude of perturbations; (2) Targeted Perturb.: performs targeted perturbation of images; (3) Saliency Check: saliency assessment of fake images; (4) Data Collection: data collection strategy, Generated or Sourced from publicly available portals; (5) Generation: generator category and number of methods used (TWIGMA's method was unknown since its images were sourced from Twitter); (6) Data Distribution: scene variation and diversity of real benchmarks; (7) Scale: number of real and fake images.
+
+However, existing datasets for training and evaluating AI-generated image detectors primarily consist of fully synthesized images, often limited to human faces [16, 71, 47, 39, 14]. This narrow focus fails to capture the diversity of real-world perturbations and does not reveal model biases toward different degrees of change. To address this, we introduce SEMI-TRUTHS, which includes AI-augmented images with varying levels of perturbation (detailed comparison in Tab. 1), enabling the evaluation of detectors against more realistic and diverse attacks like the "Sleepy Joe" video [72].
+
+To develop a resource for stress-testing specific biases in AI-generated image detectors, precise control over the nature and extent of changes is essential. To achieve this, we source images from 6 popular semantic segmentation datasets, which contain salient images and labeled entity masks across a range of subjects. We categorize the magnitude of change in SEMI-TRUTHS based on two criteria: (1) the size of the augmented region, and (2) the semantic change achieved. The size of the augmented region is captured by surface area of the mask, structural similarity index measure (SSIM), mean squared error (MSE), and a custom metric derived from MSE (see Algo. 2). Semantic changes are quantified using metrics such as Semantic Magnitude, Learned Perceptual Image Patch Similarity (LPIPS) [98], DreamSim [22], and Sentence Similarity [78] (see Sec. 3.1). Perturbed images in SEMI-TRUTHS are generated through diffusion inpainting and prompt-based-editing [28, 54] techniques using 5 different diffusion algorithms [63, 74, 61, 70].
+
+Our approach to curating SEMI-TRUTHS employs a flexible, plug-and-play framework for human-guidance-free image editing. This pipeline is designed to ensure reusability and adaptability to new data distributions, large language models for prompt perturbation, diffusion models, and various image synthesis techniques. By releasing this framework, we aim to empower the community to create customized stress tests to evaluate AI-generated image detectors for specific use cases.
+
+Finally, we demonstrate how the knowledge abstractions in SEMI-TRUTHS can be used to identify the sensitivities of existing detectors. By stress-testing 6 models, we reveal unique sensitivities to different data distributions, diffusion models, and perturbation degrees. Our goal is to offer a resource for targeted, interpretable, and standardized evaluation of AI-generated image detection systems, and to provide a customizable evaluation pipeline for the community.
+
+# 2 Related Work
+
+AI Augmented Image dataset The field of AI-based image generation and perturbation has rapidly evolved from autoencoders [20] and graphics-based techniques [81] to GANs [101, 58, 2, 49, 6] and, more recently, diffusion models [57, 70, 61, 24]. These advancements have heightened ethical concerns regarding identity theft and misinformation, [3, 27, 31] necessitating robust datasets for AI-generated image detection. While most research has focused on GAN-generated human faces [16, 71, 47, 39, 14], there is a growing emphasis on diffusion-based techniques for detection of deepfakes [76], digital forgery [75] and generic AI-generated content [102, 5, 83, 90]. However, existing datasets face several limitations that restrict their applicability as a benchmark for developing robust detection systems. They often come from a single model [83, 90] or source data distribution [102, 5], lack detailed generation and image metadata [10], and provide limited control over degree and quality of perturbations [83, 90, 102, 5, 76, 10, 66]. Furthermore, they do not offer scalable pipelines for integrating future image generation and perturbation techniques and are limited in their analysis of detection methods. Recognizing these gaps, we introduce SEMI-TRUTHS that incorporates multiple model variations, perturbing techniques, and source data distributions, provides comprehensive metadata, and offers fine-grained control over the quality and degree of perturbations (Tab. 1 summarizes SEMI-TRUTHS's contributions).
+
+Image editing pipelines With the advent of diffusion models, the field of image editing has seen tremendous advancements [33]. Recent developments in image inpainting, both in text-conditioned [92, 93, 87, 97] and unconditioned [51] settings, have enabled fine-grained control over image editing significantly enhancing precision and quality. While image inpainting requires the use of masks, prompt-based image editing [28, 54] performs targeted perturbations conditioned solely on text prompts. Existing frameworks like LANCE [62] and InstructPix2Pix [7] leverage this capability to develop automated image perturbation pipelines. LANCE [62], leveraging large language models (LLMs)[82] and image captioning[45], enables human-supervision-free image editing across diverse perturbations. Building on this, we extend LANCE [62] to handle a broader range of perturbation magnitudes, guided by semantic change definitions [8, 36]. Our approach integrates LlaVA [50] and LLAMA [82] models, combining inpainting and prompt-based techniques for precise, contextually informed perturbations.
+
+Stress Testing Pipelines Stress testing pipelines, crucial in software engineering, remain underutilized in machine learning. While various metrics exist for performance assessment and model comparison [67], they often lack the depth to fully capture model robustness and explain failure cases adequately. While initiatives like Stress Test NLI [56] focus on generating adversarial examples to evaluate models' inferential capabilities across six tasks, DynaBench [40] and CheckList [68] take a different approach by employing human-in-the-loop systems to dynamically benchmark and assess the robustness of natural language models in real-world scenarios. Simultaneously, in the vision community, Li et al. [46] utilize diffusion models to create ImageNet-E, honing in on assessing classifier robustness through object attributes, while Luo et al. [52] explore model sensitivity to user-defined text attributes using StyleGAN [2]. Building upon these endeavors, LANCE [62] advances the field by extracting insights from failures via a targeted perturbation algorithm, enabling stress testing across diverse attributes. Our work extends this paradigm to AI-generated image detection, presenting a versatile pipeline capable of performing image augmentation with varying magnitudes of perturbations across any diffusion model for a given set of image data points, facilitating evaluation and bias discovery in detector architectures through a comprehensive range of stress tests.
+
+# 3 SEMI-TRUTHS
+
+To precisely evaluate a detector's ability to distinguish between AI-generated and real images, we curate SEMI-TRUTHS, consisting of over 27,600 real images and 1,329,155 fake images. We
+
+![](images/026a710a3e5a16c09220c4f8733a7828009f5c169397ad741b768c90c75e22f4.jpg)  
+Figure 2: End-to-end pipeline for SEMI-TRUTHS curation and detector stress testing. The SEMI-TRUTHS pipeline sources data from 6 benchmarks and uses 2 perturbation techniques to perturb images. These images undergo saliency checks, metric computation, and stress testing of detectors across our curated tests based on the computed change metrics.
+
+consider several crucial factors: (1) strategies for targeted augmentation at varying magnitudes, (2) diversification of scene distributions, (3) caption perturbation methods, (4) image augmentation techniques, and (5) the saliency of augmented images. The imbalance in our SEMI-TRUTHS dataset arises from pairing each real image with multiple augmented variations, a vital requirement for the benchmarking scheme as it enables a comprehensive exploration of model sensitivities across various dimensions (such magnitude of augmentation, change in subject matter, and augmentation technique). This section details methods to quantify augmentation magnitudes, followed by our generation and saliency check pipeline, and an overview of key dataset attributes.
+
+# Small Changes
+
+Do not significantly alter the overall meaning or context of the image. This could include changing the color of a specific object, adding or removing a minor detail, adjusting the composition or perspective of the image, or slightly adjusting the color distribution of the image.
+
+# Medium Changes
+
+Slightly alter the viewer's perception of the image and its subject. They could involve minor changes to an object or its setting, like altering a background element, moving an object or person to another location within the frame, or changing the emotions of the people in the frame.
+
+# Large Changes
+
+Involve substantial modifications to the image that fundamentally transform its interpretation or message. It may even appear surprising or strange to an audience. This could include altering, adding or removing major elements of the image background and making changes to the subject of the image.
+
+Table 2: Semantic Taxonomy. Definitions of the magnitudes of semantic change, used to guide the perturbation of image captions (for prompt-based-editing) and mask labels(for inpainting) using LLMs for targeted image perturbation.
+
+# 3.1 Magnitudes of Augmentation
+
+The alteration made to an image can be quantified in two ways: (1) the proportion of the image area that has been altered (area ratio of change), and (2) the degree to which the semantics of the image were altered (semantic change). To control the degree of alteration along these axioms, we start with an initial description of the image. This description is obtained by either selecting a segmentation mask and the corresponding class label or, in the absence of mask information, by generating a caption for the image using BLIP [45].
+
+Introducing Perturbations Motivated by the categorization of semantic and abstract content from visual semantics research [8], we create a taxonomy for small, medium, and large semantic changes (see Tab. 2). This taxonomy is used to guide the perturbation of an image caption or mask label using LLaVA-Mistral-7B [50] or LLAMA-7B [82] (see Sec. E). As shown in Fig.3, the model is provided with a semantic magnitude category, its definition, a caption to perturb, and the image (if using LLaVA-Mistral-7b). For prompt-based-editing, a diffusion model augments images based on perturbed captions, introducing semantic changes. In conditional inpainting, the perturbed mask label enables precise control over changes in the masked image region.
+
+Measuring Surface Area Change While segmentation masks help localize augmentations to an image, providing an area ratio of change, diffusion model imprecision can compromise their accuracy.
