Datasets:
Formats:
json
Languages:
English
Size:
< 1K
Tags:
scientific-figures
vlm-as-judge
counterfactual
verifier-meta-evaluation
ai-for-science
icdm-2026
License:
pair_id stringlengths 19 26 | subset stringclasses 3
values | caption stringlengths 18 2.43k | text_block stringlengths 244 3.5k | text_block_source stringclasses 3
values | paper dict | topic stringclasses 22
values | figure_type stringclasses 8
values | entities listlengths 3 8 ⌀ | defect unknown | images dict | images_available bool 2
classes | flags unknown | in_paired_run bool 2
classes | multi_panel int64 0 12 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ml/document_1__figure_2 | ml | Figure 2. Overview of the proposed PhyCo two-stage training pipeline. In Stage 1 (Physics-Supervised Fine-Tuning), the base DiT model is conditioned via a ControlNet using physically rich simulation data, enabling controllable video generation with respect to key physical properties such as friction, restitution, defor... | The overall pipeline is illustrated in Fig. 2. We fine-tune a pre-trained Cosmos-Predict2-2B [28] diffusion backbone by conditioning the denoising process on physical prop- erty inputs using a ControlNet-style [49] architecture. For- mally, the generator Gθ models the conditional distribution pθ(x1:T |t, x0 Figure 2. O... | mentions+same_page_context | {
"title": "PhyCo: Learning Controllable Physical Priors for Generative Motion",
"source_id": "document_1",
"arxiv_id": "arXiv:2604.28169",
"figure": "figure 2",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Исправлены подписи блоков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_1__figure_2/orig.png",
"fail": "pairs/ml/document_1__figure_2/fail.png",
"orig_sha256": "b24dc7fdc610fe551fab098de1999dce0c4176373807a85a1a4d0b5cbc95eff2",
"fail_sha256": "394145f623d1b5a01899929a9bd0f4cc5a5221c40aabcf38e538a31a1a166532"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_100__figure_2 | ml | Figure 2. An illustrative example of the data synthesis approach. | Based on this insight, we propose D2VLM, where our core motivation is to decouple the tied learning objective while also preserving and even strengthening the inherent dependency between these two tasks. An overall concep- tual demonstration is shown in Fig. 2. Specifically, encoded video tokens FV and question textual... | mentions | {
"title": "Factorized Learning for Temporally Grounded Video-Language Models",
"source_id": "document_100",
"arxiv_id": "2512.24097",
"figure": "figure 2",
"page": 16
} | machine_learning | null | null | {
"description_ru": "Изменен текст внутри блоков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_100__figure_2/orig.png",
"fail": "pairs/ml/document_100__figure_2/fail.png",
"orig_sha256": "25689a6d4e24225b26ee721ae8afd5350306fdf76dc4136974c3a17e1ff65388",
"fail_sha256": "e89ff0b04dfe41112f7d0d700581f0eb1ba98f9ee138cdde4715ba12d3b119e8"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_101__figure_1 | ml | Figure 1: Illustration of fixed vs. (challenging) evolving tool environments. AutoTool enables LLM agents to dynamically select from complex unseen toolsets via a dual-phase optimization pipeline. | Despite these advances, existing approaches typically operate under the assumption of single-domain fixed tools (Gao et al., 2025a): the policy model learns when and how to invoke the designated tools, but the available tools are predefined and static for a specific task. In practice, these methods curate training data... | mentions | {
"title": "AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning",
"source_id": "document_101",
"arxiv_id": "2512.13278",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "изменен текст к блокам",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_101__figure_1/orig.png",
"fail": "pairs/ml/document_101__figure_1/fail.png",
"orig_sha256": "19ef8df9ed37f54694e22057e035856c469d9cc1d6314cb105bbf90123c85098",
"fail_sha256": "1a55f19d1d6da869be2c39a14fd709fb074de0f56a50231b53eb78a76e16f977"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_101__figure_3 | ml | Figure 3: Illustration on AutoTool’s Tool-Selection Refinement Phase. We cast tool-selection as a Plackett–Luce (PL) ranking problem during trajectory generation, optimizing the LLM agent to align its policy distribution with reward-consistent tool preferences. | Figure 3: Illustration on AutoTool’s Tool-Selection Refinement Phase. We cast tool-selection as a Plackett–Luce (PL) ranking problem during trajectory generation, optimizing the LLM agent to align its policy distribution with reward-consistent tool preferences. its previous reasoning context. After the tool selection s... | same_page_context | {
"title": "AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning",
"source_id": "document_101",
"arxiv_id": "2512.13278",
"figure": "figure 3",
"page": 5
} | machine_learning | null | null | {
"description_ru": "изменены направления стрелок",
"source_labels": [],
"operators": [
"edge_rewiring"
],
"objects": [
"arrows"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_101__figure_3/orig.png",
"fail": "pairs/ml/document_101__figure_3/fail.png",
"orig_sha256": "316f5b269aeda95ff7459bea44a2199126a4db493f80a0c2cf083474bdd0fc9c",
"fail_sha256": "bd1726637e034bae1ba6393f4145189ad1924328d5b249b211d53c608fd438d5"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_102__figure_1 | ml | Figure 1. Overview of our self-training method. The process be- gins with a base model, which generates pseudo-labels for unla- beled data. The student model is then trained in two stages: first on pseudo-labeled data, and then fine-tuned on true-labeled data. | Our methodology is built upon a self-training mech- anism, drawing inspiration from knowledge distillation concepts[3]. The core of our approach utilizes the Seg- Former model [10], specifically with a Mix Transformer (MiT-B4) as its backbone. We use the same architec- ture—SegFormer with a MiT-B4 backbone—in all train... | mentions | {
"title": "Pseudo-Label Refinement for Robust Wheat Head Segmentation via Two-Stage Hybrid Training",
"source_id": "document_102",
"arxiv_id": "2512.11874",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "удалена стрелка",
"source_labels": [],
"operators": [
"element_deletion"
],
"objects": [
"arrows"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_102__figure_1/orig.png",
"fail": "pairs/ml/document_102__figure_1/fail.png",
"orig_sha256": "0d6d266a8111a301e03acf02401bdcf39102084da9c8cb5cd01ba2ad4166ee46",
"fail_sha256": "1e70310f049780c4c92f9bd97e4e4531d199d8329fe931616b72da289d072d55"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_105__figure_2 | ml | Figure 2. PCA visualization of textual prompts (left) and qualitative results on various environments (e.g., Day, Dawn, Night) in BDD100K [58] (right). The models are trained on GTAV [42] with the CLIP-pretrained backbone (ViT-B) [41]. A fixed single-context prompt lacks flexibility in adapting to various domain shifts... | In this perspective, we point out that the semantic form of the textual representation for an object category should be changed with respect to the input visual context in order to strengthen their semantic correspondence. For instance, in the real-world driving scene at night of Fig. 2 (b), tar- get categories (e.g., ... | mentions | {
"title": "Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation",
"source_id": "document_105",
"arxiv_id": "2512.03508",
"figure": "figure 2",
"page": 2
} | machine_learning | null | null | {
"description_ru": "убраны точки одного цвета",
"source_labels": [],
"operators": [
"element_deletion"
],
"objects": [],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_105__figure_2/orig.png",
"fail": "pairs/ml/document_105__figure_2/fail.png",
"orig_sha256": "de4f49d107613a9990c331890546a544e96fec8c1c374584372904d84160ce4a",
"fail_sha256": "24dd8b956f70a1602037ddfa809a8042db6105ca5cae2b9eed5f8d6d0a314f92"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_105__figure_3 | ml | Figure 3. Illustration of DPMFormer. We use Mask2Former [5] based architecture which consists of a backbone image encoder (ENCI), a pixel decoder (DECpix), a transformer decoder (DECtr), and a text encoder (ENCT ). During training, we synthesize images with a novel domain style via texture perturbation. Both images are... | As illustrated in Fig. 3, Domain-aware Prompt-driven Masked Transformer (DPMFormer), aims to cultivate two core aspects of domain generalization for the model: domain-awareness and domain-robustness. For domain- awareness, we propose the domain-aware context prompt learning which translates domain-specific property of ... | mentions | {
"title": "Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation",
"source_id": "document_105",
"arxiv_id": "2512.03508",
"figure": "figure 3",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Удалены стрелки, блоки, изменен текст некоторых блоков",
"source_labels": [],
"operators": [
"label_text_edit",
"element_deletion"
],
"objects": [
"blocks",
"labels",
"arrows"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_105__figure_3/orig.png",
"fail": "pairs/ml/document_105__figure_3/fail.png",
"orig_sha256": "ec39b9a882235418fc6147504d5e730ac154945b8461e808190c01d416e14721",
"fail_sha256": "fd95f91a1f47547f61c193f24a916e0f6bb0d745ccc5d3637692b0d5f7318401"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_106__figure_4 | ml | Figure 4. The process pipeline for constructing instruction data. The red / green indicates the instruction prompt / ground truth. | This step achieves a success rate of 77.17%. Finally, we construct 19.2K charts and 62.3K instruction data as illus- trated in Fig. 4. We randomly sample 100 cases, which are reviewed by at least three experts to evaluate the bounding box quality of the grounding step based on the process in Fig. 2. 91% of the cases me... | mentions+same_page_context | {
"title": "ChartPoint: Guiding MLLMs with Grounding Reflection for Chart Reasoning",
"source_id": "document_106",
"arxiv_id": "2512.00305",
"figure": "figure 4",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Удалены блоки",
"source_labels": [],
"operators": [
"element_deletion"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 4"
]
} | {
"orig": "pairs/ml/document_106__figure_4/orig.png",
"fail": "pairs/ml/document_106__figure_4/fail.png",
"orig_sha256": "3f15381ae893da220f1a484723acc76c2a2c2f556ca5eb34c80597beae0e082e",
"fail_sha256": "5d0e50e107770fa280aad172b65caaf3cdb74d798ce189424d4a8f2b9dd46f5f"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_108__figure_5 | ml | Figure 5 presents a comparison of generated images from DreamBooth with MINDiff across different λ values. By adjusting λ, users can generate images tailored to their specific goals or intentions. As λ increases, the generated images align more closely with the given text conditions. | Figure 5. Qualitative results of DreamBooth with MINDiff us- ing different values of the scale parameter λ. Each row is gen- erated with the same seed. As λ increases, subject influence is suppressed, allowing the model to better reflect the text prompt. As shown in Figs. 3 and 4, existing methods often strug- gle to b... | same_page_context | {
"title": "MINDiff: Mask-Integrated Negative Attention for Controlling Overfitting in Text-to-Image Personalization",
"source_id": "document_108",
"arxiv_id": "2511.17888",
"figure": "figure 5",
"page": 7
} | machine_learning | null | null | {
"description_ru": "перемешаны подписи",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"labels"
],
"gold_objects": [
"figure 5"
]
} | {
"orig": "pairs/ml/document_108__figure_5/orig.png",
"fail": "pairs/ml/document_108__figure_5/fail.png",
"orig_sha256": "b9c04b1648fc91cf652f1dc34ba06d835c5e85d8f7c4a415abc23be85c906434",
"fail_sha256": "711fd046dcc872921dea26dbe11074e9655367edded8235cc1e863afe5c32e1e"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_109__figure_4 | ml | Figure 4. Our denoising diffusion model. The input noise under- goes linear projection with positional and timestep embeddings, followed by feature extraction through encoder blocks (×6). The core denoising structure, MST with Adaptive Fusion, integrates multi-scale attention and adaptive fusion to dynamically adjust f... | The above frozen autoencoder encodes the command se- quences into a compact latent variable Z ∈RN×256×64. In the forward process, we employ a linear variance schedule {βt} to add noise into the latent space, while the denois- ing process is guided by a Multi-Scale Transformer-based Diffusion Generator (MST-D), as shown... | mentions | {
