Add Batch fb509b9e-bd8d-4c45-a092-34eb667cdd0a data

.gitattributes

@@ -2770,3 +2770,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/MIND_[[:space:]]Material[[:space:]]Interface[[:space:]]Generation[[:space:]]from[[:space:]]UDFs[[:space:]]for[[:space:]]Non-Manifold[[:space:]]Surface[[:space:]]Reconstruction/626e84f7-6277-4f91-86d4-c776860f97ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/MINGLE_[[:space:]]Mixture[[:space:]]of[[:space:]]Null-Space[[:space:]]Gated[[:space:]]Low-Rank[[:space:]]Experts[[:space:]]for[[:space:]]Test-Time[[:space:]]Continual[[:space:]]Model[[:space:]]Merging/cfc47abc-9866-4e17-9828-7fcf518af0ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/MINT-CoT_[[:space:]]Enabling[[:space:]]Interleaved[[:space:]]Visual[[:space:]]Tokens[[:space:]]in[[:space:]]Mathematical[[:space:]]Chain-of-Thought[[:space:]]Reasoning/a84227dc-4b99-465b-a2c9-e8c3b5152749_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MIP[[:space:]]against[[:space:]]Agent_[[:space:]]Malicious[[:space:]]Image[[:space:]]Patches[[:space:]]Hijacking[[:space:]]Multimodal[[:space:]]OS[[:space:]]Agents/43fb50da-abb0-459a-84f9-6eb54300f801_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MIRAGE_[[:space:]]Assessing[[:space:]]Hallucination[[:space:]]in[[:space:]]Multimodal[[:space:]]Reasoning[[:space:]]Chains[[:space:]]of[[:space:]]MLLM/e1962758-57a9-4a65-bf06-70a9a1d03773_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MIRA_[[:space:]]Medical[[:space:]]Time[[:space:]]Series[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Real-World[[:space:]]Health[[:space:]]Data/096f291f-d88f-4d99-a860-e85f50b78659_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MISA_[[:space:]]Memory-Efficient[[:space:]]LLMs[[:space:]]Optimization[[:space:]]with[[:space:]]Module-wise[[:space:]]Importance[[:space:]]Sampling/a0bc146a-07c1-4929-ab2d-3262a208ae00_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MIX_[[:space:]]A[[:space:]]Multi-view[[:space:]]Time-Frequency[[:space:]]Interactive[[:space:]]Explanation[[:space:]]Framework[[:space:]]for[[:space:]]Time[[:space:]]Series[[:space:]]Classification/14368126-9db9-4d48-9dc8-76610a833915_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MJ-Video_[[:space:]]Benchmarking[[:space:]]and[[:space:]]Rewarding[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Fine-Grained[[:space:]]Video[[:space:]]Preference/2f52434d-1371-4258-a666-0b6db8fbc560_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MLE-STAR_[[:space:]]Machine[[:space:]]Learning[[:space:]]Engineering[[:space:]]Agent[[:space:]]via[[:space:]]Search[[:space:]]and[[:space:]]Targeted[[:space:]]Refinement/232fb48b-02fa-4183-bc16-7c1632450ef2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MLEP_[[:space:]]Multi-granularity[[:space:]]Local[[:space:]]Entropy[[:space:]]Patterns[[:space:]]for[[:space:]]Generalized[[:space:]]AI-generated[[:space:]]Image[[:space:]]Detection/b8823069-4da9-4af8-a7f8-9ed80f65b494_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MLLM-For3D_[[:space:]]Adapting[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]3D[[:space:]]Reasoning[[:space:]]Segmentation/fd01ffb9-f31a-45ef-97a8-a69ef261552f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MLZero_[[:space:]]A[[:space:]]Multi-Agent[[:space:]]System[[:space:]]for[[:space:]]End-to-end[[:space:]]Machine[[:space:]]Learning[[:space:]]Automation/6fafb518-c26c-4cca-a1a2-f935e73e7321_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MM-Agent_[[:space:]]LLM[[:space:]]as[[:space:]]Agents[[:space:]]for[[:space:]]Real-world[[:space:]]Mathematical[[:space:]]Modeling[[:space:]]Problem/716db960-dcc6-4ea1-aa20-7942502605ce_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MMaDA_[[:space:]]Multimodal[[:space:]]Large[[:space:]]Diffusion[[:space:]]Language[[:space:]]Models/4b81fed3-62d3-4e96-8fc4-7cf31340ecb2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MOBO-OSD_[[:space:]]Batch[[:space:]]Multi-Objective[[:space:]]Bayesian[[:space:]]Optimization[[:space:]]via[[:space:]]Orthogonal[[:space:]]Search[[:space:]]Directions/c7ab3076-b2bf-4c1b-bc5b-35c281bd9bd6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MODEL[[:space:]]SHAPLEY_[[:space:]]Find[[:space:]]Your[[:space:]]Ideal[[:space:]]Parameter[[:space:]]Player[[:space:]]via[[:space:]]One[[:space:]]Gradient[[:space:]]Backpropagation/af7ca4c0-4472-4e08-8b99-c3ce380aacee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MODEM_[[:space:]]A[[:space:]]Morton-Order[[:space:]]Degradation[[:space:]]Estimation[[:space:]]Mechanism[[:space:]]for[[:space:]]Adverse[[:space:]]Weather[[:space:]]Image[[:space:]]Recovery/249afe7e-b57c-4e7d-b1d7-ecf688983e47_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MOF-BFN_[[:space:]]Metal-Organic[[:space:]]Frameworks[[:space:]]Structure[[:space:]]Prediction[[:space:]]via[[:space:]]Bayesian[[:space:]]Flow[[:space:]]Networks/6e588b57-8afd-4f2c-acaf-6d33a96de0f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MOOSE-Chem2_[[:space:]]Exploring[[:space:]]LLM[[:space:]]Limits[[:space:]]in[[:space:]]Fine-Grained[[:space:]]Scientific[[:space:]]Hypothesis[[:space:]]Discovery[[:space:]]via[[:space:]]Hierarchical[[:space:]]Search/d4b1b561-c2be-46ea-87dd-bcf37ed88972_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MOSDT_[[:space:]]Self-Distillation-Based[[:space:]]Decision[[:space:]]Transformer[[:space:]]for[[:space:]]Multi-Agent[[:space:]]Offline[[:space:]]Safe[[:space:]]Reinforcement[[:space:]]Learning/1d9624dd-1328-422a-9c0e-386b6526c787_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MOTION_[[:space:]]Multi-Sculpt[[:space:]]Evolutionary[[:space:]]Coarsening[[:space:]]for[[:space:]]Federated[[:space:]]Continual[[:space:]]Graph[[:space:]]Learning/0d554546-67a8-40d2-8d8c-dd14a20724c2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MPCache_[[:space:]]MPC-Friendly[[:space:]]KV[[:space:]]Cache[[:space:]]Eviction[[:space:]]for[[:space:]]Efficient[[:space:]]Private[[:space:]]LLM[[:space:]]Inference/0c1cb223-ee07-4202-a348-f091fac1e4a4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MPMAvatar_[[:space:]]Learning[[:space:]]3D[[:space:]]Gaussian[[:space:]]Avatars[[:space:]]with[[:space:]]Accurate[[:space:]]and[[:space:]]Robust[[:space:]]Physics-Based[[:space:]]Dynamics/e0a3390c-2910-4856-838f-457864bfd0fa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MPS-Prover_[[:space:]]Advancing[[:space:]]Stepwise[[:space:]]Theorem[[:space:]]Proving[[:space:]]by[[:space:]]Multi-Perspective[[:space:]]Search[[:space:]]and[[:space:]]Data[[:space:]]Curation/a3c222b3-28ad-4ed4-ab01-42c4c02784f6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MR.[[:space:]]Video_[[:space:]]MapReduce[[:space:]]as[[:space:]]an[[:space:]]Effective[[:space:]]Principle[[:space:]]for[[:space:]]Long[[:space:]]Video[[:space:]]Understanding/b5679aaf-8f92-44e7-ad18-ef1411981417_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MRO_[[:space:]]Enhancing[[:space:]]Reasoning[[:space:]]in[[:space:]]Diffusion[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Multi-Reward[[:space:]]Optimization/ee797bae-0b4f-4d6f-9450-81984deea22d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MS-BART_[[:space:]]Unified[[:space:]]Modeling[[:space:]]of[[:space:]]Mass[[:space:]]Spectra[[:space:]]and[[:space:]]Molecules[[:space:]]for[[:space:]]Structure[[:space:]]Elucidation/41bca0a8-0351-42e9-9ae3-dab6926f7156_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MS-GS_[[:space:]]Multi-Appearance[[:space:]]Sparse-View[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]in[[:space:]]the[[:space:]]Wild/b7178bcc-1fd9-4ecc-9b5f-d7568030d183_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MSTAR_[[:space:]]Box-free[[:space:]]Multi-query[[:space:]]Scene[[:space:]]Text[[:space:]]Retrieval[[:space:]]with[[:space:]]Attention[[:space:]]Recycling/78cf04bc-f140-456e-a8ad-ba5ad60ac9b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MTL-KD_[[:space:]]Multi-Task[[:space:]]Learning[[:space:]]Via[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]for[[:space:]]Generalizable[[:space:]]Neural[[:space:]]Vehicle[[:space:]]Routing[[:space:]]Solver/37105c69-f8cd-4f04-8c16-5c69384fa122_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MTRec_[[:space:]]Learning[[:space:]]to[[:space:]]Align[[:space:]]with[[:space:]]User[[:space:]]Preferences[[:space:]]via[[:space:]]Mental[[:space:]]Reward[[:space:]]Models/d4fbfe80-94dc-4d5c-bd43-c5618dff1452_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MURKA_[[:space:]]Multi-Reward[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]with[[:space:]]Knowledge[[:space:]]Alignment[[:space:]]for[[:space:]]Optimization[[:space:]]Tasks/d229f26c-d00a-4774-b784-3e3498223f93_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MUSTAFAR_[[:space:]]Promoting[[:space:]]Unstructured[[:space:]]Sparsity[[:space:]]for[[:space:]]KV[[:space:]]Cache[[:space:]]Pruning[[:space:]]in[[:space:]]LLM[[:space:]]Inference/4c7c94cb-3ec5-4348-92d5-3908b0f9a38c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MV-CoLight_[[:space:]]Efficient[[:space:]]Object[[:space:]]Compositing[[:space:]]with[[:space:]]Consistent[[:space:]]Lighting[[:space:]]and[[:space:]]Shadow[[:space:]]Generation/df996b8f-6500-4123-b3d6-01c2f176c7b3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MVSMamba_[[:space:]]Multi-View[[:space:]]Stereo[[:space:]]with[[:space:]]State[[:space:]]Space[[:space:]]Model/26579436-5b72-48e4-befb-907981acc4e7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MaNGO[[:space:]]—[[:space:]]Adaptable[[:space:]]Graph[[:space:]]Network[[:space:]]Simulators[[:space:]]via[[:space:]]Meta-Learning/23beac39-bb42-4069-96a2-ed0c1a47e9e0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Machine[[:space:]]Unlearning[[:space:]]in[[:space:]]3D[[:space:]]Generation_[[:space:]]A[[:space:]]Perspective-Coherent[[:space:]]Acceleration[[:space:]]Framework/79c926a2-050b-402f-88ee-c907d453017d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Machine[[:space:]]Unlearning[[:space:]]under[[:space:]]Overparameterization/9663d7e9-785d-4290-ac7b-ebaeb9e8b38f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Machine[[:space:]]Unlearning[[:space:]]via[[:space:]]Task[[:space:]]Simplex[[:space:]]Arithmetic/3fce9140-4488-46b6-917d-b2a3e1f8c280_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MagCache_[[:space:]]Fast[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Magnitude-Aware[[:space:]]Cache/c27479b9-fd50-418d-b64d-490c7e94679d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Magical_[[:space:]]Medical[[:space:]]Lay[[:space:]]Language[[:space:]]Generation[[:space:]]via[[:space:]]Semantic[[:space:]]Invariance[[:space:]]and[[:space:]]Layperson-tailored[[:space:]]Adaptation/26a1622a-487a-4cab-9331-a38261d9f8ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MaintainCoder_[[:space:]]Maintainable[[:space:]]Code[[:space:]]Generation[[:space:]]Under[[:space:]]Dynamic[[:space:]]Requirements/65428d54-0ad9-4a98-aaeb-ed4a86334290_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Majority[[:space:]]of[[:space:]]the[[:space:]]Bests_[[:space:]]Improving[[:space:]]Best-of-N[[:space:]]via[[:space:]]Bootstrapping/7af26d99-d9c0-49ff-9744-77b9ed5762fc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Make[[:space:]]Information[[:space:]]Diffusion[[:space:]]Explainable_[[:space:]]LLM-based[[:space:]]Causal[[:space:]]Framework[[:space:]]for[[:space:]]Diffusion[[:space:]]Prediction/76865b58-a0ed-4b03-a650-ce66b9847355_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Making[[:space:]]Classic[[:space:]]GNNs[[:space:]]Strong[[:space:]]Baselines[[:space:]]Across[[:space:]]Varying[[:space:]]Homophily_[[:space:]]A[[:space:]]Smoothness–Generalization[[:space:]]Perspective/6dfa996a-9119-4451-8fc2-49cfb7710832_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mamba[[:space:]]Goes[[:space:]]HoME_[[:space:]]Hierarchical[[:space:]]Soft[[:space:]]Mixture-of-Experts[[:space:]]for[[:space:]]3D[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/cf9aa5d0-bfc4-4b70-816d-3ac7c980fd46_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mamba[[:space:]]Modulation_[[:space:]]On[[:space:]]the[[:space:]]Length[[:space:]]Generalization[[:space:]]of[[:space:]]Mamba[[:space:]]Models/d23efcaa-7c36-46f8-a843-874b29b7e508_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mamba[[:space:]]Only[[:space:]]Glances[[:space:]]Once[[:space:]](MOGO)_[[:space:]]A[[:space:]]Lightweight[[:space:]]Framework[[:space:]]for[[:space:]]Efficient[[:space:]]Video[[:space:]]Action[[:space:]]Detection/1efe64a6-ca9b-4ac3-a51a-d2810439367d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Manipulating[[:space:]]3D[[:space:]]Molecules[[:space:]]in[[:space:]]a[[:space:]]Fixed-Dimensional[[:space:]]E(3)-Equivariant[[:space:]]Latent[[:space:]]Space/4eb911b1-f738-4e5b-8af3-a0006ae64c43_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Manipulating[[:space:]]Feature[[:space:]]Visualizations[[:space:]]with[[:space:]]Gradient[[:space:]]Slingshots/8bb8e87e-47bf-456f-a9c0-3c7793ba6e03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Many[[:space:]]LLMs[[:space:]]Are[[:space:]]More[[:space:]]Utilitarian[[:space:]]Than[[:space:]]One/2a906fcc-4744-44d4-ab77-d705507f0e26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Many[[:space:]]Minds,[[:space:]]One[[:space:]]Goal_[[:space:]]Time[[:space:]]Series[[:space:]]Forecasting[[:space:]]via[[:space:]]Sub-task[[:space:]]Specialization[[:space:]]and[[:space:]]Inter-agent[[:space:]]Cooperation/c1cf017e-410f-4491-92b6-9f6329e36d16_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Marginal-Nonuniform[[:space:]]PAC[[:space:]]Learnability/33d41061-d3fe-4c3e-a97e-112cd2cadd34_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Markov[[:space:]]Persuasion[[:space:]]Processes_[[:space:]]Learning[[:space:]]to[[:space:]]Persuade[[:space:]]From[[:space:]]Scratch/b3be7fe1-1241-4215-9aa6-773b82413147_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Martingale[[:space:]]Posterior[[:space:]]Neural[[:space:]]Networks[[:space:]]for[[:space:]]Fast[[:space:]]Sequential[[:space:]]Decision[[:space:]]Making/2760ef61-f720-44c5-8cbb-e4b1fa0d3d50_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Martingale[[:space:]]Score_[[:space:]]An[[:space:]]Unsupervised[[:space:]]Metric[[:space:]]for[[:space:]]Bayesian[[:space:]]Rationality[[:space:]]in[[:space:]]LLM[[:space:]]Reasoning/899f47a8-d58c-4f6b-8264-8f2b289988d4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Mask[[:space:]]Image[[:space:]]Watermarking/f42f0264-905f-43cd-8c73-4722b759be2c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Masked[[:space:]]Diffusion[[:space:]]Models[[:space:]]as[[:space:]]Energy[[:space:]]Minimization/6f704894-6154-43bb-a08b-147fb8bf0c98_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Masked[[:space:]]Gated[[:space:]]Linear[[:space:]]Unit/ede6bb4f-1501-46e7-acb3-c2f12a467704_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Matching[[:space:]]Markets[[:space:]]Meet[[:space:]]LLMs_[[:space:]]Algorithmic[[:space:]]Reasoning[[:space:]]with[[:space:]]Ranked[[:space:]]Preferences/10877c65-3853-4f9a-8f77-de8c757842ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Matchings[[:space:]]Under[[:space:]]Biased[[:space:]]and[[:space:]]Correlated[[:space:]]Evaluations/63f2d449-d093-482c-a68b-80dbf0397d96_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MaterialRefGS_[[:space:]]Reflective[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]with[[:space:]]Multi-view[[:space:]]Consistent[[:space:]]Material[[:space:]]Inference/d2a9b180-162e-4e5a-8b17-154e288e88dc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Matryoshka[[:space:]]Pilot_[[:space:]]Learning[[:space:]]to[[:space:]]Drive[[:space:]]Black-Box[[:space:]]LLMs[[:space:]]with[[:space:]]LLMs/ccc15afc-13a0-4635-9442-d5fb3e970d0f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Max[[:space:]]Entropy[[:space:]]Moment[[:space:]]Kalman[[:space:]]Filter[[:space:]]for[[:space:]]Polynomial[[:space:]]Systems[[:space:]]with[[:space:]]Arbitrary[[:space:]]Noise/79f15fb0-0de1-4861-8e2b-52f7b91b56bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/MaxSup_[[:space:]]Overcoming[[:space:]]Representation[[:space:]]Collapse[[:space:]]in[[:space:]]Label[[:space:]]Smoothing/8f906648-4fb0-4aa0-94ae-a7b6b8f6e2fd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Maximizing[[:space:]]the[[:space:]]Value[[:space:]]of[[:space:]]Predictions[[:space:]]in[[:space:]]Control_[[:space:]]Accuracy[[:space:]]Is[[:space:]]Not[[:space:]]Enough/88e36641-7df3-4635-bc2a-efe84bc80d5e_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/MIP against Agent_ Malicious Image Patches Hijacking Multimodal OS Agents/43fb50da-abb0-459a-84f9-6eb54300f801_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c241356bddde51515727b114321c520e52cb5ad0c5eb0f7cc0a21f4dd71b8349
+size 10763725

