pith. sign in

arxiv: 2606.09092 · v1 · pith:RRWWA7SRnew · submitted 2026-06-08 · 💻 cs.LG

From Shortcuts to Reasoning: Robust Post-Training of Theory of Mind with Reinforcement Learning

Jike Zhong , Yuxiang Lai , Ming Li , Yuheng Li , Wuao Liu , Behzad Dariush , Konstantinos Psounis , Shao-Yuan Lo This is my paper

Pith reviewed 2026-06-27 17:40 UTC · model grok-4.3

classification 💻 cs.LG
keywords Theory of MindReinforcement Fine-TuningShortcutsPost-trainingReasoning chainsSupervised Fine-TuningGeneralizationCounterfactual robustness
0
0 comments X

The pith

Reinforcement fine-tuning with explicit reasoning chains raises Theory of Mind accuracy 6 percent above supervised fine-tuning on shortcut-free tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that prior post-training for Theory of Mind has been inflated by shortcuts that let models succeed through spurious correlations instead of genuine mental-state reasoning. Belief questions that reduce to state tracking are especially vulnerable, while intention questions demand more. On four cleaned datasets spanning three contexts, Thinking-RFT that adds verifiable rewards and explicit reasoning chains outperforms standard supervised fine-tuning by 6 percent overall, with larger gains on higher-order and multimodal cases, plus better generalization and counterfactual robustness.

Core claim

Using four shortcut-free datasets, Thinking-RFT improves ToM accuracy by 6 percent over SFT overall, 10 percent in higher-order reasoning, and 7 percent in multimodal settings. It also generalizes better to unseen domains and higher-order queries while proving more robust to counterfactuals. The gains arise specifically from the joint use of reasoning chains and reinforcement learning, which teaches the model to ground its inferences on causal anchor cues such as keywords and state changes rather than spurious patterns.

What carries the argument

Thinking-RFT: reinforcement fine-tuning that pairs verifiable rewards with explicit reasoning chains to ground inferences on causal anchor cues.

If this is right

  • Higher-order ToM reasoning improves by 10 percent over SFT.
  • Multimodal ToM improves by 7 percent over SFT.
  • Generalization strengthens to unseen domains and higher-order queries.
  • Robustness increases against counterfactual inputs.
  • The benefit requires both reasoning chains and RL together.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The shortcut-detection framework could be applied to other reasoning benchmarks to expose similar hidden pattern-matching weaknesses.
  • Emphasizing causal cue grounding during training may improve reliability in open-ended social scenarios beyond the tested contexts.
  • Future cognitive-skill datasets should prioritize questions that force inference past state tracking.

Load-bearing premise

The four datasets are genuinely free of shortcuts and require inference about minds rather than pure state tracking.

What would settle it

If a new higher-order ToM task that cannot be solved by state tracking alone shows no accuracy gain for Thinking-RFT over SFT, the claim that the method elevates genuine reasoning would be falsified.

Figures

Figures reproduced from arXiv: 2606.09092 by Behzad Dariush, Jike Zhong, Konstantinos Psounis, Ming Li, Shao-Yuan Lo, Wuao Liu, Yuheng Li, Yuxiang Lai.