+
+Dong et al. [17] demonstrate diffusion models can "color outside the box" during inpainting. Furthermore, the lack of mask guidance in prompt-based-editing necessitates the use of post-augmentation metrics to capture the size of alteration. Therefore we employ SSIM [89], MSE, and a custom metric that assesses the extent to which the structural components differ between the original and augmented images in pixel space. Our custom metric, derived from MSE, uses thresholding to remove noisy components followed by connected component analysis to generate masks indicating areas of change. Similar to the area ratio computed using the mask and the image, we compute a ratio using the generated mask to quantify the surface area of change. Each of these metrics are normalized between 0 and 1 and categorized into small, medium, and large changes based on percentiles: the bottom  $25^{th}$  percentile for small, the  $25^{th}$  to  $75^{th}$  percentile for medium and anything beyond the  $75^{th}$  percentile for large.
+
+Measuring Semantic Change As mentioned previously, large language models (LLMs) are used to perturb image captions and mask labels with respect to the taxonomy of semantic change shown in Tab. 2. However, the stochasticity of LLMs and diffusion models necessitates the implementation of post-augmentation metrics that provide a quantitative measure of semantic change achieved. We use three different scores for this task: LPIPS [98], DreamSim [22] (both computed between the original and augmented images), and Sentence Similarity [78] (calculated between the original and perturbed captions/mask labels; see Sec. C.2). These metrics are normalized and categorized like Surface Area Change metrics, indicating small, medium, and large augmentations.
+
+Additional Metrics In addition to surface area and semantic change, we incorporate metrics that provide a richer description of the underlying distribution, such as scene diversity and scene complexity. Scene diversity is defined by the number of unique elements within the original image, while scene complexity measures the quantity of each unique element. Both of these metrics are derived from instance segmentation maps (see Sec. C.2). Additionally, we distinguish changes by their spatial context - diffused changes are dispersed throughout the augmented image, whereas localized changes are concentrated in a specific area (see Algo. 2).
+
+# 3.2 Image Augmentation Pipeline
+
+![](images/ae4ff0ddde369850b76dd1afa8133f11650d6a7c61f3e5488eeebf0e2dc4a09a.jpg)  
+Figure 3: Image Augmentation Pipeline. Components of the image augmentation process for SEMI-TRUTHS curation using inpainting and prompt-based-editing methods.
+
+Our image augmentation pipeline, delineated in Fig. 3, expands upon the work of LANCE [62] by integrating two distinct image augmentation techniques: (1) conditional inpainting and (2) prompt-based-editing. Both approaches leverage linguistic signal as guidance in image augmentation: prompt-based-editing requires a perturbation to the image caption using LLAMA-7B [82], and conditional inpainting relies on zero-shot mask label perturbation produced by LlaVA-Mistral-7B [50]. Furthermore, the complexity of this pipeline demands comprehensive saliency checks at various stages to ensure that augmented images maintain structural integrity and align with the specified directions of change. To this end, we implement two rounds of saliency evaluation within our image augmentation pipeline to identify instances of high-quality text and image augmentations.
+
+Caption Filtering The first saliency check protocol evaluates LLM-perturbed captions to ensure two key aspects: (1) accuracy of generated BLIP [45] captions for prompt-based-editing in representing relevant image information, and (2) coherence and desirability of image edits produced
+
+by perturbed captions/labels, ensuring semantic alignment with original content. For the former, CLIPScore [29] measures the difference between embeddings of the original image and its generated caption, filtering out the lowest 5th percentile values. For the latter, cosine similarity between CLIP [65] text embeddings of the perturbed caption/mask label and the original is calculated, removing values above the  $95^{th}$  percentile (negligible change) and below the  $5^{th}$  percentile (semantic incoherence). Additional details are mentioned in Sec.B
+
+Image Saliency Check In the second stage of the saliency check pipeline, we aim to evaluate the (1) structural integrity of augmentations in the image while retaining resemblance to the original, and (2) semantic alignment of image augmentations with the perturbed captions/labels that were used to produce them. Since we lack reference images for direct comparison in image augmentations, conventional metrics like PSNR and SSIM are not suitable. Instead, we explored metrics that assess the structural integrity of each image individually. We use BRISQUE [53], a reference-free metric which quantifies the perceptual quality of an image, labeling images with a score under 70 as highly salient $^{2}$ . Similarly, we use CLIP similarity [65] between original and augmented images to ensure the diffusion model performed substantial enough augmentations on the original. We also employ CLIP directional similarity [23] to confirm that changes in images align with the changes in captions/labels. Images between the  $20^{th}$  and  $80^{th}$  percentile are considered highly salient. Additional details are mentioned in Sec. B
+
+# 3.3 SEMI-TRUTHS Details
+
+Directional Semantic Edit  
+Figure 4: SEMI-TRUTHS details and metadata. Each augmented image in SEMI-TRUTHS is accompanied by metadata detailing properties related to the native data distribution, change magnitude (both area and semantics), and directional semantic edits. Attributes highlighted in yellow are novel contributions presented in this work.  
+![](images/d1ed5bb8d3aa1bdd26013a5da97cd1e52897b62a24b35caa6c0ef364fba41522.jpg)  
+Original Caption
+
+Native Data Distribution  
+Image Augmentation Process  
+Magnitude of Change: Area  
+Magnitude of Change: Semantic  
+![](images/3bc00a9dc16c603db8a2ae70a48a6c6aeeab41d20f3a567397eea190734b9906.jpg)  
+Perturbed Caption  
+Scene Complexity  
+Scene Diversity  
+Benchmark  
+Method  
+Diffusion Model  
+Area Ratio  
+SSIM  
+Post-Edit Ratio  
+MSE  
+Semantic Mag.  
+DreamSim  
+LPIPS  
+Sentence Sim.
+
+Data Distribution We collect data from 6 semantic segmentation benchmarks representing various data distributions: CityScapes [12] for outdoor urban scenes, SUN RGBD [77] for indoor scenes, CelebA HQ [37] for human faces, Human Parsing [48] for full-body human images, and ADE20K [99] and OpenImages [43] for diverse themes. This combined dataset comprises 27,600 real images and 223,400 masks. Using conditional inpainting and prompt-based-editing techniques across 5 [61, 63, 74, 70] diffusion models for inpainting and 3 [63, 70] diffusion models for prompt-based-editing, with LlaVA-Mistral-7B [50] and LLAMA-7B [82] for prompt perturbation, we create 325,718 prompt-based-editing datapoints and 1,003,437 inpainting datapoints. After post-perturbation saliency checks,  $\sim 74\%$  of images from inpainting and  $\sim 55\%$  from prompt-based-editing techniques are labeled as highly salient, totaling  $\sim 915$ , 445 images.
+
+Metadata SEMI-TRUTHS encompasses extensive metadata accompanying both real and fake image pairs and masks, offering insights into every facet of the perturbation process (see Fig. 4). This metadata includes details about the source data distribution, such as the original benchmark from which the image was sourced, scene complexity and diversity (defined by the number and variety of scene elements), a list of unique entities present in each image, and the ratio of mask-occupied area. Additionally, it provides information about the diffusion model, perturbation technique, and language model utilized for each perturbation, alongside the original and perturbed caption/label. Furthermore,
+
+each perturbed image is accompanied by quantitative and qualitative measures of change categorized across semantic and surface area-based metrics, as outlined in Sec. 3.1. The metadata also indicates whether the change is categorized as diffused or localized, determined using a custom algorithm (detailed in Algo. 2). All of this information is crucial for testing the effectiveness of detectors across various axes, as demonstrated in Sec. 4.
+
+# 4 Experiments
+
+<table><tr><td rowspan="2">Detector</td><td rowspan="2">Backbone</td><td colspan="3">Training Data Distribution</td><td colspan="3">Precision(↑)</td><td colspan="3">Recall(↑)</td></tr><tr><td>Scene</td><td>GANs</td><td>Diffusion</td><td>All</td><td>Real</td><td>Fake</td><td>All</td><td>Real</td><td>Fake</td></tr><tr><td>1 DINOv2 [60]</td><td>ViT [18] + ResNet-50 [25]</td><td>General</td><td>X</td><td>X</td><td>29.30</td><td>37.17</td><td>21.43</td><td>49.99</td><td>99.96</td><td>00.01</td></tr><tr><td>2 CNNSpot [86]</td><td>ResNet-50 [25]</td><td>General</td><td>✓</td><td>X</td><td>30.13</td><td>35.27</td><td>25.00</td><td>49.99</td><td>99.99</td><td>00.00</td></tr><tr><td>3 DIRE [88]</td><td>ResNet-50 [25]</td><td>General</td><td>X</td><td>✓</td><td>31.09</td><td>37.18</td><td>25.00</td><td>49.99</td><td>99.99</td><td>00.00</td></tr><tr><td>4 CrossEfficientViT [11]</td><td>EfficientNet-B0 [80] + ViT [18]</td><td>Face</td><td>✓</td><td>X</td><td>46.37</td><td>34.89</td><td>57.85</td><td>46.58</td><td>62.87</td><td>30.28</td></tr><tr><td>5 UniversalFakeDetect [59]</td><td>CLIP [65]-ViT [18]</td><td>General</td><td>✓</td><td>✓</td><td>64.84</td><td>58.89</td><td>70.79</td><td>60.57</td><td>34.11</td><td>87.03</td></tr><tr><td>6 DE-FAKE [73]</td><td>CLIP [65]</td><td>General</td><td>✓</td><td>✓</td><td>61.65</td><td>49.97</td><td>73.33</td><td>61.88</td><td>52.28</td><td>71.48</td></tr></table>
+
+Table 3: AI-generated Image Detectors evaluated with SEMI-TRUTHS. We evaluated 6 detectors with varying backbones and training data distributions. Models that performed satisfactorily, highlighted in green, were selected for additional testing.