2025/MIP against Agent_ Malicious Image Patches Hijacking Multimodal OS Agents/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f0a6eec397a4d2bde85abf87f41071a609580c6766aa9f60bf5d70d5b754401
+size 1623184

2025/MIRAGE_ Assessing Hallucination in Multimodal Reasoning Chains of MLLM/e1962758-57a9-4a65-bf06-70a9a1d03773_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae0bade758c1131b299c3be357e0d5ec2134476af98b7ba7b0998b826efc5c8e
+size 3944711

2025/MIRAGE_ Assessing Hallucination in Multimodal Reasoning Chains of MLLM/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83b8c80373823245720f3079a055e720f8e375c5e34e6b896c5c7904796b0a56
+size 1426279

2025/MIRA_ Medical Time Series Foundation Model for Real-World Health Data/096f291f-d88f-4d99-a860-e85f50b78659_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cce1c04b6e1b52669026b17886fb31a2ee43e43ba692249d569e41be2ba4096
+size 1477107

2025/MIRA_ Medical Time Series Foundation Model for Real-World Health Data/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e720f18848d8262f64e65cf7bf5b50569f5ca2787fac594c8466f7c86b9312c
+size 850264

2025/MISA_ Memory-Efficient LLMs Optimization with Module-wise Importance Sampling/a0bc146a-07c1-4929-ab2d-3262a208ae00_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fd362e8f76d39db4308190a1b90cc9be895810d72969da4875d9685578e99b3
+size 942646

2025/MISA_ Memory-Efficient LLMs Optimization with Module-wise Importance Sampling/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fcd6740dbf2ba23f3b8befc69ba4f1e71a1e58f2228be6d3fcdf200e0f71b82
+size 2535368

2025/MIX_ A Multi-view Time-Frequency Interactive Explanation Framework for Time Series Classification/14368126-9db9-4d48-9dc8-76610a833915_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be8451ef3fe4d6d66315b00d41f9e324c2e1b1a989b3677a302f4158c74cd45f
+size 39322056

2025/MIX_ A Multi-view Time-Frequency Interactive Explanation Framework for Time Series Classification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72d163dfed41b2b1080bc2631974b557c00c408821fff7f1ca583003fe4dcdd5
+size 2012656

2025/MJ-Video_ Benchmarking and Rewarding Video Generation with Fine-Grained Video Preference/2f52434d-1371-4258-a666-0b6db8fbc560_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db8ad514e18db869fddf682214e2f6e9bdd49b90d342e57033ce29e0916d1d0c
+size 6963239

2025/MJ-Video_ Benchmarking and Rewarding Video Generation with Fine-Grained Video Preference/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cb497e882da3ed2d636059ad3d783769a5832449756f64d3d3a917e8597c2db
+size 1891075

2025/MLE-STAR_ Machine Learning Engineering Agent via Search and Targeted Refinement/232fb48b-02fa-4183-bc16-7c1632450ef2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13e220e43a16b3290c5b34df84c7fb912c5be654b0f21c5b13cf9a4c6b1f3dc5
+size 2049050

2025/MLE-STAR_ Machine Learning Engineering Agent via Search and Targeted Refinement/full.md

