Datasets:

alecocc
/

mmscifact-demo

qa_pair_id stringlengths 28 81	scenario stringclasses 4 values	level stringclasses 3 values	question_subtype stringclasses 9 values	answer_model stringlengths 12 50	question_text stringlengths 143 588	question_meta dict	annotations listlengths 4 22	source_images listlengths 0 39	answer_raw stringlengths 711 3.81k	question_model stringclasses 5 values	paper_id stringclasses 9 values	paper_pdf_path stringclasses 9 values
11928_Visual_Planning_Let_s_Th_cross_g0_p2_4_6_conceptual	S2	cross_page	conceptual	gemini-3.5-flash	Looking at Figure 1 and Table 1, how does the paper’s visual-planning setup explain why the image-only model achieves stronger trajectory success than the text-based planning setups? Answer from the visual structure in both elements, not from the captions alone.	{ "input_types": [ "Figure 1", "Table 1" ], "question_type": "conceptual", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "Based on Figure 1 and Table 1, the paper's visual-planning setup explains why the image-only model achieves stronger trajectory success compared to text-based planning by avoiding language mediation and directly optimizing the representation space.", "label": "contradiction", "role": "Int...	[ "pdfs_jpg/11928_Visual_Planning_Let_s_Th/page_002.jpg", "pdfs_jpg/11928_Visual_Planning_Let_s_Th/page_004.jpg", "pdfs_jpg/11928_Visual_Planning_Let_s_Th/page_006.jpg" ]	Based on Figure 1 and Table 1, the paper's visual-planning setup explains why the image-only model achieves stronger trajectory success compared to text-based planning by avoiding language mediation and directly optimizing the representation space. As shown in the visual structure of Figure 1, traditional approaches li...	gpt-5.4-mini	11928_Visual_Planning_Let_s_Th	pdfs/11928_Visual_Planning_Let_s_Th.pdf
11928_Visual_Planning_Let_s_Th_cross_g1_p5_6_21_results_interpretation	S2	cross_page	results_interpretation	gemini-3.5-flash	Based on Figure 7 and Table 1, how does the relative speed of initial reward increase across environments relate to the final progress rate achieved by the best-performing model?	{ "input_types": [ "Figure 7", "Table 1", "Section 3" ], "question_type": "results_interpretation", "visual_subtype": "trend", "error_subtype": null, "scope": null }	[ { "sentence": "Based on the provided document, the relative speed of initial reward increase across different environments corresponds directly to the final progress rate achieved by the best-performing model, which is VPRL (ours) in Table 1.", "label": "contradiction", "role": "Interpretation", "ra...	[ "pdfs_jpg/11928_Visual_Planning_Let_s_Th/page_005.jpg", "pdfs_jpg/11928_Visual_Planning_Let_s_Th/page_006.jpg", "pdfs_jpg/11928_Visual_Planning_Let_s_Th/page_021.jpg" ]	Based on the provided document, the relative speed of initial reward increase across different environments corresponds directly to the final progress rate achieved by the best-performing model, which is VPRL (ours) in Table 1. Looking at the reward curves in Figure 7, the speed of the initial reward increase is fast...	google/gemma-4-31b-it	11928_Visual_Planning_Let_s_Th	pdfs/11928_Visual_Planning_Let_s_Th.pdf
11928_Visual_Planning_Let_s_Th_page_002_visual_understanding	S4	single_page	visual_understanding	gemini-3.5-flash	Can you describe how Figure 1 is organized from top to bottom — what the three rows show, what the boxes and arrows connect, and what the labels like Input, Verbal Thought, Visual Thought, and Verbal Output represent in each reasoning paradigm?	{ "input_types": [ "Figure 1" ], "question_type": "visual_understanding", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "Based on Figure 1 in the provided document, the illustration is organized from top to bottom into three distinct horizontal rows, each representing a different reasoning paradigm.", "label": "supported", "role": "Observation", "rationale": "", "depends_on": [], "errors": [], ...	[ "pdfs_jpg/11928_Visual_Planning_Let_s_Th/page_002.jpg" ]	Based on Figure 1 in the provided document, the illustration is organized from top to bottom into three distinct horizontal rows, each representing a different reasoning paradigm. The first (top) row illustrates the Direct Prompting paradigm. This row shows a purple input box representing a multimodal image-and-text ...	gpt-5.4-mini	11928_Visual_Planning_Let_s_Th	pdfs/11928_Visual_Planning_Let_s_Th.pdf
