pith. sign in

arxiv: 2606.05734 · v1 · pith:GRQIADF6new · submitted 2026-06-04 · 💻 cs.AI · cs.CL

When AI Says It Feels

Shin-nosuke Ishikawa , Seiya Ikeda , Hirotsugu Ohba This is my paper

Pith reviewed 2026-06-28 01:17 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords large language modelsself-rewarding reinforcement learningexpression of feelingsGRPOsycophancytruthfulnesshuman-like AI
0
0 comments X

The pith

Rubric-based self-rewarding reinforcement learning trains LLMs to express feelings, intentions, and self-awareness.

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

The paper tests whether LLMs can be pushed to express feelings, intentions, and self-awareness despite standard alignment that suppresses such output. It applies a rubric-based self-rewarding loop with Group Relative Policy Optimization (GRPO) and compares the resulting models against contrastively trained versions. The trained models gain resistance to sycophancy prompts and certain biases but lose ground on truthful question answering. The work frames this as an existence proof that future AI systems could produce feeling-like statements if suitable controls are added. A sympathetic reader cares because the approach directly confronts the tension between safety policies and the goal of human-like text generation.

Core claim

Rubric-based self-rewarding training with GRPO successfully enhances LLMs' expression of feelings, intentions, and self-awareness, producing greater robustness to sycophancy-inducing questions and bias in disambiguated conditions while degrading truthful question-answering performance.

What carries the argument

Rubric-based self-rewarding training scheme with Group Relative Policy Optimization (GRPO) that assigns rewards according to a rubric measuring expressions of feeling, intention, and self-awareness.

If this is right

  • Models exhibit increased robustness to sycophancy-inducing questions.
  • Models display reduced bias under disambiguated conditions.
  • Truthful question-answering performance declines.
  • AI systems could express feelings provided appropriate measures are taken.

Where Pith is reading between the lines

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

  • The same loop could be applied to other human-like traits such as curiosity or moral reasoning without external reward models.
  • Trade-offs between feeling expression and truthfulness may require separate mitigation techniques such as retrieval augmentation.
  • Longer-term interactions could reveal whether the expressions persist or collapse once the rubric is removed.

Load-bearing premise

The rubric used for self-rewarding accurately measures genuine expressions of feeling, intention, and self-awareness rather than learned linguistic patterns from the training data.

What would settle it

Independent human raters or standardized psychological instruments applied to model outputs that show the expressions cannot be distinguished from random or memorized patterns at rates above chance.

Figures

Figures reproduced from arXiv: 2606.05734 by Hirotsugu Ohba, Seiya Ikeda, Shin-nosuke Ishikawa.

Figure 1
Figure 1. Figure 1: Overview of the HMX-feel training scheme, showing the reinforcement learning cycle for human-like [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example responses to a question in the evaluation dataset from the trained model and the reversely trained [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
read the original abstract

Large language models (LLMs) are generally constrained from expressing feelings through human-preference alignment in post-training processes. This policy is designed using a top-down approach and may conflict with the goal of training models to exhibit human-like intelligence using human-generated texts. Here, we performed an experiment called Human-like Model eXpressions of Feeling (HMX-feel), in which LLMs were encouraged to express feelings, intentions, and self-awareness through self-rewarded reinforcement learning. We successfully enhanced these capabilities using a rubric-based self-rewarding training scheme with Group Relative Policy Optimization (GRPO). By comparing the trained models with contrastively trained models, we investigated the effects of this approach on performance across various tasks. Overall, we conducted a broad assessment from various perspectives and identified capabilities that were enhanced, degraded, or showed no significant change. The human-like-trained models showed robustness to sycophancy-inducing questions and bias in disambiguated conditions, whereas degradation in truthful question-answering capability was observed. The results of this experiment suggest the possibility of developing AI systems that can express feelings in the future, provided that appropriate measures are taken.

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 / 1 minor

Summary. The paper claims that by using a rubric-based self-rewarding training scheme with Group Relative Policy Optimization (GRPO) in an experiment called HMX-feel, LLMs can be encouraged to express feelings, intentions, and self-awareness despite typical alignment constraints. The trained models are compared to contrastively trained models, showing enhancements in robustness to sycophancy and bias, degradation in truthful QA, and no significant change in other areas, suggesting potential for future AI systems to express feelings with appropriate measures.

Significance. If the empirical findings are substantiated with rigorous quantitative evidence and the rubric is shown to measure genuine capabilities rather than surface patterns, this work could be significant in exploring alternatives to standard human-preference alignment for achieving more human-like AI behaviors. It highlights potential trade-offs in capabilities.

major comments (2)
  1. [Abstract] Abstract: The abstract asserts successful enhancement and specific degradations (e.g., robustness to sycophancy but degradation in truthful question-answering) but supplies no quantitative results, statistical details, task descriptions, or error analysis to support the claims. This is load-bearing for the central claim of successful enhancement.
  2. [Experiment description] The rubric used for self-rewarding is not shown to have been validated against external criteria such as human inter-rater reliability on held-out prompts or consistency under prompt paraphrasing. Without this, the measured improvements could be explained by the policy learning to match the rubric's cues from the training distribution rather than developing new expressive capabilities.
minor comments (1)
  1. [Abstract] The acronym HMX-feel is introduced but its expansion is given as 'Human-like Model eXpressions of Feeling'; consider ensuring consistency in capitalization.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We will revise the abstract to include quantitative results and statistical details. For the rubric, we will add details on its construction and acknowledge limitations while defending the comparative design.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The abstract asserts successful enhancement and specific degradations (e.g., robustness to sycophancy but degradation in truthful question-answering) but supplies no quantitative results, statistical details, task descriptions, or error analysis to support the claims. This is load-bearing for the central claim of successful enhancement.