@@ -0,0 +1,598 @@
+# MLE-STAR: Machine Learning Engineering Agent via Search and Targeted Refinement
+
+Jaehyun  $\mathbf{Nam}^{12*}$ , Jinsung Yoon $^{1}$ , Jiefeng Chen $^{1}$
+
+Jinwoo Shin $^{2}$ , Sercan Ö. Arik $^{1}$ , Tomas Pfister $^{1}$
+
+$^{1}$ Google Cloud,  $^{2}$ KAIST
+
+jaehyun.nam@kaist.ac.kr, jinsungyoon@google.com
+
+# Abstract
+
+Agents based on large language models (LLMs) for machine learning engineering (MLE) can automatically implement ML models via code generation. However, existing approaches to build such agents often rely heavily on inherent LLM knowledge and employ coarse exploration strategies that modify the entire code structure at once. This limits their ability to select effective task-specific models and perform deep exploration within specific components, such as experimenting extensively with feature engineering options. To overcome these, we propose MLE-STAR, a novel approach to build MLE agents. MLE-STAR first leverages external knowledge by using a search engine to retrieve effective models from the web, forming an initial solution, then iteratively refines it by exploring various strategies targeting specific ML components. This exploration is guided by ablation studies analyzing the impact of individual code blocks. Furthermore, we introduce a novel ensembling method using an effective strategy suggested by MLE-STAR. Our experimental results show that MLE-STAR achieves medals in  $64\%$  of the Kaggle competitions on the MLE-bench, significantly outperforming the best alternative. $^{1}$
+
+# 1 Introduction
+
+The proliferation of machine learning (ML) has driven high-performance applications across diverse real-world scenarios, from fundamental tasks like tabular classification [1, 2, 3] to complex ones such as image denoising [4]. Despite these advances, developing such models remains a labor-intensive process for data scientists, involving extensive iterative experimentation and data engineering [5, 6]. To streamline such intensive workflows, recent research has focused on employing large language models (LLMs) [7, 8, 9] as machine learning engineering (MLE) agents [10, 11, 12]. By harnessing the coding and reasoning capabilities inherent in LLMs [13, 14], these agents conceptualize ML tasks as code optimization problems. They then navigate the potential code solutions ultimately producing executable code (e.g., a Python script) based on a provided task description and dataset (see Figure 1).
+
+Despite their promise as pioneering efforts, current MLE agents face several obstacles that limit their effectiveness. First, due to their strong reliance on inherent LLM knowledge, they are often biased toward familiar and frequently used methods (e.g., the scikit-learn library [15] for tabular data), neglecting potentially promising task-specific methods. Additionally, these agents [10, 12] typically employ an exploration strategy that modifies the entire code structure at once in each iteration. This often results in agents pivoting prematurely to other steps (e.g., model selection or hyperparameter tuning) because they lack the ability to perform deep, iterative exploration within specific pipeline components, such as experimenting different feature engineering options extensively.
+
+![](images/c0378da2d57477c35e60ff3ca229386ab313b1b0127019212ce4c5e6f52872aa.jpg)  
+Figure 1: Problem setup. ML Engineering agents are designed to process a task description and datasets across various modalities (e.g., tabular, text, image, audio, etc.) with the objective of determining the optimal solution for a given machine learning problem, such as classification, regression, sequence-to-sequence generation, image denoising, text normalization, etc.
+
+Contributions. We propose MLE-STAR, a novel ML Engineering agent that integrates web Search and TA Targeted code block Refinement (see Figure 2 for an overview). Specifically, generating initial solution code, MLE-STAR utilizes Google Search to retrieve relevant and potentially state-of-the-art approaches that could be effective towards building a model. Moreover, to improve the solution, MLE-STAR extracts a specific code block that represents a distinct ML pipeline component, such as feature engineering or ensemble building, and then concentrates on exploring strategies that are targeted to that component, using previous attempts as feedback to reflect on. Here, to identify the code block that has the greatest impact on performance, MLE-STAR performs an ablation study that evaluates the contribution of each ML component. This refinement process is repeated, modifying various code blocks (i.e., other ML components). In addition, we introduce a novel method to generate ensembles. MLE-STAR first proposes multiple candidate solutions. Then, instead of relying on a simple voting based on validation scores, MLE-STAR merges these candidates into a single improved solution using an ensemble strategy proposed by the agent itself. This ensemble strategy is iteratively refined based on the performance of the previous strategies.
+
+To verify the effectiveness, we conduct comprehensive evaluations of MLE-STAR using the MLE-bench's Kaggle competitions [16]. The experimental results demonstrate that MLE-STAR, requiring only minimal human effort (e.g., defining initial prompts that are generalizable to any tasks), significantly outperforms previous methods [12], including those requiring manual labor to collect strategies from Kaggle [10]. In particular, MLE-STAR achieves a substantial gain in medal achievement, improving it from  $36.6\%$  to  $63.6\%$  when compared to the top-performing baseline. Additionally, we show that our proposed ensemble technique provides a meaningful improvement to MLE-STAR.
+
+# 2 Related work
+
+LLM agents. Recent advances in LLMs have led to an active research in autonomous agents. General-purpose agents like ReAct [17] and HuggingGPT [18] typically use external tools to analyze various problems. Specialized agents, such as Voyager [19] for Minecraft or AlphaCode [20] for code generation, excel in specific domains, often using execution feedback to iteratively improve their approach. Extending these, we introduce MLE-STAR, an LLM agent that specialized in ML tasks.
+
+Automated machine learning. Automated machine learning (AutoML) aims to reduce reliance on human experts by automating end-to-end ML pipelines [21, 22, 23]. Auto-WEKA [24], TPOT [25], and recent advances such as AutoGluon [26], have made progress through exploring within predefined model or hyperparameter spaces. AutoML research also specializes in areas such as neural network design [27, 28, 29, 30], and feature engineering [31, 32, 33, 34, 35]. However, these methods rely on predefined search spaces, which often require domain expertise to define. To address this, LLM-based MLE agents [10, 12], including MLE-STAR, are emerging, since they employ effective exploration strategies directly in the code space, without the need of manually-curated search spaces.
+
+MLE agents. Leveraging coding and reasoning capabilities of LLMs [13, 14], research has been conducted on use of LLMs as MLE agents [11, 36, 37], which generate solution code, to automate ML workflows. While MLAB [38] and OpenHands [39] take general actions by calling tools to perform ML tasks, several studies specialize in ML automation. AIDE [12] generates candidate solutions in a tree structure to facilitate code space exploration. However, its heavy reliance on the LLM's internal knowledge can lead to outdated or overly simple model choices, and its refinement may prematurely shift focus between pipeline stages. DS-Agent [10] uses case-based reasoning [40, 41] to discover strategies for solution generation by utilizing manually curated cases (primarily from
+
+![](images/b44d5c9c13b562224e43ad7802c58ece29b617a991df8f4cf53a9ce5de0dd6ed.jpg)  
+Figure 2: Overview of MLE-STAR. (a) Using search as a tool, MLE-STAR retrieves task-specific models and uses them to generate an initial solution. (b) In each refinement step, MLE-STAR performs an ablation study to extract the code block that have the greatest impact. Previously modified code blocks are also provided as feedback for diversity. (c) The extracted code block is iteratively refined based on plans suggested by the LLM, which explores various plans using previous experiments as feedback (i.e., inner loop), and the target code block is also selected repeatedly (i.e., outer loop, where the improved solution of (c) becomes the previous solution in (b)).
+
+Kaggle). However, DS-Agent suffers from scalability issues due to its reliance on a manually built case bank, which requires significant human effort and can lead to solutions that are overfit to the source patterns. Also, it restricts applicability to novel task types (like complex multi-modal problems). Our method addresses these limitations. Instead of attempting to explore the broader code space or relying on a static case bank, MLE-STAR strategically explores implementation options for specific ML pipeline components. It also improves scalability by using LLMs with search as tool to retrieve effective models that fit the task beyond the constraints of a fixed case bank.
+
+# 3 MLE-STAR
+
+We introduce the proposed framework for MLE agents, MLE-STAR, that effectively leverages the coding and reasoning capabilities of LLMs to solve ML tasks. In a nutshell, our approach is based on first generating an initial solution by using web search as a tool (Section 3.1), and then refining solutions via nested loops. The outer loop targets one code block, which corresponds to the specific ML component extracted through an ablation study. The inner loop iteratively refines only this block until the outer loop moves to the next target (Section 3.2). We propose a novel ensemble method that improves the performance using the plan proposed by LLMs, which is iteratively refined (Section 3.3). To mitigate potential undesirable behaviors from LLMs, such as using test sample statistics for missing value imputation, we introduce specific modules (detailed in Section 3.4). The prompts and algorithms used in each step can be found in Appendix A and B, respectively.
+
+Problem setup. Formally, our goal is to find an optimal solution  $s^* = \arg \max_{s \in S} h(s)$ , where  $S$  is the space of possible solutions (i.e., Python scripts) and  $h: S \to \mathbb{R}$  is a score function (e.g., validation accuracy) [12]. To obtain  $s^*$ , we propose a multi-agent framework  $\mathcal{A}$ , which takes datasets  $\mathcal{D}$  (that might contain multiple files) and a task description  $\mathcal{T}_{\text{task}}$  (which includes task types, data
+
+modalities, score functions, etc.) as input.2 Here,  $\mathcal{A}$  consists of  $n$  LLM agents  $(\mathcal{A}_1,\dots ,\mathcal{A}_n)$ . Each agent  $\mathcal{A}_i$  possesses specific functionalities, which are elaborated upon in following sections.
+
+# 3.1 Generating an initial solution using web search as a tool
+
+Candidate model search. MLE-STAR starts by generating an initial solution. For high performance in ML tasks, selecting the appropriate model is paramount. However, relying solely on an LLM for model suggestions can lead to suboptimal choices. For instance, we observe that LLMs propose models like logistic regression [15] even for competitions like jigsaw-toxic-comment-classification, which is a text classification task, potentially because LLMs favor familiar patterns from their pretraining data over up-to-date information. To mitigate this, we propose using web search as a tool for MLE-STAR first to retrieve  $M$  effective, state-of-the-art models for the given task. This retrieved context is then used to guide the LLM in generating a more informed initial solution. Formally:
+
+$$
+\left\{\mathcal {T} _ {\text {m o d e l}} ^ {i}, \mathcal {T} _ {\text {c o d e}} ^ {i} \right\} _ {i = 1} ^ {M} = \mathcal {A} _ {\text {r e t r i e v e r}} \left(\mathcal {T} _ {\text {t a s k}}\right), \tag {1}
+$$
+
+where  $\mathcal{T}_{\mathrm{model}}$  represents the description of a retrieved model, while  $\mathcal{T}_{\mathrm{code}}$  provides corresponding example code. This example code is needed since the LLM can be unfamiliar with the model and cannot generate the executable code without proper guidance. Then, MLE-STAR involves evaluating of the performance of model  $i$ . To achieve this, candidate evaluation agent  $\mathcal{A}_{\mathrm{init}}$  first generates code,  $s_{\mathrm{init}}^i$ , using the retrieved model to solve the given ML task. This process is formally defined as:
+
+$$
+s _ {\text {i n i t}} ^ {i} = \mathcal {A} _ {\text {i n i t}} \left(\mathcal {T} _ {\text {t a s k}}, \mathcal {T} _ {\text {m o d e l}} ^ {i}, \mathcal {T} _ {\text {c o d e}} ^ {i}\right). \tag {2}
+$$
+
+We evaluate the performance of each  $s$  using a task-specific metric  $h$  on dataset  $\mathcal{D}$ . We denote the resulting score by  $h(s)$ , which encapsulates the entire process done in  $s$ : splitting  $\mathcal{D}$  into training and validation sets, training the model specified in  $s$  using the training data, and calculating  $h$  on the validation data. The performance for  $s_{\mathrm{init}}^i$  is thus  $h(s_{\mathrm{init}}^i)$ . As a result, a set of code scripts  $S_{\mathrm{init}} = \{s_{\mathrm{init}}^1,\dots ,s_{\mathrm{init}}^M\}$  and their performance scores  $\{h(s_{\mathrm{init}}^1),\dots ,h(s_{\mathrm{init}}^M)\}$  are obtained.
+
+Merging candidate models for initial solution. After the evaluation of the  $M$  retrieved models, a consolidated initial solution  $s_0$  is constructed through an iterative merging procedure. Specifically, we first define  $\pi$  be a permutation of the indices such that the scores are sorted in descending order:  $h(s_{\mathrm{init}}^{\pi(1)}) \geq h(s_{\mathrm{init}}^{\pi(2)}) \geq \dots \geq h(s_{\mathrm{init}}^{\pi(M)})$ . Then, we initialize the initial solution  $s_0$  with the top-performing script, and record the current best score, i.e.,  $s_0 \gets s_{(1)}$ ,  $h_{\mathrm{best}} \gets h(s_0)$ , where  $s_{(k)}$  denote the script  $s_{\mathrm{init}}^{\pi(k)}$  for simplicity. Finally, we sequentially attempt to incorporate the remaining scripts  $s_{(k)}$  for  $k = 2, \dots, M$  into  $s_0$ . For each  $k$ , MLE-STAR creates a candidate merged script by leveraging an agent  $\mathcal{A}_{\mathrm{merge}}$  that attempts to integrate  $s_{(k)}$  into the current  $s_0$ . Formally,
+
+$$
+s _ {0} \leftarrow \mathcal {A} _ {\text {m e r g e r}} \left(s _ {0}, s _ {(k)}\right), h _ {\text {b e s t}} \leftarrow h \left(s _ {0}\right) \tag {3}
+$$
+
+where,  $\mathcal{A}_{\mathrm{merge}}$  is guided to introduce a simple average ensemble to merge multiple models. Finally, we merge the models until the validation score  $h_{\mathrm{best}}$  no longer improves (see Appendix B).
+
+# 3.2 Refining a code block for solution improvement
+
+The iterative refinement phase begins with an initial solution  $s_0$  and proceeds for a predetermined number of  $T$  outer loop steps, indexed by  $t = 0,1,\dots ,T - 1$ . At each step  $t$ , the goal is to improve the current solution  $s_t$  to obtain  $s_{t + 1}$ , optimizing for a performance metric  $h$ . This process involves two main stages: targeted code block extraction and code block refinement.
+
+Targeted code block extraction. To effectively explore specialized improvement strategies, MLE-STAR identifies and targets specific code blocks within the ML pipeline represented by  $s_t$ . This selection is guided by an ablation study performed by an agent  $\mathcal{A}_{\mathrm{ab1}}$ . Specifically, the agent  $\mathcal{A}_{\mathrm{ab1}}$  generates a code  $a_t$  designed to perform an ablation study on  $s_t$ . This script creates variations of  $s_t$  by modifying or disabling specific components. To encourage exploration of different pipeline parts across iterations,  $\mathcal{A}_{\mathrm{ab1}}$  receives the summaries of previous ablation studies  $\{\mathcal{T}_{\mathrm{ab1}}^i\}_{i=0}^{t-1}$  as input:
+
+$$
+a _ {t} = \mathcal {A} _ {\mathrm {a b l}} \left(s _ {t}, \left\{\mathcal {T} _ {\mathrm {a b l}} ^ {i} \right\} _ {i = 0} ^ {t - 1}\right). \tag {4}
+$$
+
+![](images/2cc4d9d12ad810bc91576618aeb88c79337e732bdd4108d453dde362facb0e00.jpg)  
+Figure 3: Ensembling solutions. MLE-STAR iteratively proposes effective ensemble strategies based on previous attempts, integrating multiple solutions generated in parallel into a single solution.
+
+Then,  $a_{t}$  is executed, producing output results  $r_{t}$  (see examples in Appendix C). A summarization module  $\mathcal{A}_{\text{summarize}}$  processes the script and its results to generate a concise ablation summary  $T_{\text{ab1}}^{t}$ :
+
+$$
+\mathcal {T} _ {\mathrm {a b l}} ^ {t} = \mathcal {A} _ {\text {s u m m a r i z e}} \left(a _ {t}, r _ {t}\right). \tag {5}
+$$
+
+Finally, MLE-STAR selects a critical block which has the most significant impact on performance. To achieve this, an extractor module  $\mathcal{A}_{\mathrm{extractor}}$  analyzes the  $\mathcal{T}_{\mathrm{ab1}}^t$  to identify the code block  $c_{t}$  within  $s_t$  whose modification had the most significant impact on performance. To prioritize refinement of blocks not previously targeted, the set of already refined blocks  $\{c_i\}_{i = 0}^{t - 1}$  is provided as context:
+
+$$
+c _ {t}, p _ {0} = \mathcal {A} _ {\text {e x t r a c t o r}} \left(\mathcal {T} _ {\mathrm {a b l}} ^ {t}, s _ {t}, \left\{c _ {i} \right\} _ {i = 0} ^ {t - 1}\right). \tag {6}
+$$
+
+Here, MLE-STAR also generates the initial plan  $p_0$  for code block refinement at the same time, since  $T_{\mathrm{ab1}}$  can provide a good starting point by analyzing the modification of corresponding component.
+
+Code block refinement. Once the targeted code block  $c_{t}$  is defined, MLE-STAR explores various refinement strategies to improve the metric  $h$ . This involves an inner loop exploring  $K$  potential refinement for  $c_{t}$ . An agent  $\mathcal{A}_{\mathrm{coder}}$  first implements  $p_{0}$ , transforming  $c_{t}$  into a refined block  $c_{t}^{0}$ , i.e.,  $c_{t}^{0} = \mathcal{A}_{\mathrm{coder}}(c_{t}, p_{0})$ . A candidate solution  $s_{t}^{0}$  is formed by substituting  $c_{t}^{0}$  into  $s_{t}$ :
+
+$$
+s _ {t} ^ {0} = s ^ {t}. \operatorname {r e p l a c e} \left(c _ {t}, c _ {t} ^ {0}\right), \tag {7}
+$$
+
+where, replace denotes the code replacement operation. Finally, the performance  $h(s_t^0)$  is evaluated.
+