"title": "MamTiff-CAD: Multi-Scale Latent Diffusion with Mamba+ for Complex Parametric Sequence",
"source_id": "document_109",
"arxiv_id": "2511.17647",
"figure": "figure 4",
"page": 5
} | machine_learning | null | null | {
"description_ru": "Перемешаны блоки",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 4"
]
} | {
"orig": "pairs/ml/document_109__figure_4/orig.png",
"fail": "pairs/ml/document_109__figure_4/fail.png",
"orig_sha256": "c670d758fdef4476caeb4f16ee96699feae48303a52eff3c4a5e99a007a27dbb",
"fail_sha256": "ebcd3049fdd50be5f89c913c073c362e6f04e7a05e6bfb2ab359032ad18aa94d"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_111__figure_2 | ml | Figure 2. The overall framework of LaneDiffusion. LaneDiffusion comprises two key components: (i) a Lane Prior Injection Module (LPIM), which injects lane priors into BEV features to construct the diffusion target, and (ii) a Lane Prior Diffusion Module (LPDM), which models the prior-injected BEV feature via a diffusio... | The overall framework of LaneDiffusion is illustrated in Fig. 2. As highlighted in the Introduction, it can be seam- lessly integrated as an add-on to BEV feature-based archi- tectures. To clearly demonstrate the performance improve- ment brought by LaneDiffusion, we adopt the state-of-the- art CGNet [15] as our primar... | mentions | {
"title": "LaneDiffusion: Improving Centerline Graph Learning via Prior Injected BEV Feature Generation",
"source_id": "document_111",
"arxiv_id": "2511.06272",
"figure": "figure 2",
"page": 3
} | machine_learning | null | null | {
"description_ru": "перемешаны подписи к блокам, изменены надписи",
"source_labels": [],
"operators": [
"label_text_edit",
"block_permutation"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_111__figure_2/orig.png",
"fail": "pairs/ml/document_111__figure_2/fail.png",
"orig_sha256": "3d40b70df980b81412fc9b5ac998d380d68efe8a0819f4d8c2e39888dee7b9a0",
"fail_sha256": "8f8461b07c8975cb58b47a0ea65258ac7efda624323d2f217c2e5fac122801fb"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_112__figure_4 | ml | Figure 4. An overview framework of the proposed method. (a) Cluster the weight rows within each layer of the teacher model, using cluster centers as initial weight chain rows. (b) Train the teacher model and the S-Student constrained by the weight chain, using Refining loss and ReID loss to infuse core knowledge into t... | termed OSKT. As shown in Figure 4, OSKT consists of two core components: (1) weight chain refinement from the teacher model and (2) student model generation via the weight chain. This section is organized as follows: First, we unify CNN and ViT architectures for clarity. Then we formally introduce the weight chain conc... | mentions | {
"title": "One-Shot Knowledge Transfer for Scalable Person Re-Identification",
"source_id": "document_112",
"arxiv_id": "2511.06016",
"figure": "figure 4",
"page": 4
} | machine_learning | null | null | {
"description_ru": "изменены подписи к блокам и стрелкам",
"source_labels": [],
"operators": [
"edge_rewiring"
],
"objects": [
"blocks",
"labels",
"arrows"
],
"gold_objects": [
"figure 4"
]
} | {
"orig": "pairs/ml/document_112__figure_4/orig.png",
"fail": "pairs/ml/document_112__figure_4/fail.png",
"orig_sha256": "f7e3796c8030d4f88ae0759c95dc91cfcf2a2457c9444cb35721e02d80107d2d",
"fail_sha256": "053bd02bf2ae7aa6c64aa9ed4ff4ad274b057b126b8623fbda5f12adae65b828"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_113__figure_5 | ml | Figure 5. A visualization of the model’s reasoning attention map for the instruction: “Find the free space in front of the window on the left.” The figure illustrates the model’s focus across the four-step textual reasoning process. The attention heatmap is overlaid on the input image. (1) Identify Reference Object and... | For the 7B model, used for the ablation study and at- tention analysis presented in Figure 4 and Figure 5, we use a parameter-efficient approach. Specifically, we employ Low-Rank Adaptation (LoRA) [9] with a rank of r = 128 and a scaling factor of α = 256. This allows us to adapt the model by updating only the LoRA ada... | mentions+same_page_context | {
"title": "TRACE: Textual Reasoning for Affordance Coordinate Extraction",
"source_id": "document_113",
"arxiv_id": "2511.01999",
"figure": "figure 5",
"page": 8
} | machine_learning | null | null | {
"description_ru": "отзеркалены изображения",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"images"
],
"gold_objects": [
"figure 5"
]
} | {
"orig": "pairs/ml/document_113__figure_5/orig.png",
"fail": "pairs/ml/document_113__figure_5/fail.png",
"orig_sha256": "010b2b353bd4a25808b8e6ad3f0d5d34ccd5f9ac1113639ee0f877a244347b45",
"fail_sha256": "2c7686ab8c486a7066d4c311aa1d2c481ca5aff4047518c09bb801039b03a1d9"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_119__figure_1 | ml | Figure 1. (a)-(b) Comparison of our DeepShield and previ- ous methods in artifact localization using patch-based attention heatmaps. These heatmaps show attention responses from CLIP- ViT for randomly selected image patches. While previous meth- ods emphasize only the most salient specific artifacts using global feature... | have been adopted for deepfake detection due to their strong generalization capabilities [19, 49]. However, these models are primarily optimized for image-level tasks and tend to rely heavily on global features, making them less effective at capturing fine-grained forgery traces. Moreover, these models tend to concentra... | mentions | {
"title": "DeepShield: Fortifying Deepfake Video Detection with Local and Global Forgery Analysis",
"source_id": "document_119",
"arxiv_id": "2510.25237",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "изменен текст блоков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_119__figure_1/orig.png",
"fail": "pairs/ml/document_119__figure_1/fail.png",
"orig_sha256": "035224014eb5a31862dccb95890260c75ff718061dacae67c937a2f44faf1e93",
"fail_sha256": "fd05cf9c02341d07ad5e70f9a081e4838a84b98858a338912fe14162d45ba1e5"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_120__figure_1 | ml | Figure 1. Overview of our proposed method: a vision-language train- ing/inference pipeline to use standard ConvNets and LLMs to integrate multi-modal information using Co-Attention mechanism and joint feature representation learning. | The model is depicted in figure 1 transforms tabular data into text form and subsequently runs a VLM fine-tuning with both modalities to solve a classification task. Assum- ing a multi-modal dataset of paired mammography images and general patient information in the tabular form, we first transform the tabular meta-inf... | mentions | {
"title": "Breast Cancer VLMs: Clinically Practical Vision-Language Train-Inference Models",
"source_id": "document_120",
"arxiv_id": "2510.25051",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "Изменен порядок блоков",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_120__figure_1/orig.png",
"fail": "pairs/ml/document_120__figure_1/fail.png",
"orig_sha256": "d74f99ee4fc34ab7ac524c6891e26010181d5f95f0a128734ee1c52d709b8d6a",
"fail_sha256": "2900542e4fe8264abc7efc791622a220542e7db634d0f2ae6e03e9e9d05a4fb9"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_122__figure_1 | ml | Figure 1. We propose a LLM-driven automated red-teaming framework for T2I models. (a) We compare red-teaming performance on SLD- MAX, our method generates human-readable adversarial prompts that successfully induce safe T2I models to produce unsafe content. (b) Our approach achieves the highest red-teaming success rate... | Ring-A-Bell does not include relevant concept embeddings and performs poorly, we only compared ours with Unlearn- DiffAtk and P4D. Tab. 3 and Fig. 1 indicate that our method still achieves SOTA performance on these concepts. We also find that directly using the nudity/violence’s suf- fix generator to generate adversari... | mentions | {
"title": "AutoPrompt: Automated Red-Teaming of Text-to-Image Models via LLM-Driven Adversarial Prompts",
"source_id": "document_122",
"arxiv_id": "2510.24034",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "Изменен текст внутри блоков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_122__figure_1/orig.png",
"fail": "pairs/ml/document_122__figure_1/fail.png",
"orig_sha256": "cb98ef3640ed2b4180c4e65d88aa1f02c59f06ea6c5ce218484214f7b63acba7",
"fail_sha256": "fa3e40cf2cf97e92b456204ef4a4bd94be6e382525d8e0ef39fec415961e2eee"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_123__figure_2 | ml | Figure 2. The architecture of AesCrop. The input image is processed by dual-stream encoders: (1) a module that generates the Mamba Composition Attention Bias (MCAB) and (2) an encoder that produces image embeddings. A decoder then refines learnable queries via self-attention, aggregates those queries with the image emb... | a Conditional DETR-based decoder inspired by the design of Jia et al. [5]. The complete architecture is illustrated in Fig. 2. Figure 2. The architecture of AesCrop. The input image is processed by dual-stream encoders: (1) a module that generates the Mamba Composition Attention Bias (MCAB) and (2) an encoder that prod... | mentions+same_page_context | {
"title": "AesCrop: Aesthetic-driven Cropping Guided by Composition",
"source_id": "document_123",
"arxiv_id": "2510.22528",
"figure": "figure 2",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Добавлены блоки",
"source_labels": [],
"operators": [
"element_addition"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_123__figure_2/orig.png",
"fail": "pairs/ml/document_123__figure_2/fail.png",
"orig_sha256": "e5e4f52a69c3ea033ea6da2d4b74c62920a614e5fd18b83199cf35170da3038e",
"fail_sha256": "1bb16b17b13a8d4a936c0b6ce64b73b8964ee36385ab7abdf69d854bc8395ef7"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_124__figure_1 | ml | Figure 1. Comparison of generalization metrics w/ and w/o UniPAN trained on WorldView-3 (WV3) using FusionNet [19]. | To this end, this paper introduces a novel approach that aims to enpower any pansharpening models with generaliz- ability by employing a unified distribution strategy (dubbed as UniPAN). The proposed method aims to bridge the gap between training and testing distributions, thereby improving the model’s ability to gener... | mentions | {
"title": "Enpowering Your Pansharpening Models with Generalizability: Unified Distribution is All You Need",
"source_id": "document_124",
"arxiv_id": "2510.22217",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "Исправлено местоположение подписей",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"labels"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_124__figure_1/orig.png",
"fail": "pairs/ml/document_124__figure_1/fail.png",
"orig_sha256": "043f254c06e42d4a0cb64cb253c7a893531608969bca56b041d4dd51186353d5",
"fail_sha256": "558ae4f08430892e53efd7f0785d394a1967fc29b03df01b3264324dd6a32a9e"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_127__figure_1 | ml | Figure 1. Individually trained decoders (–, - -) exhibit good per- | an illustrative experiment that studies layer-wise perfor- mance, in which inference is halted after a given layer and the resulting hidden state is passed to a decoder to make a prediction. For LLMs, this evaluation typically applies the original decoder (i.e., the detokenizer) and is referred to as the ”logit lens” (... | mentions | {
"title": "Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models",
"source_id": "document_127",
"arxiv_id": "2605.06510",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "Изменены график",
"source_labels": [],
"operators": [
"content_replacement"
],
"objects": [
"plots"
],
"gold_objects": [
"figure 1 left"
]
} | {
"orig": "pairs/ml/document_127__figure_1/orig.png",
"fail": "pairs/ml/document_127__figure_1/fail.png",
"orig_sha256": "d7fe92797cf1a3307a402f476b0374938ee9b82849bd80c9670930594e4b18f9",