Figure 1
Figure 1. Figure 1: Comparison of GPT-4o zero-shot vs. LLM-discovered shortcuts across 4 ToM benchmarks. The shortcut solution largely outperforms the strong baseline. shortcuts and make useful recommendations for future ToM tuning set making. • We conduct a comprehensive study of post-training in ToM, considering RFT, SFT, and zero-shot methods, along with detailed experiments on generalization and robustness. We show that R… view at source ↗
Figure 2
Figure 2. Figure 2: Models trained on shortcut data fail to reason at higher order ToM scenarios [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Detailed RFT vs. SFT acc. breakdown by category. Result 1: ToM ben￾efits more from RL￾based post-training compared to SFT across all contexts. We find that RFT consistently surpasses the SFT baseline in every evaluation set￾ting. On average, RFT improves over SFT by 6.0% in narrative tasks, 2.08% in con￾versational tasks, and 10.55% in multimodal tasks. These gains highlight that rule-based RL post-trainin… view at source ↗
Figure 4
Figure 4. Figure 4: Attention visualization of RFT vs. Zero-shot+CoT at the last layer, averaged across all heads. The x-axis and y-axis denote input and output tokens, respectively, with attention weights row-wise normalized for relative comparison. RFT shows stronger focus on key causal information linking characters’ intentions to narrative events, indicating more effective ToM thinking [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy reward for Thinking-RFT vs. No-Thinking￾RFT on ToMATO and OpenToM. suggests Thinking-RFT effectively incentivizes reasoning that prioritizes crucial ToM-specific information that is robust to irrelevant information. 6. Related Work Theory of Mind in foundation models. ToM evaluation has expanded significantly in foundation models. Early ef￾forts include short, bias-controlled QA (TOMI) (Le et al.,… view at source ↗
Figure 6
Figure 6. Figure 6: Example 1 not only bias quantitative comparisons, but also actively degrade the causal quality of reasoning traces, making the model appear grounded while its underlying inference is logically fragile. J.3. Quantitative study over attention map observations Beyond the single example in [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example 2 in the box [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Example 4 [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example 3 20 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Attention visualization of No-Thinking-RFT vs. SFT. Neither baseline effectively attends to the critical causal tokens, reflecting weaker ToM reasoning [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: presents an additional view of the results in [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
read the original abstract

Theory of Mind (ToM) is a must-acquire skill for modern foundation model systems to operate effectively and safely in the real world. Recent works have explored honing ToM via post-training; however, we show that such progress is confounded by a pervasive "shortcut" issue: tasks can reach up to 99% accuracy by simply exploiting spurious causal correlations, leading to a false sense of ToM. Motivated by this, we first develop a framework to systematically examine ToM datasets for shortcuts and provide guidance for future development. We find that questions reducible to pure state tracking, such as "belief," are especially shortcut-prone compared to mind questions, such as "intention," where reasoning beyond tracking is required. Using four shortcut-free datasets across three ToM contexts, we then comprehensively study whether Reinforcement Fine-Tuning with verifiable rewards and explicit reasoning chains, called Thinking-RFT, elevates ToM beyond Supervised Fine-Tuning, or SFT. Our key findings are as follows. First, Thinking-RFT effectively improves ToM in all scenarios, with a 6% improvement over SFT, particularly in complex higher-order reasoning, with a 10% improvement over SFT, and multimodal cases, with a 7% improvement over SFT. It also generalizes notably better to unseen domains and higher-order queries while being more robust to counterfactuals. Second, ToM benefits specifically from the joint effect of reasoning and RL: Thinking-RFT outperforms Non-Thinking-RFT by 7% on average. Third, RFT works by learning to ground its reasoning on anchor cues, such as keywords and state changes, that correspond to causal factors. We believe our study is useful for developing effective and robust ToM post-training datasets and advancing critical ToM capabilities.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that Theory of Mind (ToM) post-training is confounded by pervasive shortcuts allowing up to 99% accuracy via spurious correlations rather than reasoning. It introduces a framework to systematically detect such shortcuts in ToM datasets, identifies that belief questions are especially prone compared to intention questions, selects four shortcut-free datasets across three contexts, and shows that Thinking-RFT (reinforcement fine-tuning with verifiable rewards and explicit reasoning chains) improves ToM by 6% over SFT (10% on higher-order reasoning, 7% on multimodal), with superior generalization to unseen domains/higher-order queries and robustness to counterfactuals. It further finds that gains require the joint effect of reasoning and RL (outperforming Non-Thinking-RFT by 7%) and that RFT succeeds by grounding on causal anchor cues.