+To discover potentially more effective or novel refinement strategies, MLE-STAR iteratively generates and evaluates further plans. For  $k = 1,\dots ,K - 1$ , a planning agent  $\mathcal{A}_{\mathrm{planner}}$  proposes the next plan  $p_k$ . This agent leverages the previous attempts within the current outer step  $t$  as feedback:
+
+$$
+p _ {k} = \mathcal {A} _ {\text {p l a n n e r}} \left(c _ {t}, \left\{\left(p _ {j}, h \left(s _ {t} ^ {j}\right)\right) \right\} _ {j = 0} ^ {k - 1}\right). \tag {8}
+$$
+
+For each plan  $p_k$ , the coding agent generates the corresponding refined block, i.e.,  $c_t^k = \mathcal{A}_{\mathrm{coder}}(c_t,p_k)$ , creates the candidate solution  $s_t^k = s_t$ . replace  $(c_t,c_t^k)$ , and evaluates its performance  $h(s_t^k)$ . After exploring  $K$  refinement strategies (indexed  $k = 0,\dots ,K - 1$ ), the best-performing candidate solution is identified:  $k^{*} = \arg \max_{k\in \{0,\dots ,K - 1\}}h(s_t^k)$ . The solution for the next outer step,  $s_{t + 1}$ , is updated to  $s_t^{k^*}$  only if an improvement over  $s_t$  is found. This iterative process continues until  $t = T$ .
+
+# 3.3 Further improvement by exploring ensemble strategies
+
+To further improve upon the best single solution generated, we introduce a novel ensembling procedure (Figure 3). Standard practice might involve generating multiple candidate solutions and selecting the one with the highest score [42] according to metric  $h$ . However, analogous to model ensembling, we posit that suboptimal solutions might contain complementary strengths, and combining multiple solutions could lead to superior performance compared to relying on any single one. Therefore, we employ the planning capabilities of MLE-STAR to automatically discover effective strategies for ensembling. Specifically, let  $\{s_l\}_{l=1}^L$  be a set of  $L$  distinct solutions obtained (e.g., from parallel runs of the process described earlier). Our goal is to find an effective ensemble plan  $e$  that merges these solutions, which mirrors the structure of the targeted code block refinement stage. We start with an initial ensemble plan  $e_0$  (e.g., a simple strategy like averaging the final predictions obtained from the models trained using each solution  $s_l$ ), proposed by MLE-STAR itself. After the performance  $h(s_{\mathrm{ens}}^0)$  for the initial plan  $e_0$  is calculated, for a fixed number of iterations,  $r = 1, \dots, R$ , the planning agent  $\mathcal{A}_{\mathrm{ens\_planner}}$ , specialized in suggesting ensemble plans, proposes subsequent ensemble plans  $e_r$ .
+
+This agent uses the history of previously attempted ensemble plans and their resulting performance as feedback, i.e.,  $e_r = \mathcal{A}_{\text{ens_planner}}(\{s_l\}_{l=1}^L, \{(e_j, h(s_{\text{ens}}^j))\}_{j=0}^{r-1})$ . Each  $e_r$  is implemented via  $\mathcal{A}_{\text{ensemble}}$  to obtain  $s_{\text{ens}}^r$ :
+
+$$
+s _ {\mathrm {e n s}} ^ {r} = \mathcal {A} _ {\text {e n s e m b l e r}} \left(e _ {r}, \{s _ {l} \} _ {l = 1} ^ {L}\right). \tag {9}
+$$
+
+Finally, after exploring  $R$  ensemble strategies, the ensemble result that achieves the highest performance is selected as the final output, yielding the final ensembled result  $s_{\mathrm{ens}}^{*} = s_{\mathrm{ens}}^{r^{*}}$ :  $r^{*} = \arg \max_{r \in \{0, \dots, R\}} h(s_{\mathrm{ens}}^{r})$ . This procedure allows MLE-STAR to autonomously explore and identify potentially novel and effective ways to combine multiple complex solutions.
+
+# 3.4 Additional modules for robust MLE agents
+
+Debugging agent. We detail the design of our debugging agent within MLE-STAR. If the execution of a Python script  $s$  triggers an error, resulting in a record  $T_{\text{bug}}$  (e.g., a traceback), MLE-STAR employs a debugging module  $A_{\text{debugger}}$  to attempt correction. This process iteratively updates the script:
+
+$$
+s \leftarrow \mathcal {A} _ {\text {d e b u g g e r}} (s, \mathcal {T} _ {\text {b u g}}). \tag {10}
+$$
+
+The debugging step is repeated until either the script executes successfully, or a predefined maximum number of debugging rounds is reached. If the bug cannot be resolved, MLE-STAR proceeds to the next task using the latest version of the script that is known to be executable.
+
+Data leakage checker. We observe that LLM-generated Python scripts might have the risk of introducing data leakage, for example, by improperly accessing information from a test dataset during training dataset preparation (see Figure 6). To address this, we introduce a checker agent,  $\mathcal{A}_{\mathrm{leakage}}$ , which analyzes the solution script  $s$  prior to its execution. Recognizing that full-script analysis can be inefficient for lengthy code, we adopt a targeted approach. First, we extract the code block  $c_{\mathrm{data}}$  where data preprocessing is done. Second,  $c_{\mathrm{data}}$  is passed to the checker. If  $\mathcal{A}_{\mathrm{leakage}}$  detects potential data leakage, it generates a corrected version  $c_{\mathrm{data}}^{*}$ :  $c_{\mathrm{data}}^{*} = \mathcal{A}_{\mathrm{leakage}}(c_{\mathrm{data}})$ . Finally, the original script  $s$  is updated by replacing the identified segment with its corrected version:  $s \gets s.\mathrm{replace}(c_{\mathrm{data}}, c_{\mathrm{data}}^{*})$ . If no leakage is detected in  $c_{\mathrm{data}}$  by  $\mathcal{A}_{\mathrm{leakage}}$ , the script  $s$  remains unmodified. All generated solutions are passed through a data leakage checker,  $\mathcal{A}_{\mathrm{leakage}}$ , prior to their execution for evaluation.
+
+Data usage checker. We observe that LLM-generated scripts sometimes neglect using provided data sources, focusing solely on simple formats like CSVs (see Figure 7). To ensure the utilization of all relevant provided data, MLE-STAR introduces a data usage checker agent,  $\mathcal{A}_{\mathrm{data}}$ . Specifically, before MLE-STAR starts refinement,  $\mathcal{A}_{\mathrm{data}}$  checks the initial solution  $s_0$  along with the task description  $\mathcal{T}_{\mathrm{task}}$ . If relevant provided data is not adequately used,  $\mathcal{A}_{\mathrm{data}}$  revises the initial script as:
+
+$$
+s _ {0} \leftarrow \mathcal {A} _ {\text {d a t a}} \left(s _ {0}, \mathcal {T} _ {\text {t a s k}}\right). \tag {11}
+$$
+
+# 4 Experiments
+
+In this section, we validate the effectiveness of MLE-STAR using 22 Kaggle competitions from MLE-bench Lite [16]. Our results demonstrate that MLE-STAR significantly outperforms baselines, including those employing various LLMs (Section 4.1). Furthermore, we show that using better models and leveraging our proposed ensemble strategy effectively improves performance (Section 4.2). We also provide the example solutions generated by MLE-STAR, in Appendix D.
+
+Common setup. All experiments are conducted on 22 Kaggle competitions from MLE-bench Lite [16] using three random seeds, unless otherwise specified. Here, we use an agent  $\mathcal{A}_{\mathrm{test}}$ , which takes the task description and the final solution as input, and outputs the code that incorporates loading test sample and creating a submission file (see Appendix E for details). MLE-STAR begins by retrieving four model candidates. MLE-STAR refines for four inner loops, while exploring four outer loops. For ensemble, MLE-STAR generates two solutions in parallel, and explore ensemble strategies for five rounds. Following the MLE-bench's setup, we set a maximum time limit of 24 hours for a fair comparison (see computation analysis in Appendix F). We primarily consider AIDE [12] as our main baseline, given its state-of-the-art performance on MLE-bench. It is important to note that other baselines often limit their generalizability across various task types (e.g., audio classification, sequence-to-sequence), frequently showcasing results only on simpler modalities like tabular [11, 37]. For instance, DS-Agent [10] requires a manually constructed case bank, and their current GitHub repository lacks cases for audio classification, sequence-to-sequence, image classification, etc.
+
+Table 1: Main results from MLE-bench Lite. Each experiment is repeated using three seeds, except for o1-preview (AIDE) and GPT-4o (AIDE), which use 16 and 36 seeds, respectively. All results are taken from the GitHub repository of MLE-bench paper [16], except for the model using Gemini-2.0-Flash and Gemini-2.5-Pro. Scores represent the mean and one standard error of the mean.  
+
+<table><tr><td>Model</td><td>Made Submission (%)</td><td>Valid Submission (%)</td><td>Above Median (%)</td><td>Bronze (%)</td><td>Silver (%)</td><td>Gold (%)</td><td>Any Medal (%)</td></tr><tr><td colspan="8">MLE-STAR (Ours)</td></tr><tr><td>gemini-2.5-pro</td><td>100.0±0.0</td><td>100.0±0.0</td><td>83.3±4.6</td><td>6.1±3.0</td><td>21.2±5.1</td><td>36.4±6.0</td><td>63.6±6.0</td></tr><tr><td>gemini-2.0-flash</td><td>95.5±2.6</td><td>95.5±2.6</td><td>63.6±6.0</td><td>9.1±3.6</td><td>4.5±2.6</td><td>30.3±5.7</td><td>43.9±6.2</td></tr><tr><td colspan="8">AIDE [12]</td></tr><tr><td>gemini-2.0-flash</td><td>87.9±4.0</td><td>78.8±5.0</td><td>39.4±6.0</td><td>4.5±2.6</td><td>9.1±3.5</td><td>12.1±4.0</td><td>25.8±5.4</td></tr><tr><td>o1-preview</td><td>99.7±0.3</td><td>90.3±1.6</td><td>58.2±2.6</td><td>4.8±1.1</td><td>11.1±1.7</td><td>20.7±2.2</td><td>36.6±2.6</td></tr><tr><td>gpt-4o</td><td>82.1±1.4</td><td>65.7±1.7</td><td>29.9±1.6</td><td>3.4±0.6</td><td>5.8±0.8</td><td>9.3±1.0</td><td>18.6±1.4</td></tr><tr><td>llama-3.1-405b-instruct</td><td>72.7±5.5</td><td>51.5±6.2</td><td>18.2±4.7</td><td>0.0±0.0</td><td>4.5±2.6</td><td>6.1±2.9</td><td>10.6±3.8</td></tr><tr><td>claude-3-5-sonnet</td><td>81.8±4.7</td><td>66.7±5.8</td><td>33.3±5.8</td><td>3.0±2.1</td><td>6.1±2.9</td><td>10.6±3.8</td><td>19.7±4.9</td></tr><tr><td colspan="8">MLAB [38]</td></tr><tr><td>gpt-4o</td><td>84.8±4.4</td><td>63.6±5.9</td><td>7.6±3.3</td><td>3.0±2.1</td><td>1.5±1.5</td><td>1.5±1.5</td><td>6.1±2.9</td></tr><tr><td colspan="8">OpenHands [39]</td></tr><tr><td>gpt-4o</td><td>81.8±4.7</td><td>71.2±5.6</td><td>16.7±4.6</td><td>3.0±2.1</td><td>3.0±2.1</td><td>6.1±2.9</td><td>12.1±4.0</td></tr></table>
+
+Table 2: Comparison with DS-Agent.  
+
+<table><tr><td>Task</td><td>Metric</td><td>DS-Agent</td><td>MLE-STAR</td></tr><tr><td>WBY</td><td>MAE (↓)</td><td>213</td><td>166</td></tr><tr><td>MCC</td><td>RMLSE (↓)</td><td>0.2964</td><td>0.2911</td></tr><tr><td>ST</td><td>Accuracy (↑)</td><td>0.7982</td><td>0.8091</td></tr><tr><td>ES</td><td>AUROC (↑)</td><td>0.8727</td><td>0.9101</td></tr></table>
+
+Table 3: Performance with Claude-Sonnet-4.  
+
+<table><tr><td>Task</td><td>Metric</td><td>2.0-Flash</td><td>Sonnet-4</td></tr><tr><td>DDD</td><td>RMSE (↓)</td><td>0.0681</td><td>0.0155</td></tr><tr><td>DBI</td><td>Log Loss (↓)</td><td>0.4535</td><td>0.3114</td></tr><tr><td>SAI</td><td>Log Loss (↓)</td><td>0.2797</td><td>0.2610</td></tr><tr><td>WCR</td><td>AUROC (↑)</td><td>0.9903</td><td>0.9888</td></tr></table>
+
+# 4.1 Main results
+
+Quantitative results. As demonstrated in Table 1, MLE-STAR significantly enhances the performance of various baseline models. For instance, when applied to Gemini-2.0-Flash, MLE-STAR improves AIDE's any medal achieving rates in Kaggle competitions from  $25.8\%$  to  $43.9\%$ , representing an improvement of over 18 percentage points, and rate of above median from  $39.4\%$  to  $63.6\%$ . Notably, MLE-STAR with Gemini-2.0-Flash also substantially outperforms AIDE using a powerful reasoning model (i.e., o1-preview) in terms of achieving gold medals in  $10\%$  more tasks. Moreover, using Gemini-2.5-Pro, MLE-STAR shows a medal achievement of over  $60\%$ .
+
+Comparison to DS-Agent. While DS-Agent [10] shows competitive results on ML tasks, it necessitates human effort to curate its case bank from Kaggle. Consequently, a direct comparison between DS-Agent and AIDE or our method is not feasible, as collecting tasks across diverse modalities, such as audio classification or image denoising, requires additional effort. Nevertheless, we utilize four tabular classification tasks, i.e., wild-blueberry-yield (WBY), media-campaign-cost (MCC), spaceship-titanic (ST), and enzyme-substrate (ES), the same ones employed during DS-Agent's development stage [10], for a comparison. All experiments are done for 5 seeds following the original setup. As shown in Table 2, MLE-STAR significantly outperforms DS-Agent even without human efforts. See Appendix G for additional results, including comparison with AutoGluon [26].
+
+# 4.2 Ablation studies
+
+Performance with an advanced reasoning model. To assess if a more advanced reasoning model could enhance MLE-STAR's performance, we conduct an experiment with the recently released advanced reasoning models. First of all, as shown in Table 1, Gemini-2.5-Pro [43] yields better performance than using Gemini-2.0-Flash. For example, in denoising-dirty-documents competition, MLE-STAR with Gemini-2.0-Flash scored above the median across all three seeds, failing to achieve any medals. However, when using Gemini-2.5-Pro, MLE-STAR achieves two gold medals and one silver medal. These results demonstrate that MLE-STAR is designed to harness the advancements of rapidly improving reasoning-based LLMs.
+
+Table 4: Ablation on ensemble strategy. Experiment results on MLE-bench Lite, repeated three seeds using Gemini-2.0-Flash. Scores represent the mean and one standard error of the mean.  
+
+<table><tr><td>Ensemble strategy</td><td>Made Submission (%)</td><td>Valid Submission (%)</td><td>Above Median (%)</td><td>Bronze (%)</td><td>Silver (%)</td><td>Gold (%)</td><td>Any Medal (%)</td></tr><tr><td colspan="8">AIDE [12]</td></tr><tr><td>None</td><td>87.9±4.0</td><td>78.8±5.0</td><td>39.4±6.0</td><td>4.5±2.6</td><td>9.1±3.5</td><td>12.1±4.0</td><td>25.8±5.4</td></tr><tr><td colspan="8">MLE-STAR (Ours)</td></tr><tr><td>None</td><td>95.5±2.6</td><td>95.5±2.6</td><td>57.6±6.1</td><td>7.6±3.3</td><td>4.5±2.6</td><td>25.8±5.4</td><td>37.9±6.0</td></tr><tr><td>Best-of-N</td><td>95.5±2.6</td><td>95.5±2.6</td><td>62.1±6.0</td><td>6.1±3.0</td><td>7.6±3.3</td><td>28.8±5.6</td><td>42.4±6.1</td></tr><tr><td>Average ensemble</td><td>95.5±2.6</td><td>95.5±2.6</td><td>60.6±6.1</td><td>6.1±3.0</td><td>12.1±4.0</td><td>25.8±9.4</td><td>43.9±6.2</td></tr><tr><td>Ours</td><td>95.5±2.6</td><td>95.5±2.6</td><td>63.6±6.0</td><td>9.1±3.6</td><td>4.5±2.6</td><td>30.3±5.7</td><td>43.9±6.2</td></tr></table>
+
+Table 5: Sensitivity analysis on the number of ensemble rounds. Experiment results on 4 tasks from MLE-bench Lite, repeated three seeds using Gemini-2.0-Flash. We report the mean score.  
+
+<table><tr><td>Ensemble Round</td><td>DDD (RMSE; ↓)</td><td>DBI (Log Loss; ↓)</td><td>SAI (Log Loss; ↓)</td><td>WCR (AUROC; ↑)</td></tr><tr><td>1</td><td>0.07147</td><td>0.45351</td><td>0.28164</td><td>0.98943</td></tr><tr><td>3</td><td>0.06805</td><td>0.45351</td><td>0.27967</td><td>0.98898</td></tr><tr><td>5</td><td>0.06805</td><td>0.45351</td><td>0.27967</td><td>0.99028</td></tr></table>
+
+Table 6: Ablation on proposed components. Experiment results on 4 tasks from MLE-bench Lite, repeated three seeds using Gemini-2.0-Flash. We report the mean score and bold the best one.  
+
+<table><tr><td>Targeted Refinement</td><td>Search Tool</td><td>DDD (RMSE; ↓)</td><td>DBI (Log Loss; ↓)</td><td>SAI (Log Loss; ↓)</td><td>WCR (AUROC; ↑)</td></tr><tr><td>X</td><td>✓</td><td>0.10818</td><td>0.45689</td><td>0.29141</td><td>0.98532</td></tr><tr><td>✓</td><td>X</td><td>0.09303</td><td>0.65242</td><td>0.30529</td><td>0.96509</td></tr><tr><td>✓</td><td>✓</td><td>0.06805</td><td>0.45351</td><td>0.27967</td><td>0.99028</td></tr></table>
+
+In addition, we conduct additional experiments using Claude-Sonnet-4. Here, we select four different type of competitions: image-to-image (denoising-dirty-documents; DDD), image classification (dog-breed-identification; DBI), text classification (spooky-author-identification, SAI), and audio classification (the-icml-2013-whale-challenge-right-whaleRedux; WCR). We run each competition for three seeds. As shown in Table 3, Claude-Sonnet-4 also shows promising results, indicating that our framework is also compatible and generalizable in terms of LLM type.
+
+Effectiveness of proposed ensemble method. As highlighted in Table 4, MLE-STAR demonstrates a significant performance improvement over the competing baseline, i.e., AIDE, achieving over a  $12\%$  higher rate of obtaining any medal even without additional ensemble strategy. Notably, by ensembling multiple solution candidates, our approach yields even greater performance gains, i.e., MLE-STAR consistently improves the success rate for achieving any medal (and specifically gold medals), also surpassing the median human expert's performance by a larger margin compared to scenarios where this ensembling method is not used. While simpler strategies, such as selecting the solution with the best validation score or averaging final submissions, also offer benefits, MLE-STAR shows stronger effectiveness, e.g., leading to a higher number of gold medals.
+
+Furthermore, we conduct a sensitivity analysis on the number of ensemble rounds. Here, we utilize four datasets as same as Table 3. Table 5 indicates that while we utilize five rounds for ensemble strategy exploration, comparable performance can be achieved with fewer rounds.
+
+Effectiveness of proposed components. Here, we focus on two key components of the proposed approach: targeted refinement strategy and integrating search as a tool. As shown in Table 6, we verify the effectiveness of our targeted refinement and the benefits of the search tool integration. Specifically, our findings indicate that refining only the specific code block identified by the ablation study agent is more effective than refining the entire codebase. Moreover, our experiments consistently show that utilizing search as a tool significantly enhances the overall performance.
+
+![](images/7eca026cd92570620358828b7d0604649ab57d5c331b52a9fd4137d476ee1258.jpg)  
+Figure 4: Model usage  $(\%)$  on image classification competitions. Other models  $(11.7\%)$ , which are used by MLE-STAR, are omitted.
+
+![](images/0a01a3be8434a326d007738c348db4d27ec3b4b42df7efec8983ce83a70bcc6d.jpg)  
+Figure 5: Human intervention. By manually adding a model description, MLE-STAR integrates its training into the framework.  
+Figure 6: MLE-STAR's data leakage checker introduces appropriate preprocessing.  
+Figure 7: MLE-STAR's data usage checker captures previously unused information.
+
+```python
+# Combined DataFrame for Consistent Preprocessing
+all_col = pd.concat([train_col, test_col])
+# Impute Numerical features with median
+num cols = ['Age', 'RoomService', 'FoodCourt']
+for col in num cols:
+    if all_col[X_train].isnull().any():
+        median_val = all_col[X_train].median()
+        all_col[X_train] = all_col[X_train].fillna(median_val)