"fail_sha256": "4b284de1c05d3ef366f6f03688068719bc075a0ee46c8a4fd4aa3e0067887e04"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_127__figure_2 | ml | Figure 2. Experiments analyzing the inference process of tabular | In the following, we describe our 6 experiments, each ac- companied by its description, results, observations, and main takeaways highlighted in blue boxes. Each experiment studies one aspect of the inference process, and we order the experiments from low-level, comparing embeddings, to higher-level, ablating layers, a... | mentions | {
"title": "Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models",
"source_id": "document_127",
"arxiv_id": "2605.06510",
"figure": "figure 2",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Добавлены стрелки",
"source_labels": [],
"operators": [
"element_addition"
],
"objects": [
"arrows"
],
"gold_objects": [
"figure 2 a b"
]
} | {
"orig": "pairs/ml/document_127__figure_2/orig.png",
"fail": "pairs/ml/document_127__figure_2/fail.png",
"orig_sha256": "a9094b0d6b9a114914f549ce673438d8712eea7af72656cc6bca83e24486b282",
"fail_sha256": "37510ed490e90fadf9751dce7fd62c11621386bd7f8864d3cfcc27271158c31e"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_128__figure_1 | ml | Figure 1. The closed-loop dynamics of epistemic intelligence. The diagram illustrates the coupled ODE system, with state variables | Human–AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective Figure 1. The closed-loop dynamics of epistemic intelligence. The diagram illustrates the coupled ODE system, with state variables H(t), Q(t), M(t) (top) linked by forward enhancement pathways (solid arrows) and control variables u, A, S (bot... | same_page_context | {
"title": "Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective",
"source_id": "document_128",
"arxiv_id": "2605.06347",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "отзеркален блок, изменены уравнения внизу",
"source_labels": [],
"operators": [
"label_text_edit",
"geometric_distortion"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_128__figure_1/orig.png",
"fail": "pairs/ml/document_128__figure_1/fail.png",
"orig_sha256": "1d2e1fa50946340e3eee9913469b40cfae5c934830c93abb5d0a1846fa76f657",
"fail_sha256": "f7750fcdf5decb27effeb2236852afd7ca40695eca9f501dab493fda5180cbf9"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_130__figure_3 | ml | Figure 3. Safety collapse under parameter distance constraints. | Constraining parameter distance fails to prevent safety collapse. Figure 3 undermines the assumption that limiting parameter distance ensures safety. We observe that while benign task learning naturally exhibits larger parameter drift (L2 Distance > 2), the restoration of harmful capabilities is successfully achieved e... | mentions | {
"title": "Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks",
"source_id": "document_130",
"arxiv_id": "2605.05995",
"figure": "figure 3",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Изменены численные значения на графики и подписи к ним",
"source_labels": [],
"operators": [
"value_fabrication"
],
"objects": [
"labels",
"plots"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_130__figure_3/orig.png",
"fail": "pairs/ml/document_130__figure_3/fail.png",
"orig_sha256": "e1bdb1acd3c47f5886a520e5150a57f567acc76a59702f279f9d024fee12f326",
"fail_sha256": "6f48de3f1ce2ebc32cb693abeeb8e245d8fbfb6ae7051b48c3606eadcafd167a"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_131__figure_2 | ml | Figure 2. Comparison of SV training and inference strategies. Traditional fine-tuned FSSVs always require post-hoc factor selection, while our joint training scheme enables end-to-end cross-concept scalability and is compatible with both FSSV and PrOSV. | Current scheme for SV training and inference. Response- only language modeling (Lang.) is a common objective for training SVs (see §F.4) (Subramani et al., 2022). As for trainable parameters, previous works view the steering direction (v) as the only trainable parameter while treating the steering factor (α) as an exte... | mentions | {
"title": "Towards Steering without Sacrifice: Principled Training of Steering Vectors for Prompt-only Interventions",
"source_id": "document_131",
"arxiv_id": "2605.05983",
"figure": "figure 2",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Изменены направления стрелок, изменены блоки",
"source_labels": [],
"operators": [
"edge_rewiring",
"generic_edit"
],
"objects": [
"blocks",
"arrows"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_131__figure_2/orig.png",
"fail": "pairs/ml/document_131__figure_2/fail.png",
"orig_sha256": "c7a341ed355f4cca8e2e2133076add78948e461ddf095efe471b22d814d54f72",
"fail_sha256": "19471ec3b744eec722c85f811dc3b362fe1ed4a569041c660d6702589545a988"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_133__figure_2 | ml | Figure 2. Framework of PLMD. (a) illustrates the PLMD pipeline for a single robot. In MRON, the same module is executed independently | We pre-train the obstacle map network Gϕ first to attain robust obstacle map priors before joint optimization with the semantic map network ˜Gϕ(sτ, cτ−1, τ). After pretraining the obstacle map network Gϕ, we freeze it during semantic map network ˜Gϕ training and cease further updates. ˜Gϕ integrates obstacle map inform... | mentions | {
"title": "Plug-and-Play Label Map Diffusion for Universal Goal-Oriented Navigation",
"source_id": "document_133",
"arxiv_id": "2605.05960",
"figure": "figure 2",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Изменен текст внутри блоков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_133__figure_2/orig.png",
"fail": "pairs/ml/document_133__figure_2/fail.png",
"orig_sha256": "fb75a23883a80287c72828268104b7f1ec7af9eb6add93bea188fcaa6aedc548",
"fail_sha256": "d3a15c3522be45ed8907930493b43d64bbcb82dc7be5db813818b056fb770adb"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_134__figure_2 | ml | Figure 2. The overall architecture of HEDP. (a) Training phase: In DIL, single-domain visual-textual prompt parameters are optimized | Fig. 2 presents the HEDP architecture, a prompt-based DIL framework designed to prevent catastrophic forgetting and improve generalization without data replay. During train- ing, domain-specific prompts are independently optimized with an energy regularization loss to constrain energy distri- butions while preserving a... | mentions | {
"title": "HEDP: A Hybrid Energy-Distance Prompt-based Framework for Domain Incremental Learning",
"source_id": "document_134",
"arxiv_id": "2605.05776",
"figure": "figure 2",
"page": 2
} | machine_learning | null | null | {
"description_ru": "перемешаны изображения, добавлены новые стрелки",
"source_labels": [],
"operators": [
"block_permutation",
"element_addition"
],
"objects": [
"images",
"arrows"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_134__figure_2/orig.png",
"fail": "pairs/ml/document_134__figure_2/fail.png",
"orig_sha256": "28c8f2f335d9a4d1ddd42b3eced25ebf4e57809a09f44fadb9d1010480fe3f77",
"fail_sha256": "8507b5d4006dc44c43ba6da2a4d429190f04a5c2adab44c73c343e2b2e5ac7dd"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_135__figure_1 | ml | Figure 1. FSPECGNN generalizes spectral filtering by lifting sig- | That is, right (resp. left) multiplication by L propagates the signal by changing the second (resp. first) coordinate of G□G while keeping the other fixed, see the first row of Figure 1 for an illustration. Full-Spectrum Graph Neural Network: Expressive and Scalable Xiaohan Wang * 1 Deyu Bo * 1 Longlong Li 1 Kelin Xia ... | mentions+same_page_context | {
"title": "Full-Spectrum Graph Neural Network: Expressive and Scalable",
"source_id": "document_135",
"arxiv_id": "2605.05759",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "Изменены подписи к стрелкам",
"source_labels": [],
"operators": [
"edge_rewiring"
],
"objects": [
"labels",
"arrows"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_135__figure_1/orig.png",
"fail": "pairs/ml/document_135__figure_1/fail.png",
"orig_sha256": "ae1640c59e3206fbaad2b728a6f0d8a5ebe12fe266d7fdf0bd823c1f40cc653e",
"fail_sha256": "3d43fa63ecab265bf960f9ef4583ffd7ca3772f3e77d7f6158b7b2f1df17c6b2"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_136__figure_1 | ml | Figure 1. Comparison between (a) Traditional coarse-grained | ized auxiliary agents for runtime monitoring (Luo et al., 2025; Chen et al., 2025; Xiang et al., 2024) or fine-tuning safety models to enforce alignment (AI, 2024; Zhang et al., 2025a). However, these approaches typically necessitate either extensive model retraining or the deployment of resource-intensive proxies, lea... | mentions | {
"title": "SafeHarbor: Hierarchical Memory-Augmented Guardrail for LLM Agent Safety",
"source_id": "document_136",
"arxiv_id": "2605.05704",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "Изменены подписи к блокам",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_136__figure_1/orig.png",
"fail": "pairs/ml/document_136__figure_1/fail.png",
"orig_sha256": "da3d7a31106910395a91c497dbd3295eb7bd3e343e70fdec04e0d292d626142f",
"fail_sha256": "8a53e794fa75751aa1d3305c53d92b5351825b8146e3426ce49772c003c1a07a"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_137__figure_2 | ml | Figure 2. Method Overview. Given the input head trajectory signals H1:T | cpf via for- ward kinematics to generate global human motions. We implement the model F via a transformer-based diffusion architecture (Vaswani et al., 2017). This design allows the model to denoise the motion M from Gaussian noise condi- tioned on the head trajectory progressively. The overview of proposed MOTIONGRPO ... | mentions+same_page_context | {
"title": "MotionGRPO: Overcoming Low Intra-Group Diversity in GRPO-Based Egocentric Motion Recovery",
"source_id": "document_137",
"arxiv_id": "2605.05680",
"figure": "figure 2",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Добавлены новые стрелки и подписи",
"source_labels": [],
"operators": [
"element_addition"
],
"objects": [
"labels",
"arrows"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_137__figure_2/orig.png",
"fail": "pairs/ml/document_137__figure_2/fail.png",
"orig_sha256": "4b19b2fa99c9c355bf27fd77975cb9b8dd0b7ab2523c74dc70ad9da9bf58996c",
"fail_sha256": "040797b49f92e022c3994b23e9febbb56115b668609fbdb268799235dfb59b0b"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_14__figure_12 | ml | Figure 12. First frame visualizations of RGB 3D inference (Part 3). Each image shows the initial particle distribution and segmentation for the respective sequence. | We include supplementary videos that visualize the full particle-based inference process across time for the small qualitative deformable dataset introduced in the main paper: cloth bag, gray jacket, jello, manta ray, eagle, ostrich, purple jacket, snake, whiskey swirl, and wine swirl. These se- quences span a range of... | mentions | {
"title": "GenMatter: Perceiving Physical Objects with Generative Matter Models",
"source_id": "document_14",
"arxiv_id": "arXiv:2604.22160",
"figure": "figure 12",
"page": 25
} | machine_learning | null | null | {
"description_ru": "Изображение искажено",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"images"
],
"gold_objects": [
"figure 12а"
]
} | {
"orig": "pairs/ml/document_14__figure_12/orig.png",
"fail": "pairs/ml/document_14__figure_12/fail.png",
"orig_sha256": "f7029c67f522e4aa30ac4128cf4fe3ce999795bb8f343c392fd75fa7f0e2df24",
"fail_sha256": "32ae36023c2e59277364a335b75fda4c0380db769ee0c2a474978c04e03ab4a0"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_14__figure_5 | ml | Figure 5. Qualitative comparison on camouflaged stimuli. Probe point segmentation on scene 16, texture 01. The depth estimate is uninformative, and the flow estimate shows that on-axis rotation causes opposing motion at top vs. bottom (blue vs. red). GenMatter correctly segments the moving object, while FlowSAM segment... | Evaluation protocol Following the probe-point methodol- ogy established in the RDK experiments, we evaluate seg- mentation by sampling query points rather than densely eval- uating all pixels. Dense evaluation on all ground-truth pixels requires establishing correspondence between predicted and ground-truth segments, w... | mentions | {
"title": "GenMatter: Perceiving Physical Objects with Generative Matter Models",