+
+We conduct extensive experiments with SEMI-TRUTHS to evaluate the effectiveness of AI-generated image detectors in distinguishing real images from AI-augmented content (see Tab. 3). In the following sections, we show how knowledge abstraction over image augmentations in the dataset helps identify nuanced biases in various detectors. All evaluations are conducted on a  $10\%$  sample of SEMI-TRUTHS, totaling 87,000 images (27,000 real and 60,000 augmented). The evaluation dataset is available at: https://huggingface.co/datasets/semi-truths/Semi-Truths-Evalset. Since the real class (Real) is unaffected in the distribution-specific analysis, the key metric to observe is Recall on the augmented (Fake) class. A dip in Recall for a specific group indicates the detector's sensitivity to that augmentation. Detector default settings (provided in their respective codebases) have been used for conducting evaluations.
+
+Overall Detector Performance We select a diverse set of open-source AI-generated image detectors for stress testing. As demonstrated in Tab. 3, each model has a unique architecture and training distribution. We first evaluate these detectors in a zero-shot setting using metrics like Precision, Recall, and F1-Score to identify top performers for further analysis. Of the 6 models chosen, only half demonstrated adequate performance for continued evaluation. The underperforming models include (1) DinoV2, a foundation vision model leveraged for zero-shot AI-generated image prediction, (2) CNNSpot, a ResNet-50 trained solely on GAN-generated content, and (3) DIRE, a ResNet-50 trained on diffusion-generated content.
+
+Sensitivity to Data Distribution To assess potential biases toward specific data distributions, we inspect detector performances on various semantic segmentation benchmarks represented in SEMI-TRUTHS. Fig. 5 shows that detector performance varies significantly across data sources. Notably, CrossEfficientViT [11], which is trained on GAN-generated images of human faces, exhibits a significant performance drop on human faces sourced from benchmarks ADE20K, CityScapes [12], and SUN-RGBD [77] (CrossEfficientViT pre-emptively filters any images that do not contain a human face). In contrast, DE-FAKE [73], trained on general scene images, exhibits the worst performance on CelebA-HQ [44] and HumanParsing [48] due to limited focus on humans and portrait-like images in its training distribution. On the other hand, UniversalFakeDetect [59], trained on indoor bedroom images and other generic scenes, fails to perform well with SUN RGBD and shows a significant performance drop on CityScapes.
+
+Furthermore, we investigate the detectors' ability to handle highly complex and diverse multi-instance scenes. Their performance is evaluated across varying degrees of scene diversity (number of unique class instances in the images) and scene complexity (number of instances in total), categorized into small, medium, and large bins (additional details in Sec. C.2). We find that UniversalFakeDetect's [73] performance drops with increasing scene diversity and complexity. In contrast, DE-FAKE [73]
+
+![](images/63bdab01e8bdf975877aafd90eb1d98875fe7cdb4858866e113e418e47bcd715.jpg)
+
+![](images/fe6e7577f4a204bfb8af50c5505b2bcb90eb865f9be44d19423558703bec4638.jpg)  
+Figure 5: Detectors are sensitive to semantic aspects of data distribution. The 3 detectors, CrossEfficientViT, DE-FAKE and UniversalFakeDetect were evaluated across varying (a) data distribution, (b) scene complexity and (c) scene diversity.
+
+remains fairly robust across different scene variations. CrossEfficientViT [11] shows improved performance with increasing scene complexity and diversity, which can be attributed to human-centered benchmarks like CelebA-HQ [44] and HumanParsing [48] segmenting distinct facial features and body parts which results in lower scene complexity.
+
+These results highlight that detectors are highly sensitive to the semantic attributes of data distributions, emphasizing the importance of stress tests to identify and address distributional weaknesses.
+
+![](images/225868460865ce89f28d3b7029124a939b1d8ac229dfb9dfb9fefb89718ebc50.jpg)  
+Figure 6: Performance variation across image augmentation methods and diffusion algorithms. SEMI-TRUTHS offers data generated using various diffusion algorithms and augmentation methods facilitating detector evaluation on these aspects.
+
+![](images/0f9171b89bc36d0abaf5e905b354734cfaf6c813eac1a19a3ede57861a820586.jpg)
+
+Evaluation across Augmentation Techniques and Models SEMI-TRUTHS contains images generated using two different augmentation approaches - conditional inpainting and prompt-based-editing as well as five different diffusion algorithms: StableDiffusion v1.4, StableDiffusion v1.5, StableDiffusion XL [61], OpenJourney [63], and Kandinsky 2.2 [74]. This diversity in generated content enables investigation of detector sensitivities to different augmentation procedures. As shown in Fig.6, UniversalFakeDetect [59] performs best on images augmented with Kandinsky 2.2 [74] and worst on those augmented with StableDiffusion v1.5 [70], with a  $10\%$  difference in Recall score. The inverse is true for DE-FAKE [73]. CrossEfficientVit [11] performs best on images augmented with StableDiffusion v1.4 and worst with Kandinsky 2.2 [74], with a  $12\%$  drop in performance. Furthermore, CrossEfficientViT [11] and DE-FAKE [73] are more sensitive to inpainted images, whereas UniversalFakeDetect [59] performs worst on content augmented with prompt-based-editing.
+
+Evaluation across Varying Magnitudes of Augmentation As detailed in Sec. 3.1, each image in SEMI-TRUTHS is fitted with an array of descriptive attributes that capture the magnitude of change. In Fig.7 we examine the impact of varying degrees of augmentation on detector performance, focusing on both surface area and semantic changes. Note that CrossEfficientViT [11] performs better on smaller values of Area Ratio, where as UniversalFakeDetect [59] performs better on larger changes.
+
+![](images/871bcf21a223a4c28ab84c8c868ccb4c27870f69e05fd9a908f95ae1b2d99136.jpg)  
+Stress Testing Detectors on Various Magnitudes of Change  
+Figure 7: Performance variation of select detectors across various magnitudes of augmentation. DE-FAKE [73] is robust across the board, Area Ratio captures the sensitivity exhibited in UniversalFakeDetect [59] and CrossEfficientViT [11].
+
+<table><tr><td>Phrase(Original → Edited)</td><td>Counts</td><td>Recall</td></tr><tr><td colspan="3">Easy cases</td></tr><tr><td>1 lower lip → nose</td><td>70</td><td>66.67</td></tr><tr><td>2 left brow → left brow with slight arch</td><td>99</td><td>50.0</td></tr><tr><td>3 car → car with shiny chrome accents</td><td>59</td><td>45.16</td></tr><tr><td colspan="3">Difficult cases</td></tr><tr><td>4 lower lip → lipstick</td><td>190</td><td>15.79</td></tr><tr><td>5 skin → skin with subtle freckles</td><td>127</td><td>7.14</td></tr><tr><td>6 left ear → earring</td><td>177</td><td>6.67</td></tr><tr><td colspan="3">(a) CrossEfficientViT [11]</td></tr><tr><td colspan="3">(b) UniversalFakeDetector [59]</td></tr><tr><td colspan="3">(c) De-FAKE [73]</td></tr></table>
+
+Table 4: Directional Semantic Edits for investigating detector biases. Directional Semantic Edits provide insights on which edits to a certain entity has a higher chance of fooling detectors, we notice that patterns vary significantly across detectors.
+
+UniversalFakeDetect's [59] performance also drops as DreamSim [22] scores increase. Even though DE-FAKE [73] is not the best performing model, it appears to be the most robust against various magnitudes of change across the board.
+
+**Directional Semantic Edits** Quantitative metrics can be reductive when describing how the semantics or narrative of an image changes. Transitioning to an embedded space to assess similarity, for example, often results in significant information loss. To address this issue, we introduce "Directional Semantic Edits", which groups augmented images from SEMI-TRUTHS by distinct pairs of original and perturbed caption/mask labels. In the evaluation set, certain directional semantic edits occurred as frequently as 445 times. Each detector is evaluated on these groups, and metrics are sorted by Recall, as shown in Tab.4. Each model exhibits distinct performance variations based on specific semantic changes. Notably, UniversalFakeDetect [59] performs best on augmentations to facial features but worst on augmentations to vegetation. Conversely, DE-FAKE [73] excels at detecting augmentations to cars and vegetation but struggles with augmentations to human faces. CrossEfficientViT [11] shows varied performance with augmentations to human faces, appearing in both its highest and lowest ranks, indicating sensitivity to the magnitude of the change.
+
+Further analysis of these augmentations can maximize the potential of these algorithms by informing decisions about the most suitable ensemble techniques. For example, while UniversalFakeDetect [59] struggles with vegetation-to-tree augmentations, DE-FAKE [73] excels, suggesting a suitable combination for ensemble approaches. Such analysis reveals the most challenging directional augmentations, offering insights into detector model limitations.