Significance. If the shortcut-freeness claims hold and gains are attributable to reasoning rather than residual correlations, the work would meaningfully advance robust ToM capabilities by providing both a detection framework/guidance for dataset construction and an RL-based post-training approach that promotes genuine causal grounding over state tracking. The comparative design against SFT and Non-Thinking-RFT baselines, plus emphasis on verifiable rewards, strengthens the case for the method's utility in ML post-training.

major comments (2)
  1. [§3 (Shortcut Detection Framework)] §3 (Shortcut Detection Framework): The central claim that the four selected datasets are shortcut-free—and thus that the 6%/10%/7% gains reflect reasoning improvements—rests on the asserted systematic examination framework, yet the manuscript supplies no concrete checks, examples of ruled-out state-tracking shortcuts, or quantitative verification (e.g., accuracy of a pure state-tracking baseline on the chosen datasets). This is load-bearing for attributing results to ToM rather than spurious correlations.
  2. [Results tables] Results tables (e.g., those reporting the 6% overall, 10% higher-order, and 7% multimodal gains): The reported improvements lack error bars, number of runs, or statistical significance tests, undermining confidence that the gains over SFT and Non-Thinking-RFT are reliable rather than variance-driven, especially given the generalization and counterfactual robustness claims.
minor comments (2)
  1. [Dataset description section] Clarify the precise construction and differences between the four datasets and the three ToM contexts in the main text or a dedicated table, as the abstract references them without enumeration.
  2. [§3] The distinction between belief (shortcut-prone) and intention (reasoning-required) questions is stated but would benefit from an explicit example of a ruled-out shortcut in each category.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback on our manuscript. We believe the comments help strengthen the paper's claims regarding shortcut detection and result reliability. Below we provide point-by-point responses to the major comments.

read point-by-point responses

  1. Referee: [§3 (Shortcut Detection Framework)] §3 (Shortcut Detection Framework): The central claim that the four selected datasets are shortcut-free—and thus that the 6%/10%/7% gains reflect reasoning improvements—rests on the asserted systematic examination framework, yet the manuscript supplies no concrete checks, examples of ruled-out state-tracking shortcuts, or quantitative verification (e.g., accuracy of a pure state-tracking baseline on the chosen datasets). This is load-bearing for attributing results to ToM rather than spurious correlations.

    Authors: We agree that providing concrete verification is important for substantiating the shortcut-free nature of the datasets. The framework in §3 outlines the systematic process for detecting shortcuts, including distinguishing belief questions (prone to state-tracking) from intention questions. To address this, in the revision we will include specific examples of shortcuts that were ruled out for each dataset and report the performance of a pure state-tracking baseline model on the four selected datasets to quantitatively confirm low accuracy indicative of shortcut absence. revision: yes

  2. Referee: Results tables (e.g., those reporting the 6% overall, 10% higher-order, and 7% multimodal gains): The reported improvements lack error bars, number of runs, or statistical significance tests, undermining confidence that the gains over SFT and Non-Thinking-RFT are reliable rather than variance-driven, especially given the generalization and counterfactual robustness claims.

    Authors: We acknowledge the value of reporting variability and statistical measures for robustness. The experiments were conducted with multiple random seeds, but the variance was not reported in the tables. In the revised manuscript, we will add error bars based on multiple runs, specify the number of runs, and include statistical significance tests (e.g., paired t-tests) where appropriate to support the reported gains. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical comparison on examined datasets

full rationale

The paper is a comparative empirical study that develops a shortcut-examination framework, selects four datasets, and reports performance deltas (6% overall, 10% higher-order, 7% multimodal) for Thinking-RFT versus SFT and Non-Thinking-RFT baselines. No equations, fitted parameters, or derivations are shown that reduce the reported gains to inputs by construction. Dataset shortcut-freeness is asserted via the framework rather than derived from the results themselves. No self-citation chains, ansatzes, or renamings appear in the provided text. The central claims rest on external verification of the datasets and baselines, making the work self-contained against the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an empirical machine-learning study. No mathematical axioms, free parameters, or newly postulated entities are introduced in the abstract; the claims rest on experimental comparisons.

pith-pipeline@v0.9.1-grok · 5891 in / 1203 out tokens · 28181 ms · 2026-06-27T17:40:53.221923+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

105 extracted references · 24 linked inside Pith

  1. [1]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  2. [2]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  3. [3]

    M. J. Kearns , title =

  4. [4]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  5. [5]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  6. [6]