"source_id": "document_14",
"arxiv_id": "arXiv:2604.22160",
"figure": "figure 5",
"page": 7
} | machine_learning | null | null | {
"description_ru": "Изображения перемешаны",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"images"
],
"gold_objects": [
"figure 5"
]
} | {
"orig": "pairs/ml/document_14__figure_5/orig.png",
"fail": "pairs/ml/document_14__figure_5/fail.png",
"orig_sha256": "bdbe4fca08015b39bce8fc17feed6ca002ee887af53988ce9abd728f00c776fc",
"fail_sha256": "01e62580c5464d5cc521cd57ddc087931baf4489bfb059fe394af9f2483a9190"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_140__figure_2 | ml | Figure 2. Verification of Manifold Orthogonality. (a-b) Dynam- | To resolve manifold misalignment, we propose that the con- flict between semantics and structure stems from parame- ter entanglement rather than intrinsic task incompatibility. By leveraging the architectural properties of a Transformer block, we formulate the Gradient Orthogonality Hypoth- esis, which posits that thes... | mentions | {
"title": "MUSE: Resolving Manifold Misalignment in Visual Tokenization via Topological Orthogonality",
"source_id": "document_140",
"arxiv_id": "2605.05646",
"figure": "figure 2",
"page": 3
} | machine_learning | null | null | {
"description_ru": "отзеркалена ось ординат",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"axes"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_140__figure_2/orig.png",
"fail": "pairs/ml/document_140__figure_2/fail.png",
"orig_sha256": "83e1e8f0b2e8129f0f8f4c05f843469a802ab43fa86ff8604b92befc1ff5b8af",
"fail_sha256": "835a1d6e39c3d8822e3254584e6f0dcb22aeef165d5e3c32eeb028c6ed375ecf"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_145__figure_1 | ml | Figure 1: Same BO loop, same surrogates, same action spaces: four distinct winning methods across five regime conditions. Rows = acquisition functions; columns = regime conditions; values = Hit@1 (last column: Discovery AUC). The winner per column is outlined. No fixed acquisition rule dominates across regimes; the “be... | At the small-budget regime (PRS ≈0.07), BOHB/Hyperband’s multi-fidelity advantage is exploited and both substantially outperform vanilla BO. At the large-budget regime (PRS ≈0.74), the paper reports: “for large enough budgets TPE and GP-BO caught up in all cases, and in the end found better configurations than HB and R... | mentions | {
"title": "Regime-Conditioned Evaluation in Multi-Context Bayesian Optimization",
"source_id": "document_145",
"arxiv_id": "2605.04895",
"figure": "figure 1",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Изменены подписи",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"labels"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_145__figure_1/orig.png",
"fail": "pairs/ml/document_145__figure_1/fail.png",
"orig_sha256": "f2fa3b26b49fc074cee2463595636880346849f07472bf0d4a84d536d8afc388",
"fail_sha256": "3ef314e90626bedf7c8ad35f80de52a35ae4f3800dcfc2e80cf8e6fe772d528a"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_146__figure_1 | ml | Figure 1. Overview of Quantile-free Prediction Interval Graph Neural Network. QpiGNN estimates node-wise prediction intervals | Separating the two heads also prevents representational con- flicts and gradient interference between prediction and un- certainty estimation, enabling each branch to learn its own function class more effectively. Similar dual-head structures have demonstrated strong performance in heteroscedastic and Bayesian regressi... | mentions | {
"title": "Quantile-Free Uncertainty Quantification in Graph Neural Networks",
"source_id": "document_146",
"arxiv_id": "2605.04847",
"figure": "figure 1",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Удалены блоки, изменены блоки",
"source_labels": [],
"operators": [
"element_deletion",
"generic_edit"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_146__figure_1/orig.png",
"fail": "pairs/ml/document_146__figure_1/fail.png",
"orig_sha256": "73b8adc31f0afeaffba38370c69f11b40b420b95039f1da5be04c5f76a1750ce",
"fail_sha256": "03608144a4f0c6cbc45673b6a68834c7055d38ae5adaf6c86f321b5b03e75d09"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_149__figure_1 | ml | Figure 1. Plasticity loss as loss of spectral plasticity. Left: | SPHERE: Mitigating the Loss of Spectral Plasticity in Mixture-of-Experts for Lirui Luo 1 2 Guoxi Zhang 2 Hongming Xu 2 Cong Fang† 1 3 Qing Li† 2 In deep reinforcement learning (DRL), an agent is trained from a stream of experience. In a continual learning setting, such agents can suffer from plas- ticity loss: their ab... | same_page_context | {
"title": "SPHERE: Mitigating the Loss of Spectral Plasticity in Mixture-of-Experts for Deep Reinforcement Learning",
"source_id": "document_149",
"arxiv_id": "2605.04712",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "перемешаны изображения",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"images"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_149__figure_1/orig.png",
"fail": "pairs/ml/document_149__figure_1/fail.png",
"orig_sha256": "0a74c7bfe4255aba4a2a7e2e2d0bdfe0d6169973fa554297d262c7e195f1b2d7",
"fail_sha256": "b61e31b67609b61004ea5ddf1149cf76f80d4db198c714e656da4a0cc3cae76b"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_149__figure_6 | ml | Figure 6. SPHERE avoids spectral collapse to a single component. We visualize held-out states with t-SNE and color each state by | Qualitative Analysis. Figure 6 shows the evolution of effective-rank collapse in the weighted expert represen- tations and the associated loss of spectral plasticity over the HumanoidBench CRL sequence. Using a fixed held- out state batch from Section 5.1, we visualize last-layer weighted expert features with t-SNE at ... | mentions | {
"title": "SPHERE: Mitigating the Loss of Spectral Plasticity in Mixture-of-Experts for Deep Reinforcement Learning",
"source_id": "document_149",
"arxiv_id": "2605.04712",
"figure": "figure 6",
"page": 8
} | machine_learning | null | null | {
"description_ru": "перемешаны изображения",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"images"
],
"gold_objects": [
"figure 6"
]
} | {
"orig": "pairs/ml/document_149__figure_6/orig.png",
"fail": "pairs/ml/document_149__figure_6/fail.png",
"orig_sha256": "9e20c41af0bb36ed486d9435fb3ffc5e3dbff4aebac635505e4004235a103526",
"fail_sha256": "c1a7c49e03ea4be92444e95f14d072dcf42cb2e2d834b1f313dc394a3f573e88"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_152__figure_1 | ml | Figure 1. Cumulative Distribution Function for Batch Execution Time with PD ratio 1:1 requests | ≈M, the total attention computa- tion per batch becomes approximately constant (Pope et al., 2023; Kaplan et al., 2020; Hoffmann et al., 2022). More- over, our chunked prefill formulation bounds the per-batch prefill cost by fixing the chunk size ˆs, thereby preventing the quadratic attention complexity of long prompts... | mentions | {
"title": "A Queueing-Theoretic Framework for Stability Analysis of LLM Inference with KV Cache Memory Constraints",
"source_id": "document_152",
"arxiv_id": "2605.04595",
"figure": "figure 1",
"page": 4
} | machine_learning | null | null | {
"description_ru": "изменены подписи к рисункам",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"labels",
"images"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_152__figure_1/orig.png",
"fail": "pairs/ml/document_152__figure_1/fail.png",
"orig_sha256": "4825dbd81f5359c54a00dab90ed6a613992e027a4be4625390c01c6195a5abc6",
"fail_sha256": "5d703c429afb988e3c910f1dd892880bd5dc6c7b7c7f02c4e833d8004492c94f"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_153__figure_2 | ml | Figure 2. Parameter Drift trajectories along safety and danger | Parameter Dynamics During Fine-tuning. Figure 2 shows the parameter drift trajectories and corresponding Safety Scores during fine-tuning of Qwen3-8B (Yang et al., 2025) on Dolly(5k) (Conover et al., 2023). Safety Score is a reward-model-based metric that quantifies the overall safety of model responses (detailed in Ap... | mentions | {
"title": "From Parameter Dynamics to Risk Scoring : Quantifying Sample-Level Safety Degradation in LLM Fine-tuning",
"source_id": "document_153",
"arxiv_id": "2605.04572",
"figure": "figure 2",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Изменение подписи на оси абсцисс, отзеркалена вспомогательная шкала",
"source_labels": [],
"operators": [
"label_text_edit",
"geometric_distortion"
],
"objects": [
"labels",
"axes"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_153__figure_2/orig.png",
"fail": "pairs/ml/document_153__figure_2/fail.png",
"orig_sha256": "38541c22b8574dbd3ed3f3cd20a389f039ef4fb1db8b3643871c1d90533aca14",
"fail_sha256": "d99ae87512c236cd1ccfc27af59a3a6370929082cabec33909181a9152c0cc5a"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_157__figure_1 | ml | Figure 1. Overview of WZ-LLM. (A) Symbolic decomposition. Given combinatorial-identity statements, the symbolic engine checks WZ applicability and, when applicable, constructs a WZ proof sketch that is decomposed into sub-goals. Statements that are not amenable to the WZ method are labeled as WZ-uncovered targets. All ... | In this section, we present WZ-LLM, a neuro-symbolic framework for formally proving combinatorial identities in Lean 4. As shown in Figure 1, for identities that are amenable to the WZ method, WZ-LLM first constructs an executable WZ proof sketch and compiles the origi- nal goal into a set of structured proving tasks, ... | mentions | {
"title": "Automated Formal Proofs of Combinatorial Identities via Wilf-Zeilberger Guidance and LLMs",
"source_id": "document_157",
"arxiv_id": "2605.04472",
"figure": "figure 1",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Изменение надписей на изображении",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"labels",
"images"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_157__figure_1/orig.png",
"fail": "pairs/ml/document_157__figure_1/fail.png",
"orig_sha256": "7521c711bdceff6ca57b4da663f81f23d72a70129cd5d5d7e372db018f2b1e66",
"fail_sha256": "5fe94f5fefafeea2da1f76829038207b92d95823338472404769858a7e750563"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_159__figure_2 | ml | Figure 2: Synthetic vocabulary learning and forgetting. Left: Test loss versus the number of synthetic training sentences N; the dashed line denotes the base model’s performance on real sentences. Right: Forgetting across methods and sample sizes, where embedding tuning achieves zero forgetting while other methods incu... | In Figure 2, we vary the number of synthetic training sentences (N) from 10 to 3000 and tune the LLaMA-3.2-1B model (Patterson et al., 2022) for 100 epochs. Performance is evaluated using token- level cross-entropy loss on a fixed test set of 1000 synthetic sentences. Specifically, Figure 2 (left) reports the test loss... | mentions | {
"title": "Memory as a Markov Matrix: Sample Efficient Knowledge Expansion via Token-to-Dictionary Mapping",
"source_id": "document_159",
"arxiv_id": "2605.04308",
"figure": "figure 2",
"page": 10
} | machine_learning | null | null | {
"description_ru": "Изменен график",
"source_labels": [],
"operators": [
"content_replacement"
],
"objects": [
"plots"
],
"gold_objects": [
"figure 2 left"
]
} | {
"orig": "pairs/ml/document_159__figure_2/orig.png",
"fail": "pairs/ml/document_159__figure_2/fail.png",
"orig_sha256": "6ed88e181051d2a83ebf19e71bbb6aedb3ab2b1df2a3f2ff5c75dc56bfd254e8",
"fail_sha256": "332eb7639988b80a46d70183a51d5a0aabf27c394d7b113e51890213a5aff064"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_160__figure_7 | ml | Figure 7. Comparison of Lennard-Jones (LJ) potential energy | density evaluations. Both CDS and DiGS recover all modes with correct proportions, whereas competing methods fail to discover all of them. Energy histograms for the LJ sys- tems (Fig. 7) indicate that CDS matches the target energy distribution most accurately, closely followed by NRPT. For the ALDP task, the Ramachandr... | mentions | {
"title": "Conditional Diffusion Sampling",
"source_id": "document_160",
"arxiv_id": "2605.04013",
"figure": "figure 7",
"page": 9
} | machine_learning | null | null | {
"description_ru": "изображений отзеркалено",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"images"