+
+Surveying Human Perception of Magnitudes of Change To build intuition about the algorithms we use to quantify the degree of visual and semantic change achieved during image augmentation, we conduct a user study to evaluate if any metrics align with human perception. Annotators are asked to categorize changes between original and augmented images as "not much," "some," or "a lot," corresponding to our "small," "medium," and "large" change bins. We then compute correlation coefficients (Pearson [41], Kendall Tau [64], and Spearman [1]) between human scores and quantitative measures in SEMI-TRUTHS. The results in Tab.5 show that Area Ratio, a novel metric presented in this work, demonstrates the highest correlation with human perception, whereas other metrics demonstrate little to no correlation. It is important to note, however, that some changes may be imperceptible to the human eye but appear drastic in pixel space (additional discussion in Sec. D).
+
+<table><tr><td rowspan="2">Correlation Coeff.</td><td colspan="3">Change Metrics(↑)</td></tr><tr><td>Area Ratio</td><td>LPIPS Score</td><td>SSIM</td></tr><tr><td>1 Pearson</td><td>0.46</td><td>0.14</td><td>-0.16</td></tr><tr><td>2 Kendall-Tau</td><td>0.40</td><td>0.15</td><td>-0.14</td></tr><tr><td>3 Spearman</td><td>0.50</td><td>0.19</td><td>-0.17</td></tr></table>
+
+Table 5: Correlation between quantitative measures of change and Human Perception. Correlation coefficients computed between human-annotated magnitudes of change and quantitative metrics available in the dataset. Quantitative metrics not displayed here had coefficients  $< {0.10}$  .
+
+![](images/6c13542154a91a7cf7c6874fe423c4ab1e494e921600d6acc31caf6d2dcd7572.jpg)  
+Figure 8: Relationship between quantitative change metrics and Human Perception of change (small, medium, large) in SEMI-TRUTHS. Each violin plot shows the distribution of metric values for a change category.
+
+# 5 Discussion
+
+**Limitations and Future Work** Our inpainting pipeline currently relies on manual semantic mask input from existing semantic segmentation benchmarks, limiting usability. To improve, automatic mask generation methods like SAM [42] can be embedded into the augmentation pipeline, similar to InstructEdit [85]. Additionally, using LLAMA-7B [82] and LlaVA [50] models for zero-shot perturbation has led to many poor-quality outputs, requiring filtering. Future iterations will involve fine-tuning these models. We are also aware of potential biases in metrics like LPIPS [98], Sentence Similarity [78], and DreamSim [22], which may impact evaluations. To mitigate this issue, we will incorporate a combination of multiple open-source LLMs to compute semantic change metrics, thereby reducing the inherent biases associated with any single model, a process facilitated by our modular pipeline which enables easy switching between different LLMs of the user's choosing.
+
+Ethical Issues and Bias Mitigation While our project aims to create a test suite for evaluating and improving detector robustness, it can also be used to create fake images capable of deceiving AI-generated image detectors, potentially facilitating the spread of misinformation. Hence, we curated SEMI-TRUTHS by sourcing images from publicly available datasets with minimal potential for harm, and any manipulations on such images should not serve as a potential threat to society as per our knowledge. Additionally, despite our efforts at diversification of data and models, inherent biases from these modules may persist, potentially perpetuating or exacerbating existing inequalities, resulting in uneven performance across different contexts and types of images. To minimize additional bias, we employ a diverse range of perturbation techniques, diffusion models, LLMs, and source benchmarks. This diversity, along with comprehensive metadata—including original and perturbed captions/labels and images—enables users to analyze perturbation styles and identify existing biases in the generative models.
+
+# 6 Conclusion
+
+To address the rising threat of misinformation from AI-augmented images, we introduce SEMI-TRUTHS: a comprehensive dataset of 1,329,155 AI-augmented images, with detailed metadata on source distribution, augmentation techniques, change magnitudes, diffusion models, and directional edits (paired original and perturbed captions). Our plug-and-play image perturbation pipeline enables easy generation of additional augmentations and offers a standardized platform to test detector robustness across curated scenarios. Our analysis reveals that state-of-the-art detectors exhibit varying sensitivity to perturbation levels, data distributions, and augmentation methods, providing valuable insights into detector functionality. With a semantic taxonomy for defining change types and a quality-check pipeline, SEMI-TRUTHS also serves as a robust training and testing resource, enhancing the resilience of AI-generated image detectors. Furthermore, its diverse metadata enables bias analysis, supporting research into model fairness. We believe the user-friendly design of SEMI-TRUTHS will facilitate ongoing research into robustness against evolving generative models, helping combat misinformation effectively.
+
+# References
+
+[1] Spearman Rank Correlation Coefficient, pages 502-505. Springer New York, New York, NY,
+
+2008. 9  
+[2] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space?, 2019. 3  
+[3] C. Avey. Ethical pros and cons of ai image generation. IEEE Computer Society/Tech News/Community Voices, December 27 2023. 3  
+[4] Jianmin Bao, Dong Chen, Fang Wen, Houqiang Li, and Gang Hua. Cvae-gan: fine-grained image generation through asymmetric training. In Proceedings of the IEEE international conference on computer vision, pages 2745-2754, 2017. 2  
+[5] Jordan J. Bird and Ahmad Lotfi. Cifake: Image classification and explainable identification of ai-generated synthetic images, 2023. 2, 3  
+[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis, 2019. 3  
+[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 3  
+[8] B. Burford, P. Briggs, and J. P. Eakens. A taxonomy of the image: On the classification of content for image retrieval. Visual Communication, 2(2):123-161, 2003. 3, 4  
+[9] Wayne Chang. Ai is the final blow for an id system whose time has passed. Forbes, August 2024. 2  
+[10] Yiqun Chen and James Zou. Twigma: A dataset of ai-generated images with metadata from twitter, 2023. 2, 3  
+[11] Davide Alessandro Coccomini, Nicola Messina, Claudio Gennaro, and Fabrizio Falchi. Combining EfficientNet and Vision Transformers for Video Deepfake Detection, page 219-229. Springer International Publishing, 2022. 7, 8, 9, 27  
+[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding, 2016. 6, 7  
+[13] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2  
+[14] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil Jain. On the detection of digital face manipulation, 2020. 2, 3  
+[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2  
+[16] Brian Dolhansky, Ben Pflaum Joanna Bitton, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge dataset, 2020. 2, 3  
+[17] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models, 2023. 5  
+[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 7  
+[19] Simon Ellery. Fake photos of pope francis in a puffer jacket go viral, highlighting the power and peril of ai, March 2023. 2  
+[20]faceswap.dev.Welcome -faceswap，March 2024.3
+
+[21] Hannah Fry. Laguna beach high school investigates ai-generated images of students - los angeles times. Los Angeles Times, April 2024. 2  
+[22] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023. 2, 5, 9, 10  
+[23] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators, 2021. 6  
+[24] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis, 2022. 3  
+[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 7  
+[26] Kathleen Magramo Heather Chen. Finance worker pays out $25 million after video call with deepfake 'chief financial officer' | CNN — cnn.com. https://www.cnn.com/2024/02/04/ asia/deepfake-cfo-scam-hong-kong-intl-hnk/index.html. 2  
+[27] M. Heikkila. The algorithm: Ai-generated art raises tricky questions about ethics, copyright, and security. MIT Technology Review, September 20 2022. 3  
+[28] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. 2, 3, 8  
+[29] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 6  
+[30] R Devon Hjelm, Athul Paul Jacob, Tong Che, Adam Trischler, Kyunghyun Cho, and Yoshua Bengio. Boundary-seeking generative adversarial networks, 2017. 2  
+[31] Tiffany Hsu and Stuart A Thompson. The new york times company. A.I. Muddies Israel-Hamas War in Unexpected Way, Oct 2023. 3  
+[32] Huaibo Huang, Ran He, Zhenan Sun, Tieniu Tan, et al. Introvae: Introspective variational autoencoders for photographic image synthesis. Advances in neural information processing systems, 31, 2018. 2  
+[33] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey, 2024. 3  
+[34] Rowan T Hughes, Liming Zhu, and Tomasz Bednarz. Generative adversarial networks-enabled human-artificial intelligence collaborative applications for creative and design industries: A systematic review of current approaches and trends. Frontiers in artificial intelligence, 4:604234, 2021. 2  
+[35] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, July 2017. 2  
+[36] Alejandro Jaimes and Shih-Fu Chang. Conceptual framework for indexing visual information at multiple levels. In Electronic imaging, 1999. 3  
+[37] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. 2018. 6  
+[38] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, June 2019. 2  
+[39] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S. Woo. Fakeavceleb: A novel audio-video multimodal deepfake dataset, 2022. 2, 3
+
+[40] Douwe Kiela, Max Bartolo, Yixin Nie, Divyansh Kaushik, Atticus Geiger, Zhengxuan Wu, Bertie Vidgen, Grusha Prasad, Amanpreet Singh, Pratik Ringshia, Zhiyi Ma, Tristan Thrush, Sebastian Riedel, Zeerak Waseem, Pontus Stenetorp, Robin Jia, Mohit Bansal, Christopher Potts, and Adina Williams. Dynabench: Rethinking benchmarking in nlp, 2021. 3  
+[41] Wilhelm Kirch, editor. Pearson's Correlation Coefficient, pages 1090-1091. Springer Netherlands, Dordrecht, 2008. 9  
+[42] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 10  
+[43] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 6  
+[44] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7, 8  
+[45] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 3, 4, 5  