+Python script after refined by data leakage checker
+# Preprocess Training Set and get fit statistics
+X_train processed, stats =
+preprocess_data(X_train, fit.stats=None)
+# Preprocess Test Set using statistics from Train Set
+X_test processed =
+preprocess_data(X_test, fit.stats=stats)
+```
+
+```python
+Python script before refined by data usage checker  
+data = pd.read_csv('./input/train.csv')  
+Python script after refined by data leakage checker  
+def process_xyz_files(df): # Features using XYZ files  
+def volume(filepath):  
+    lines = open(filepath, 'r').readlines()  
+    a_vec = list(map(float, lines[2].split()[1:])  
+    b_vec = list(map(float, lines[3].split()[1:])  
+    c_vec = list(map(float, lines[4].split()[1:])  
+    volume = np.dot(a_vec, np.cross(b_vec, c_vec))  
+    return abs(volume)  
+for index, row in df.iterrrows():  
+    file_path = os.path.join(row['id'], 'geometry.xyz')  
+df.loc[index, 'atomic_volume'] = volume(file_path)  
+return df  
+data = process_xyz_files(data) # Process the train data
+```
+
+# 5 Discussion
+
+Qualitative observations on selected models. Figure 4 illustrates the model usage of two MLE agents: AIDE and MLE-STAR. AIDE primarily employs ResNet [44] for image classification. However, ResNet, released in 2015, is now considered outdated and can result in suboptimal performance. In contrast, our MLE-STAR primarily utilizes more recent and competitive models like Efficient-Net [45] or ViT [46], leading to the performance gain, winning  $37\%$  of the medals, more than AIDE, which wins  $26\%$  of the image classification challenges.
+
+Human intervention. MLE-STAR readily adopts even more recent models with minimal human intervention. While MLE-STAR automatically constructs a model description  $\{\mathcal{T}_{\mathrm{model}},\mathcal{T}_{\mathrm{code}}\}$  using search as tool, a natural extension involves leveraging human expertise for this construction. As shown in Figure 5, by manually adding a model description for RealMLP [47], MLE-STAR successfully integrates its training into the framework, a model not previously retrieved. In addition, users can also specify the target code blocks by replacing the ablation summary with manually written instructions.
+
+Misbehavior of LLMs and corrections. We observe that while the code generated by the LLM executed correctly, their content is sometime unrealistic, exhibiting hallucination. For example, Figure 6 illustrates an impractical approach where test data is preprocessed using its own statistics. Since test data must remain unseen, correction in the code is necessitated, for which, MLE-STAR employs a data leakage checker  $\mathcal{A}_{\mathrm{leakage}}$  to identify such issues in the generated Python script. If a problem is detected, MLE-STAR refines the code. As shown in the Figure, MLE-STAR successfully identifies the issue and modifies the code by, first extracting statistics from the training data and then preprocessing the test data using these calculated statistics. In addition, the improvement process
+
+Table 7: Improvement failure when not using data leakage checker  $\mathcal{A}_{\mathrm{leakage}}$  on spaceship-titanic competition.  
+
+<table><tr><td>Metric</td><td>Accuracy (↑)</td></tr><tr><td>Validation</td><td>0.8188 → 0.8677</td></tr><tr><td>Test</td><td>0.8033 → 0.7343</td></tr></table>
+
+Table 8: Ablation study of data usage checker  $\mathcal{A}_{\mathrm{data}}$  on nomad2018-predicting competition.  
+
+<table><tr><td>Model</td><td>Adata</td><td>RMSLE (↓)</td></tr><tr><td>MLE-STAR</td><td>X</td><td>0.0591</td></tr><tr><td>MLE-STAR</td><td>✓</td><td>0.0559</td></tr></table>
+
+![](images/9f6d730b5ade4e92941106e5ba367352db7ee0fea86650418fbd6ba709cb1f74.jpg)  
+Figure 8: Solution refinement trajectory.
+
+can fail to generalize when  $\mathcal{A}_{\mathrm{leakage}}$  is not employed, as exemplified in Table 7. In this example, the validation accuracy (i.e., the target objective) improves, but the test accuracy drops significantly. This is attributed to the LLM performing feature engineering using the target variable Transported, which is not accessible in the test set, leading to data leakage and subsequently, poor test performance.
+
+We also observe that LLMs often generate Python scripts that overlook some of the provided data sources. For example, in the nomad2018-predicting competition, Gemini-2.0-Flash solely loads train.csv, neglecting the use of geometry.xyz (see Figure 7). To address this, MLE-STAR employs  $\mathcal{A}_{\mathrm{data}}$ , which reexamines the task description to ensure that all given data is utilized. As shown in Figure 7, this design enables MLE-STAR to incorporate previously neglected data. As a result, performance is significantly improved, as shown in Table 8.
+
+Progressive improvement via MLE-STAR refinement. This section details the progressive improvement of solutions achieved by MLE-STAR, as measured by validation metrics. Given the task-specific nature of evaluation metrics, we report the average relative error reduction (\%) across the all 22 challenges in MLE-bench Lite [16]. This metric measures the extent to which MLE-STAR reduces the error of an initial solution. Figure 8 demonstrates a consistent improvement as MLE-STAR proceeds through its refinement steps, which each step focusing on refining a single code block via an inner loop. Significantly, the magnitude of improvement is notable in the early refinement stages. We posit that this stems from MLE-STAR's ablation study module which helps to target the most influential code blocks for modification first.
+
+Discussion on potential plagiarism. Following MLE-bench, we utilize Dolos [48], a source code plagiarism detection tool, to analyze generated solution code by MLE-STAR, against the top associated notebooks (i.e., Jupyter notebook) from each Kaggle competition. Our analysis, summarized in Table 13 (see Appendix L), shows that no final solution code and notebook pair exceeded a  $60\%$  similarity score (i.e., a criteria suggested from the MLE-bench paper), indicating no detected instances of plagiarism. This also shows that MLE-STAR's solution is sufficiently new compared to the existing solutions in Kaggle.
+
+# 6 Conclusion
+
+We propose MLE-STAR, a novel MLE agent designed for various ML tasks. Our key idea is to utilize a search engine to retrieve effective models and then explore various strategies targeting specific ML pipeline components to improve the solution. The effectiveness of MLE-STAR is validated by winning medals in  $64\%$  (where  $36\%$  are gold medals) of the MLE-bench Kaggle competitions.
+
+Limitation. We acknowledge that MLE-STAR requires higher cost due to increased token usage. We include corresponding cost analysis in Appendix K. However, it is worth to note that still, with Gemini-2.0-Flash, the cost of MLE-STAR is only about \(0.24 per each ML challenge. In addition, since Kaggle competitions are publicly accessible, there is a potential risk that LLMs might have been trained with the relevant discussions about the challenge. Nevertheless, we show that MLE-STAR's solution is sufficiently novel (using LLM as a judge) compared to the discussions on Kaggle (see Appendix H), and also show that its similarity compared to notebooks on Kaggle does not exceed \(60\%\), alleviating such plagiarism issue (see Appendix L).
+
+# Acknowledgements and disclosure of funding
+
+We would like to thank Raj Sinha, Subin Kim, Changyeon Kim, Dongjun Lee, Jihoon Tack, and anonymous reviewers for their helpful feedback and discussions. This work was partly supported by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program (KAIST)) and ITRC (Information Technology Research Center) grant funded by the Korea government (MSIT) (IITP-2025-RS-2024-00436857, 50%).
+
+# References
+
+[1] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016.  
+[2] Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. Catboost: unbiased boosting with categorical features. Advances in Neural Information Processing Systems, 2018.  
+[3] Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model. Nature, 2025.  
+[4] Linwei Fan, Fan Zhang, Hui Fan, and Caiming Zhang. Brief review of image denoising techniques. Visual computing for industry, biomedicine, and art, 2019.  
+[5] Noah Hollmann, Samuel Müller, and Frank Hutter. Large language models for automated data science: Introducing caafe for context-aware automated feature engineering. Advances in Neural Information Processing Systems, 2023.  
+[6] Jaehyun Nam, Kyuyoung Kim, Seunghyuk Oh, Jihoon Tack, Jaehung Kim, and Jinwoo Shin. Optimized feature generation for tabular data via llms with decision tree reasoning. Advances in Neural Information Processing Systems, 2024.  
+[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 2020.  
+[8] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.  
+[9] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024.  
+[10] Siyuan Guo, Cheng Deng, Ying Wen, Hechang Chen, Yi Chang, and Jun Wang. DS-agent: Automated data science by empowering large language models with case-based reasoning. International Conference on Machine Learning, 2024.  
+[11] Sirui Hong, Yizhang Lin, Bang Liu, Bangbang Liu, Binhao Wu, Ceyao Zhang, Chenxing Wei, Danyang Li, Jiaqi Chen, Jiayi Zhang, et al. Data interpreter: An llm agent for data science. arXiv preprint arXiv:2402.18679, 2024.  
+[12] Zhengyao Jiang, Dominik Schmidt, Dhruv Srikanth, Dixing Xu, Ian Kaplan, Deniss Jacenko, and Yuxiang Wu. Aide: AI-driven exploration in the space of code. arXiv preprint arXiv:2502.13138, 2025.  
+[13] Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. Swe-bench: Can language models resolve real-world github issues? International Conference on Learning Representations, 2024.
+
+[14] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livocodebench: Holistic and contamination free evaluation of large language models for code. International Conference on Learning Representations, 2025.  
+[15] Fabian Pedregosa, Géel Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of Machine Learning Research, 2011.  
+[16] Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, et al. Mle-bench: Evaluating machine learning agents on machine learning engineering. International Conference on Learning Representations, 2025.  
+[17] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. International Conference on Learning Representations, 2023.  
+[18] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 2023.  
+[19] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv: Arxiv-2305.16291, 2023.  
+[20] Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Leblond, Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. Competition-level code generation with alphabet. Science, 2022.  
+[21] Matthias Feurer, Katharina Eggensperger, Stefan Falkner, Marius Lindauer, and Frank Hutter. Auto-sklearn 2.0: Hands-free automl via meta-learning. Journal of Machine Learning Research, 2022.  
+[22] Erin LeDell and Sebastien Poirier. H2O AutoML: Scalable automatic machine learning. ICML Workshop on AutoML, 2020.  
+[23] Haifeng Jin, Qingquan Song, and Xia Hu. Auto-keras: An efficient neural architecture search system. ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2019.  
+[24] Lars Kotthoff, Chris Thornton, Holger H Hoos, Frank Hutter, and Kevin Leyton-Brown. Autoweka 2.0: Automatic model selection and hyperparameter optimization in weka. Journal of Machine Learning Research, 2017.  
+[25] Randal S Olson and Jason H Moore. Tpot: A tree-based pipeline optimization tool for automating machine learning. ICML Workshop on AutoML, 2016.  
+[26] Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. Autogluon-tabular: Robust and accurate automl for structured data. arXiv preprint arXiv:2003.06505, 2020.  
+[27] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. International Conference on Machine Learning, 2018.  
+[28] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. International Conference on Learning Representations, 2017.  
+[29] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. AAAI Conference on Artificial Intelligence, 2019.  
+[30] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. Journal of Machine Learning Research, 2019.
+
+[31] Wei Fan, Erheng Zhong, Jing Peng, Olivier Verscheure, Kun Zhang, Jiangtao Ren, Rong Yan, and Qiang Yang. Generalized and heuristic-free feature construction for improved accuracy. SIAM International Conference on Data Mining, 2010.  
+[32] James Max Kanter and Kalyan Veeramachaneni. Deep feature synthesis: Towards automating data science endeavors. IEEE International Conference on Data Science and Advanced Analytics, 2015.  
+[33] Liyao Li, Haobo Wang, Liangyu Zha, Qingyi Huang, Sai Wu, Gang Chen, and Junbo Zhao. Learning a data-driven policy network for pre-training automated feature engineering. International Conference on Learning Representations, 2023.  
+[34] Franziska Horn, Robert Pack, and Michael Rieger. The autofocus python library for automated feature engineering and selection. Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2019.  
+[35] Tianping Zhang, Zheyu Aqa Zhang, Zhiyuan Fan, Haoyan Luo, Fengyuan Liu, Qian Liu, Wei Cao, and Li Jian. Openfe: Automated feature generation with expert-level performance. International Conference on Machine Learning, 2023.  
+[36] Samuel Schmidgall, Yusheng Su, Ze Wang, Xineng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Zicheng Liu, and Emad Barsoum. Agent laboratory: Using llm agents as research assistants. arXiv preprint arXiv:2501.04227, 2025.  
+[37] Ziming Li, Qianbo Zang, David Ma, Jiawei Guo, Tuney Zheng, Minghao Liu, Xinyao Niu, Yue Wang, Jian Yang, Jiaheng Liu, et al. Autokaggle: A multi-agent framework for autonomous data science competitions. arXiv preprint arXiv:2410.20424, 2024.  
+[38] Qian Huang, Jian Vora, Percy Liang, and Jure Leskovec. Mlagentbench: Evaluating language agents on machine learning experimentation. International Conference on Machine Learning, 2024.  
+[39] Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. Openhands: An open platform for ai software developers as generalist agents. International Conference on Learning Representations, 2024.  
+[40] Janet L Kolodner. An introduction to case-based reasoning. Artificial intelligence review, 1992.  
+[41] Ian Watson and Farhi Marir. Case-based reasoning: A review. The knowledge engineering review, 1994.  
+[42] Yuki Ichihara, Yuu Jinnai, Tetsuro Morimura, Kenshi Abe, Kaito Ariu, Mitsuki Sakamoto, and Eiji Uchibe. Evaluation of best-of-n sampling strategies for language model alignment. Transactions on Machine Learning Research, 2025.  
+[43] Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025.  
+[44] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2016.  
+[45] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. International Conference on Machine Learning, 2019.  
+[46] 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. International Conference on Learning Representations, 2021.  
+[47] David Holzmüller, Léo Grinsztajn, and Ingo Steinwart. Better by default: Strong pre-tuned mlp's and boosted trees on tabular data. Advances in Neural Information Processing Systems, 2024.
+
+[48] Rien Maertens, Charlotte Van Petegem, Niko Strijbol, Toon Baeyens, Arne Carla Jacobs, Peter Dawyndt, and Bart Mesuere. Dolos: Language-agnostic plagiarism detection in source code. Journal of Computer Assisted Learning, 2022.  
+[49] Liqiang Jing, Zhehui Huang, Xiaoyang Wang, Wenlin Yao, Henhao Yu, Kaixin Ma, Hongming Zhang, Xinyau Du, and Dong Yu. Dsbench: How far are data science agents to becoming data science experts? International Conference on Learning Representations, 2025.  
+[50] Yiming Huang, Jianwen Luo, Yan Yu, Yitong Zhang, Fangyu Lei, Yifan Wei, Shizhu He, Lifu Huang, Xiao Liu, Jun Zhao, et al. Da-code: Agent data science code generation benchmark for large language models. arXiv preprint arXiv:2410.07331, 2024.  
+[51] Xueyu Hu, Ziyu Zhao, Shuang Wei, Ziwei Chai, Qianli Ma, Guoyin Wang, Xuwu Wang, Jing Su, Jingjing Xu, Ming Zhu, et al. Infiagent-dabench: Evaluating agents on data analysis tasks. arXiv preprint arXiv:2401.05507, 2024.  
+[52] Ziming You, Yumiao Zhang, Dexuan Xu, Yiwei Lou, Yandong Yan, Wei Wang, Huaming Zhang, and Yu Huang. Databwiseagent: A notebook-centric llm agent framework for automated data science. arXiv preprint arXiv:2503.07044, 2025.
+
+# 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: All claims in the introduction and abstract accurately reflect the contribution and scope.
+
+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: We discuss in Section 6.
+
+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: We do not have a theory in this paper.
+
+# 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: We provide implementation details in Section 4. We also provide all prompts we use in Appendix A.
+
+# 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: [No]
+
+Justification: The benchmark is already open-sourced, but we do not currently submit code when submitting.
+
+# 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: We provide the details in Section 4 and Appendix F.
+
+# 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: All experiments are conducted with multiple seeds, and we report one standard error of the mean.
+
+# 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: We provide compute resources we used in Appendix F.
+
+# 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: We do not have any ethical concerns.
+
+# 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: [Yes]
+
+Justification: We discuss the broader impacts in Appendix I.
+
+# 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: [NA]
+
+Justification: Our framework does not introduce 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: We have cited all papers and datasets in Reference.
+
+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 URL.  
+- 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: [NA]
+
+Justification: We do not release new assets.
+
+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: We do 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: We do not have human subject.
+
+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.
+
+# 16. Declaration of LLM usage
+
+Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required.
+
+# Answer: [NA]
+
+Justification: The core method development does not involve LLMs as any important, original, or non-standard components.
+
+# Guidelines:
+
+- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.  
+- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described.