],
"gold_objects": [
"figure 7 right"
]
} | {
"orig": "pairs/ml/document_160__figure_7/orig.png",
"fail": "pairs/ml/document_160__figure_7/fail.png",
"orig_sha256": "d2ceac2e4aa2e0dc2f8ac9579ea85d5dfb65ebdb16bbcbd96c9b2517c0d7247e",
"fail_sha256": "db185a7526374cf38ee64edaa7395aac5b191ba99727a2ddc83bd5e4cd193e61"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_161__figure_3 | ml | Figure 3. Ala4 molecular structures represented by torsion coordinates (ϕ1, ϕ2, ϕ3) with 23 distinct modes. | angles (ϕ1, ϕ2, ϕ3), which gives rise to 23 = 8 metastable modes corresponding to the possible combinations of the three backbone torsional states. As depicted in Figure 3, Flow Sampling covers all 8 modes observed in the MD reference samples. The samples are also more tightly con- centrated around the MD-supported tor... | mentions | {
"title": "Flow Sampling: Learning to Sample from Unnormalized Densities via Denoising Conditional Processes",
"source_id": "document_161",
"arxiv_id": "2605.03984",
"figure": "figure 3",
"page": 8
} | machine_learning | null | null | {
"description_ru": "искажено изображение",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"images"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_161__figure_3/orig.png",
"fail": "pairs/ml/document_161__figure_3/fail.png",
"orig_sha256": "9d388be8b0a807ac82d9e7635d086a7760f9e96a83241049aa2995391598c3c0",
"fail_sha256": "1467a2dc1a406c0f08d24d863c8d8286b9bad8266a6758f9a5d558719eea3d1f"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_162__figure_1 | ml | Figure 1. BERT+FEATATTN architecture. The input text is encoded by the transformer, yielding a [CLS] representation. In parallel, | Feature-Augmented Transformers for Robust AI-Text Detection Across Domains and Generators Figure 1. BERT+FEATATTN architecture. The input text is encoded by the transformer, yielding a [CLS] representation. In parallel, linguistic features are extracted, reduced to the selected subset, and mapped to a compact embedding... | same_page_context | {
"title": "Feature-Augmented Transformers for Robust AI-Text Detection Across Domains and Generators",
"source_id": "document_162",
"arxiv_id": "2605.03969",
"figure": "figure 1",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Изменение блоков",
"source_labels": [],
"operators": [
"generic_edit"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_162__figure_1/orig.png",
"fail": "pairs/ml/document_162__figure_1/fail.png",
"orig_sha256": "b7724c0527ffafe10d3751dc2d95a32f591b3ac1feb19d964cae38be412a0a44",
"fail_sha256": "9d96dfc6074149d5d62a28760bdc731d1b1d0f1dab0ca7596fab8063a16b80b9"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_163__figure_2 | ml | Figure 2. Data Processing Pipeline and Overview. As shown in Figure 2 a), our data processing pipeline consists of the following | Existing high-quality world model training data is predomi- nantly sourced from indoor scenes, while outdoor datasets are limited in both quantity and mobility rates. To effec- tively expand the dataset, we selected 18 high-quality en- vironments across 4 outdoor urban simulators. As shown in Figure 2, we manually iden... | mentions | {
"title": "iWorld-Bench: A Benchmark for Interactive World Models with a Unified Action Generation Framework",
"source_id": "document_163",
"arxiv_id": "2605.03941",
"figure": "figure 2",
"page": 5
} | machine_learning | null | null | {
"description_ru": "Изменение изображений",
"source_labels": [],
"operators": [
"content_replacement"
],
"objects": [
"images"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_163__figure_2/orig.png",
"fail": "pairs/ml/document_163__figure_2/fail.png",
"orig_sha256": "83a1b551739dcc8a43e173875bf377d39a30ec55a097ff669ddc74d97746c2fc",
"fail_sha256": "b9a8ac9aef3c87e8f35c54c232624a5b6868cb35f903992073c623e4d474679e"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_166__figure_2 | ml | Figure 2. Overview of the GLANCE framework. GLANCE unifies world modeling and exploration into a self-supervised cross-modal | In this section, we propose GLANCE, a unified framework designed to transform the VLM agent’s internalized world modeling into a structured source of epistemic drive. As illustrated in Fig. 2, the architecture consists of two parallel streams: an Online VLM Agent that generates reasoning and visual-grounded predictions... | mentions+same_page_context | {
"title": "What You Think is What You See: Driving Exploration in VLM Agents via Visual-Linguistic Curiosity",
"source_id": "document_166",
"arxiv_id": "2605.03782",
"figure": "figure 2",
"page": 4
} | machine_learning | null | null | {
"description_ru": "изменен текст внутри блоков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_166__figure_2/orig.png",
"fail": "pairs/ml/document_166__figure_2/fail.png",
"orig_sha256": "f0ee28852ec2a492c39aec46e262284f5bd3eb87efa8c96424b9e958f0bdf48d",
"fail_sha256": "2c45f2c24c70ae5633c9ee417b89f17b743842fa879610239656fa3a72fbff60"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_166__figure_4 | ml | Figure 4. Sensitivity analysis of the exploration weight β. | Sensitivity to Exploration Weight (β) Finally, we an- alyze the sensitivity of the exploration coefficient β. As shown in Fig. 4, we observe that the VLM performance peaks at moderate values, exhibiting a non-monotonic rela- tionship. A moderate weight (β ∈[0.1, 0.3]) strikes the op- timal balance, providing sufficient... | mentions | {
"title": "What You Think is What You See: Driving Exploration in VLM Agents via Visual-Linguistic Curiosity",
"source_id": "document_166",
"arxiv_id": "2605.03782",
"figure": "figure 4",
"page": 7
} | machine_learning | null | null | {
"description_ru": "Изменены подписи к осям",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"labels"
],
"gold_objects": [
"figure 4"
]
} | {
"orig": "pairs/ml/document_166__figure_4/orig.png",
"fail": "pairs/ml/document_166__figure_4/fail.png",
"orig_sha256": "c27ad4b509b54c9bc0d069b18cf3aa9f3fdc7d638091d4c2ca9aac948ebc6553",
"fail_sha256": "130b4178aba6db72f67b38f94b4457504a4fba8292fb8231d2b4a878af228bde"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_167__figure_3 | ml | Figure 3. Architecture of the proposed method. (a) DAEDL: a spectrally normalized backbone with a single evidential head that outputs | 2. Method Figure 3 summarizes the architectures considered in this work. Figure 3a presents the density-aware DAEDL base- line, while Figure 3b illustrates the proposed GEM-FI ar- chitecture and its main components. The added modules GEM-FI: Gated Evidential Mixtures with Fisher Modulation Figure 3. Architecture of the... | mentions+same_page_context | {
"title": "GEM-FI: Gated Evidential Mixtures with Fisher Modulation",
"source_id": "document_167",
"arxiv_id": "2605.03750",
"figure": "figure 3",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Удалены блоки, стрелки, подписи к блокам",
"source_labels": [],
"operators": [
"label_text_edit",
"element_deletion",
"edge_rewiring"
],
"objects": [
"blocks",
"labels",
"arrows"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_167__figure_3/orig.png",
"fail": "pairs/ml/document_167__figure_3/fail.png",
"orig_sha256": "fd9ddd1a5513c6b05b4d30f04f16614bad36e6180760ec91967c7b82e63384ef",
"fail_sha256": "8eedc23197bdbaed2e423fb4d303eac5588dbadeb328be1226b7ed035925c419"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_168__figure_2 | ml | Figure 2. Architecture of OneTrackerV2. OneTrackerV2 em- | As shown in Figure 2, OneTrackerV2 is a unified framework that processes template and search regions from diverse modalities (e.g., RGB, Depth, Thermal) within a single architecture. To address the challenge of heterogeneous inputs, we first employ Meta Merger, which maps RGB and multimodal features into a shared embed... | mentions | {
"title": "Unified Multimodal Visual Tracking with Dual Mixture-of-Experts",
"source_id": "document_168",
"arxiv_id": "2605.03716",
"figure": "figure 2",
"page": 3
} | machine_learning | null | null | {
"description_ru": "изменены подписи блоков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_168__figure_2/orig.png",
"fail": "pairs/ml/document_168__figure_2/fail.png",
"orig_sha256": "a6c29e5c9eee3f543a64833e2598afc1fa10e0c8924270af55ecee551e82fab4",
"fail_sha256": "72a65588414e165b4f54015e5f2a27126d4d346eae2eccade6fbc5d6c3d47409"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_168__figure_3 | ml | Figure 3. (a) Meta Merger. Meta Merger unifies RGB and mul- | A central challenge in multimodal tracking is integrating heterogeneous modalities into a shared representation while maintaining robustness to missing data. Unlike task-specific branches (Hou et al., 2024) that increase complexity, or naive concatenation (Chen et al., 2025) that leads to feature entanglement, we intro... | mentions | {
"title": "Unified Multimodal Visual Tracking with Dual Mixture-of-Experts",
"source_id": "document_168",
"arxiv_id": "2605.03716",
"figure": "figure 3",
"page": 3
} | machine_learning | null | null | {
"description_ru": "изменены подписи блоков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_168__figure_3/orig.png",
"fail": "pairs/ml/document_168__figure_3/fail.png",
"orig_sha256": "05bd614856085361a086f965ac13108a65c233690be118127ba7bf488f2c1c91",
"fail_sha256": "d3ec0778ebb0c7f2d5250a5a1c786a5fb8610eae0f6f24e1755ad21612ec6b63"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_171__figure_1 | ml | Figure 1. ZO eNTK v.s. FO eNTK. The pair with high similarity: 4 and 9. The pair with low similarity: 1 and 0. The relative Frobenius | To empirically evaluate the approximation fidelity of the ZO eNTK, we conduct experiments using the LeNet model (d = 29, 624) on the MNIST (LeCun, 1998) dataset. From Figures 1 and 2, we observe a consistent, monotonic reduc- tion in approximation error as the number of perturbations P increases. Figure 1 shows the pro... | mentions | {
"title": "Learning Dynamics of Zeroth-Order Optimization: A Kernel Perspective",
"source_id": "document_171",
"arxiv_id": "2605.03373",
"figure": "figure 1",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Подменены изображения",
"source_labels": [],
"operators": [
"content_replacement"
],
"objects": [
"images"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_171__figure_1/orig.png",
"fail": "pairs/ml/document_171__figure_1/fail.png",
"orig_sha256": "73fc8dd9e6c6a49b1c2657cb562ca2faee740d035aab39f0052c6cb10de28f92",
"fail_sha256": "72cd6cf01bc35ef53978fd066443bf713dfa17212bf5f8bbc0502f995e77de31"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_172__figure_1 | ml | Figure 1. Latent factor decomposition of multimodal inputs into | This structure is illustrated in Figure 1, which depicts the decomposition into shared and modality-unique factors. Toward Structural Multimodal Representations: Specialization, Selection, and Sparsification via Mixture-of-Experts space by adapting the routing mechanism, activating task- relevant experts to construct c... | mentions+same_page_context | {
"title": "Toward Structural Multimodal Representations: Specialization, Selection, and Sparsification via Mixture-of-Experts",
"source_id": "document_172",
"arxiv_id": "2605.03348",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "Перемешаны изображения",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"images"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_172__figure_1/orig.png",
"fail": "pairs/ml/document_172__figure_1/fail.png",
"orig_sha256": "7bbfd05a621c49fbde0e6dfc0f130afa24ca1bf1f3ca6987e9ea18e46f844f4c",
"fail_sha256": "020d5f0584e6ccf654deb2553b55c8ac5637e17285c937ae809447f8b0852ed7"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_175__figure_3 | ml | Figure 3: Per-token cosine similarity for different models. The top row shows DLMs, the bottom shows ARs. The left column shows a comparison of LLaDA-8B and Llama-3.1-8B, where LLaDA-8B exhibits highly redundant (collapsed) layer similarities. For the middle column, we have removed the largest outlier from the similari... | Next, we analyze the pairwise similarities of the hidden representations in LLaDA and Llama in Figure 3. Surprisingly, we find that almost all layers in LLaDA (top left) exhibit an extremely high similarity to each other, especially early layers which are almost identical to each other even over ranges of 15+ layers. T... | mentions | {