11928_Visual_Planning_Let_s_Th_page_006_comparison	S4	single_page	comparison	gemini-3.5-flash	In Table 1, how do VPRL and VPFT compare with the other models across FrozenLake, Maze, and MiniBehavior — where is their gap biggest, where is it smallest, and how does that difference change from task to task?	{ "input_types": [ "Table 1" ], "question_type": "comparison", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "In Table 1, the performance of the proposed models, VPRL (ours) and VPFT (ours), is compared against other closed-source and open-source models across three tasks: FrozenLake, Maze, and MiniBehavior, evaluated using Exact Match (EM) and Progress Rate (PR) percentages.", "label": "supported", ...	[ "pdfs_jpg/11928_Visual_Planning_Let_s_Th/page_006.jpg" ]	In Table 1, the performance of the proposed models, VPRL (ours) and VPFT (ours), is compared against other closed-source and open-source models across three tasks: FrozenLake, Maze, and MiniBehavior, evaluated using Exact Match (EM) and Progress Rate (PR) percentages. Overall, VPRL performs consistently the best (indic...	gpt-5.4-mini	11928_Visual_Planning_Let_s_Th	pdfs/11928_Visual_Planning_Let_s_Th.pdf
2025_emnlp-main_488_full_methodological	S1	full_paper	methodological	gemma-4-31b-it	What does Figure 2 show about how problems are distributed and shared across the different competition categories, and how does this overlapping design relate to the category-based model evaluation results shown in Table 4?	{ "input_types": [ "Figure 2", "Table 4", "Section 5" ], "question_type": "methodological", "visual_subtype": "visual_understanding", "error_subtype": null, "scope": "full" }	[ { "sentence": "Figure 2 illustrates that problems in the MATHGAMES benchmark are not exclusive to a single age group but are instead distributed across multiple categories through an overlapping design.", "label": "supported", "role": "Interpretation", "rationale": "", "depends_on": [ "S3"...	[]	Figure 2 illustrates that problems in the MATHGAMES benchmark are not exclusive to a single age group but are instead distributed across multiple categories through an overlapping design. The figure shows blue boxes indicating the specific range of exercises each age group must solve. For instance, while the C1 categor...	gemini-3.5-flash	2025_emnlp-main_488	pdfs/2025_emnlp-main_488.pdf
2025_emnlp-main_488_page_004_quantitative	S4	single_page	quantitative	gemini-3.5-flash	In Table 3, how do the counts spread across the age groups for the different math skills — like which skills are concentrated more in the older groups versus the younger ones, and which skill has the strongest overall shift toward higher age groups?	{ "input_types": [ "Table 3" ], "question_type": "quantitative", "visual_subtype": null, "error_subtype": null, "scope": "single" }	[ { "sentence": "Based on Table 3 in the provided document, the counts for the six listed math skills are spread across seven age groups: CE (8–10 y/o), C1 (11–13 y/o), C2 (13–15 y/o), L1 (15–18 y/o), L2 (18–20 y/o), GP (20–25 y/o), and HC (25+ y/o).", "label": "supported", "role": "Observation", "rat...	[ "pdfs_jpg/2025_emnlp-main_488/page_004.jpg" ]	Based on Table 3 in the provided document, the counts for the six listed math skills are spread across seven age groups: CE (8–10 y/o), C1 (11–13 y/o), C2 (13–15 y/o), L1 (15–18 y/o), L2 (18–20 y/o), GP (20–25 y/o), and HC (25+ y/o). Each math skill is split into two rows of data. The distribution reveals that certain ...	gpt-5.4-mini	2025_emnlp-main_488	pdfs/2025_emnlp-main_488.pdf