2025/MLE-STAR_ Machine Learning Engineering Agent via Search and Targeted Refinement/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca39ea2750659bc3c2e3b9ca97c13f4a0db9e6c6d58c9c6b06b6009f5fa24a90
+size 499793

2025/MLEP_ Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection/b8823069-4da9-4af8-a7f8-9ed80f65b494_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e7ce01250ea1a4e24af0f8590aace35868ef6eb42e50092cf58b030f71d820a
+size 31169158

2025/MLEP_ Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection/full.md

@@ -0,0 +1,525 @@
+# MLEP: Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection
+
+Lin Yuan, Xiaowan Li, Yan Zhang* Jiawei Zhang, Hongbo Li, Xinbo Gao* Chongqing Key Laboratory of Image Cognition, Chongqing University of Posts and Telecommunications, Chongqing 400065, China yuanlin@cqupt.edu.cn, s230201063@stu.cqupt.edu.cn, {yanzhang1991, zhangjw, lihongbo, gaoxb}@cqupt.edu.cn
+
+# Abstract
+
+Advances in image generation technologies have raised growing concerns about their potential misuse, particularly in producing misinformation and deepfakes. This creates an urgent demand for effective methods to detect AI-generated images (AIGIs). While progress has been made, achieving reliable performance across diverse generative models and scenarios remains challenging due to the absence of source-invariant features and the limited generalization of existing approaches. In this study, we investigate the potential of using image entropy as a discriminative cue for AIGI detection and propose Multi-granularity Local Entropy Patterns (MLEP), a set of feature maps computed based on Shannon entropy from shuffled small patches at multiple image scales. MLEP effectively captures pixel dependencies across scales and dimensions while disrupting semantic content, thereby reducing potential content bias. Based on MLEP, we can easily build a robust CNN-based classifier capable of detecting AIGIs with enhanced reliability. Extensive experiments in an open-world setting, involving images synthesized by 32 distinct generative models, demonstrate that our approach achieves substantial improvements over state-of-the-art methods in both accuracy and generalization.
+
+# 1 Introduction
+
+The rapid development of generative technologies has transformed image synthesis, with models like GAN [1], diffusion model [2], and their variants achieving impressive realism. While enabling new applications in creative industries, these advancements have also raised concerns over misuse in misinformation and deepfakes [3, 4], prompting an urgent need for reliable AI-generated image (AIGI) detection methods. Researchers have leveraged spatial [5, 6, 7, 8] and frequency-domain cues [9, 10, 11, 12], as well as high-level knowledge from pretrained diffusion models [13, 14] and LLMs [15, 16, 17] for AIGI detection. Yet, the lack of source-invariant representations still limits the cross-domain detection robustness, especially when across different models and content types.
+
+To address this challenge, we aim to identify a generalized, content-agnostic pattern that can reliably distinguish AIGIs from real photographs. Inspired by recent studies [7, 8], our work builds on two key observations. Tan et al.[7] found that generative models typically involve internal upsampling operations and propose Neighboring Pixel Relationships (NPR) to capture resulting structural artifacts. However, NPR operates on small local patches and retains visible semantic structures, introducing bias that may hinder generalization. Zheng et al.[8] emphasized the impact of "semantic artifacts" on detection and propose disrupting image semantics by shuffling  $32 \times 32$  patches. While this improves cross-scene generalization, the relatively large patch size still preserves semantic information. We argue that such artifacts persist and continue to limit content-agnostic detection.
+
+![](images/3b04b7f195200a7add7d29355f30dc449cefd05df01342e24daeac11a67dd7c0.jpg)  
+Figure 1: Comparison of local entropy distributions between real and AI-generated images using  $2 \times 2$  patches, with entropy values from  $\{0, 0.8, 1.0, 1.5, 2.0\}$ . Real images consistently show a higher likelihood of entropy reaching 2.0.
+
+Through large-scale subjective observation, we noticed a distinct "glossy and smooth" texture in AI-generated images, prompting an investigation into their entropy characteristics, a statistical measure of pixel randomness [18]. We conducted a preliminary study comparing local entropy distributions (using  $2 \times 2$  patches) between real and AI-generated images. As shown in Fig. 1, real images consistently exhibit a higher probability of maximum entropy (2.0), suggesting the potential of entropy as a discriminative feature for AIGI detection. Motivated by this, we propose using image entropy as an alternative to pixel differences as proposed by NPR [7]. Entropy captures pixel relationships while reducing semantic dependency by focusing on pixel value distributions rather than contrasts. To further suppress semantic artifacts, we adopt fine-grained patch shuffling (smaller than the  $32 \times 32$  patches used in [8]), which also reduces the computational overhead of entropy computation. Additionally, we incorporate multi-scale resampling and an overlapping sliding window to enhance the granularity of entropy patterns. Our contributions are summarized as follows:
+
+- To the best of our knowledge, it is the first attempt to explore the potential of image entropy as a cue for detecting AI-generated images. Using image entropy not only enhances detection accuracy and generalization compared to state-of-the-art methods but also highlights intrinsic differences between real and AI-generated images in terms of pixel randomness, as quantified by image entropy.  
+- We propose Multi-granularity Local Entropy Patterns (MLEP), a set of feature maps with entropy computed from shuffled small patches across multiple resampling scales. MLEP effectively disrupts image semantics to mitigate content bias, while capturing pixel relationships across both spatial and scale dimensions. Using MLEP as input, a standard CNN classifier can be trained for robust and generalized AIGI detection.  
+- Extensive quantitative and qualitative analyses validate the effectiveness of the MLEP design, showing significant improvements over state-of-the-art methods across multiple AI-generated image datasets.
+
+# 2 Related Work
+
+Spatial-domain Detection Spatial-domain methods typically rely on handcrafted spatial features, local patterns, or pixel statistics to distinguish between real and generated images. The representative methods include generalized feature extraction from CNN-based model [5] and inter-pixel correlation between rich and poor texture regions [6]. Tan et al. [7] observed that upsampling operations are prevalent in image generation models and proposed utilizing neighboring pixel relationships (NPR), computed through local pixel differences, as a simple yet effective cue for AIGI detection. Zheng et al. [8] discovered that image semantic information negatively impacts detection performance and proposed a simple linear classifier that utilizes image patch shuffling to disrupt the original semantic artifacts. Cozzolino et al. [19] proposed a zero-shot detection method that models the distribution of real images using lossless coding with multi-resolution prediction, identifying AI-generated images by detecting higher-than-expected coding costs that indicate deviations from real-image statistics. Yang et al. [20] proposed Discrepancy Deepfake Detector  $(\mathrm{D}^3)$ , which enhances cross-generator generalization by introducing a parallel branch that extracts a discrepancy signal from
+
+distorted features to complement the original representation, achieving better robustness without compromising in-domain performance.
+
+Frequency-domain Detection To tackle the subtlety of spatial artifacts in AI-generated images, frequency-domain methods analyze image frequency components, enabling more effective real vs. fake differentiation. The study in [9] found that GAN-generated images exhibit generalized artifacts in discrete cosine transform (DCT) spectrum, which can be readily identified. Qian et al. [10] proposed a face forgery detection network based on frequency-aware decomposed image components and local frequency statistics. Luo et al. [11] proposed a feature representation based on high-frequency noise at multiple scales and enhanced detection performance by integrating it with an attention module. Liu et al. [12] utilized noise patterns in the frequency domain as feature representations for detecting AI-generated images. Tan et al. [21] proposed FreqNet toward detection generalizability, which focuses on high-frequency components of images, exploiting high-frequency representation across spatial and channel dimension.
+
+Detection leveraging Pretrained Models This group of methods aims to derive generalized features for AIGI detection by leveraging the knowledge learned by large models pretrained on extensive datasets. Wang et al. [13] proposed an artifact representation named DDiffusion Reconstruction Error (DIRE), which obtains the difference between the input image and its reconstructed object through a pretrained diffusion model. Chen et al. [14] proposed utilizing pretrained diffusion models to generate high-quality synthesized images, serving as challenging samples to enhance the detector's performance. Ojha et al. [15] utilized representations from a fixed pretrained CLIP model as generalized features for detection. Chen et al. [22] introduced ForgeLens, a data-efficient CLIP-ViT framework that enhances generalization to unseen forgeries by using proposed lightweight weight shared guidance module (WSGM) and forgery-aware feature integrator (FAFormer) to guide frozen features toward forgery-relevant information. Zhang et al. [23] proposed VIB-Net, which employs variational information bottlenecks to enforce authentication task-related feature learned from pretrained CLIP encoder, significantly improving generalization across different generative model types. Similar methods such as [16, 17] also leveraged textual information from vision-language models to further enhance detection performance.
+
+# 3 The Approach
+
+The proposed approach leverages entropy-based feature extraction to analyze local pixel randomness in a multi-granularity, semantic-agnostic manner. It begins by dividing the image into small patches and applying random shuffling to reduce semantic bias. A multi-scale pyramid is then constructed by downsampling and upsampling the scrambled image, introducing resampling artifacts. Local entropy is computed using a  $2 \times 2$  sliding window across the entire image, capturing complexity across intra-block, inter-block, and inter-scale levels. The resulting multi-granularity local entropy patterns (MLEP) are used as input to a standard CNN classifier for distinguishing AI-generated from real images. An overview of the method is shown in Fig. 2, with key components detailed below.
+
+# 3.1 Semantic Suppression via Patch Shuffling
+
+Inspired by previous work [6, 8] that mitigates semantic bias via patch-based processing, we adopt finer patch shuffling to further disrupt image content. Given an input image  $X \in \mathbb{R}^{H \times W \times C}$ , we first partition it into patches of uniform size of  $L \times L$ :
+
+$$
+X = \left\{X _ {i, j} \in \mathbb {R} ^ {L \times L \times C} \right\} _ {1 \leq i \leq \frac {H}{L}, 1 \leq j \leq \frac {W}{L}}, \tag {1}
+$$
+
+where  $L$  is a small integer (typically  $< 8$ ), and  $H, W$  are assumed divisible by  $L$ . The patches are then randomly permuted, resulting in a visually scrambled image denoted as  $\tilde{X}$ :
+
+$$
+\tilde {X} = \left\{\tilde {X} _ {\pi (i, j)} = X _ {i, j} \right\} _ {1 \leq i \leq \frac {H}{L}, 1 \leq j \leq \frac {W}{L}}, \tag {2}
+$$
+
+where  $\pi$  is a bijection defining the patch permutation. Note that partitioning and shuffling are applied independently to each color channel.
+
+![](images/ee8dee0deca2eff7d288d4716907897a0fae1c28f774ee4541e4cb94ae92dae7.jpg)  
+Figure 2: Illustration of AI-generated image detection using multi-granularity local entropy patterns (MLEP), which involves three core steps to obtain the MLEP feature: Patch Shuffling, Multi-Scale Resampling, and Local Entropy Pattern computation. The resulting MLEP features are then fed into a CNN-based classifier (e.g., ResNet) to effectively identify AI-generated images.
+
+# 3.2 Multi-Scale Resampling
+
+Inspired by [7] showing that generative models often use upsampling to produce high-resolution outputs, we propose detecting generation artifacts via multi-scale analysis. We hypothesize that resampling generated images reveals distinctive patterns useful for detection. To this end, we first construct a multi-scale pyramid by resampling the scrambled image  $\tilde{X}$  with scale factors  $\mathbb{S} = \{s_1,s_2,\dots ,s_K\}$ , with each scale  $s_k\in (0,1]$  applied using an interpolation function  $\mathrm{Down}(\cdot ,s_k)$ :
+
+$$
+\tilde {X} _ {\vee} ^ {(k)} = \operatorname {D o w n} (\tilde {X}, s _ {k}), \quad \tilde {X} _ {\vee} ^ {(k)} \in \mathbb {R} ^ {\lfloor s _ {k} \cdot H \rfloor \times \lfloor s _ {k} \cdot W \rfloor \times C}, \tag {3}
+$$
+
+which are then upsampled back to its original shape using an interpolation function  $\mathrm{Up}(\cdot ,H,W)$
+
+$$
+\tilde {X} _ {\wedge} ^ {(k)} = \operatorname {U p} \left(\tilde {X} _ {\vee} ^ {(k)}, H, W\right), \quad \tilde {X} _ {\wedge} ^ {(k)} \in \mathbb {R} ^ {H \times W \times C}. \tag {4}
+$$
+
+The resulting multi-scale resampling image  $\hat{X}$  is created by concatenating all the upsampled images along the channel dimension:
+
+$$
+\hat {X} = \operatorname {C o n c a t} \left(\tilde {X} _ {\wedge} ^ {(1)}, \tilde {X} _ {\wedge} ^ {(2)}, \dots , \tilde {X} _ {\wedge} ^ {(K)}\right), \quad \hat {X} \in \mathbb {R} ^ {H \times W \times (C \cdot K)}. \tag {5}
+$$
+
+# 3.3 Multi-granularity Local Entropy Patterns
+
+The core of our approach is the design of Local Entropy Patterns (LEP), which quantify textural randomness using a  $2 \times 2$  sliding window over pixel sets  $\hat{X}_{i,j} = \{x_{m,n}\}_{m\in \{i,i + 1\},n\in \{j,j + 1\}}$ , based on Shannon's definition of information entropy [18]:
+
+$$
+\operatorname {L E P} \left(\hat {X} _ {i, j}\right) = - \sum_ {m, n} p \left(x _ {m, n}\right) \cdot \log_ {2} p \left(x _ {m, n}\right), \tag {6}
+$$
+
+where  $p(x_{m,n})$  represents the probability of occurrence of the pixel value  $x_{m,n}$  within that specific patch  $\hat{X}_{i,j}$ . By restricting the sliding window to  $2 \times 2$  (four pixels), entropy values are confined to five discrete levels:  $\mathbb{V} = \{0,0.8,1.0,1.5,2\}$ , as shown in Fig. 3c. The proof and an efficient computation algorithm for LEP on a  $2 \times 2$  window are provided in the supplementary material. With a stride of 1, the  $2 \times 2$  sliding window introduces overlap in LEP computation. Due to patch shuffling, this captures both intra-patch and inter-patch entropy—reflecting local randomness within and across original image regions—as illustrated in Fig. 3b. Applied across multiple scales, LEP further captures inter-scale entropy, forming the basis of the final Multi-granularity Local Entropy Patterns (MLEP).
+
+Given the computed MLEP feature maps denoted as  $\bar{X} \in \mathbb{V}^{(H - 1) \times (W - 1) \times (C \cdot K)}$ , a representative CNN-based classifier can be trained to differentiate between photographic and AI-generated images. Denoting the classifier as  $f$ , the training objective is defined using the binary cross-entropy loss:
+
+$$
+\mathcal {L} _ {B C E} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ y _ {i} \log \left(f \left(\bar {X} _ {i}\right)\right) + \left(1 - y _ {i}\right) \log \left(1 - f \left(\bar {X} _ {i}\right)\right) \right], \tag {7}
+$$
+
+where  $y_{i}$  represents the true labels,  $f(\bar{X}_i)$  the predictions, and  $N$  the number of training samples.
+
+![](images/8c28e7cc9f173de8aa2d6ff625a4e1ab4552acf8b0c89c8d92fbd1576ac5c41b.jpg)  
+(a) LEP computation based a  $2 \times 2$  window sliding over the entire image.
+
+![](images/92f22d8413cd1ee0cb0f876ba5cbaa1a6e848282744c0723c581824bbc1426d3.jpg)  
+(b) Illustration of intra-block and interblock LEP due to image patch shuffling.
+
+![](images/dede735ac553f0e0c21058a7c1e7b37bd5c0815e8a6a0649b38ec88299cf0827.jpg)  
+(c) Five possible LEP values corresponding to different pixel occurrences within a  $2 \times 2$  window.  
+Figure 3: Illustration of the MLEP computation.
+
+# 4 Experiments
+
+# 4.1 Experimental Settings
+
+Datasets We adopt the cross-dataset setup from [7], using the ForenSynths [5] dataset for training, which includes 20 content categories with 18,000 ProGAN [24] generated images and an equal number of real images from LSUN [25]. Following [7], we train only on four categories: cars, cats, chairs, and horses, posing a challenging cross-scene setting. Following [5, 13, 7, 8], we evaluate on synthesized images from 32 image generation models (16 GAN-based and 16 Diffusion-based, including variants). The GAN-Set includes ProGAN [24], StyleGAN [26], StyleGAN2 [27], BigGAN [28], CycleGAN [29], StarGAN [30], GauGAN [31], AttGAN [32], BEGAN [33], CramerGAN [34], InfoMaxGAN [35], MMDGAN [36], RelGAN [37], S3GAN [38], SNGAN [39], and STGAN [40], with the former seven obtained from the dataset ForenSynths [5] and the latter nine from the dataset GANGen-Detection [41]. The Diffusion-Set contains DDPM [2], IDDPM [42], ADM [43], LDM [44], PNDM [45], VQ-Diffusion [46], Stable Diffusion (SD) v1/v2 [44], DALL-E mini [47], three Glide [48] variants², and two LDM [44] variants³. Of these models, the first eight are sourced from the DiffusionForensics dataset [13], while the remainder are from the UniversalFakeDetect dataset [15]. Furthermore, we include images from two commercial models, Midjourney and DALL-E 2, sourced from the social platform Discord⁴ as provided by [7]. Each above AIGI subset comprises an equal number of real samples paired with the corresponding generative counterparts. All test images were obtained according to the instructions provided by [7].
+
+Implementation details All images were resized to  $224 \times 224$ , with random cropping for training and center cropping for testing. Multiple variants of the patch size  $(L)$ , resampling scales  $(\mathbb{S})$ , and the classifier backbone were tested with results shown in the ablation study. The training was performed using the Adam optimizer (learning rate of 0.002, batch size of 64). All experiments ran on a server with two NVIDIA RTX A5000 GPUs.
+
+Baseline methods We compare against representative baselines, including CNNDet [5], F3Net [10], LGrad [49], UnivFD [15], CLIPping [16], NPR [7], Zheng [8], FreqNet [21], FatFormer [17],
+