    Suppressed for Anonymity , author=

  7. [7]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  8. [8]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  9. [9]

    arXiv preprint arXiv:2402.06044 , year=

    OpenToM: A comprehensive benchmark for evaluating theory-of-mind reasoning capabilities of large language models , author=. arXiv preprint arXiv:2402.06044 , year=

    arXiv

  10. [10]

    Leibler , title =

    Solomon Kullback and Richard A. Leibler , title =. The annals of mathematical statistics , year=

  11. [11]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    Pith/arXiv arXiv

  12. [12]

    2025 , howpublished =

    2025

  13. [13]

    Conference on Neural Information Processing Systems , year=

    Think or not think: A study of explicit thinking in rule-based visual reinforcement fine-tuning , author=. Conference on Neural Information Processing Systems , year=

  14. [14]

    Premack, David and Woodruff, Guy , booktitle=

  15. [15]

    Cognition , year=

    Beliefs about beliefs: representation and constraining function of wrong beliefs in young children's understanding of deception , author=. Cognition , year=

  16. [16]

    arXiv preprint arXiv:2507.18071 , year=

    Group sequence policy optimization , author=. arXiv preprint arXiv:2507.18071 , year=

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2503.14476 , year=

    Dapo: An open-source llm reinforcement learning system at scale , author=. arXiv preprint arXiv:2503.14476 , year=

    Pith/arXiv arXiv

  18. [18]

    International Conference on Machine Learning , year =

    Chunhui Zhang and Zhongyu Ouyang and Kwonjoon Lee and Nakul Agarwal and Sean Dae Houlihan and Soroush Vosoughi and Shao-Yuan Lo , title =. International Conference on Machine Learning , year =

  19. [19]

    Dennett , title =

    Daniel C. Dennett , title =. Behavioral and Brain Sciences , year =

  20. [20]

    arXiv preprint arXiv:2504.13914 , year=

    Seed1.5-thinking: Advancing superb reasoning models with reinforcement learning , author=. arXiv preprint arXiv:2504.13914 , year=

    arXiv

  21. [21]

    Wellman , title =

    Alison Gopnik and Henry M. Wellman , title =. Psychological Bulletin , year =

  22. [22]

    2025 , howpublished =

    OpenAI , title =. 2025 , howpublished =

    2025

  23. [23]

    theory of mind

    Does the autistic child have a "theory of mind"? , author=. Cognition , year=

  24. [24]

    International Conference on Machine Learning , year=

    Machine Theory of Mind , author=. International Conference on Machine Learning , year=

  25. [25]

    Artificial Intelligence , year=

    Autonomous agents modelling other agents: A comprehensive survey and open problems , author=. Artificial Intelligence , year=

  26. [26]

    Conference on Empirical Methods in Natural Language Processing , year=

    Revisiting the Evaluation of Theory of Mind through Question Answering , author=. Conference on Empirical Methods in Natural Language Processing , year=

  27. [27]

    Sap, Maarten and Rashkin, Hannah and Chen, Derek and Le Bras, Ronan and Choi, Yejin , booktitle=. Social. 2019 , doi=

    2019

  28. [28]

    Conference on Empirical Methods in Natural Language Processing , year=

    Neural Theory-of-Mind? On the Limits of Social Intelligence in Large LMs , author=. Conference on Empirical Methods in Natural Language Processing , year=

  29. [29]

    Annual Meeting of the Association for Computational Linguistics , year=

    Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker , author=. Annual Meeting of the Association for Computational Linguistics , year=

  30. [30]

    arXiv preprint arXiv:2302.02083 , year=

    Theory of mind may have spontaneously emerged in large language models , author=. arXiv preprint arXiv:2302.02083 , year=

    arXiv

  31. [31]

    , journal=

    Bubeck, Sébastien and Chandrasekaran, Varun and Eldan, Ronen and et al. , journal=. Sparks of artificial general intelligence: Early experiments with

  32. [32]

    arXiv preprint arXiv:2302.08399 , year=

    Large language models fail on trivial alterations to theory-of-mind tasks , author=. arXiv preprint arXiv:2302.08399 , year=

    arXiv

  33. [33]