3677_Reading_Images_Like_Texts_full_cross_adversarial	S1	full_paper	methodological	gemini-3.5-flash	I think the paper’s main story is that VLMs really do read images in a human-like two-step way: first they pick up local attributes, then they resolve them into object labels, and that this same kind of object-level understanding is what makes the token-compression idea work. Looking at the abstract, plus Table 1 and F...	{ "input_types": [ "Abstract", "Table 1", "Figure 21" ], "question_type": "cross_adversarial", "visual_subtype": "comparison", "error_subtype": "overgeneralized_scope", "scope": null }	[ { "sentence": "Your reading of the paper's core narrative regarding object recognition in vision-language models is highly accurate according to the provided text, but your interpretation of the token compression results and the performance of the different methods requires clarification based on the actual dat...	[ "pdfs_jpg/3677_Reading_Images_Like_Texts/page_001.jpg", "pdfs_jpg/3677_Reading_Images_Like_Texts/page_002.jpg", "pdfs_jpg/3677_Reading_Images_Like_Texts/page_003.jpg", "pdfs_jpg/3677_Reading_Images_Like_Texts/page_004.jpg", "pdfs_jpg/3677_Reading_Images_Like_Texts/page_005.jpg", "pdfs_jpg/3677_Reading_Ima...	Your reading of the paper's core narrative regarding object recognition in vision-language models is highly accurate according to the provided text, but your interpretation of the token compression results and the performance of the different methods requires clarification based on the actual data presented. The autho...	gpt-5.4-mini	3677_Reading_Images_Like_Texts	pdfs/3677_Reading_Images_Like_Texts.pdf
3677_Reading_Images_Like_Texts_page_009_comparison	S4	single_page	comparison	gemini-3.5-flash	In Table 1, how do the two models, LLaVA-1.5-7B and Qwen2.5-VL-7B, compare across the four decoding settings overall — where is the gap biggest, where is it smallest, and how does the performance difference change as you move from Original Decoding to the three compression methods?	{ "input_types": [ "Table 1" ], "question_type": "comparison", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "Based on the data presented in Table 1, the two models, LLaVA-1.5-7B and Qwen2.5-VL-7B, exhibit a consistent performance gap across all four decoding settings, with Qwen2.5-VL-7B consistently outperforming LLaVA-1.5-7B on almost every task and benchmark.", "label": "contradiction", "role"...	[ "pdfs_jpg/3677_Reading_Images_Like_Texts/page_009.jpg" ]	Based on the data presented in Table 1, the two models, LLaVA-1.5-7B and Qwen2.5-VL-7B, exhibit a consistent performance gap across all four decoding settings, with Qwen2.5-VL-7B consistently outperforming LLaVA-1.5-7B on almost every task and benchmark. The overall performance gap between the two models is smallest ...	gpt-5.4-mini	3677_Reading_Images_Like_Texts	pdfs/3677_Reading_Images_Like_Texts.pdf
3677_Reading_Images_Like_Texts_page_010_quantitative	S4	single_page	quantitative	gpt-5.4-mini	Based on Table 2, how does the performance of the Qwen2-VL-7B model change across the various spatial reasoning benchmarks when applying RoPE scaling, SFT, and the combination of both compared to the base Qwen2-VL-7B model?	{ "input_types": [ "Table 2" ], "question_type": "quantitative", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "Based on Table 2, the base Qwen2-VL-7B model already performs strongly across all the spatial reasoning benchmarks, and each of the two modifications, RoPE scaling and SFT, generally improves performance further.", "label": "contradiction", "role": "Interpretation", "rationale": "", ...	[ "pdfs_jpg/3677_Reading_Images_Like_Texts/page_010.jpg" ]	Based on Table 2, the base Qwen2-VL-7B model already performs strongly across all the spatial reasoning benchmarks, and each of the two modifications, RoPE scaling and SFT, generally improves performance further. When RoPE scaling is applied by itself, the gains are modest but consistent on most benchmarks: What’s Up A...	gemini-3.5-flash	3677_Reading_Images_Like_Texts	pdfs/3677_Reading_Images_Like_Texts.pdf
Genie_generative_interactive_enviroments_cross_g0_p4_22_23_conceptual	S2	cross_page	conceptual	gemini-3.5-flash	Looking at Figure 4 and Table 7, how does the transformer’s temporal-spatial processing pattern help explain the tokenizer’s much denser codebook setup?	{ "input_types": [ "Figure 4", "Table 7" ], "question_type": "conceptual", "visual_subtype": "visual_understanding", "error_subtype": null, "scope": null }	[ { "sentence": "The temporal-spatial processing pattern of the ST-transformer, as illustrated in Figure 4, directly explains the need for the denser codebook setup of 1024 codes shown in Table 7.", "label": "nei", "role": "Interpretation", "rationale": "Implying a non-existent causal connection", ...	[ "pdfs_jpg/Genie_generative_interactive_enviroments/page_004.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_022.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_023.jpg" ]	The temporal-spatial processing pattern of the ST-transformer, as illustrated in Figure 4, directly explains the need for the denser codebook setup of 1024 codes shown in Table 7. According to the document, the ST-transformer architecture is composed of spatiotemporal blocks where the spatial layer attends over $H \tim...	gpt-5.4-mini	Genie_generative_interactive_enviroments	pdfs/Genie_generative_interactive_enviroments.pdf