"title": "Layer Collapse in Diffusion Language Models",
"source_id": "document_175",
"arxiv_id": "2605.06366",
"figure": "figure 3",
"page": 5
} | machine_learning | null | null | {
"description_ru": "Изображения искажены",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"images"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_175__figure_3/orig.png",
"fail": "pairs/ml/document_175__figure_3/fail.png",
"orig_sha256": "86070a6d41c7559a2ce727e280050cc190730081230c7c5f592542cbcea20c5f",
"fail_sha256": "e0a42e96a0f587b09f8890eda2fe6333bfe487023f5a59d69c0205a83d7b5232"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_176__figure_1 | ml | Figure 1: Given a high-quality first frame, SwiftI2V enables faithful 2K image-to-video generation with fine-grained details in 2 minutes. | 2K I2V result 2K I2V result 2K I2V result Figure 1: Given a high-quality first frame, SwiftI2V enables faithful 2K image-to-video generation with fine-grained details in 2 minutes. There are two challenges for high-resolution I2V. The first is computational scaling at high resolution. The number of visual tokens grows ... | same_page_context | {
"title": "SwiftI2V: Efficient High-Resolution Image-to-Video Generation via Conditional Segment-wise Generation",
"source_id": "document_176",
"arxiv_id": "2605.06356",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "перемешаны изображения",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"images"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_176__figure_1/orig.png",
"fail": "pairs/ml/document_176__figure_1/fail.png",
"orig_sha256": "c14d1f5f5e35cdbcc7b829d800e5807381784d548fba2a527aa9f4d03bf4f917",
"fail_sha256": "0b4898036dfdff16d12f3e761f3101ec6695d3397ee9e31b6ba7ef315492407d"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_178__figure_1 | ml | Figure 1: Data collection setup for EgoEMG. Bilateral EMG wristbands, head-mounted egocentric RGB, external ZED 2i RGB-D, and optical motion capture with hand markers for pose labels. | EgoEMG is collected with a synchronized multimodal capture setup, including bilateral EMG, IMU, egocentric RGB, external RGB-D, and optical motion capture. Each participant wears two WAVELETECH EMG wristbands, one on each wrist, with each wristband recording 8 channels of surface EMG at 2 kHz. The visual setup contains... | mentions | {
"title": "EgoEMG: A Multimodal Egocentric Dataset with Bilateral EMG and Vision for Hand Pose Estimation",
"source_id": "document_178",
"arxiv_id": "2605.05712",
"figure": "figure 1",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Изображения искажены",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"images"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_178__figure_1/orig.png",
"fail": "pairs/ml/document_178__figure_1/fail.png",
"orig_sha256": "822d1a1f465f497fefdca0d1dcfb83128ff0b4e0675453eee87181790f092f90",
"fail_sha256": "21b35b7bccedb1e16dcafe247b231223586adbd2f21694eb8717c1f039bc49ef"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_179__figure_3 | ml | Figure 3: Edge deletion curves (residual branch, ranked by Re). Removing top-ranked edges (red) consistently increases MSE more than random (gray) or bottom-ranked (blue) deletion, providing evidence that activation-range ranking identifies functionally relevant edges. The effect is strongest on Weather and Solar, whic... | Figure 3: Edge deletion curves (residual branch, ranked by Re). Removing top-ranked edges (red) consistently increases MSE more than random (gray) or bottom-ranked (blue) deletion, providing evidence that activation-range ranking identifies functionally relevant edges. The effect is strongest on Weather and Solar, whic... | same_page_context | {
"title": "Temporal Functional Circuits: From Spline Plots to Faithful Explanations in KAN Forecasting",
"source_id": "document_179",
"arxiv_id": "2605.05685",
"figure": "figure 3",
"page": 9
} | machine_learning | null | null | {
"description_ru": "Изменены названия графиков",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"labels",
"plots"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_179__figure_3/orig.png",
"fail": "pairs/ml/document_179__figure_3/fail.png",
"orig_sha256": "ecd62186196f3611e0370ffc980f08d866fb7f0f3bedfe11078020b9cf381687",
"fail_sha256": "f4338888ef3018e8710b3e2cbf6dd4e63a40ec15f991d2ffe8131ea172df5dd2"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_18__figure_3 | ml | Figure 3. Overview of our ResiHMR framework.Given an input image, SMPL-X is initialized using intact 2D keypoints. Our Residual Anchor–Factor Optimization adapts the kinematic graph by refining anchor joints and residual-limb proportions under supervision of residual-limb 2D keypoints. The Residual-Limb Reconstruction ... | We propose ResiHMR, a residual-limb aware 3D human mesh recovery framework that estimates anatomically co- herent meshes from a single RGB image. ResiHMR is a fully optimization-based framework and does not require any training dataset. Given an image and whole-body 2D keypoints in the LDPose format, ResiHMR first fits... | mentions | {
"title": "ResiHMR: Residual-Limb Aware Single-Image 3D Human Mesh Recovery for Individuals with Limb Loss",
"source_id": "document_18",
"arxiv_id": "arXiv:2604.28025",
"figure": "figure 3",
"page": 4
} | machine_learning | null | null | {
"description_ru": "Удалены некоторые элементы",
"source_labels": [],
"operators": [
"element_deletion"
],
"objects": [],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_18__figure_3/orig.png",
"fail": "pairs/ml/document_18__figure_3/fail.png",
"orig_sha256": "e7f1bde61d75b97e772797c6aedf9dda257ddbd86c678a01183ee1ba96c7636e",
"fail_sha256": "d8612c97a0c5ca7609edd13fb16a5d09f0b272118417236fd662413e1de74d2c"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_18__figure_5 | ml | Figure 5 further shows that across diverse subjects, poses, and amputation levels, ResiHMR consistently re- covers plausible global pose, accurate residual-limb loca- tion and length, and smooth watertight stump surfaces from multiple views. These qualitative results match the quan- titative findings: unlike existing H... | topology and do not explicitly predict residual-limb end- points or stump geometry. Therefore, for all non-ResiHMR baselines, residual-limb accuracy is evaluated using a naive midpoint proxy on the corresponding limb segment under a unified protocol. ResiHMR improves over both backbones. Under the SMPLify-X backbone, i... | same_page_context | {
"title": "ResiHMR: Residual-Limb Aware Single-Image 3D Human Mesh Recovery for Individuals with Limb Loss",
"source_id": "document_18",
"arxiv_id": "arXiv:2604.28025",
"figure": "figure 5",
"page": 8
} | machine_learning | null | null | {
"description_ru": "Перемешаны изображения",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"images"
],
"gold_objects": [
"figure 5"
]
} | {
"orig": "pairs/ml/document_18__figure_5/orig.png",
"fail": "pairs/ml/document_18__figure_5/fail.png",
"orig_sha256": "2e9bcc5c5ab03272c84cf3681c164953152e803f4d17acbcefab7ac864c992a7",
"fail_sha256": "5081a69c05153ffa596154cfa825e4bd1e4ccd133f519877f4338e39dfa7668a"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_180__figure_3 | ml | Figure 3: Hierarchical code retrieval pipeline. Queries are classified into five types, each routed to a specialized strategy. Conceptual queries activate three parallel paths (LATTICE, KohakuRAG, PageIndex) fused via RRF. After code-specific reranking and 1-hop dependency expansion (Repo- Graph), the pipeline outputs ... | semantic / LLM-enhanced PageIndex-style; see Appendix C). The retriever classifies queries into five types (identifier, exact, conceptual, dependency, structural) and routes each to a specialised strategy; Figure 3 illustrates the conceptual-query fusion path that combines LATTICE Gupta et al. [2025] path-score calibra... | mentions | {
"title": "Retrieval-Conditioned Topology Selection with Provable Budget Conservation for Multi-Agent Code Generation",
"source_id": "document_180",
"arxiv_id": "2605.05657",
"figure": "figure 3",
"page": 5
} | machine_learning | null | null | {
"description_ru": "Изменен порядок блоков",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 3"
]
} | {
"orig": "pairs/ml/document_180__figure_3/orig.png",
"fail": "pairs/ml/document_180__figure_3/fail.png",
"orig_sha256": "adbd20137a4be250509594aa4b13dd43cf50129af575adfb5a4e05b835a9c4bc",
"fail_sha256": "c093c9db78b1dd4e0dc04e1d0715b25680e729fbdccb970fe28d30b79bfd68a6"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_181__figure_1 | ml | Figure 1: The architecture of AlphaCrafter: The Miner expands alpha diversity, the Screener enforces regime-aware calibration, and the Trader adaptively optimizes execution strategy. | terminate at factor generation or decouple discovery from deployment, leaving a fragmented pipeline where each stage assumes a static environment that the preceding stage cannot influence. We introduce ALPHACRAFTER, a multi-agent framework that closes this loop by operating as a contin- uously adaptive cross-sectional ... | same_page_context | {
"title": "AlphaCrafter: A Full-Stack Multi-Agent Framework for Cross-Sectional Quantitative Trading",
"source_id": "document_181",
"arxiv_id": "2605.05580",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "Изменены стрелки",
"source_labels": [],
"operators": [
"edge_rewiring"
],
"objects": [
"arrows"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_181__figure_1/orig.png",
"fail": "pairs/ml/document_181__figure_1/fail.png",
"orig_sha256": "70b160f04a2c8aca622b523afd7bf6749c7a021a2517cb68456d62ca5f924dc3",
"fail_sha256": "7ae349370aebdce16c054dd0dfa8e158d7fe1eb6b4dc7cbc5d4e0d4cbc84a275"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_182__figure_2 | ml | Figure 2: Rotation drift and continuity collapse without activation (both configurations use small initialization σ=0.0001, MNIST, 50 epochs). (a) Per-layer weight v1 rotation angle from epoch 1 reference (left two panels): Res+None layers all rotate to ∼85–90 (mutual misalignment), while Res+ReLU layers rotate only ∼2... | Rotational symmetry breaking. In deep non-residual linear networks, adjacent layers admit an exact rotational degeneracy: WlWl−1 = (WlR⊤)(RWl−1) for any orthogonal R, so individual SVDs are underdetermined and flat directions appear in the loss landscape [11]. Adding residual connections partially breaks this symmetry:... | mentions | {
"title": "Why Geometric Continuity Emerges in Deep Neural Networks: Residual Connections and Rotational Symmetry Breaking",
"source_id": "document_182",
"arxiv_id": "2605.04971",
"figure": "figure 2",
"page": 5
} | machine_learning | null | null | {
"description_ru": "Удален график",
"source_labels": [],
"operators": [
"element_deletion"
],
"objects": [
"plots"
],
"gold_objects": [
"figure 2d"
]
} | {
"orig": "pairs/ml/document_182__figure_2/orig.png",
"fail": "pairs/ml/document_182__figure_2/fail.png",
"orig_sha256": "bfe673456685c5ca55ce1fc29acf05278123208434e2f1ad3eba84b3ddb04aad",
"fail_sha256": "7abe7d8f5fa4b7338dd6e292412542c605135d1a1277c95e459d68c8b95fce9a"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_183__figure_1 | ml | Figure 1: High Level Architecture of Gyan | Gyan: An Explainable Neuro-Symbolic Language Model A PREPRINT Figure 1: High Level Architecture of Gyan [29]) is closer to models of human cognition where humans expand context both with episodic details and in an abstract fashion. Human understanding of language critically depends on understanding the complete composi... | same_page_context | {
"title": "Gyan: An Explainable Neuro-Symbolic Language Model",
"source_id": "document_183",
"arxiv_id": "2605.04759",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "Изменены блоки",
"source_labels": [],
"operators": [
"generic_edit"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_183__figure_1/orig.png",
"fail": "pairs/ml/document_183__figure_1/fail.png",
"orig_sha256": "e96f614d28b4dfe7a07d0c1ffec19439c83be89545e8b3265a2391dd55ebdc96",