+Table 1: Detection performance in terms of Acc.  $(\%)$  and A.P.  $(\%)$  on the GAN-based datasets.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">ProGAN</td><td colspan="2">StyleGAN</td><td colspan="2">StyleGAN2</td><td colspan="2">BigGAN</td><td colspan="2">CycleGAN</td><td colspan="2">StarGAN</td><td colspan="2">GauGAN</td><td colspan="2">AttGAN</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>CNNDet [5]</td><td>91.4</td><td>99.4</td><td>63.8</td><td>91.4</td><td>76.4</td><td>97.5</td><td>52.9</td><td>73.3</td><td>72.7</td><td>88.6</td><td>63.8</td><td>90.8</td><td>63.9</td><td>92.2</td><td>51.1</td><td>83.7</td></tr><tr><td>F3Net [10]</td><td>99.4</td><td>100.0</td><td>92.6</td><td>99.7</td><td>88.0</td><td>99.8</td><td>65.3</td><td>69.9</td><td>76.4</td><td>84.3</td><td>100.0</td><td>100.0</td><td>58.1</td><td>56.7</td><td>85.2</td><td>94.8</td></tr><tr><td>LGrad [49]</td><td>99.0</td><td>100.0</td><td>94.8</td><td>99.9</td><td>96.0</td><td>99.9</td><td>82.9</td><td>90.7</td><td>85.3</td><td>94.0</td><td>99.6</td><td>100.0</td><td>72.4</td><td>79.3</td><td>68.6</td><td>93.8</td></tr><tr><td>Ojha [15]</td><td>99.7</td><td>100.0</td><td>89.0</td><td>98.7</td><td>83.9</td><td>98.4</td><td>90.5</td><td>99.1</td><td>87.9</td><td>99.8</td><td>91.4</td><td>100.0</td><td>89.9</td><td>100.0</td><td>78.5</td><td>91.3</td></tr><tr><td>Zheng [8]</td><td>99.7</td><td>100.0</td><td>90.7</td><td>95.3</td><td>97.6</td><td>99.7</td><td>67.0</td><td>67.6</td><td>85.2</td><td>92.6</td><td>98.7</td><td>100.0</td><td>57.1</td><td>56.8</td><td>79.4</td><td>87.7</td></tr><tr><td>CLIPping [16]</td><td>99.8</td><td>100.0</td><td>94.3</td><td>99.4</td><td>83.5</td><td>98.7</td><td>93.8</td><td>99.4</td><td>95.4</td><td>99.9</td><td>99.1</td><td>100.0</td><td>93.4</td><td>99.9</td><td>91.3</td><td>97.4</td></tr><tr><td>NPR [7]</td><td>99.8</td><td>100.0</td><td>96.3</td><td>99.8</td><td>97.3</td><td>100.0</td><td>87.5</td><td>94.5</td><td>95.0</td><td>99.5</td><td>99.7</td><td>100.0</td><td>86.6</td><td>88.8</td><td>83.0</td><td>96.2</td></tr><tr><td>FreqNet [21]</td><td>99.6</td><td>100.0</td><td>90.2</td><td>99.7</td><td>87.9</td><td>99.5</td><td>90.5</td><td>96.0</td><td>95.8</td><td>99.6</td><td>85.6</td><td>99.8</td><td>93.4</td><td>98.6</td><td>89.8</td><td>98.8</td></tr><tr><td>FatFormer [17]</td><td>99.9</td><td>100.0</td><td>97.1</td><td>99.8</td><td>98.8</td><td>99.9</td><td>99.5</td><td>100.0</td><td>99.4</td><td>100.0</td><td>99.8</td><td>100.0</td><td>99.4</td><td>100.0</td><td>99.3</td><td>100.0</td></tr><tr><td>ForgeLens [22]</td><td>99.9</td><td>100.0</td><td>90.3</td><td>98.7</td><td>94.2</td><td>98.8</td><td>98.9</td><td>99.0</td><td>99.6</td><td>99.6</td><td>99.8</td><td>100.0</td><td>99.1</td><td>99.4</td><td>90.1</td><td>90.0</td></tr><tr><td>D3[20]</td><td>99.4</td><td>100.0</td><td>94.9</td><td>99.2</td><td>95.6</td><td>99.4</td><td>99.1</td><td>100.0</td><td>92.6</td><td>98.5</td><td>95.7</td><td>99.3</td><td>97.9</td><td>99.9</td><td>84.8</td><td>92.9</td></tr><tr><td>VIBAIGC [23]</td><td>99.9</td><td>100.0</td><td>89.0</td><td>98.5</td><td>87.0</td><td>97.2</td><td>95.3</td><td>99.1</td><td>98.7</td><td>99.7</td><td>97.7</td><td>99.6</td><td>99.3</td><td>99.9</td><td>93.4</td><td>98.1</td></tr><tr><td>Ours</td><td>99.6</td><td>100.0</td><td>99.6</td><td>100.0</td><td>99.9</td><td>100.0</td><td>87.1</td><td>93.6</td><td>98.3</td><td>99.3</td><td>100.0</td><td>100.0</td><td>82.0</td><td>87.9</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan="2">Method</td><td colspan="2">BEGAN</td><td colspan="2">CramerGAN</td><td colspan="2">InfoMaxGAN</td><td colspan="2">MMDGAN</td><td colspan="2">RelGAN</td><td colspan="2">S3GAN</td><td colspan="2">SNGAN</td><td colspan="2">STGAN</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>CNNDet [5]</td><td>50.2</td><td>44.9</td><td>81.5</td><td>97.5</td><td>71.1</td><td>94.7</td><td>72.9</td><td>94.4</td><td>53.3</td><td>82.1</td><td>55.2</td><td>66.1</td><td>62.7</td><td>90.4</td><td>63.0</td><td>92.7</td></tr><tr><td>F3Net [10]</td><td>87.1</td><td>97.5</td><td>89.5</td><td>99.8</td><td>67.1</td><td>83.1</td><td>73.7</td><td>99.6</td><td>98.8</td><td>100.0</td><td>65.4</td><td>70.0</td><td>51.6</td><td>93.6</td><td>60.3</td><td>99.9</td></tr><tr><td>LGrad [49]</td><td>69.9</td><td>89.2</td><td>50.3</td><td>54.0</td><td>71.1</td><td>82.0</td><td>57.5</td><td>67.3</td><td>89.1</td><td>99.1</td><td>78.5</td><td>86.0</td><td>78.0</td><td>87.4</td><td>54.8</td><td>68.0</td></tr><tr><td>Ojha [15]</td><td>72.0</td><td>98.9</td><td>77.6</td><td>99.8</td><td>77.6</td><td>98.9</td><td>77.6</td><td>99.7</td><td>78.2</td><td>98.7</td><td>85.2</td><td>98.1</td><td>77.6</td><td>98.7</td><td>74.2</td><td>97.8</td></tr><tr><td>Zheng [8]</td><td>67.4</td><td>98.0</td><td>74.2</td><td>93.8</td><td>71.0</td><td>93.1</td><td>68.4</td><td>89.4</td><td>98.4</td><td>99.9</td><td>70.8</td><td>69.9</td><td>72.4</td><td>94.0</td><td>92.3</td><td>100.0</td></tr><tr><td>CLIPping [16]</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td><td>94.7</td><td>99.7</td><td>94.8</td><td>99.9</td><td>92.2</td><td>98.3</td><td>88.4</td><td>97.7</td><td>94.4</td><td>99.5</td><td>87.2</td><td>96.4</td></tr><tr><td>NPR [7]</td><td>99.0</td><td>99.8</td><td>98.7</td><td>99.0</td><td>94.5</td><td>98.3</td><td>98.6</td><td>99.0</td><td>99.6</td><td>100.0</td><td>79.0</td><td>80.0</td><td>88.8</td><td>97.4</td><td>98.0</td><td>100.0</td></tr><tr><td>FreqNet [21]</td><td>98.8</td><td>100.0</td><td>95.1</td><td>98.2</td><td>94.5</td><td>97.3</td><td>95.1</td><td>98.2</td><td>100.0</td><td>100.0</td><td>88.4</td><td>94.3</td><td>85.3</td><td>90.5</td><td>98.8</td><td>100.0</td></tr><tr><td>FatFormer [17]</td><td>99.9</td><td>100.0</td><td>98.4</td><td>100.0</td><td>98.4</td><td>100.0</td><td>98.4</td><td>100.0</td><td>99.5</td><td>100.0</td><td>99.0</td><td>100.0</td><td>98.3</td><td>99.9</td><td>98.8</td><td>99.8</td></tr><tr><td>ForgeLens [22]</td><td>88.4</td><td>97.0</td><td>87.2</td><td>93.1</td><td>87.6</td><td>92.6</td><td>87.5</td><td>92.3</td><td>92.6</td><td>93.1</td><td>98.7</td><td>99.3</td><td>86.7</td><td>91.7</td><td>90.0</td><td>95.2</td></tr><tr><td>D3[20]</td><td>89.5</td><td>97.3</td><td>95.2</td><td>99.2</td><td>95.7</td><td>99.2</td><td>94.6</td><td>99.0</td><td>91.4</td><td>97.5</td><td>98.4</td><td>99.9</td><td>93.8</td><td>98.7</td><td>93.0</td><td>98.5</td></tr><tr><td>VIBAIGC [23]</td><td>96.7</td><td>99.5</td><td>95.3</td><td>99.0</td><td>90.5</td><td>96.5</td><td>95.3</td><td>98.7</td><td>94.7</td><td>98.7</td><td>94.6</td><td>98.9</td><td>93.4</td><td>98.2</td><td>82.4</td><td>92.5</td></tr><tr><td>Ours</td><td>99.4</td><td>100.0</td><td>98.5</td><td>98.8</td><td>98.0</td><td>99.8</td><td>98.9</td><td>99.8</td><td>100.0</td><td>100.0</td><td>83.4</td><td>91.7</td><td>97.6</td><td>99.7</td><td>99.9</td><td>100.0</td></tr></table>
+
+Table 2: Detection performance in terms of Acc.  $(\%)$  and A.P.  $(\%)$  on the Diffusion-based datasets.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">ADM</td><td colspan="2">DDPM</td><td colspan="2">IDDPM</td><td colspan="2">LDM</td><td colspan="2">PNDM</td><td colspan="2">VQ-Diffusion</td><td colspan="2">SDv1</td><td colspan="2">SDv2</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>CNNDet [5]</td><td>53.9</td><td>71.8</td><td>62.7</td><td>76.6</td><td>50.2</td><td>82.7</td><td>50.4</td><td>78.7</td><td>50.8</td><td>90.3</td><td>50.0</td><td>71.0</td><td>38.0</td><td>76.7</td><td>52.0</td><td>90.3</td></tr><tr><td>F3Net [10]</td><td>80.9</td><td>96.9</td><td>84.7</td><td>99.4</td><td>74.7</td><td>98.9</td><td>100.0</td><td>100.0</td><td>72.8</td><td>99.5</td><td>100.0</td><td>100.0</td><td>73.4</td><td>97.2</td><td>99.8</td><td>100.0</td></tr><tr><td>LGrad [49]</td><td>86.4</td><td>97.5</td><td>99.9</td><td>100.0</td><td>66.1</td><td>92.8</td><td>99.7</td><td>100.0</td><td>69.5</td><td>98.5</td><td>96.2</td><td>100.0</td><td>90.4</td><td>99.4</td><td>97.1</td><td>100.0</td></tr><tr><td>Ojha [15]</td><td>78.4</td><td>92.1</td><td>72.9</td><td>78.8</td><td>75.0</td><td>92.8</td><td>82.2</td><td>97.1</td><td>75.3</td><td>92.5</td><td>83.5</td><td>97.7</td><td>56.4</td><td>90.4</td><td>71.5</td><td>92.4</td></tr><tr><td>Zheng [8]</td><td>72.1</td><td>78.9</td><td>78.9</td><td>80.5</td><td>49.9</td><td>52.0</td><td>99.7</td><td>100.0</td><td>90.4</td><td>96.9</td><td>99.6</td><td>100.0</td><td>94.0</td><td>99.7</td><td>87.9</td><td>96.4</td></tr><tr><td>CLIPping [16]</td><td>78.9</td><td>93.8</td><td>80.3</td><td>85.7</td><td>82.4</td><td>94.4</td><td>90.2</td><td>97.6</td><td>81.7</td><td>93.7</td><td>96.3</td><td>99.3</td><td>58.0</td><td>93.1</td><td>82.6</td><td>94.9</td></tr><tr><td>NPR [7]</td><td>88.6</td><td>98.9</td><td>99.8</td><td>100.0</td><td>91.8</td><td>99.8</td><td>100.0</td><td>100.0</td><td>91.2</td><td>100.0</td><td>100.0</td><td>100.0</td><td>97.4</td><td>99.8</td><td>93.8</td><td>100.0</td></tr><tr><td>FreqNet [21]</td><td>67.2</td><td>91.3</td><td>91.5</td><td>99.8</td><td>59.0</td><td>97.3</td><td>98.9</td><td>100.0</td><td>85.2</td><td>99.8</td><td>100.0</td><td>100.0</td><td>63.9</td><td>98.1</td><td>81.8</td><td>98.4</td></tr><tr><td>FatFormer [17]</td><td>70.8</td><td>93.4</td><td>67.2</td><td>72.5</td><td>69.3</td><td>94.3</td><td>97.3</td><td>100.0</td><td>99.3</td><td>100.0</td><td>100.0</td><td>100.0</td><td>61.7</td><td>96.8</td><td>84.4</td><td>98.2</td></tr><tr><td>ForgeLens [22]</td><td>69.8</td><td>92.4</td><td>52.1</td><td>52.3</td><td>62.1</td><td>75.8</td><td>99.6</td><td>100.0</td><td>83.4</td><td>97.1</td><td>99.5</td><td>100.0</td><td>93.2</td><td>99.8</td><td>63.2</td><td>87.6</td></tr><tr><td>D3[20]</td><td>89.0</td><td>97.8</td><td>85.2</td><td>94.4</td><td>87.7</td><td>96.2</td><td>88.2</td><td>96.4</td><td>90.0</td><td>96.8</td><td>96.1</td><td>99.9</td><td>98.0</td><td>99.8</td><td>93.6</td><td>98.8</td></tr><tr><td>VIBAIGC [23]</td><td>69.3</td><td>81.8</td><td>90.2</td><td>97.7</td><td>84.6</td><td>97.1</td><td>56.8</td><td>86.6</td><td>94.8</td><td>99.3</td><td>94.2</td><td>99.5</td><td>60.0</td><td>88.7</td><td>58.3</td><td>83.2</td></tr><tr><td>Ours</td><td>97.0</td><td>99.8</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td><td>99.8</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td><td>98.5</td><td>99.9</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan="2">Method</td><td colspan="2">DALL-E mini</td><td colspan="2">Glide-100-10</td><td colspan="2">Glide-100-27</td><td colspan="2">Glide-50-27</td><td colspan="2">LDM-200</td><td colspan="2">LDM-200cfg</td><td colspan="2">Midjourney</td><td colspan="2">DALL-E 2</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>CNNDet [5]</td><td>51.8</td><td>61.3</td><td>53.3</td><td>72.9</td><td>53.0</td><td>71.3</td><td>54.2</td><td>76.0</td><td>52.0</td><td>64.5</td><td>51.6</td><td>63.1</td><td>48.6</td><td>38.5</td><td>49.3</td><td>44.7</td></tr><tr><td>F3Net [10]</td><td>71.6</td><td>79.9</td><td>88.3</td><td>95.4</td><td>87.0</td><td>94.5</td><td>88.5</td><td>95.4</td><td>73.4</td><td>83.3</td><td>80.7</td><td>89.1</td><td>73.2</td><td>80.4</td><td>79.6</td><td>87.3</td></tr><tr><td>LGrad [49]</td><td>88.5</td><td>97.3</td><td>89.4</td><td>94.9</td><td>87.4</td><td>93.2</td><td>90.7</td><td>95.1</td><td>94.2</td><td>99.1</td><td>95.9</td><td>99.2</td><td>68.3</td><td>76.0</td><td>75.1</td><td>80.9</td></tr><tr><td>Ojha [15]</td><td>89.5</td><td>96.8</td><td>90.1</td><td>97.0</td><td>90.7</td><td>97.2</td><td>91.1</td><td>97.4</td><td>90.2</td><td>97.1</td><td>77.3</td><td>88.6</td><td>50.0</td><td>49.8</td><td>66.3</td><td>74.6</td></tr><tr><td>Zheng [8]</td><td>67.9</td><td>72.2</td><td>79.4</td><td>87.8</td><td>76.8</td><td>84.5</td><td>78.2</td><td>85.9</td><td>81.3</td><td>90.1</td><td>84.0</td><td>91.7</td><td>73.2</td><td>78.5</td><td>81.4</td><td>89.2</td></tr><tr><td>CLIPping [16]</td><td>91.1</td><td>98.6</td><td>92.0</td><td>98.6</td><td>91.2</td><td>98.8</td><td>94.3</td><td>99.3</td><td>92.8</td><td>98.9</td><td>77.4</td><td>94.3</td><td>51.1</td><td>50.7</td><td>62.6</td><td>72.3</td></tr><tr><td>NPR [7]</td><td>94.5</td><td>99.5</td><td>98.2</td><td>99.8</td><td>97.8</td><td>99.7</td><td>98.2</td><td>99.8</td><td>99.1</td><td>99.9</td><td>99.0</td><td>99.8</td><td>77.4</td><td>81.9</td><td>80.7</td><td>83.0</td></tr><tr><td>FreqNet [21]</td><td>97.4</td><td>99.8</td><td>88.1</td><td>96.4</td><td>84.5</td><td>96.1</td><td>86.7</td><td>96.3</td><td>97.5</td><td>99.9</td><td>97.4</td><td>99.9</td><td>55.5</td><td>65.3</td><td>52.9</td><td>61.8</td></tr><tr><td>FatFormer [17]</td><td>98.8</td><td>99.8</td><td>94.2</td><td>99.2</td><td>94.4</td><td>99.1</td><td>94.7</td><td>99.4</td><td>98.6</td><td>99.8</td><td>94.9</td><td>99.1</td><td>62.8</td><td>85.4</td><td>68.8</td><td>93.2</td></tr><tr><td>ForgeLens [22]</td><td>99.0</td><td>100.0</td><td>98.0</td><td>99.9</td><td>97.5</td><td>99.7</td><td>98.5</td><td>99.8</td><td>99.5</td><td>99.9</td><td>97.7</td><td>99.4</td><td>77.6</td><td>91.7</td><td>76.8</td><td>94.9</td></tr><tr><td>D3[20]</td><td>92.8</td><td>98.3</td><td>94.4</td><td>98.8</td><td>94.7</td><td>98.8</td><td>94.9</td><td>98.9</td><td>94.8</td><td>99.4</td><td>88.3</td><td>95.9</td><td>92.5</td><td>98.6</td><td>78.0</td><td>94.6</td></tr><tr><td>VIBAIGC [23]</td><td>87.8</td><td>96.9</td><td>87.2</td><td>97.6</td><td>86.4</td><td>97.5</td><td>89.2</td><td>98.0</td><td>95.9</td><td>99.4</td><td>77.4</td><td>92.8</td><td>50.4</td><td>47.2</td><td>55.9</td><td>69.4</td></tr><tr><td>Ours</td><td>95.7</td><td>99.9</td><td>99.9</td><td>100.0</td><td>100.0</td><td>100.0</td><td>99.8</td><td>100.0</td><td>99.9</td><td>100.0</td><td>99.8</td><td>100.0</td><td>87.5</td><td>97.1</td><td>87.3</td><td>97.4</td></tr></table>
+
+ForgeLens [22],  $\mathrm{D}^3$  [20] and VIBAIGC [23]. Accuracy (Acc.) and average precision (A.P.) are used as metrics. Following the same protocol, we re-evaluated CLIPping [16], Zheng [8], FreqNet [21], FatFormer [17], ForgeLens [22],  $\mathrm{D}^3$  [20], and VIBAIGC [23] using their official open-source implementations, while results for the remaining baselines were taken from [7].
+
+# 4.2 Overall Evaluation of Detection Generalizability
+
+We evaluated the generalization performance of our AIGI detection method across datasets. Accuracy (Acc.) and average precision (A.P.) compared to state-of-the-art GAN- and Diffusion-based methods are reported in Tables 1 and 2, using patch size  $L = 2$ , resampling scales  $\mathbb{S} = \{1,1 / 2,1 / 4\}$ , with a
+
+Table 3: Mean Acc. and A.P. over 16 GAN-based, 16 Diffusion-based, and all 32 datasets.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">GAN-Set</td><td colspan="2">Diff.-Set</td><td colspan="2">Mean</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>CNNDet [5]</td><td>65.4</td><td>86.2</td><td>51.4</td><td>70.7</td><td>58.4</td><td>78.4</td></tr><tr><td>F3Net [10]</td><td>78.7</td><td>90.6</td><td>83.0</td><td>93.6</td><td>80.8</td><td>92.1</td></tr><tr><td>LGrad [49]</td><td>78.0</td><td>86.9</td><td>87.2</td><td>95.2</td><td>82.6</td><td>91.1</td></tr><tr><td>Ojha [15]</td><td>83.2</td><td>98.6</td><td>77.5</td><td>89.5</td><td>80.4</td><td>94.1</td></tr><tr><td>Zheng [8]</td><td>80.6</td><td>89.9</td><td>80.9</td><td>86.5</td><td>80.8</td><td>88.2</td></tr><tr><td>CLIPping [16]</td><td>93.9</td><td>99.1</td><td>81.4</td><td>91.5</td><td>87.7</td><td>95.3</td></tr><tr><td>NPR [7]</td><td>93.8</td><td>97.0</td><td>94.2</td><td>97.6</td><td>94.0</td><td>97.3</td></tr><tr><td>FreqNet [21]</td><td>93.1</td><td>98.2</td><td>81.7</td><td>93.8</td><td>87.4</td><td>96.0</td></tr><tr><td>FatFormer [17]</td><td>99.0</td><td>100.0</td><td>84.8</td><td>95.6</td><td>91.9</td><td>97.8</td></tr><tr><td>ForgeLens [22]</td><td>93.2</td><td>96.2</td><td>85.5</td><td>93.0</td><td>89.4</td><td>94.6</td></tr><tr><td>D3[20]</td><td>94.5</td><td>98.7</td><td>91.1</td><td>97.7</td><td>92.8</td><td>98.2</td></tr><tr><td>VIBAIGC [23]</td><td>93.5</td><td>98.1</td><td>77.2</td><td>89.8</td><td>85.4</td><td>94.0</td></tr><tr><td>Ours</td><td>96.4</td><td>98.2</td><td>97.8</td><td>99.6</td><td>97.1</td><td>98.9</td></tr></table>
+
+ResNet-50 backbone, which yield the optimal validation results. MLEP consistently achieves top performance across most datasets. Remarkably, it generalizes well to diffusion-generated images, despite being trained solely on GAN-based data (ProGAN [24]). Even on datasets with entirely different content (e.g., face-centric sets like StarGAN, InfoMaxGAN, and AttGAN), MLEP maintains strong performance, underscoring its cross-scene robustness. Table 3 further shows that MLEP outperforms NPR [7], with average gains of  $3.1\%$  in Acc. and  $1.6\%$  in A.P., despite NPR's already strong results.
+
+# 4.3 Ablation Study
+
+We next conducted a series of ablation studies to evaluate the effectiveness of key components and hyperparameters in the proposed approach.
+
+Effectiveness of patch shuffling and multi-scale resampling We first assessed the impact of two key components: patch shuffling and multiscale resampling. Ablation results in Table 4 show that removing either component noticeably reduces performance, with patch shuffling contributing more. Even without both, LEP alone achieves over  $94.3\%$  accuracy, higher than NPR  $(94.0\%)$  [7], highlighting the effectiveness of entropy-based features.
+
+Table 4: Ablation study on the impact of key components, where PS represents patch shuffling and MR denotes multiscale resampling.  
+
+<table><tr><td rowspan="2">LEP</td><td rowspan="2">PS</td><td rowspan="2">MR</td><td colspan="2">GAN-set</td><td colspan="2">Diff.-set</td><td colspan="2">Mean</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>✓</td><td></td><td></td><td>93.6</td><td>94.1</td><td>94.9</td><td>95.7</td><td>94.3</td><td>94.9</td></tr><tr><td>✓</td><td></td><td>✓</td><td>93.4</td><td>94.1</td><td>95.8</td><td>96.9</td><td>94.6</td><td>95.5</td></tr><tr><td>✓</td><td>✓</td><td></td><td>95.7</td><td>98.2</td><td>97.5</td><td>99.6</td><td>96.6</td><td>98.9</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>96.4</td><td>98.2</td><td>97.8</td><td>99.6</td><td>97.1</td><td>98.9</td></tr></table>
+
+Impact of the resampling interpolation method We also evaluated the impact of interpolation methods, comparing bilinear, bicubic, and nearest-neighbor (Table 5). Bilinear outperforms nearest-neighbor and performs comparably to bicubic. This might be because bilinear and bicubic blend neighboring pixel values, introducing richer entropy variations, while
+
+Table 5: Impact of the interpolation method.  
+
+<table><tr><td rowspan="2">Interp.</td><td colspan="2">GAN-set</td><td colspan="2">Diff.-set</td><td colspan="2">Mean</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>Bilinear</td><td>96.4</td><td>98.2</td><td>97.8</td><td>99.6</td><td>97.1</td><td>98.9</td></tr><tr><td>Bicubic</td><td>96.2</td><td>97.9</td><td>97.9</td><td>99.3</td><td>96.9</td><td>99.1</td></tr><tr><td>Nearest</td><td>94.6</td><td>97.4</td><td>96.8</td><td>99.2</td><td>95.7</td><td>98.3</td></tr></table>
+
+nearest-neighbor simply copies pixel values, resulting in limited entropy diversity.
+
+Impact of patch size and scale factors We further examined the effects of patch size and resampling scales by testing different hyperparameter settings, as shown in Table 6. The best performance was achieved with the smallest patch size  $(l = 2)$ , indicating that stronger semantic scrambling improves detection. Moderate multi-scale fusion  $(\mathbb{S} = \{1,1 / 2,1 / 4\})$  also led to optimal results, confirming the benefit of incorporating resampling artif
+
+Table 6: Impact of patch size  $L$  and resampling scaling factors  $\mathbb{S}$ .  
+
+<table><tr><td rowspan="2">L</td><td rowspan="2">S</td><td colspan="2">GAN-set</td><td colspan="2">Diff.-set</td><td colspan="2">Mean</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td rowspan="3">2</td><td>{1,1/2}</td><td>95.8</td><td>97.8</td><td>97.5</td><td>99.6</td><td>96.6</td><td>98.7</td></tr><tr><td>{1,1/2,1/4}</td><td>96.4</td><td>98.2</td><td>97.8</td><td>99.6</td><td>97.1</td><td>98.9</td></tr><tr><td>{1,1/2,1/4,1/8}</td><td>91.7</td><td>97.9</td><td>95.3</td><td>99.5</td><td>93.5</td><td>98.7</td></tr><tr><td rowspan="3">4</td><td>{1,1/2}</td><td>94.5</td><td>96.6</td><td>95.5</td><td>98.8</td><td>95.0</td><td>97.7</td></tr><tr><td>{1,1/2,1/4}</td><td>94.5</td><td>96.8</td><td>96.6</td><td>99.1</td><td>95.5</td><td>97.9</td></tr><tr><td>{1,1/2,1/4,1/8}</td><td>94.2</td><td>96.4</td><td>96.5</td><td>98.8</td><td>95.4</td><td>97.6</td></tr><tr><td rowspan="3">8</td><td>{1,1/2}</td><td>93.9</td><td>95.8</td><td>95.4</td><td>97.7</td><td>94.7</td><td>96.7</td></tr><tr><td>{1,1/2,1/4}</td><td>94.0</td><td>96.5</td><td>95.8</td><td>99.1</td><td>94.9</td><td>97.8</td></tr><tr><td>{1,1/2,1/4,1/8}</td><td>94.4</td><td>96.0</td><td>95.8</td><td>98.1</td><td>95.1</td><td>97.0</td></tr></table>
+
+Impact of sliding window stride We evaluated the effect of stride in the  $2 \times 2$  sliding window for LEP computation, testing strides of 1 and 2 (Table 7). A stride of 1 significantly outperforms 2, highlighting the importance of interblock entropy in MLEP. This supports our multi-granularity design, which captures both intra- and inter-block texture patterns as depicted in Fig. 3.
+
+Table 7: Impact of sliding window stride.  
+
+<table><tr><td rowspan="2">Stride</td><td colspan="2">GAN-set</td><td colspan="2">Diff.-set</td><td colspan="2">Mean</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>1</td><td>96.4</td><td>98.2</td><td>97.8</td><td>99.6</td><td>97.1</td><td>98.9</td></tr><tr><td>2</td><td>94.9</td><td>97.3</td><td>97.0</td><td>99.5</td><td>95.9</td><td>98.4</td></tr></table>
+