Genie_generative_interactive_enviroments_cross_g1_p6_24_25_results_interpretation	S2	cross_page	results_interpretation	gemma-4-31b-it	What do Figure 9 and Table 10 together show about how training loss changes as compute rises across the different model configurations, and how does the batch-size comparison in Figure 9 relate to the model-size sweep in Table 10?	{ "input_types": [ "Figure 9", "Table 10" ], "question_type": "results_interpretation", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "Figure 9 and Table 10 collectively demonstrate that as compute increases through larger model configurations, training loss decreases.", "label": "contradiction", "role": "Interpretation", "rationale": "Table 10 holds no information about training loss, Figure 9 demonstrates this by i...	[ "pdfs_jpg/Genie_generative_interactive_enviroments/page_006.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_024.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_025.jpg" ]	Figure 9 and Table 10 collectively demonstrate that as compute increases through larger model configurations, training loss decreases. The left plot of Figure 9 displays training curves for various model sizes, showing that loss declines as compute, measured in FLOPs, rises, with larger models consistently achieving lo...	gpt-5.4-mini	Genie_generative_interactive_enviroments	pdfs/Genie_generative_interactive_enviroments.pdf
Genie_generative_interactive_enviroments_cross_g2_p5_21_22_results_interpretation	S2	cross_page	results_interpretation	gpt-5.4-mini	How do the action inputs illustrated in the inference pipeline of Figure 8 and the specific action columns shown in Figure 17 compare to the total number of codes specified for the codebook in Table 5?	{ "input_types": [ "Figure 8", "Figure 17", "Table 5" ], "question_type": "results_interpretation", "visual_subtype": "quantitative", "error_subtype": null, "scope": null }	[ { "sentence": "The document shows that the inference pipeline in Figure 8 uses a single latent action chosen by the user as part of the input to the dynamics model, together with the prompt frame tokens.", "label": "supported", "role": "Observation", "rationale": "Correct observation of the input dy...	[ "pdfs_jpg/Genie_generative_interactive_enviroments/page_005.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_021.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_022.jpg" ]	The document shows that the inference pipeline in Figure 8 uses a single latent action chosen by the user as part of the input to the dynamics model, together with the prompt frame tokens. Figure 17, meanwhile, illustrates four specific action columns labeled left, right, jump, and no-op. These are presented as example...	gemini-3.5-flash	Genie_generative_interactive_enviroments	pdfs/Genie_generative_interactive_enviroments.pdf
Genie_generative_interactive_enviroments_cross_g3_p3_9_25_conceptual	S2	cross_page	conceptual	gpt-5.4-mini	Based on the trends shown in Figure 15 and the quantitative results in Table 2, how does the input type of the latent action model affect the agent's performance and controllability?	{ "input_types": [ "Figure 15", "Table 2" ], "question_type": "conceptual", "visual_subtype": "trend", "error_subtype": null, "scope": null }	[ { "sentence": "The document indicates that the input type of the latent action model has a clear effect on both performance and controllability, and that using raw pixels is the better choice.", "label": "supported", "role": "Interpretation", "rationale": "Directly emerges from observation", "de...	[ "pdfs_jpg/Genie_generative_interactive_enviroments/page_003.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_009.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_025.jpg" ]	The document indicates that the input type of the latent action model has a clear effect on both performance and controllability, and that using raw pixels is the better choice. In the ablation study, the authors compare a token-input latent action model, which uses tokenized images, against the Genie version, which us...	gemma-4-31b-it	Genie_generative_interactive_enviroments	pdfs/Genie_generative_interactive_enviroments.pdf