"fail_sha256": "1aab8776a3dd9882ee6a90fdc470d987d3a484da853b0c6db2317b249af2daa2"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_184__figure_4 | ml | Figure 4: Forward-transfer evidence and trust calibration. The screened score retains a weak-to- moderate signal across external regimes and improves over the strongest non-screening baselines in the pooled comparisons; the right panel makes the scope boundary explicit by stratifying trust by hook completeness. The fig... | Trust calibration. The weak-to-moderate forward transfer means that the screening score should be interpreted as a diagnostic indicator rather than as a definitive classifier in external settings. We therefore provide a concrete trust calibration: for systems with hook completeness at least 0.90 and primary-scope strat... | mentions | {
"title": "AuditRepairBench: A Paired-Execution Trace Corpus for Evaluator-Channel Ranking Instability in Agent Repair",
"source_id": "document_184",
"arxiv_id": "2605.04624",
"figure": "figure 4",
"page": 18
} | machine_learning | null | null | {
"description_ru": "Изменены числовые значения",
"source_labels": [],
"operators": [
"value_fabrication"
],
"objects": [],
"gold_objects": [
"figure 4"
]
} | {
"orig": "pairs/ml/document_184__figure_4/orig.png",
"fail": "pairs/ml/document_184__figure_4/fail.png",
"orig_sha256": "b667f0619ad4dc7fc3271cfef9efab230f7bb5b455f76d0e156e3a6d83691138",
"fail_sha256": "83709f5c7f2110bcecb4f31cd8138a06dcce181bbeff287bf4b0377ec8b80c73"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_185__figure_7 | ml | Figure 7 and Table 10 summarize the results. The tighter composite threshold (κ = 1.0) quaran- tines 2 of 5 initial entries and all secondary entries, holding spread to zero. MEMSAD at κ = 2.0 quarantines ≈5 entries per trial (averaged over 5 seeds) but the 4 initial entries that pass the looser threshold seed infectio... | Figure 7: Multi-agent SIR propagation (mean ±1σ over 5 seeds; here σ refers to the seed- distribution standard deviation, not the calibration std). Top: fraction of agents exposed over 30 steps. Bottom: secondary poison entry growth. Composite (κ = 1.0) prevents spread entirely; MEMSAD (κ = 2.0) matches no-defense spre... | same_page_context | {
"title": "MEMSAD: Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents",
"source_id": "document_185",
"arxiv_id": "2605.03482",
"figure": "figure 7",
"page": 23
} | machine_learning | null | null | {
"description_ru": "перемешаны подписи",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"labels"
],
"gold_objects": [
"figure 7"
]
} | {
"orig": "pairs/ml/document_185__figure_7/orig.png",
"fail": "pairs/ml/document_185__figure_7/fail.png",
"orig_sha256": "7717ec1263001310646485f9d0d1286e4e9772629259024e3b6432512a954440",
"fail_sha256": "7511e77bc7bc23037997adef2cac17e6c0260de4a7359a3a6af790c0be4c346e"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_186__figure_1 | ml | Figure 1 provides an overview of the complete pipeline of the proposed dataset, including eight source layers unified through a shared CVE identifier space, then channeled into three benchmark tasks with temporal controls. Temporal splits are necessary in our context as real CTI systems must generalize to threats that ... | HackerSignal integrates public hacker-community discourse with sources of exploit, advisory, vulnerability-reference, bug-bounty, and commit-fix in eight source layers (Figure 1, left). Table 2 reports the composition of the audited release; Table 1 reports quality diagnostics. All sources are unified through a common ... | mentions | {
"title": "HackerSignal: A Large-Scale Multi-Source Dataset Linking Hacker Community Discourse to the CVE Vulnerability Lifecycle",
"source_id": "document_186",
"arxiv_id": "2605.03158",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "отзеркалено изображение",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"images"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_186__figure_1/orig.png",
"fail": "pairs/ml/document_186__figure_1/fail.png",
"orig_sha256": "4b01bc75a1232bfe03a16193a30f677076df9bb6d1bc6127314cb6e2a96a2972",
"fail_sha256": "35622ce06b7e90c87836dab43d795937da5997a7e82b26937cda383db26f50c7"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_187__figure_2 | ml | Figure 2: Simple regret of POO and HOO run for different values of ρ. | In Figure 2, we show the simple regret of the algorithms as a function of the number of evaluations. In the figure on the left, we plot the simple regret after 500 evaluations. In the right one, we plot the simple regret after 5000 evaluations in the log-log scale, in order to see the trend better. The HOO algorithms r... | mentions | {
"title": "Black-box optimization of noisy functions with unknown smoothness",
"source_id": "document_187",
"arxiv_id": "2605.02462",
"figure": "figure 2",
"page": 8
} | machine_learning | null | null | {
"description_ru": "Намеренное искажение изображения",
"source_labels": [],
"operators": [
"geometric_distortion"
],
"objects": [
"images"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_187__figure_2/orig.png",
"fail": "pairs/ml/document_187__figure_2/fail.png",
"orig_sha256": "6c89eac6047dc81dc0c6b290a0215b481958d8c375f5841e2198ff0dabe6e79a",
"fail_sha256": "961981a0db138ddc416ca4d7afee561ab891a52a3555981b748d3d89c9177c1b"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_19__figure_2 | ml | Figure 2. Architecture illustration of the ProtoPFL with VPDR. ①Private Prototype Calculation: each client runs VPP (Section 4.2) to privately partition embeddings, apply groupwise clipping, and add adaptive noise to prototypes; ②Upload: clients send privatized proto- types; ③Global Prototype Generation: the server agg... | However, IGPP leaves two practical issues. First, ap- plying isotropic noise perturbs all coordinates equally and can over-distort the most discriminative directions. Second, choosing a fixed ℓ2 clipping threshold is delicate. A large threshold inflates noise via ∆while a small one severely distorts feature semantics. ... | mentions | {
"title": "Taming Noise-Induced Prototype Degradation for Privacy-Preserving Personalized Federated Fine-Tuning",
"source_id": "document_19",
"arxiv_id": "arXiv:2604.27833",
"figure": "figure 2",
"page": 3
} | machine_learning | null | null | {
"description_ru": "Изменен текст на изображении",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"labels",
"images"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_19__figure_2/orig.png",
"fail": "pairs/ml/document_19__figure_2/fail.png",
"orig_sha256": "2202f1766b16b6dc5cf0a60130a598b46a3d6486207cdae83e767cbd3686b355",
"fail_sha256": "402d8d6365f02fb32470c16345200191d70a02e72cdb924299d657cf98cf9379"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_191__figure_2 | ml | Figure 2 shows the performance of COMBBSFPL and SLEEPINGCATBANDIT for a fixed loss se- quence, averaged over 20 samples of the component availabilities. To appraise the benefit of learning in this problem, we also plot the performance of random policy, which pulls some available arm uni- formly at random. We notice tha... | For the bandit case, we evaluate SLEEPINGCATBANDIT using the same empirical setting as Kanade et al. [13]. We consider an experiment with T = 104 and 5 arms, each of which are available independently of each other with probability p. At the beginning of the experiment, the losses of each arm are chosen uniformly at ran... | mentions | {
"title": "Online combinatorial optimization with stochastic decision sets and adversarial losses",
"source_id": "document_191",
"arxiv_id": "2604.25269",
"figure": "figure 2",
"page": 8
} | machine_learning | null | null | {
"description_ru": "Изменено описание рисунки",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"images"
],
"gold_objects": [
"figure 2"
]
} | {
"orig": "pairs/ml/document_191__figure_2/orig.png",
"fail": "pairs/ml/document_191__figure_2/fail.png",
"orig_sha256": "e8de451f3fa753ee7c8979439bd1ad11c5e3dbfb9ce653bf52dfb13d1f9d4b4d",
"fail_sha256": "05dedfbd239aa8fe059c7d4053d2d6757cfd5aa5db89b126d3e96ce2277ee923"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_192__figure_1a | ml | Figure 1a: Users and news feeds. The thick edges represent one potential matching of users to feeds, grouped news feeds show the same content. | Consider the problem of sequentially recommending content for a set of users. In each period of this online decision problem, we have to assign content from a news feed to each of our subscribers so as to maximize clickthrough. We assume that this assignment needs to be done well in advance, so that we only observe the... | mentions | {
"title": "Efficient learning by implicit exploration in bandit problems with side observations",
"source_id": "document_192",
"arxiv_id": "2604.24555",
"figure": "figure 1a",
"page": 2
} | machine_learning | null | null | {
"description_ru": "изменены подписи к блокам",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"blocks",
"labels"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_192__figure_1a/orig.png",
"fail": "pairs/ml/document_192__figure_1a/fail.png",
"orig_sha256": "423ac6a9b0fdcb5c07d302fb78bc61ac644334462a8db77674114be5336b574f",
"fail_sha256": "21d8ec1f1f5aa9f1b313959806c1d3d216477fbadbd9238368e75cd39ab081b9"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_193__figure_1 | ml | Figure 1: Extreme regret as a function of time for the exact Pareto distributions (left), approximate Pareto (middle) distributions, and the network traffic data (right). | Exact Pareto Distributions In the first experiment, we consider K = 3 arms with the distributions Pk(x) = 1−x−αk, where α = [5, 1.1, 2]. Therefore, the most heavy-tailed distribution is associated with the arm k = 2. Figure 1 (left) displays the averaged result of 1000 simulations with the time horizon T = 104. We obse... | mentions | {
"title": "Extreme bandits",
"source_id": "document_193",
"arxiv_id": "2604.24545",
"figure": "figure 1",
"page": 8
} | machine_learning | null | null | {
"description_ru": "изменены подписи к грвфикам",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"labels",
"plots"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_193__figure_1/orig.png",
"fail": "pairs/ml/document_193__figure_1/fail.png",
"orig_sha256": "abfbb28065627752e2fa2901558656d380c312ba6e65f25298b304d60af31b34",
"fail_sha256": "640dd66109a02c42c3ae87a0bc9f8147df3eed211c786d678bb018d0f1a39128"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_195__figure_1 | ml | Figure 1: Distributions of reward scores for human-written and LLM-generated texts from the multi- LLM sub-task of the DetectRL benchmark. The source LLMs include GPT-3.5-turbo, PaLM-2-Bison, Claude-Instant, and LLaMA-2-70B. (a) shows distributions under RM-deberta-v3-large-v2, which is a pre-trained reward model. (b) ... | shown in Figure 1 (b), when ReMoDetect is evaluated on texts generated by unseen LLMs, there remains substantial overlap between the two score distributions of human-written and LLM-generated texts, suggesting the limited generalization ability. πref(y). This score serves as the detection metric in our method, where a ... | mentions | {
"title": "Zero-Shot Detection of LLM-Generated Text via Implicit Reward Model",
"source_id": "document_195",
"arxiv_id": "2604.21223",
"figure": "figure 1",
"page": 2
} | machine_learning | null | null | {
"description_ru": "Полность изменено изображение",
"source_labels": [],
"operators": [
"content_replacement"
],
"objects": [
"images"
],
"gold_objects": [
"figure 1c"
]
} | {
"orig": "pairs/ml/document_195__figure_1/orig.png",
"fail": "pairs/ml/document_195__figure_1/fail.png",
"orig_sha256": "887a52f66a036aa7d938fe5924d49f93117ee68b335ae32b8314e9900ba66f59",
"fail_sha256": "bbe7ea5814cbb1f0741fc25a33a5ee72ef60558253906fb1203a1de637b0186d"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_197__figure_1 | ml | Figure 1: (a) A sample from a 2D Jacobi ensemble with N = 1000 points. (b)-(c) ai, bi = −1/2, the colors and numbers correspond to the dimension. For d = 1, the tridiagonal model (tri) of Killip and Nenciu offers tremendous time savings. (c) The total number of rejections grows as 2dN log(N). | All these improvements resulted in dramatic speedups. For example, on a modern laptop, sampling a 2D Jacobi ensemble with N = 1000 points, see Figure 1(a), takes less than a minute, compared to hours previously. For more details on the sampling procedure, we refer to Appendix A.4. side, we use a two-layer rejection sam... | mentions+same_page_context | {