+[46] Xiaodan Li, Yuefeng Chen, Yao Zhu, Shuhui Wang, Rong Zhang, and Hui Xue. Imagenet-e: Benchmarking neural network robustness via attribute editing, 2023. 3  
+[47] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3  
+[48] Xiaodan Liang, Si Liu, Xiaohui Shen, Jianchao Yang, Luoqi Liu, Jian Dong, Liang Lin, and Shuicheng Yan. Deep human parsing with active template regression. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 37(12):2402-2414, Dec 2015. 6, 7, 8  
+[49] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing, 2021. 3  
+[50] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 4, 5, 6, 10, 27  
+[51] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022. 3  
+[52] Jinqi Luo, Zhaoning Wang, Chen Henry Wu, Dong Huang, and Fernando De la Torre. Zeroshot model diagnosis, 2023. 3  
+[53] Anish Mittal, Anush K. Moorthy, and Alan C. Bovik. Blind/referenceless image spatial quality evaluator. In 2011 Conference Record of the Forty Fifth Asilomar Conference on Signals, Systems and Computers (ASILOMAR), pages 723-727, 2011. 6  
+[54] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models, 2022. 2, 3, 8  
+[55] Scott Monteith, Tasha Glenn, John R Geddes, Peter C Whybrow, Eric Achtyes, and Michael Bauer. Artificial intelligence and increasing misinformation. The British Journal of Psychiatry, 224(2):33-35, 2024. 2  
+[56] Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Rose, and Graham Neubig. Stress test evaluation for natural language inference, 2018. 3  
+[57] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022. 3
+
+[58] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 3  
+[59] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models, 2023. 7, 8, 9, 27  
+[60] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 7, 27  
+[61] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 3, 6, 8  
+[62] Viraj Prabhu, Sriram Yenamandra, Prithvijit Chattopadhyay, and Judy Hoffman. Lance: Stress-testing visual models by generating language-guided counterfactual images, 2023. 3, 5, 27  
+[63] prompthero. Openjourney. 2, 6, 8  
+[64] Llukan Puka. Kendall's Tau, pages 713-715. Springer Berlin Heidelberg, Berlin, Heidelberg, 2011. 9  
+[65] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 6, 7  
+[66] Md Awsafur Rahman, Bishmoy Paul, Najibul Haque Sarker, Zaber Ibn Abdul Hakim, and Shaikh Anowarul Fattah. Artifact: A large-scale dataset with artificial and factual images for generalizable and robust synthetic image detection, 2023. 3  
+[67] O. Rainio, J. Teuho, and R. Klen. Evaluation metrics and statistical tests for machine learning. Scientific Reports, 14:6086, 2024. 3  
+[68] Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902-4912, Online, July 2020. Association for Computational Linguistics. 3  
+[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2  
+[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 3, 6, 8  
+[71] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In International Conference on Computer Vision (ICCV), 2019. 2, 3  
+[72] Elyse Samuels. White House social media director tweets manipulated video of Biden. 9 2020. 2  
+[73] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. De-fake: Detection and attribution of fake images generated by text-to-image generation models, 2023. 7, 8, 9, 27  
+[74] Arseniy Shakhmatov, Anton Razzhigaev, Aleksandr Nikolich, Vladimir Arkhipkin, Igor Pavlov, Andrey Kuznetsov, and Denis Dimitrov. Ikandinsky 2.2, 2023. 2, 6, 8
+
+[75] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models, 2022. 3  
+[76] Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models, 2023. 2, 3  
+[77] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 6, 7  
+[78] Xiaofei Sun, Yuxian Meng, Xiang Ao, Fei Wu, Tianwei Zhang, Jiwei Li, and Chun Fan. Sentence similarity based on contexts, 2022. 2, 5, 10  
+[79] Natnicha Surasit. Criminal exploitation of deepfakes in South East Asia globalinitiative.net. https://globalinitiative.net/analysis/deepfakes-ai-cyber-scam-south-east-asia-organized-crime/.2  
+[80] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. ArXiv, abs/1905.11946, 2019. 7  
+[81] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2016. 3  
+[82] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 3, 4, 5, 6, 10, 27  
+[83] Iulia Turc and Gaurav Nemade. Midjourney user prompts &amp; generated images (250k), 2022. 2, 3  
+[84] Arash Vahdat and Jan Kautz. Nvae: A deep hierarchical variational autoencoder. Advances in neural information processing systems, 33:19667-19679, 2020. 2  
+[85] Qian Wang, Biao Zhang, Michael Birsak, and Peter Wonka. Instructedit: Improving automatic masks for diffusion-based image editing with user instructions, 2023. 10  
+[86] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnngenerated images are surprisingly easy to spot...for now. In CVPR, 2020. 7, 27  
+[87] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting, 2023. 3  
+[88] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. arXiv preprint arXiv:2303.09295, 2023.7, 27  
+[89] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5  
+[90] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs], 2022. 2, 3  
+[91] By Mike Wendling. Ai can be easily used to make fake election photos - report, March 2024. 2  
+[92] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model, 2022. 3
+
+[93] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin C. K. Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models, 2023. 3  
+[94] Danni Xu, Shaojing Fan, and Mohan Kankanhalli. Combating misinformation in the era of generative ai models. In Proceedings of the 31st ACM International Conference on Multimedia, MM '23, page 9291–9298, New York, NY, USA, 2023. Association for Computing Machinery. 2  
+[95] Angela Yang. Even katy perry's mom was fooled by what appeared to be ai-generated met gala pics, May 2024. 2  
+[96] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv., 56(4), nov 2023. 2  
+[97] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting, 2023. 3  
+[98] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2, 5, 10  
+[99] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6  
+[100] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, October 2017. 2  
+[101] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Computer Vision (ICCV), 2017 IEEE International Conference on, 2017. 3  
+[102] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image, 2023. 2, 3
+
+# Checklist
+
+1. For all authors...
+
+(a) Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope? [Yes] See Sec. 3, 4.  
+(b) Did you describe the limitations of your work? [Yes] See Sec. 5.  
+(c) Did you discuss any potential negative societal impacts of your work? [Yes] See Sec. 5.  
+(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] We have discussed it further in Sec. 5 and Sec. D.
+
+2. If you are including theoretical results...
+
+(a) Did you state the full set of assumptions of all theoretical results? [NA]  
+(b) Did you include complete proofs of all theoretical results? [NA]
+
+3. If you ran experiments (e.g. for benchmarks)...
+
+(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] The URLs are mentioned in the appendix and in main text: (code) https://github.com/J-Kruk/SemiTruths, (data) https://huggingface.co/datasets/semi-truths/Semi-Truths.  
+(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] We have described the split of the main dataset used for evaluation in Sec. 4. We have also mentioned that we have performed out of the box evaluation of the detectors using the default settings provided in their respective codebase.  
+(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [NA] We use the default codebase provided by the detectors as mentioned in Sec. 4 hence we do not report any custom seeds.  
+(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] These details have been provided in the appendix.
+
+4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
+
+(a) If your work uses existing assets, did you cite the creators? [Yes] We have cited all the existing work we have built up on, such as LANCE (Sec. 2), diffusion models (Sec. 4), and AI-generated image detectors evaluated (Sec. 4).  
+(b) Did you mention the license of the assets? [Yes] See Sec. H.  
+(c) Did you include any new assets either in the supplemental material or as a URL? [No]  
+(d) Did you discuss whether and how consent was obtained from people whose data you're using/curating? [NA] All the existing assets we have used are open source, publicly available for use.  
+(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] We have included a discussion in Sec. 5.
+
+5. If you used crowdsourcing or conducted research with human subjects...
+
+(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [Yes] see Sec. D.  
+(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [Yes] A discussion on participant risks and the study's IRB Exempt status can be found in Sec.D.  
+(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [Yes] Annotator compensation can also be found in Sec.D.
+
+# A Image Augmentation
+
+The following augmentations were performed to improve diffusion inpainted image quality.
+
+![](images/82a2da7addfb6389b5e3bbafd883bf23624aa142252e753649f6faf5a32abbc4.jpg)  
+Original Mask
+
+![](images/79cfd13d72beaf27bfc12b03e33d3ed2fcac18979a0b7fb41ebe985d5cb36b47.jpg)  
++ Mask Filter (x2)
+
+![](images/ee2fa9921b537826ec3ce3454ad7409750a10a1906708850b2d59bcfa1ef2431.jpg)  
+Inpainting Mask Preprocessing  
++ Mode Filter
+
+![](images/da8a5acdb5cf24a010da8235429b212e10b34a25d359ee88a8765d960319caf3.jpg)  
++ Gaussian Blur  
+Figure 9: This figure demonstrates how each mask is preprocessed before it is used to augment images through diffusion inpainting. Two Max Filters are applied with width 9, following a mode filter to smooth edges, and finally Gaussian blur.
+
+# A.1 Data Preprocessing
+
+We preprocess and condense all benchmarks into one by standardizing segmentation masks, sizes of visual media, and metadata. Area ratios are computed for every mask, this is a value that denotes the percentage of the original image highlighted by the segmentation. Masks that are too small, i.e. denote a region that is less than  $5\%$  of the image area, are excluded. We process all masks by applying two Max Filters (size 9), one Mode Filter (size 9) to smooth edges, followed by Gaussian blur (radius 16), as demonstrated by Fig.9. This procedure is optimized by inspecting the quality of image augmentations produced. Fig.10 highlights how diffusion inpainting outputs are affected by various magnitudes of Gaussian blur and Max filtering.