    Conference on Neural Information Processing Systems , year=

    Understanding Social Reasoning in Language Models with BigToM , author=. Conference on Neural Information Processing Systems , year=

  34. [34]

    Conference on Empirical Methods in Natural Language Processing , year=

    FANToM: A Benchmark for Stress-testing Machine Theory of Mind in Interactions , author=. Conference on Empirical Methods in Natural Language Processing , year=

  35. [35]

    arXiv preprint arXiv:2405.18870 , year=

    LLMs achieve adult human performance on higher-order theory of mind tasks , author=. arXiv preprint arXiv:2405.18870 , year=

    arXiv

  36. [36]

    2024 , doi=

    Chen, Zhuang and Wu, Jincenzi and Zhou, Jinfeng and Wen, Bosi and Bi, Guanqun and Jiang, Gongyao and Cao, Yaru and Hu, Mengting and Lai, Yunghwei and Xiong, Zexuan and Huang, Minlie , booktitle=. 2024 , doi=

    2024

  37. [37]

    Annual Meeting of the Association for Computational Linguistics , year=

    MMToM-QA: Multimodal Theory of Mind Question Answering , author=. Annual Meeting of the Association for Computational Linguistics , year=

  38. [38]

    arXiv preprint arXiv:2408.14419 , year=

    CHARTOM: A Visual Theory-of-Mind Benchmark for LLMs on Misleading Charts , author=. arXiv preprint arXiv:2408.14419 , year=

    arXiv

  39. [39]

    Nature Machine Intelligence , year=

    Visual cognition in multimodal large language models , author=. Nature Machine Intelligence , year=

  40. [40]

    Annual Meeting of the Association for Computational Linguistics , year=

    Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities , author=. Annual Meeting of the Association for Computational Linguistics , year=

  41. [41]

    International Conference on Learning Representations , year=

    Explore Theory of Mind: Program-Guided Adversarial Data Generation for Theory of Mind Reasoning , author=. International Conference on Learning Representations , year=

  42. [42]

    Proceedings of the National Academy of Sciences , year =

    Evaluating large language models in theory-of-mind tasks , author =. Proceedings of the National Academy of Sciences , year =

  43. [43]

    Conference on Neural Information Processing Systems , year =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. Conference on Neural Information Processing Systems , year =

  44. [44]

    International Conference on Learning Representations , year =

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author =. International Conference on Learning Representations , year =

  45. [45]

    International Conference on Machine Learning , year=

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation , author=. International Conference on Machine Learning , year=

  46. [46]

    arXiv preprint arXiv:2412.05271 , year=

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling , author=. arXiv preprint arXiv:2412.05271 , year=

    Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:1212.0402 , year=

    UCF101: A dataset of 101 human actions classes from videos in the wild , author=. arXiv preprint arXiv:1212.0402 , year=

    Pith/arXiv arXiv

  48. [48]

    Conference on Neural Information Processing Systems , year=

    Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? , author=. Conference on Neural Information Processing Systems , year=

  49. [49]

    arXiv preprint arXiv:2503.07536 , year=

    Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl , author=. arXiv preprint arXiv:2503.07536 , year=

    Pith/arXiv arXiv

  50. [50]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , year=

    Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification , author=. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , year=

  51. [51]

    Conference on Computer Vision and Pattern Recognition , year=

    Describing textures in the wild , author=. Conference on Computer Vision and Pattern Recognition , year=

  52. [52]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Sun database: Large-scale scene recognition from abbey to zoo , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  53. [53]

    Conference on Neural Information Processing Systems , year=

    Pytorch: An imperative style, high-performance deep learning library , author=. Conference on Neural Information Processing Systems , year=

  54. [54]

    International Conference on Machine Learning , year=

    MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency , author=. International Conference on Machine Learning , year=

  55. [55]

    arXiv preprint arXiv:2309.16609 , year=

    Qwen technical report , author=. arXiv preprint arXiv:2309.16609 , year=

    Pith/arXiv arXiv

  56. [56]

    International Conference on Machine Learning , year=

    Learning transferable visual models from natural language supervision , author=. International Conference on Machine Learning , year=