"title": "On two ways to use determinantal point processes for Monte Carlo integration",
"source_id": "document_197",
"arxiv_id": "2604.19698",
"figure": "figure 1",
"page": 7
} | machine_learning | null | null | {
"description_ru": "изменена цветовая палитра",
"source_labels": [],
"operators": [
"restyling"
],
"objects": [],
"gold_objects": [
"figure 1b"
]
} | {
"orig": "pairs/ml/document_197__figure_1/orig.png",
"fail": "pairs/ml/document_197__figure_1/fail.png",
"orig_sha256": "abc1d43521be799f4ea1e18891078d0ea30d5018e3bf6d1287e09ea23218146c",
"fail_sha256": "9954bea1400d6c98ba4e57765c37445040fdc3ac1245cc7a66759bd645787c6f"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_197__figure_2 | ml | Figure 2: (a)-(d) cf. Section 4.1, the numbers in the legend are the slope and R2 (e)-(h) cf. Section 4.2. | reproduce their experiment in Figure 2 for larger N, and compare with the behavior of bI EZ N is an exact estimator, see Figure 2(e)-(h). The second setting has M = N + 1, so that the number of points N is never enough for the variance in (11) to be zero. The results of both settings are to be found in Appendix B.2. si... | mentions+same_page_context | {
"title": "On two ways to use determinantal point processes for Monte Carlo integration",
"source_id": "document_197",
"arxiv_id": "2604.19698",
"figure": "figure 2",
"page": 7
} | machine_learning | null | null | {
"description_ru": "Изменены числовой диапазон оси ординат",
"source_labels": [],
"operators": [
"value_fabrication"
],
"objects": [
"axes"
],
"gold_objects": [
"figure 2h"
]
} | {
"orig": "pairs/ml/document_197__figure_2/orig.png",
"fail": "pairs/ml/document_197__figure_2/fail.png",
"orig_sha256": "7d214e983af9e52ca2d9897d70b0fc2f62df1998b51ddb8743d148dfb9dede86",
"fail_sha256": "6163b45462a43dc0fac0f288bb7d759ca907e3be42c9fd114638292e52264827"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_198__figure_1 | ml | Figure 1: Visualization of a call to SmoothCruiser(s0, ε0, δ′). | We now describe our planning algorithm. Its building blocks are two procedures, sampleV (Algo- rithm 2) and estimateQ (Algorithm 3) that recursively call each other. The procedure sampleV returns a noisy estimate of V (s) with a bias bounded by ε. The procedure estimateQ averages the outputs of several calls to sampleV... | mentions | {
"title": "Planning in entropy-regularized Markov decision processes and games",
"source_id": "document_198",
"arxiv_id": "2604.19695",
"figure": "figure 1",
"page": 5
} | machine_learning | null | null | {
"description_ru": "Изменены индексы элементов",
"source_labels": [],
"operators": [
"label_text_edit"
],
"objects": [
"labels"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_198__figure_1/orig.png",
"fail": "pairs/ml/document_198__figure_1/fail.png",
"orig_sha256": "ebadd636615e9a58b7b945eecde570a457eaffa865fd929d97bff7de90d243e8",
"fail_sha256": "b9a4e80ad651c0133b23eaa8cc9ec862b34e07cb9306f2d84e24cf627094ab18"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_2__figure_1 | ml | Figure 1. The figure shows the impact of integrating self- discovered implicit concepts via our Differentiable Knowledge Unit. Shown: input image, ground truth, initial probabilities p(yk|z), refined probabilities p∆(yk|zˆ), and visualizations of con- cept activations to provide insight into the learned semantics. | Learning to Reason: Targeted Knowledge Discovery and Fuzzy Logic Update for Gurucharan Srinivas*∗ Joshua Niemeijer∗ Frank Ko¨ster German Aerospace Center (DLR) {Gurucharan.Srinivas, Joshua.Niemeijer, Frank.Koester}@dlr.de Integrating domain knowledge into deep neural networks is a promising way to improve generalizatio... | same_page_context | {
"title": "Learning to Reason: Targeted Knowledge Discovery and Fuzzy Logic Update for Robust Image Recognition",
"source_id": "document_2",
"arxiv_id": "arXiv:2604.27759",
"figure": "figure 1",
"page": 1
} | machine_learning | null | null | {
"description_ru": "Исправлена блок схема",
"source_labels": [],
"operators": [
"generic_edit"
],
"objects": [
"blocks"
],
"gold_objects": [
"figure 1 right"
]
} | {
"orig": "pairs/ml/document_2__figure_1/orig.png",
"fail": "pairs/ml/document_2__figure_1/fail.png",
"orig_sha256": "f3039a1250ef9831ee98427e755607e1384fee36f44070fa711d814ac54bcf52",
"fail_sha256": "4fe4906467076d16d96c375a8306d76f19324b139178ca07e9759043361a8c2f"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_201__figure_6 | ml | Figure 6: Final detector performance curves for the best configuration (Laplacian, top-k fusion, plain z-score). (a) The ROC curve shows consistent separation above the diagonal across all oper- ating points. (b) The PR curve maintains high precision at low recall, with AUPRC well above the class-prior baseline. | Four-dataset snapshot. Read in isolation, the RecodAI-LUC headline (AUROC = 0.606) can undersell the method: the same code, the same Stable Diffusion v1.5 backbone, and the same authentic-only calibration protocol produce AUROC = 0.752 on MICC-F220 (n = 44), AUROC = 0.774 with AUPRC = 0.833, MCC = 0.499, and TPR = 32.5... | mentions | {
"title": "Spectral Forensics of Diffusion Attention Graphs for Copy-Move Forgery Detection",
"source_id": "document_201",
"arxiv_id": "2604.17287",
"figure": "figure 6",
"page": 10
} | machine_learning | null | null | {
"description_ru": "Удален график",
"source_labels": [],
"operators": [
"element_deletion"
],
"objects": [
"plots"
],
"gold_objects": [
"figure 6"
]
} | {
"orig": "pairs/ml/document_201__figure_6/orig.png",
"fail": "pairs/ml/document_201__figure_6/fail.png",
"orig_sha256": "c29fc835f232d2094bac38377bffb216985ee6f25e7e4245306bede780c993f8",
"fail_sha256": "4f6375a7e188c810c8a79dd2895ae519773ed8073152b0fa09659d633c584686"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
ml/document_203__figure_1 | ml | Figure 1: KEEP rates on correct versus incorrect items for three metacognitive profiles. Points on the diagonal indicate identical KEEP rates regardless of correctness (no monitoring-control coupling). Points below the diagonal indicate selective withdrawal of incorrect items (coupled monitoring-control). Left: Profile... | Figure 1: KEEP rates on correct versus incorrect items for three metacognitive profiles. Points on the diagonal indicate identical KEEP rates regardless of correctness (no monitoring-control coupling). Points below the diagonal indicate selective withdrawal of incorrect items (coupled monitoring-control). Left: Profile... | same_page_context | {
"title": "The Metacognitive Monitoring Battery: A Cross-Domain Benchmark for LLM Self-Monitoring",
"source_id": "document_203",
"arxiv_id": "2604.15702",
"figure": "figure 1",
"page": 6
} | machine_learning | null | null | {
"description_ru": "перемещаны графики",
"source_labels": [],
"operators": [
"block_permutation"
],
"objects": [
"plots"
],
"gold_objects": [
"figure 1"
]
} | {
"orig": "pairs/ml/document_203__figure_1/orig.png",
"fail": "pairs/ml/document_203__figure_1/fail.png",
"orig_sha256": "235d0eabcd69db2f6414e4a6618b36ce05f1cf055484da0d35012530586327ed",
"fail_sha256": "c36a26e16c07fe2363bd361d8513ef606a40b2064262f124460e77d858721181"
} | true | {
"out_of_region": false,
"selection": "approximate_paper_201"
} | true | null |
End of preview. Expand in Data Studio
Verifying the Verifier — Controlled-Counterfactual Benchmark
Companion dataset for the paper Verifying the Verifier: A Controlled-Counterfactual Benchmark for Scientific Figure Evaluators, accepted at the ICDM 2026 Workshop AI4S-Bench 2026.
- Code and evaluation harness: github.com/Saintghetto17/Verifying-the-Verifier
Each pair holds a faithful reference figure (orig.png) and a
counterfactual twin (fail.png) that carries exactly one injected semantic
defect of a recorded type, together with the figure's caption and the
source-text block that discusses it. A verifier scores each image on its own,
from the caption, the source text and that single image, and should rank the
reference above its corrupted twin. Because the defect is injected, the correct
order is known before any model is called.
Release status
| Subset | Domain | Pairs | Images released | Paired run (paper) |
|---|---|---|---|---|
ml |
Machine learning | 201 | 201 | 198 |
bt |
Biotechnology, 16 topics | 212 | 212 | 204 (208 billed) |
b75 |
Climate modeling, materials discovery, drug discovery, protein structure prediction, Earth system science | 75 | 15 | 75 |
| Total | 488 | 428 | 477 |
B75 images: 15 of 75 are released now; the other 60 image pairs will be uploaded soon. All 75 B75 records are already present in
metadata/pairs.jsonlwith caption, source text, entities and defect label; records without images have"images_available": falseand"images": null.
Layout
pairs/<subset>/<local_id>/orig.png reference figure V+
pairs/<subset>/<local_id>/fail.png counterfactual twin V-
metadata/pairs.jsonl one record per pair (all 488)
metadata/ml_candidates.jsonl every ML candidate and the filter that removed it
metadata/summary.json counts for the corpus and taxonomy tables
splits/ml.txt, bt.txt, b75.txt all pairs of a subset
splits/bt_run.txt 208 BT pairs of the paired run (out-of-region pairs held out)
splits/b75_images.txt B75 pairs whose images are released
splits/balanced100.txt operator-balanced 100-pair ML sample (ablations, Tables III, IV, VI)
Record fields
| Field | Meaning |
|---|---|
pair_id |
<subset>/<local_id> |
subset |
ml, bt or b75 |
caption, text_block |
the verifier's text inputs |
text_block_source |
ML only: mentions, same_page_context or both |
paper |
title, source document, figure label, page |
topic, figure_type, multi_panel, entities |
descriptive metadata (where available) |
defect.description_ru |
free-form annotator description of the injected defect (Russian) |
defect.source_labels |
BT / B75: free-form English labels of the source annotation |
defect.operators |
normalized operator tags (10 tags, not mutually exclusive) |
defect.objects |
ML: affected-object tags (blocks, labels, images, plots, arrows, axes) |
images |
relative paths and SHA-256 of orig.png / fail.png, or null |
flags.out_of_region |
BT: the stated corruption falls outside the figure region (4 pairs) |
flags.selection |
ML: approximate_paper_201 (see below) |
in_paired_run |
whether the pair enters the paired run |
The defect fields must never be shown to the verifier; they exist for scoring and analysis only.
Construction notes
- ML. Candidates come from the ML split of the
Verification Before Assembly
corpus. Pairs that are marked corrupted but byte-identical, with source text
under 40 words, or with a caption under five words are removed (235 remain).
The paper's manual review then reduced the set to 201; the manual exclusion
list was not preserved, so this release selects 201 pairs deterministically,
keeping the paper's text-source strata (134 mentions, 32 mentions + page
context, 35 page context) and, within each stratum, the pairs with the
longest source text. The selection is therefore an approximation of the
paper's ML subset and is marked as such in
flags.selection. - BT. 212 biotechnology pairs with originals and corruptions cropped from
the same page rectangle. Four pairs are flagged
out_of_regionand are held out of the paired run. - B75. 75 cross-domain pairs: 64 hand-made corruptions on new papers and
11 drug-discovery pairs from the biotechnology corpus (ids prefixed
bt_).
Usage
from huggingface_hub import snapshot_download
root = snapshot_download("Saintghetto17/verifying-the-verifier", repo_type="dataset")
or, inside the code repository:
python -m vtv download # installs into data/
python -m vtv validate
License
Metadata, defect annotations and split lists are released under CC BY 4.0. Figure images are crops of published articles and remain subject to the licenses of their source publications; they are provided for research on verifier evaluation only.
Citation
@inproceedings{verifying_the_verifier_2026,
title = {Verifying the Verifier: A Controlled-Counterfactual Benchmark for Scientific Figure Evaluators},
author = {Gromov, Egor and Novitskii, Ilia and Malashenko, Boris and Mironov, Kirill and
Gavrishok, Nikolay and Rutkovskii, Evgenii and Efimova, Valeria},
booktitle = {ICDM 2026 Workshop AI4S-Bench 2026},
year = {2026}
}
- Downloads last month
- 69