+
+![](images/219c6f94331989255906af355f4146b14e9efbec39cfd382a23fcd920a19fdd3.jpg)  
+Mask Blur Optimization for Inpainting  
+Figure 10: This figure demonstrates how each mask is preprocessed before image augmentation through diffusion inpainting. Two Max Filters are applied with width 9, a Mode Filter to smooth edges, followed by Gaussian blur.
+
+![](images/59fb518de6100fbaf37e3cd8e7dc83bd514423739868d5e4b7437bcc7c6b120a.jpg)  
+Mask Size Optimization for Inpainting
+
+# A.2 Visualizing Capabilities of Various Diffusion Models
+
+To select the diffusion models used for image augmentation in SEMI-TRUTHS, a qualitative investigation is conducted to evaluate the quality of the generated content. Fig.11 displays one such exploration, in which we compare the inpainting capabilities of each diffusion model on a number of image, mask pairs.
+
+![](images/1291d60227d6e7099cd1aa3c223c42082780b018940b3e558e66ea1a489587f5.jpg)  
+Figure 11: Exploring capabilities of diffusion models for inpainting image augmentation.
+
+# B Quality Verification
+
+# B.1 Caption Filtering
+
+The first step in ensuring high quality generations is to select high quality prompts that will be provided as to diffusion models. In addition to the steps mentioned in the paper, we also filter the generated captions based on the following criteria:
+
+1. Length of perturbed text: To ensure high quality generation, we prune all generated text. In the case of inpainting, we discard any perturbed label larger than 5 words. For prompt-tuning we discard any perturbed captions larger than 30 words.  
+2. Special characters: We observe that models used for image augmentation do not respond well to the inclusion of special characters in the prompts. Hence, any caption or label containing a special character is discarded.
+
+# B.2 Image Saliency Check
+
+Additional details regarding quality check metrics mentioned in Sec.3.2, as well as metrics used after the image perturbing phase to determine the quality of images, are given below:
+
+1. Directional Similarity score: This metric ensures the change in the images is reflective of the change in the captions/label, and the generated image aligns with the perturbed caption/label. Directional similarity is calculated as follows:
+
+![](images/e5e27727f89aadec425e2bbf6989e60dcb594487a8e7747e9ac1f94c1ee22ca5.jpg)  
+Figure 12: Overview of the quality check pipeline: Low-quality captions or label perturbations are rejected before image generation, and low-quality images are rejected at the end of the pipeline. Only generations that maintain high quality throughout the process are accepted into the final dataset.
+
+$$
+\operatorname {C o s i n e} \left(\operatorname {C L I P} _ {\text {o r i g i n a l i m a g e}} - \operatorname {C L I P} _ {\text {p e r t u r b e d i m a g e}}, \right.
+$$
+
+$$
+C L I P _ {\text {o r i g i n a l t e x t}} - C L I P _ {\text {p e r t u r b e d t e x t}})
+$$
+
+We retain the images that correspond to the values lying above the  $23^{rd}$  percentile of the distribution.
+
+2. Remove completely black images: In certain cases, if the prompt is complex or considered potentially harmful, the diffusion models mentioned in Sec. 3.3 generate completely black images with no information. We remove all such generations from the dataset.
+
+Examples of the images corresponding to varying distribution of metrics are shown in Fig.13.
+
+Through our quality check process, we assign a True or False label depending on whether they pass the saliency check. For an image to pass the saliency check it must simultaneously satisfy all
+
+![](images/54ea6966dd96fde0ca8a9674646e903a731cddac28884b5f80eb0bbd54856cfa.jpg)  
+Figure 13: Examples of original and perturbed images throughout the spectrum of each quality check metric, from minimum to maximum
+
+the following conditions: (1) have a Brisque score below 70 (2) have the CLIP image similarity the original and augmented image lie between the 20th an 80th percentile values, and (3) have the directional similarity score lie above the 23rd percentile value. Examples of filtered out images are shown in ??
+
+![](images/8f6f0648f5827c61f50020255f0a016b3b610e3cc3c827aa6f8021e34e3a8924.jpg)  
+Brisque Score  
+All images with low scores kept
+
+![](images/ba97c22d8361df911771ca47ee779b628e6bd7dd95cd1b06c43aec0a7e92966e.jpg)  
+Filtered out: Score too high  
+Original  
+Augmented
+
+![](images/b068c2ae8e4b85139cf936237adee8b2c8be0d8ade34e250163a928c45bd929c.jpg)
+
+![](images/324f1b810925a05d2e70ec1cae2582e9a74cdb30600328e74d8814c6b1a76ced.jpg)  
+Filtered out: Score too low  
+Original
+
+![](images/34d5b685a3ae8c1e6bf054663535ab365017b433134b69560227f242b406a112.jpg)  
+Augmented
+
+![](images/694d1ce8855d8dc27d86d6776558c0a3ffa42e5be0fc83c1ed31536122504f4d.jpg)  
+CLIP Image Similarity Score
+
+![](images/beea1f546bd0c9db3a1200de6c0e63788ae2b857a3dc34e119dbb31c22db7ed8.jpg)  
+Filtered out: Score too high  
+Original
+
+![](images/c02daa26cdd697d6b0516d1464a63e4d6e0b205e68419ef5aace2a42ce609e07.jpg)  
+Augmented
+
+![](images/89d2689cc1757539441fb42ee064f0bd6c44a6dd351834e8de9e4035c8303449.jpg)  
+Filtered out: Score too low  
+Original
+
+![](images/d5d28ec500174cc6a3719e2d562f8e2d1e0d65f2ca6a6b41ba0f120d49644fd1.jpg)  
+Augmented
+
+![](images/93f6e47e78407ef5e731283260e7f0a2391caef6aa2ec3395516a80e70ef9d3f.jpg)  
+Directional Similarity Score  
+Figure 14: Distribution of quality check metrics across prompt based perturbations. The images lying in the grayed out areas are discarded during quality check.
+
+![](images/8e9c495b8e6e6d97d86c88a8830e1b2bf8c7e8b4c74d9f1fd06402cc4b5410ea.jpg)  
+Filtered out: Score too high  
+Original
+
+![](images/2197a30aeb9149e6f1e186f2b8720f10346727aa7f79b5f503980e4a3568b2cb.jpg)
+
+![](images/9befd92ad5b696ad37badbc4bf1cae11ec8eb0fd75468883849b4ce3c7a0644b.jpg)  
+Augmented
+
+# C Experimental Design
+
+# C.1 Cropping of Detector Input Images
+
+Most AI-generated image detection models evaluated in our experiments prepare input images by resizing and taking a  $224 \times 224$  center-crop. For our experimental setup, we acknowledge that this method of pre-processing may lead to cases where the perturbations made to the image lie outside the cropped area. To address this concern, we implemented a method of cropping the input image in a way that does not exclude the augmentation using the image perturbation masks. For inpainting the generation masks are used, and for prompt tuning we use the masks generated after image perturbing.
+
+The part of the image to be perturbed is masked using white pixels. This original mask is center-cropped, and the surface area of white pixels in the cropped mask is compared with the surface area of white pixels in the original mask. If this surface area ratio passes a specified threshold, it is
+
+Algorithm 1 Detector Input Image Cropping Method  
+1: Input: input image img, perturbation mask mask, initial transformation involving center crop  $T$ , threshold for good crop thresh, modified transformation  $T'$   
+2: Initialize: goodCrop  $\leftarrow$  False  
+3: Initialize: centerPixel  $\leftarrow$  (maskHeight/2, maskWidth/2)  
+4: while goodCrop  $\neq$  True do  
+5: maskTrans  $\leftarrow$  T(mask)  
+6: maskSA  $\leftarrow$  numWhitePixels(mask)  
+7: maskTransSA  $\leftarrow$  numWhitePixels(maskTrans)  
+8: maskRatio = maskSA/maskTransSA  
+9: if maskRatio  $\geq$  thresh then ▷ Mask surface area preserved after resize and cropping  
+10: goodCrop  $\leftarrow$  True  
+11: centerPixel  $\leftarrow$  updateCenterPixel() ▷ Save new center pixel to crop input image  
+12: else  
+13: centerPixel  $\leftarrow$  center(randomCrop(mask)) ▷ Recrop randomly and save center pixel  
+14: continue  
+15: end if  
+16: end while  
+17:  $T'\gets$  transform(centerPixel) ▷ Use the new center pixel for custom cropping  
+18: croppedImage  $\leftarrow$ $T'(img)$   
+19: Return: croppedImage ▷ Final Cropped Image Input
+
+determined that the cropped area includes a sufficient amount of the perturbed area for the detectors to identify, and this transformed input is passed in through the detector.
+
+# C.2 Metric Binning Strategy
+
+![](images/a6aa03a38adaa0c482b659f62ffb296ee92460b70347164709b254db640c5f8f.jpg)
+
+![](images/ede648f4ec591332204c89e7b57eb4f416139745e85a896909ff1bb8c4a5c2dc.jpg)
+
+![](images/5e91723450775749ff0ad66b5063cbaab2d164d593b809f7f7dac6cf5fcb4c0b.jpg)
+
+![](images/b9b7a997af27655e105751649067938bcf75248563f256fc09fdc594753b9f13.jpg)  
+Figure 15: Scene diversity visualized across the datasets with cutoffs for binning
+
+![](images/3232fe44ad65f2b63c6c58450cfebe0d884eeb4dddedc0d30e15d36d71d918ad.jpg)
+
+![](images/0606477f6ab337ec334b52c6e12a9cf6c49de01f1350fc0ae2ae28864a2bd96d.jpg)
+
+Scene Diversity We use scene diversity of the original image to determine how the diversity of the scene plays a role in misleading AI-generated image detectors. To calculate scene diversity, we use class labels provided by the segmentation masks to understand the composition of the scene by finding how many unique classes are present in each scene. We then divide the values into three bins of small, medium and large where the values below 1 standard deviation from the mean are assigned the bin small, the values above 1 standard deviation from the mean are assigned the bin large and the
+
+rest are assigned the bin medium. Fig. 15 illustrates the scene diversity distributions for the entire source dataset as well as for each of the individual source datasets.