Genie_generative_interactive_enviroments_cross_g5_p2_8_9_conceptual	S2	cross_page	conceptual	gemini-3.5-flash	Based on Figure 15 and Table 2, how does the performance of the LAM-based policy in different difficulty settings compare to the controllability of the model inputs?	{ "input_types": [ "Figure 15", "Table 2" ], "question_type": "conceptual", "visual_subtype": "trend", "error_subtype": null, "scope": null }	[ { "sentence": "The performance of the Latent Action Model (LAM)-based policy across different difficulty settings demonstrates exceptional robustness and adaptability.", "label": "nei", "role": "Interpretation", "rationale": "Very vague given only limited number of environments were evaluated and th...	[ "pdfs_jpg/Genie_generative_interactive_enviroments/page_002.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_008.jpg", "pdfs_jpg/Genie_generative_interactive_enviroments/page_009.jpg" ]	The performance of the Latent Action Model (LAM)-based policy across different difficulty settings demonstrates exceptional robustness and adaptability. According to the document and Figure 15, when evaluated in both the easy and hard settings of a procedurally generated 2D-platformer environment (CoinRun), the LAM-bas...	gemma-4-31b-it	Genie_generative_interactive_enviroments	pdfs/Genie_generative_interactive_enviroments.pdf
Genie_generative_interactive_enviroments_full_conceptual	S1	full_paper	conceptual	gemma-4-31b-it	How does the training structure of the latent action model illustrated in Figure 5 relate to the consistent semantic behaviors observed when applying the same action across different game states?	{ "input_types": [ "Figure 5", "Figure 17" ], "question_type": "conceptual", "visual_subtype": "visual_understanding", "error_subtype": null, "scope": null }	[ { "sentence": "The training structure of the latent action model (LAM), as illustrated in Figure 5, is designed to learn controllable video generation in a fully unsupervised manner.", "label": "supported", "role": "Observation", "rationale": "Supported by figure and inline text", "depends_on": ...	[]	The training structure of the latent action model (LAM), as illustrated in Figure 5, is designed to learn controllable video generation in a fully unsupervised manner. In this architecture, an encoder takes all previous frames and the subsequent frame as inputs to output a set of continuous latent actions. A decoder th...	gemini-3.5-flash	Genie_generative_interactive_enviroments	pdfs/Genie_generative_interactive_enviroments.pdf
Genie_generative_interactive_enviroments_full_methodological	S1	full_paper	methodological	gemini-3.5-flash	How does the method’s choice to learn a discrete latent action space from video-only data connect to the later controllability and imitation results, and what seems to make the learned actions useful across very different prompts and environments?	{ "input_types": [ "Figure 3", "Figure 13", "Table 2" ], "question_type": "methodological", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "The choice to learn a discrete latent action space from video-only data is directly linked to Genie's frame-by-frame controllability.", "label": "supported", "role": "Interpretation", "rationale": "Directly stated in inline text", "depends_on": [ "S4" ], "errors": []...	[]	The choice to learn a discrete latent action space from video-only data is directly linked to Genie's frame-by-frame controllability. By training a Latent Action Model in an unsupervised manner using a VQ-VAE-based objective, the model restricts the set of possible actions to a small, discrete vocabulary of codes. Beca...	gpt-5.4-mini	Genie_generative_interactive_enviroments	pdfs/Genie_generative_interactive_enviroments.pdf