+Compatibility with various backbones Lastly, we assessed the compatibility of MLEP with various ResNet backbones [50], including ResNet-18, 34, 50, and 101. As shown in Table 8, all variants achieved strong performance, with slight gains from larger models. This confirms the generality and scalability of the proposed feature extraction method.
+
+Table 8: Evaluation on various ResNets.  
+
+<table><tr><td rowspan="2">Backbone</td><td colspan="2">GAN-set</td><td colspan="2">Diff.-set</td><td colspan="2">Mean</td></tr><tr><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td><td>Acc.</td><td>A.P.</td></tr><tr><td>ResNet-18</td><td>96.0</td><td>97.9</td><td>97.7</td><td>99.5</td><td>96.8</td><td>98.7</td></tr><tr><td>ResNet-34</td><td>96.1</td><td>98.2</td><td>97.7</td><td>99.7</td><td>96.9</td><td>98.9</td></tr><tr><td>ResNet-50</td><td>96.4</td><td>98.2</td><td>97.8</td><td>99.6</td><td>97.1</td><td>98.9</td></tr><tr><td>ResNet-101</td><td>96.4</td><td>98.3</td><td>97.8</td><td>99.6</td><td>97.1</td><td>99.0</td></tr></table>
+
+Influence of generation model's text prompts In text-to-image diffusion models, the specificity of input prompts may greatly affect the visual quality and details of generated images. To inspect the influence of text prompts on AIGI detection performance, we conducted an additional experiment using DiffusionDB[51], a large dataset with 14 million Stable Diffusion images generated from 1.8 million unique prompts. We randomly selected two subsets of 3,000 images, one set generated from complex prompts (over 200 characters with keywords like "high quality," "detailed," and "realistic") and the other from simple prompts (under 100 characters and without those keywords). We evaluated our trained detector on both subsets and found almost no difference in detection accuracy:  $99.65\%$  accuracy on the simple set and  $99.62\%$  accuracy on the complex set. This further demonstrates the generalizability of the proposed method over different types of AI-generated content.
+
+# 4.4 Interpretability of MLEP
+
+To illustrate the effectiveness of MLEP for AIGI detection, we conducted a set of qualitative analysis detailed as follows.
+
+Entropy patterns between real and AI-generated images We first visualize LEP maps for several real-fake image pairs, along with their differences in the pixel, entropy, and Fourier domains. Here, "fake" refers to AI-reconstructed images resembling the originals. Since LEP values are sparse and capped at 2, we normalize them to [0, 255] for visualization. As shown in Fig. 4, LEP differences are
+
+![](images/8213dc4c974b6c19cd2093ecfa70eb12a8e28c44101c9c4f653b04fe41d6b3ae.jpg)  
+Figure 4: Visualization of local entropy patterns for several real-fake image pairs, along with their differences in the pixel, entropy, and Fourier domains.
+
+![](images/5cc4d77c0876abb10ccb22a5d72120a534d1392b6f1349ea45b59eb5fdd5a3d6.jpg)  
+Figure 5: Qualitative comparison among Zheng [8], NPR [7], and our method. LEP preserves minimal visible semantics, while MLEP (without resampling) further suppresses semantic content.
+
+far more pronounced than pixel-level differences, especially for high-quality generations like Stable Diffusion v2, where pixel differences are visually negligible. In the frequency domain, real-fake differences show more consistent patterns than in the pixel space, supporting content-agnostic detection. These results highlight LEP's ability to amplify real-fake discrepancies while minimizing semantic interference.
+
+Semantic suppression capability of MLEP We further examine the semantic suppression capability of MLEP compared to two competitive methods: Zheng [8] and NPR [7]. Fig.5 visualizes feature maps of LEP and MLEP (without multi-scale resampling), alongside those from Zheng [8] and NPR [7]. The  $32 \times 32$  shuffled patches in Zheng[8] still retain noticeable semantic cues both locally and globally. NPR [7] produces edge-like features by computing pixel differences, leaving much of the original semantics intact. In contrast, LEP substantially suppresses semantic content by highlighting pixel-level randomness, and MLEP further eliminates it through fine-grained patch shuffling, enabling learning content-agnostic representation for AIGI detection.
+
+![](images/4e511b976779090864b356680049a5b7876d200b21b2764088b678fa8000c4f5.jpg)  
+Figure 6: t-SNE visualization of real vs. fake samples.
+
+Feature distribution of real and AI-generated images Finally, Fig.6 visualizes the t-SNE distribution [52] of real and fake samples based on the final feature layer of a ResNet-50 classifier, comparing our method with two competitive baselines—NPR [7] and FreqNet [21—which also use ResNet-50. We showcase results on four generative models: StarGAN [30], AttGAN [32], ADM [43], and SDv2 [44]. The proposed local entropy patterns (LEP) achieve noticeably cleaner real-fake separation than the baselines, and MLEP further enhances this distinction, demonstrating stronger discriminative capability for AIGI detection.
+
+# 5 Conclusion and Limitations
+
+This paper explores the use of entropy as a cue for detecting AI-generated images (AIGI) and introduces Multi-granularity Local Entropy Patterns (MLEP), a set of entropy-based feature maps derived from shuffled small patches across multiple image scales. MLEP captures pixel relationships across spatial and scale dimensions while disrupting image semantics, thereby mitigating content bias. Using MLEP as input, a CNN-based classifier (e.g., ResNet) achieves robust and highly generalizable detection performance.
+
+**Limitations** Nonetheless, limitations still remain. The paper does not explicitly address the robustness of the detector under common image post-processing operations. In fact, when applying different levels of JPEG compression, blurring, or noise, we observe a  $17\%$  to  $45\%$  drop in detection accuracy—performance that is less satisfactory compared to methods explicitly optimized for robustness. This limitation stems from the fact that MLEP was not specifically designed to handle such transformations, and no special data augmentation techniques were employed during training. Instead, the paper is focused on exploring the potential of using information entropy as a discriminative signal for AIGI detection and on revealing the intrinsic differences in local entropy patterns between real and AI-generated images. Interestingly, the proposed method shows strong robustness to image rescaling: when images are downsampled to half their original resolution, the mean detection accuracy remains above  $92.4\%$  (only  $4.7\%$  drop). We attribute this to the multi-scale resampling strategy used during training, which effectively introduced resolution variability as a form of implicit data augmentation. Moreover, entropy is computed only within small  $2 \times 2$  windows, as using larger windows would exponentially increase computational complexity. In the future, more efforts could be devoted to improve the robustness and computation efficiency of entropy-based approach.
+
+# Acknowledgements
+
+This work is supported by the National Natural Science Foundation of China under grants 62201107, U22A2096, 62502060, 62402073, and 62221005, in part by the Natural Science Foundation of Chongqing under grant CSTB2023NSCQ-LZX0061, in part by the Science and Technology Innovation Key R&D Program of Chongqing under grant CSTB2023TIAD-STX0016, and in part by the Science and Technology Research Program of Chongqing Municipal Education Commission under grants KJQN202300606, KJQN202300619, and KJQN202500649. Special thanks are extended to Prof. Nannan Wang, Prof. Xiuli Bi, Prof. Gwanggil Jeon, and Prof. Touradj Ebrahimi for their invaluable guidance, insightful feedback, and continuous encouragement throughout this research.
+
+# References
+
+[1] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, "Generative Adversarial Nets," in Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, ser. NIPS'14, 2014, p. 2672-2680.  
+[2] J. Ho, A. Jain, and P. Abbeel, “Denoising Diffusion Probabilistic Models,” Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020.  
+[3] N. P. Howe and B. Thompson, “This isn’t the Nature Podcast-how deepfakes are distorting reality.” Nature, 2023.  
+[4] D. Xu, S. Fan, and M. Kankanhalli, “Combating Misinformation in the Era of Generative AI Models,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, p. 9291–9298.  
+[5] S.-Y. Wang, O. Wang, R. Zhang, A. Owens, and A. A. Efros, "Cnn-generated images are surprisingly easy to spot... for now," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 8695-8704.  
+[6] N. Zhong, Y. Xu, S. Li, Z. Qian, and X. Zhang, "Patchcraft: Exploring Texture Patch for Efficient AI-generated Image Detection," arXiv preprint arXiv:2311.12397, 2024.  
+[7] C. Tan, Y. Zhao, S. Wei, G. Gu, P. Liu, and Y. Wei, “Rethinking the Up-Sampling Operations in CNN-Based Generative Network for Generalizable Deepfake Detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 28-32, 2024.  
+[8] C. Zheng, C. Lin, Z. Zhao, H. Wang, X. Guo, S. Liu, and C. Shen, “Breaking Semantic Artifacts for Generalized AI-generated Image Detection,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.  
+[9] J. Frank, T. Eisenhofer, L. Schonherr, A. Fischer, D. Kolossa, and T. Holz, “Leveraging Frequency Analysis for Deep Fake Image Recognition,” in International Conference on Machine Learning. PMLR, 2020, pp. 3247–3258.  
+[10] Y. Qian, G. Yin, L. Sheng, Z. Chen, and J. Shao, “Thinking in Frequency: Face Forgery Detection by Mining Frequency-Aware Clues,” in European conference on computer vision. Springer, 2020, pp. 86–103.  
+[11] Y. Luo, Y. Zhang, J. Yan, and W. Liu, "Generalizing Face Forgery Detection with High-Frequency Features," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 16317-16326.  
+[12] B. Liu, F. Yang, X. Bi, B. Xiao, W. Li, and X. Gao, “Detecting Generated Images by Real Images,” in European Conference on Computer Vision. Springer, 2022, pp. 95–110.  
+[13] Z. Wang, J. Bao, W. Zhou, W. Wang, H. Hu, H. Chen, and H. Li, “DIRE for Diffusion-Generated Image Detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 22 445–22 455.  
+[14] B. Chen, J. Zeng, J. Yang, and R. Yang, “DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection of Diffusion Generated Images,” in *Forty-first International Conference on Machine Learning*, 2024.  
+[15] U. Ojha, Y. Li, and Y. J. Lee, "Towards Universal Fake Image Detectors that Generalize Across Generative Models," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 24480-24489.
+
+[16] S. A. Khan and D.-T. Dang-Nguyen, "CLIPping the Deception: Adapting Vision-Language Models for Universal deepfake detection," in Proceedings of the 2024 International Conference on Multimedia Retrieval, 2024, pp. 1006-1015.  
+[17] H. Liu, Z. Tan, C. Tan, Y. Wei, J. Wang, and Y. Zhao, "Forgery-aware Adaptive Transformer for Generalizable Synthetic Image Detection," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10770-10780.  
+[18] C. E. Shannon, “A mathematical theory of communication,” The Bell system technical journal, vol. 27, no. 3, pp. 379–423, 1948.  
+[19] D. Cozzolino, G. Poggi, M. Nießner, and L. Verdoliva, “Zero-shot Detection of AI-generated Images,” in European Conference on Computer Vision. Springer, 2024, pp. 54–72.  
+[20] Y. Yang, Z. Qian, Y. Zhu, O. Russakovsky, and Y. Wu, “D^3: Scaling Up Deepfake Detection by Learning from Discrepancy,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 23850-23859.  
+[21] C. Tan, Y. Zhao, S. Wei, G. Gu, P. Liu, and Y. Wei, "Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning," in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 5052-5060.  
+[22] Y. Chen, L. Zhang, and Y. Niu, "Forgelens: Data-Efficient Forgery Focus for Generalizable Forgery Image Detection," 2025.  
+[23] H. Zhang, Q. He, X. Bi, W. Li, B. Liu, and B. Xiao, "Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 23828-23837.  
+[24] T. Karras, T. Aila, S. Laine, and J. Lehtinen, "Progressive Growing of Gans for Improved Quality, Stability, and Variation," in International Conference on Learning Representations, 2018.  
+[25] F. Yu, A. Seff, Y. Zhang, S. Song, T. Funkhouser, and J. Xiao, "LSUN: Construction of a Large-Scale Image Dataset using Deep Learning with Humans in the Loop," arXiv preprint arXiv:1506.03365, 2015.  
+[26] T. Karras, S. Laine, and T. Aila, "A Style-Based Generator Architecture for Generative Adversarial Networks," in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 4396-4405.  
+[27] T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, "Analyzing and Improving the Image Quality of StyleGAN," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 8110-8119.  
+[28] A. Brock, J. Donahue, and K. Simonyan, “Large Scale GAN Training for High Fidelity Natural Image Synthesis,” in International Conference on Learning Representations, 2018.  
+[29] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks," in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2223-2232.  
+[30] Y. Choi, M. Choi, M. Kim, J.-W. Ha, S. Kim, and J. Choo, "StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 8789-8797.  
+[31] T. Park, M.-Y. Liu, T.-C. Wang, and J.-Y. Zhu, "Semantic Image Synthesis with Spatially-Adaptive Normalization," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2337-2346.  
+[32] Z. He, W. Zuo, M. Kan, S. Shan, and X. Chen, "AttGAN: Facial Attribute Editing by Only Changing What You Want," IEEE transactions on image processing, vol. 28, no. 11, pp. 5464-5478, 2019.  
+[33] D. Berthelot, “BEGAN: Boundary Equilibrium Generative Adversarial Networks,” arXiv preprint arXiv:1703.10717, 2017.  
+[34] M. G. Bellemare, I. Danihelka, W. Dabney, S. Mohamed, B. Lakshminarayanan, S. Hoyer, and R. Munos, “The Cramer Cistance as a Solution to Biased Wasserstein Gradients,” in International Conference on Learning Representations, 2018.
+
+[35] K. S. Lee, N.-T. Tran, and N.-M. Cheung, "InfoMax-GAN: Improved Adversarial Image Generation via Information Maximization and Contrastive Learning," in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2021, pp. 3942-3952.  
+[36] C.-L. Li, W.-C. Chang, Y. Cheng, Y. Yang, and B. Póczos, "MMD GAN: Towards Deeper Understanding of Moment Matching Network," Advances in neural information processing systems, vol. 30, 2017.  
+[37] W. Nie, N. Narodytska, and A. Patel, “RelGAN: Relational Generative Adversarial Networks for Text Generation,” in International conference on learning representations, 2018.  
+[38] M. Lučić, M. Tschannen, M. Ritter, X. Zhai, O. Bachem, and S. Gelly, “High-Fidelity Image Generation with Fewer Labels,” in International conference on machine learning. PMLR, 2019, pp. 4183–4192.  
+[39] T. Miyato, T. Kataoka, M. Koyama, and Y. Yoshida, "Spectral Normalization for Generative Adversarial Networks," in International Conference on Learning Representations, 2018.  
+[40] M. Liu, Y. Ding, M. Xia, X. Liu, E. Ding, W. Zuo, and S. Wen, “STGAN: A Unified Selective Transfer Network for Arbitrary Image Attribute Editing,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3673–3682.  
+[41] C. Tan, R. Tao, H. Liu, and Y. Zhao, "GANGen-Detection: A Dataset Generated by GANs for Generalizable Deepfake Detection," github.com/chuangchuangtan/GANGen-Detection, 2024.  
+[42] A. Q. Nichol and P. Dhariwal, “Improved Denoising Fiffusion Probabilistic Models,” in International conference on machine learning. PMLR, 2021, pp. 8162–8171.  
+[43] P. Dhariwal and A. Nichol, "Diffusion Models Beat GANs on Image Synthesis," Advances in neural information processing systems, vol. 34, pp. 8780-8794, 2021.  
+[44] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-Resolution Image Synthesis with Latent Diffusion Models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684–10695.  
+[45] L. Liu, Y. Ren, Z. Lin, and Z. Zhao, “Pseudo Numerical Methods for Diffusion Models on Manifolds,” in International Conference on Learning Representations, 2022.  
+[46] S. Gu, D. Chen, J. Bao, F. Wen, B. Zhang, D. Chen, L. Yuan, and B. Guo, "Vector Quantized Diffusion Model for Text-to-Image Synthesis," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10696-10706.  
+[47] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever, “Zero-Shot Text-to-Image Generation,” in International conference on machine learning. PMLR, 2021, pp. 8821–8831.  
+[48] A. Q. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen, "Glide: Towards photorealistic image generation and editing with text-guided diffusion models," in International Conference on Machine Learning. PMLR, 2022, pp. 16784-16804.  
+[49] C. Tan, Y. Zhao, S. Wei, G. Gu, and Y. Wei, "Learning on Gradients: Generalized Artifacts Representation for GAN-Generated Images Detection," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 12 105–12 114.  
+[50] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778.  
+[51] Z. J. Wang, E. Montoya, D. Munechika, H. Yang, B. Hoover, and D. H. Chau, "DiffusionDB: A Large-scale Prompt Gallery Dataset for Text-to-Image Generative Models," in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023, pp. 893–911.  
+[52] L. Van der Maaten and G. Hinton, "Visualizing data using t-SNE," Journal of Machine Learning Research, vol. 9, no. 11, 2008.
+
+# 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: The abstract and introduction can accurately reflect the paper's contributions and scope.
+
+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: In the conclusion section, we discussed the technical limitations and possible future directions of this work.
+
+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: [Yes]
+
+Justification: In this work, we not only present a complete implementation framework but also validate our hypothesis through extensive experiments and ablation studies.
+
+# 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: We detail all experimental settings, including parameters and configurations, to ensure reproducibility. The code will be publicly released upon acceptance of the paper.
+
+# 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: The entire source code and datasets for training and testing will be publicly released upon acceptance of the paper.
+
+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: In the paper, the datasets, hyperparameter selection, and optimizer selection are provided.
+
+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: [No]
+
+Justification: This study focuses on average detection performance measured by accuracy, with results reported per dataset. Therefore, including error bars is not strictly necessary.
+
+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: Section 4 provides details on the computing resources used in the experiments, including GPU type and memory specifications.
+
+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: Yes. The research fully complies with all provisions of the NeurIPS Code of Ethics.
+
+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: [Yes]
+
+Justification: The introduction outlines the social issues this work aims to address and its potential positive impact, while the conclusion discusses its limitations.
+
+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: [NA]
+
+Justification: Not applicable. Our dataset contains only public, non-sensitive images with CC licenses, and the model has no risk.
+
+# 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: Yes. All third-party assets (code, data, models) are explicitly credited with original sources.
+
+# 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 URL.  
+- 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: [NA]
+
+Justification: The paper does not release new assets.
+
+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: This 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.
+
+- 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.
+
+# 16. Declaration of LLM usage
+
+Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required.
+
+Answer: [NA]
+
+Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components.
+
+Guidelines:
+
+- The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.  
+- Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described.

2025/MLEP_ Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30b8c290637568eb89aa47fc06eb193aade2329b8add011b37a7bbec25ff4066
+size 1169085