+
+![](images/c1ba91d732f4ced58f1b4121e7fc429a0c29f7eb4e9301c3031e96a93aa296ef.jpg)  
+(a) ADE20K
+
+![](images/c64a426703ec9a2b53e0ead75f44e54ca64be49711e6fbaade4acd06e778b3e4.jpg)  
+(b) CelebAHQ
+
+![](images/96df73c4285e6bf49e3c7703afc350ec6ef3c7b3b8955999d32ce24edbf7d259.jpg)  
+(c) CitySCapes
+
+![](images/9c7b7fc9552271b3494ea49befcd0bfdff71fe5d015e34c7d5d2c71cbe52d1d9.jpg)  
+(d) HumanParsing
+
+![](images/8aadb49f7ce87752ac4000dc8ecb98cefe5d8be49846214d9be4a25fe8b59a49.jpg)  
+(e) SUN RGBD
+
+![](images/47635dee2f4e6de69e006ecb41cd8e3bd748d138e0b42e1560cc675e3617db81.jpg)  
+(f) OpenImages
+
+![](images/44fca451be6c61d0b58317f5403bf409f1faac890eac909eab40768abfbed931.jpg)  
+Figure 16: Scene complexity visualized across the datasets with cutoffs for binning  
+(a) DreamSim  
+Figure 17: Distribution across different semantic change metrics with their binning cutoffs
+
+![](images/1b4fdb202577a0c1413b5802a51719bcf48ac2a2d0f2210bf14978481efdd3c5.jpg)  
+(b) LPIPS Score
+
+![](images/1da995321251ded3b33b53fd095c7d024bae81a8fb1ab7b71d1f27fa398bba7b.jpg)  
+(c) Sentence Similarity
+
+Scene Complexity We use scene complexity of the original image to determine how the complexity of the scene plays a role in misleading AI-generated image detectors. To calculate scene complexity, we count the number of instances present in each image using the provided instance segmentation masks. We then divide the values into three bins of small, medium and large where the values below 1 standard deviation from the mean are assigned the bin small, the values above 1 standard deviation from the mean are assigned the bin large and the rest are assigned the bin medium. Fig. 16 illustrates the scene complexity distributions for the entire source dataset as well as for each of the individual source datasets.
+
+Semantic and Surface Change Metrics For semantic and surface area change, these values are categorized into 3 bins (small, medium and large), where the bottom  $25^{th}$  percentile was assigned to small changes, the middle  $25^{th}$  to  $75^{th}$  percentile for medium and above  $75^{th}$  for large. The distribution of changes across the different bins for the entire dataset along with the cutoffs for binning is illustrated in Fig. 17.
+
+# C.3 Post Perturbation Surface Area Change Metric
+
+The Algo. 2 indicates the generation of post-perturbation surface area change metric and corresponding mask. We also use these metrics to calculate additional metrics that categorize a change as
+
+![](images/19042a2d1ace82717cf1db2bd69fd82ac0aa4e3e3e03ac2718fd30f9f116f5a3.jpg)
+
+![](images/ebdb50966bc11379f03c044934fd237a27d34e0ae08d0113d0bb6558815556d5.jpg)
+
+![](images/832b40406bce1f835b0e8705167f0b5bbf009a98da407475981e90351f35b549.jpg)
+
+![](images/2d29c36466e3497b5aae33655512504053fd566f8fc14c5b11d7493958c47f20.jpg)
+
+![](images/9df1153e014b6d5182e368f01b93671763c5a191e11b94c79693442966520767.jpg)
+
+![](images/abcdd3fee9ea4c21100f0994b4a57bca13a11060952b8188f3a06af4d46dd9cc.jpg)
+
+![](images/b7707fe441883f1b23ab248f94b00901bb43b6d87b347d60905c8aa40038d460.jpg)  
+(a)
+
+![](images/acf59f56676ef99d29abe66c59be8350dbdbac9217d7a4e922940d166e5943e0.jpg)  
+Figure 18: Examples of localized changesSteps in the calculation of post-perturbation mask and corresponding localized change identification
+
+![](images/741542318880bc4127b37c03c22e157e9ed250a5da84918888db28be3f7f27ac.jpg)
+
+![](images/15187e580ca6498b209794823e2b9e03924578274923819ecd926420e2a9cc7e.jpg)
+
+![](images/582cec915569ddf828b28e7fef8d800d916fd7505da16ea2236fd521889c3941.jpg)
+
+![](images/0a77ed792b47414fd89c7b303c59ae889c8bde49d6fdd1241672b208ede77dca.jpg)
+
+![](images/4ffda3c40d47f0f7c24e9657983ec0ebe2dfdbf991cf3af610439f8a7c63c683.jpg)
+
+![](images/740e3e4b9286e184792b0d6bc5905c0b35715103759ec08a7aa8ea1f2bccf778.jpg)
+
+![](images/5ee70dbeda59c94bc310f28ebd3480c35c00f838f1e3f0daa062ef672feb1fe8.jpg)
+
+![](images/7bad35dd5db3ea0b406217552026f021822c4ec78148f53f837029583d10021d.jpg)  
+(b)
+
+![](images/a24337fa5b13c267ca48ffffd1dab0f40b41d3455533ae633ccfdf93545b5eb6.jpg)
+
+![](images/0de6189f780299b43847779437530b35d6fdea99f5c56cf51d2a9aee8386a824.jpg)
+
+![](images/7f78d675de093aa78e7ac9ff56a2772b8dd7063c4472a5ea78c298da150ff75d.jpg)
+
+![](images/2bd434274796d99d2d756861bba67e07a420401fe1310da3e3731343dc1bb1e7.jpg)
+
+![](images/84c565fbcf03d5c86351a43f63773285a3ea084f532338bb7cde3c51b21d9fd4.jpg)
+
+![](images/100750e1e0b0923010015e95590bb32902a83afc9ebc3b1602e42ab23c6a525d.jpg)
+
+![](images/9d4db86b116b70691d05fc506455f8429c9bc140e7b4f78c4e08e4de2b701522.jpg)
+
+![](images/4fe7ca87cc7e9d55edabc4d7798ce84cf486dff3bb9e6dfd903b5032f2063365.jpg)
+
+![](images/00e09e0e6e516cbdbd67fe0bb35e1794940f5fca3406b3ec962ffcc3fbd243a4.jpg)  
+(c)
+
+![](images/d81ebc1154526a5fdb8c68de2854068de33db099f039583e44102e9e918c0dd2.jpg)
+
+![](images/6295366dc06d174ab1b18331e472f97fd47cc60d9e8fda7c03a290353c5142ac.jpg)
+
+![](images/d549f8ad2050462ffeca0619c7224678981a9ee52227b509518e2ed94c3ddf3e.jpg)
+
+![](images/22b4ca25f73aa2f5c723819d74c0dcc2b990a7d206b880291bdc4734b5d075c5.jpg)
+
+![](images/6983a96235ded9fb64590cbdc49965b39fa1adae555f14e5b9e6fe841e4f3e24.jpg)
+
+![](images/00cccf2faaf69b048297f15340decf99b5fbf004ed5017976fcc4e418bb92eea.jpg)
+
+![](images/35621a732393e9e169d1019d7bf1617850b1b745d59af6ca82f43bf15c2ab014.jpg)
+
+![](images/8679f576ed61f6faff3d429b06be1d4f0eeffa55986737a78a77c60cdae8cc84.jpg)  
+(a)
+
+![](images/954ee77a1bb498a04f094076a9055fff07e0dd153f99e79858d747a453175d3c.jpg)  
+Figure 19: Examples of diffused changesSteps in the calculation of post-perturbation mask and corresponding diffused change identification
+
+![](images/2d0c04bb23f3df716d13f8de8b2b16aafba10e49c28bfb497f691a942a40f85e.jpg)
+
+![](images/e95c4976925ea15f6251cbddfad7e6691b0d729ee1ec214283716c18eae7b25d.jpg)
+
+![](images/f3a0f9cc45f54776809784b94546eb7c02cd6233fef83df78e9b9506c21124ca.jpg)
+
+![](images/c8319bb14988e4e730af8dc5999cb41826b68353bde321c3c8283fa500f5568a.jpg)
+
+![](images/bf73fb0d1d49ac041e01b5dddeb1e28a6e35ba32b406606ca342a544d998ff24.jpg)
+
+![](images/71d423c629160cb25e6792482341dd4005d330d8236b2f29bdb25e53cf4dfe2a.jpg)
+
+![](images/b80ab464a0531a3f2c8bd90b9dbd1be9d112aec7148e98c19619d131d3461269.jpg)  
+(b)
+
+![](images/74fe57e81de0d7872320b24a1f5255ec90e043ac2f34cf4ec0e36daf60b3a55b.jpg)
+
+![](images/e00a503c3e3844ebe97fe728bdce8737795c5c6266705c08b6628bf438173df1.jpg)
+
+![](images/cfb00ab66a35a161697ef3c8041d3a6557279177a7b5cc484a6bc1ed451b0ae7.jpg)
+
+![](images/43297f8e38e71a581d2f6d7366c385ef35f435e84c1a9d5ff54ac2a98e90e797.jpg)
+
+![](images/31a33bdbde957756a8c52d81b3a6d403552449c8ae08c07db8e33bf2e47a40a8.jpg)
+
+![](images/b429801e8eeb440cbb4bf30af12281c49812dd841667baee80c8f43536064e19.jpg)
+
+![](images/c629b876bf90f0649ff6eb7ecb676daf966c31a34f2577ae57990513a430fb8c.jpg)
+
+![](images/f25e3641e0ada005e0bb9cdc1d4431396df00beed5e2a4bc55e013f6db04b750.jpg)
+
+![](images/c1fef662549d9d15d0d03f88171339602b847d34e005f140fc6856ed87432ee6.jpg)  
+(c)
+
+![](images/5225c97fee657aa0ade97222fb11e7eb3e58f6d850a8720437a026bd93c13afb.jpg)
+
+![](images/be4a61256248e8ccb3ec8ee193389beb45d1e01d98284d439cb922607b595fd5.jpg)
+
+localized or diffused. Here localized means that the changes are concentrated in one area whereas diffused indicates that the change is spread out throughout the image. The algorithm also highlights how we use information gathered from the connected components to classify a change as diffused or localized. Fig. 18 and Fig. 19 illustrate a few examples of diffused and localized changes.