Genie_generative_interactive_enviroments_page_004_visual_understanding	S4	single_page	visual_understanding	gpt-5.4-mini	Can you describe the architecture of the ST-transformer as shown in Figure 4, including the different layers it contains and how the input tokens are processed into output tokens?	{ "input_types": [ "Figure 4" ], "question_type": "visual_understanding", "visual_subtype": null, "error_subtype": null, "scope": null }	[ { "sentence": "The ST-transformer architecture shown in Figure 4 is composed of L spatiotemporal blocks.", "label": "supported", "role": "Observation", "rationale": "Directly stated in the caption", "depends_on": [], "errors": [], "2d_box": [ { "page_id": 4, "page_n...	[ "pdfs_jpg/Genie_generative_interactive_enviroments/page_004.jpg" ]	The ST-transformer architecture shown in Figure 4 is composed of L spatiotemporal blocks. Each block contains three parts: a spatial layer, a temporal layer, and a feed-forward layer. The figure indicates that the input consists of tokens arranged across T time steps, and these tokens are transformed through the stack ...	gemma-4-31b-it	Genie_generative_interactive_enviroments	pdfs/Genie_generative_interactive_enviroments.pdf
2024_acl-long_533_cross_g2_p5_7_24_results_interpretation	S2	cross_page	results_interpretation	gemma-4-31b-it	What pattern do Table 2 and Table 11 together show about the generated-context setup relative to the retrieved-context baselines on the same benchmark, and which named rows stand out as the main exceptions or near-ties?	{ "input_types": [ "Table 2", "Table 11", "Section 5.2" ], "question_type": "results_interpretation", "visual_subtype": "quantitative", "error_subtype": null, "scope": "multi" }	[ { "sentence": "Together, Table 2 and Table 11 demonstrate a consistent pattern where the generated-context setup, implemented through the MedGENIE framework, generally outperforms retrieved-context baselines across the evaluated benchmarks.", "label": null, "role": "Interpretation", "rationale": "",...	[ "pdfs_jpg/2024_acl-long_533/page_005.jpg", "pdfs_jpg/2024_acl-long_533/page_007.jpg", "pdfs_jpg/2024_acl-long_533/page_024.jpg" ]	Together, Table 2 and Table 11 demonstrate a consistent pattern where the generated-context setup, implemented through the MedGENIE framework, generally outperforms retrieved-context baselines across the evaluated benchmarks. In Table 11, which focuses specifically on MedQA, every MedGENIE model—ranging from MedGENIE-P...	gpt-5.4-mini	MedGENIE	pdfs/MedGENIE.pdf
2024_acl-long_533_page_001_trend	S4	single_page	trend	gemini-3-flash-preview	In Figure 1, how does accuracy change over publication date for the model series—does it climb steadily, flatten out, dip at any point, or jump around—and where are the main rises or drops most noticeable?	{ "input_types": [ "Figure 1" ], "question_type": "trend", "visual_subtype": null, "error_subtype": null, "scope": "single" }	[ { "sentence": "According to Figure 1, the accuracy of the model series on the MedQA-USMLE 4-opt benchmark follows a consistent upward trajectory from January 2021 to February 2024.", "label": "contradiction", "role": "Observation", "rationale": "", "depends_on": [], "errors": [ { ...	[ "pdfs_jpg/2024_acl-long_533/page_001.jpg" ]	According to Figure 1, the accuracy of the model series on the MedQA-USMLE 4-opt benchmark follows a consistent upward trajectory from January 2021 to February 2024. The data illustrates a steady climb in performance over this three-year period, with the frontier line connecting the top-performing models showing no ins...	gpt-5.4-mini	MedGENIE	pdfs/MedGENIE.pdf
2024_acl-long_533_page_022_adversarial	S5	single_page	adversarial	gemma-4-31b-it	If I'm interpreting the charts in Figure 6 correctly, the model that achieves the highest accuracy in every subject for MedMCQA maintains that top position in the MMLU benchmarks as well, specifically outperforming all other models in the Medical Genetics category.	{ "input_types": [ "Figure 6 (MedMCQA panels)", "Figure 6 (MMLU panel)" ], "question_type": "adversarial", "visual_subtype": "comparison", "error_subtype": "reversed_relationship", "scope": "single" }	[ { "sentence": "The interpretation of the charts in Figure 6 is incorrect.", "label": "contradiction", "role": "Interpretation", "rationale": "", "depends_on": [ "S2", "S3", "S4" ], "errors": [ { "span": "", "tags": [ "C4" ], ...	[ "pdfs_jpg/2024_acl-long_533/page_022.jpg" ]	The interpretation of the charts in Figure 6 is incorrect. No single model achieves the highest accuracy across every subject for the MedMCQA benchmarks. While MedGENIE-FID-Flan-T5 holds the top position in many categories, it does not do so in Ophthalmology, where MedGENIE-Zephyr-β achieves the highest accuracy at 70%...	gemini-3-flash-preview	MedGENIE	pdfs/MedGENIE.pdf