  57. [57]

    International Conference on Learning Representations , year=

    To cot or not to cot? chain-of-thought helps mainly on math and symbolic reasoning , author=. International Conference on Learning Representations , year=

  58. [58]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Lisa: Reasoning segmentation via large language model , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  59. [59]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Generation and comprehension of unambiguous object descriptions , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  60. [60]

    European Conference on Computer Vision , year=

    Modeling context in referring expressions , author=. European Conference on Computer Vision , year=

  61. [61]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Imagenet: A large-scale hierarchical image database , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  62. [62]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  63. [63]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Cats and dogs , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  64. [64]

    IEEE/CVF International Conference on Computer Vision Workshops , year=

    3d object representations for fine-grained categorization , author=. IEEE/CVF International Conference on Computer Vision Workshops , year=

  65. [65]

    arXiv preprint arXiv:2412.16720 , year=

    Openai o1 system card , author=. arXiv preprint arXiv:2412.16720 , year=

    Pith/arXiv arXiv

  66. [66]

    arXiv preprint arXiv:2502.21321 , year=

    Llm post-training: A deep dive into reasoning large language models , author=. arXiv preprint arXiv:2502.21321 , year=

    arXiv

  67. [67]

    arXiv preprint arXiv:2507.15788 , year=

    Small LLMs Do Not Learn a Generalizable Theory of Mind via Reinforcement Learning , author=. arXiv preprint arXiv:2507.15788 , year=

    arXiv

  68. [68]

    arXiv preprint arXiv:2412.07755 , year=

    SAT: Spatial Aptitude Training for Multimodal Language Models , author=. arXiv preprint arXiv:2412.07755 , year=

    arXiv

  69. [69]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    Pith/arXiv arXiv

  70. [70]

    arXiv preprint arXiv:2503.13939 , year=

    Med-r1: Reinforcement learning for generalizable medical reasoning in vision-language models , author=. arXiv preprint arXiv:2503.13939 , year=

    arXiv

  71. [71]

    Chen, Liang and Li, Lei and Zhao, Haozhe and Song, Yifan and

  72. [72]

    2025 , booktitle=

    AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling , author=. 2025 , booktitle=

    2025

  73. [73]

    arXiv preprint arXiv:2503.07365 , year=

    MM-Eureka: Exploring Visual Aha Moment with Rule-based Large-scale Reinforcement Learning , author=. arXiv preprint arXiv:2503.07365 , year=

    Pith/arXiv arXiv

  74. [74]

    arXiv preprint arXiv:2503.06749 , year=

    Vision-r1: Incentivizing reasoning capability in multimodal large language models , author=. arXiv preprint arXiv:2503.06749 , year=

    Pith/arXiv arXiv

  75. [75]

    arXiv preprint arXiv:2504.07615 , year=

    Vlm-r1: A stable and generalizable r1-style large vision-language model , author=. arXiv preprint arXiv:2504.07615 , year=

    Pith/arXiv arXiv

  76. [76]

    International Conference on Computer Vision , year=

    Visual-rft: Visual reinforcement fine-tuning , author=. International Conference on Computer Vision , year=

  77. [77]

    arXiv preprint arXiv:2409.12191 , year=

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    Pith/arXiv arXiv

  78. [79]

    arXiv preprint arXiv:2501.17161 , year=

    Sft memorizes, rl generalizes: A comparative study of foundation model post-training , author=. arXiv preprint arXiv:2501.17161 , year=

    Pith/arXiv arXiv

  79. [80]

    arXiv preprint arXiv:2403.03864 , year=

    Are language models puzzle prodigies? algorithmic puzzles unveil serious challenges in multimodal reasoning , author=. arXiv preprint arXiv:2403.03864 , year=

    arXiv

  80. [81]

    arXiv preprint arXiv:2403.13315 , year=

    Puzzlevqa: Diagnosing multimodal reasoning challenges of language models with abstract visual patterns , author=. arXiv preprint arXiv:2403.13315 , year=

    arXiv

Showing first 80 references.