+
+# D Human Evaluation
+
+![](images/0a47128d6d2bb90668af873097b17712bcdfc92aeb52640337b4e35f5cdb69d7.jpg)  
+How to think about the "difference" between two images:  
+Figure 20: The instructive figure provided to annotators during the user study. The visualization was design to guide the participant on what kind of inputs are expected, without priming them on explicit definitions that we used in designing our semantic change categories.
+
+Algorithm 2 Post-Perturbation Surface Area Change  $(I_1, I_2)$  
+1: Initialize key parameters  
+2: Input:  $I_{1}$  (original image),  $I_{2}$  (perturbed image)  
+3: Threshold:  $T \gets 0.1$   
+4: Kernel:  $K \gets$  white Pixels(11, 11)  
+5: # Calculating mean squared error between the Real and Perturbed RGB images  
+6:  $mse_{rgb} \gets (I_1[(:, :, 0] - I_2[(:, :, 0])^2 + (I_1[(:, :, 1] - I_2[(:, :, 1])^2 + (I_1[(:, :, 2] - I_2[(:, :, 2])^2)$   
+7: # Normalization to (0,1)  
+8:  $mse_{rgb}^{norm} \gets \frac{mse_{rgb}}{(255 \times 255 \times 3)}$   
+9: # Thresholding to remove noise  
+10:  $mse\_thresh \gets$  Remove_Noise( $mse_{rgb}^{norm}, T$ )  
+11: # Generate binary mask  
+12: binary_mask  $\leftarrow$  mse thresh > 0  
+13: # Dilate mask  
+14: dilated_mask  $\leftarrow$  Dilate(binary_mask, K)  
+15: # Find all connected components for the mask  
+16: CC, CC_img, Stats, CC_centroids  $\leftarrow$  Get_connected_components(dilated_mask)  
+17: # Merge neighbouring components  
+18: merged_components  $\leftarrow$  Merge(CC_img, CC_centroids)  
+19: # Remove extremely small components to get final post perturbation mask  
+20: post_edit_mask  $\leftarrow$  Remove_Small(merged_components, min_size, minConnected)  
+21: # Calculate post perturbation ratio  
+22: post_edit_ratio  $\leftarrow \frac{\text{num\_whitePixels(post_edit_mask)}}{\text{total\_pixels(post_edit_mask)}}$   
+23: # Calculate the largest distance between the connected components  
+24: max_centroid_dist  $\leftarrow \max(\text{pdist}(CC_centroids))$   
+25: # Get the ratio of the largest connected component  
+26: largest_CC_ratio  $\leftarrow \frac{\text{max(calc_CC_size(Stats))}}{\text{totalPixels(post_edit_mask)}}$   
+27: # Categorize as localized or diffused change  
+28: if largest_CC_ratio  $\geq 0.2$  or (#CC>mean(#CC) and max_centroid_dist) then  
+29: change  $\leftarrow$  Diffused  
+30: else  
+31: change  $\leftarrow$  Localized  
+32: end if
+
+The annotated dataset contains 800 image pairs, 100 for each augmentation method and diffusion model combination. A total of 145 individuals contributed to this survey through Amazon Mechanical Turk, with an approximate compensation of $ 15 per hour. Quality was maintained through 25 pre-annotated image pairs that served as attention checks, as well as qualifications that were granted to a community of vetted annotators. Three unique annotations were collected per image pair, for which the Interclass Correlation (ICC2k) score across all image pairs is 0.835. The final human perception score was the mode of the set.
+
+The motivation behind this experiment is to evaluate how various means of quantifying magnitude of change in an image relate to human perception. The results show that the Area Ratio metric, a contribution of this work, had the highest correlation with human annotation. However, its important to note that specific metrics, such as MSE RGB, are incredibly sensitive to the slightest of changes. Whereas high-performing generative models can alter images in a way that would not be perceived by the human eye. Therefore, some metrics used in SEMI-TRUTHS capture important information about image augmentation even if they do not correlate with human perception.
+
+This user study was IRB Exempt, as it did not obtain/access any private, personally identifiable, or demographic data about the participants. The only information provided was a statement on perceived change between two images.
+
+# E Perturbation Model Experiments
+
+We considered different types of models to perform the perturbations in both the generated captions, and the mask label. We used both unimodal (LlaMA 2) and multimodal (LLaVA Mistral 7B, LLaVA Hermes 34B) models to perform these perturbations and compared their outputs to settle on our final pipeline.
+
+We find that the performance of these models varied significantly across different perturbation methods. Initially, we used multimodal models to suggest perturbations, as they could process images as input and thereby capture contextual information that might be missed when only captions are provided. However, we observed that the LlaVA Mistral 7B model performed poorly in caption perturbation in the prompt-perturbation method. To address this, we scaled up the model size by incorporating LlaVA Hermes 34B into our pipeline. Surprisingly, we discovered that despite scaling up the best caption perturbations for prompt-perturbation were achieved using the LlaMA model instead. A comparison of the perturbed captions and the related generated images can be found in Fig.21. However we maintain LlaVA Mistral 7B as part of the inpainting pipeline as it shows best performance and also retains image context during suggesting perturbed labels for masks.
+
+![](images/ab742ea01cb931307242eb788179cf0860042652ae9b19fd84120dc621343963.jpg)  
+Figure 21: Examples of perturbations made by LlaMA and LlaVA, and their resulting images.
+
+# F Evaluation Metrics
+
+The choice of metrics plays an essential role when evaluating on an imbalanced dataset. Therefore, we report individual class Precision, Recall and F1, along with overall Precision, Recall and F1, to provide a more interpretable overview. The majority of our experiments emphasize Recall on the fake class, as the impact of various augmentations primarily affects this class while leaving real images largely unaffected. However, we recognize that additional metrics addressing class imbalance would offer a more comprehensive evaluation. Consequently, we have included AUC-ROC and AUC-PR curves for both the original and a balanced evaluation set (containing 27,000 real and fake images) for the experiments mentioned in Tab. 3.
+
+![](images/4a0bd8c0a81e32b0a280e13f6f99b0b8b60b3c0b82c2d1f84f6c478bf3b053ac.jpg)  
+Figure 22: Precision-Recall curve with AP and AUC-ROC for each detector evaluated on Semi-Truths Eval Dataset
+
+![](images/c108deb1a05d768f849b59aa3e125a35a554f6abcff8e71b35d66a189650f7da.jpg)
+
+![](images/e4fc6b3dc71f712036580ccaa245dd27aa0b56e202764f236d4648a9af1e8f7b.jpg)  
+Figure 23: Precision-Recall curve with AP and AUC-ROC for each detector evaluated on a balanced Semi-Truths Eval Dataset
+
+![](images/9afed5f510023acff902036e4fc66ecaf7230a347d1c481ce1649dceaa71dc1a.jpg)
+
+# G Compute Requirements
+
+We used A40 GPUs from internal university cluster to run the augmentation techniques. Each dataset and diffusion model variation for each augmentation technique used 1 A40 GPU to run. Each image augmentation took 2 minutes to generate for both prompt-based-edit and inpainting techniques.
+
+# H License of Assets
+
+1. LANCE [62]: Apache 2.0 license  
+2. Diffusers library (used for setting up inpainting pipeline and for all diffusion model inference) - Apache 2.0 license  
+3. LlaVa [50] - Apache 2.0 license  
+4. Llama [82] - Apache 2.0 license  
+5. UniversalFakeDetect [59] - No license  
+6. DIRE [88] - No license  
+7. DE-FAKE [73] - No license  
+8. DinoV2 [60] - Apache 2.0 license  
+9. CrossEfficientViT [11] - MIT License  
+10. CNNSpot [86] - Creative Commons Public License

2024/Semi-Truths_ A Large-Scale Dataset of AI-Augmented Images for Evaluating Robustness of AI-Generated Image detectors/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:074fdcfc5712ac270ad46b9fc75af8aaf130aec349995cc295d45fc4fd8de35f
+size 1331856