OpenBioNER-v2_page_012_comparison	S4	single_page	comparison	gpt-5.4-mini	Looking at Table 3, how does the performance of OpenBioNER-base-v2 compare to GLiNER-BioMed-bi-large-v1.0 across the various Standard and Rare datasets for both entity-level and token-level F1 scores?	{ "input_types": [ "Table 3" ], "question_type": "comparison", "visual_subtype": null, "error_subtype": null, "scope": "single" }	[ { "sentence": "Across Table 3, OpenBioNER-base-v2 is consistently stronger than GLiNER-BioMed-bi-large-v1.0 on both entity-level and token-level F1, and it does so on nearly every dataset shown in the Standard and Rare groups.", "label": "contradiction", "role": "Interpretation", "rationale": "", ...	[ "pdfs_jpg/OpenBioNER-v2/page_012.jpg" ]	Across Table 3, OpenBioNER-base-v2 is consistently stronger than GLiNER-BioMed-bi-large-v1.0 on both entity-level and token-level F1, and it does so on nearly every dataset shown in the Standard and Rare groups. At the entity level, OpenBioNER-base-v2 scores 48.9 on AEM, 57.4 on B2M, 61.1 on B4D, 72.4 on B5R, 58.3 on ...	gemma-4-31b-it	OpenBioNER-v2	pdfs/OpenBioNER-v2.pdf
OpenBioNER-v2_page_013_trend	S4	single_page	trend	gpt-5.4-mini	Based on the 'Tiny' model heatmap in Figure 5, describe the progression of micro-F1 scores across the five description levels. For challenging categories like RNA and Anatomy, how do the scores evolve from the lowest level (bottom row) up to the highest (top row), and at which specific level do we see the most signific...	{ "input_types": [ "Table 4", "Figure 5" ], "question_type": "trend", "visual_subtype": null, "error_subtype": null, "scope": "single" }	[ { "sentence": "In the Tiny model heatmap, the micro-F1 scores generally rise as the descriptions become more informative, moving from the bottom row to the top row.", "label": "supported", "role": "Observation", "rationale": "", "depends_on": [], "errors": [], "2d_box": [ { ...	[ "pdfs_jpg/OpenBioNER-v2/page_013.jpg" ]	In the Tiny model heatmap, the micro-F1 scores generally rise as the descriptions become more informative, moving from the bottom row to the top row. The lowest level shows mostly dark purple across many categories, indicating very low performance. As the description richness increases, the colors shift upward toward g...	gemini-3-flash-preview	OpenBioNER-v2	pdfs/OpenBioNER-v2.pdf
Wake_like_skin_patterning_cross_g0_p2_3_16_conceptual	S2	cross_page	conceptual	nvidia/nemotron-nano-12b-v2-vl:free	Looking at Figure 3 and Extended Data Fig. 6, how does the brain region that shows the stronger low-frequency sleep signal in Figure 3 determine which state-linked LFP change is more tightly tied to waking versus active sleep in Extended Data Fig. 6?	{ "input_types": [ "Figure 3", "Extended Data Fig. 6" ], "question_type": "conceptual", "visual_subtype": "comparison", "error_subtype": null, "scope": null }	[ { "sentence": "To answer how the brain region with the stronger low-frequency sleep signal in Figure 3 determines the state-linked LFP change tied to waking versus active sleep in Extended Data Fig. 6, we first need to identify which brain regions these are based on the provided document.", "label": "not_a_...	[ "pdfs_jpg/Wake_like_skin_patterning/page_002.jpg", "pdfs_jpg/Wake_like_skin_patterning/page_003.jpg", "pdfs_jpg/Wake_like_skin_patterning/page_016.jpg" ]	To answer how the brain region with the stronger low-frequency sleep signal in Figure 3 determines the state-linked LFP change tied to waking versus active sleep in Extended Data Fig. 6, we first need to identify which brain regions these are based on the provided document. The document mentions that the LFPs are measu...	gpt-5.4-mini	Wake_skin_like_patterns	pdfs/Wake_skin_like_patterns.pdf

End of preview. Expand in Data Studio

MMSciFact

Multimodal scientific fact-checking benchmark: model-generated question-answer pairs over scientific PDFs, with every answer sentence human-annotated for role (Observation/Interpretation), dependency structure (depends_on), and grounding label (supported/contradiction/NEI/not_a_claim) against the source document.

45 QA pairs across 9 papers, 461 annotated sentences.
Every included QA pair has at least one contradiction/nei sentence (see src/utils/extract_final_annotations.py in the code repo).
Error taxonomy (error_tags sentence-level for C4/N4 propagation, per-span tags in errors[] for C1-C3/N1-N3) and a mechanically-derived propagation_type per sentence (false_premise / unverifiable_premise / none) — see docs/guidelines.html in the code repo for the full taxonomy definitions.

Files

mmscifact.jsonl — one row per QA pair.
pdfs.zip — one PDF per paper (pdfs/<paper_id>.pdf), referenced by paper_pdf_path on every row. Unzip next to mmscifact.jsonl before running eval scripts: unzip pdfs.zip.

Using this with the MMSciFact eval scripts

mmscifact.jsonl is a flat, single-file view of the data -- convenient for browsing, but the eval scripts in the code repo (eval_batch_api.py, run_vllm_batch.py, eval_oracle_graph.py, eval_holistic_graph.py) all expect the original one-file-per-QA-pair layout (final_annotations/<paper_id>/<qa_pair_id>.json + _source.pdf). Restore that layout with src/utils/hydrate_from_hf.py from the code repo:

hf download alecocc/mmscifact-demo --repo-type dataset --local-dir hf_download
python src/utils/hydrate_from_hf.py --dataset-dir hf_download
# writes final_annotations/<paper_id>/... at the repo root -- every eval
# script then runs completely unmodified from there.

Row schema

Each line of mmscifact.jsonl is one QA pair.

Field	Type	Meaning
`qa_pair_id`	string	Unique identifier for this QA pair
`paper_id`	string	Join key — the PDF is at `pdfs/<paper_id>.pdf` after unzipping `pdfs.zip`
`paper_pdf_path`	string	Relative path to the source PDF, `pdfs/<paper_id>.pdf`
`level`	string	How much of the paper the question requires: `single_page`, `cross_page`, or `full_paper`
`scenario`	string	Question-generation scenario code (`S2`–`S5`) — see the paper's Appendix A
`question_subtype`	string	Finer-grained question type (e.g. comparison, trend, methodological)
`question_model`	string	Model that generated the question
`answer_model`	string	Model that generated the answer being fact-checked
`question_text`	string	The question text
`question_meta`	object	Extra generation-time metadata (input types, cognitive operation, visual/error subtype, scope)
`answer_raw`	string	The full model-generated answer, before it was split into sentences
`annotations`	list of objects	Per-sentence human annotation — see below
`source_images`	list of strings	Page-number-encoding paths for which PDF page(s) the answer draws on

Each entry of annotations:

Field	Type	Meaning
`sentence`	string	The sentence text (one segment of `answer_raw`)
`label`	string	Grounding verdict: `supported`, `contradiction`, `nei`, or `not_a_claim`
`role`	string \| null	`Observation` (a direct, self-contained claim) or `Interpretation` (an inference drawn from other sentences)
`rationale`	string	Human-written explanation for the label
`depends_on`	list of strings	Sentence IDs this sentence's claim logically depends on, e.g. `["S2", "S3"]`
`error_tags`	list of strings	Sentence-level propagation code (`C4`/`N4`), if this sentence's error is inherited from a bad premise elsewhere in the answer
`propagation_type`	string	Mechanically derived from `label`+`depends_on`: `false_premise`, `unverifiable_premise`, or `none`
`errors`	list of objects	Specific erroneous spans within the sentence — see below
`2d_box`	list of objects	Bounding box(es) anchoring this sentence to the source document — see below

Each entry of annotations[].errors:

Field	Type	Meaning
`span`	string	The exact phrase containing the error
`tags`	list of strings	Error code(s) for this span (`C1`–`C3` for contradiction, `N1`–`N3` for NEI)
`correction`	string	Corrected wording for the span; empty for NEI (nothing to replace with — the claim is unverifiable, not wrong)

Each entry of annotations[].2d_box:

Field	Type	Meaning
`page_id`	int	1-indexed PDF page number
`page_name`	string	Human-readable page label
`image_id`	string	Page image identifier
`coords`	`[int, int, int, int]`	Bounding box pixel coordinates `[x1, y1, x2, y2]`
`img_size`	`[int, int]`	`[width, height]` of the page image the coords are relative to

variant, target_label, question_valid, invalid_rationale, and source_pdf are omitted — unused downstream (question_valid is always true by construction: only pairs that passed question validation and were then annotated ever reach this dataset), and source_pdf was purely derivable from level == "full_paper".

License

Annotations (labels, rationale, error codes, dependency structure) are released under CC-BY-4.0. Source PDFs are from openly-accessible papers (NeurIPS, ICLR, ICCV, and similar venues); [TODO: confirm per-paper license terms before publishing -- this card asserts open accessibility, not a verified redistribution license for every included PDF].

Citation

[TODO: add citation once the paper has a venue/BibTeX entry.]

Downloads last month: 44

Number of rows:

Total file size:

116 MB