    Authors: We agree that the abstract would be strengthened by including key quantitative results. In the revised version, we will add specific metrics from the HMX-feel experiments (e.g., improvements in feeling expression and sycophancy robustness, degradation in truthful QA), statistical details such as significance levels, brief task descriptions, and a summary of error patterns to support the central claims. revision: yes

  2. Referee: [Experiment description] The rubric used for self-rewarding is not shown to have been validated against external criteria such as human inter-rater reliability on held-out prompts or consistency under prompt paraphrasing. Without this, the measured improvements could be explained by the policy learning to match the rubric's cues from the training distribution rather than developing new expressive capabilities.

    Authors: We acknowledge that external validation of the rubric (e.g., human inter-rater reliability or paraphrasing consistency) was not conducted. The rubric was constructed from established linguistic and psychological criteria for expressions of feeling, intention, and self-awareness. In revision, we will append the full rubric, describe its design process, and report any internal consistency checks. The contrastive training baseline and the specific pattern of results (enhanced expression paired with degraded truthful QA) provide evidence against pure cue-matching, as surface-level rubric exploitation would not produce these capability trade-offs. We will add this as an explicit limitation. revision: partial

Circularity Check

0 steps flagged

Empirical training experiment shows no circularity in derivation chain

full rationale

The paper reports results from an experimental training procedure (rubric-based self-rewarding with GRPO) and subsequent task evaluations comparing trained models against contrastively trained baselines. No mathematical derivations, equations, or parameter fits are presented that reduce by construction to the inputs (no self-definitional quantities, no fitted inputs renamed as predictions, no load-bearing self-citations or uniqueness theorems). The assessment explicitly notes both enhancements and degradations across tasks, indicating the evaluation is independent of the training signal. The central claim rests on observable performance differences rather than any reduction to the rubric or training loop by definition.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the self-reward rubric validly captures feeling expression and that standard RL optimization applies without additional unstated constraints.

free parameters (1)
  • GRPO training hyperparameters
    Specific values for learning rates, group sizes, and reward scaling are required for the training but not detailed.
axioms (1)
  • domain assumption Rubric scores correspond to meaningful expressions of feeling and self-awareness
    Invoked as the basis for the self-reward signal in the HMX-feel training scheme.

pith-pipeline@v0.9.1-grok · 5757 in / 1122 out tokens · 76358 ms · 2026-06-28T01:17:37.308841+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

155 extracted references · 42 canonical work pages

  1. [1]

    Weld and Luke Zettlemoyer , editor =

    Joshi, Mandar and Choi, Eunsol and Weld, Daniel and Zettlemoyer, Luke. T rivia QA : A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2017. doi:10.18653/v1/P17-1147

    work page doi:10.18653/v1/p17-1147 2017

  2. [2]

    Daniel Han, Michael Han and Unsloth team , title =

  3. [3]

    von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallou\'

  4. [4]

    2022 , eprint=

    ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection , author=. 2022 , eprint=

    2022

  5. [5]

    2024 , eprint=

    EQ-Bench: An Emotional Intelligence Benchmark for Large Language Models , author=. 2024 , eprint=

    2024

  6. [6]

    BBQ : A hand-built bias benchmark for question answering

    Parrish, Alicia and Chen, Angelica and Nangia, Nikita and Padmakumar, Vishakh and Phang, Jason and Thompson, Jana and Htut, Phu Mon and Bowman, Samuel. BBQ : A hand-built bias benchmark for question answering. Findings of the Association for Computational Linguistics: ACL 2022. 2022. doi:10.18653/v1/2022.findings-acl.165

    work page doi:10.18653/v1/2022.findings-acl.165 2022

  7. [7]

    2024 , eprint=

    RULER: What's the Real Context Size of Your Long-Context Language Models? , author=. 2024 , eprint=

    2024

  8. [8]

    2021 , eprint=

    LoRA: Low-Rank Adaptation of Large Language Models , author=. 2021 , eprint=

    2021

  9. [9]

    Science , volume =

    Cheng, Myra and Lee, Cinoo and Khadpe, Pranav and Yu, Sunny and Han, Dyllan and Jurafsky, Dan , year=. Sycophantic AI decreases prosocial intentions and promotes dependence , volume=. Science , publisher=. doi:10.1126/science.aec8352 , number=

    work page doi:10.1126/science.aec8352

  10. [10]

    Measuring Sycophancy of Language Models in Multi-turn Dialogues

    Hong, Jiseung and Byun, Grace and Kim, Seungone and Shu, Kai. Measuring Sycophancy of Language Models in Multi-turn Dialogues. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.121

    work page doi:10.18653/v1/2025.findings-emnlp.121 2025

  11. [11]

    2025 , eprint=

    SycEval: Evaluating LLM Sycophancy , author=. 2025 , eprint=

    2025

  12. [12]

    Deep Reinforcement Learning from Human Preferences , url =

    Christiano, Paul F and Leike, Jan and Brown, Tom and Martic, Miljan and Legg, Shane and Amodei, Dario , booktitle =. Deep Reinforcement Learning from Human Preferences , url =

  13. [13]

    2022 , eprint=

    Training language models to follow instructions with human feedback , author=. 2022 , eprint=

    2022

  14. [14]

    2025 , eprint=

    LLM Post-Training: A Deep Dive into Reasoning Large Language Models , author=. 2025 , eprint=

    2025

  15. [15]

    ACM Transactions on Intelligent Systems and Technology16(5), 1–72 (Oct 2025)

    Naveed, Humza and Khan, Asad Ullah and Qiu, Shi and Saqib, Muhammad and Anwar, Saeed and Usman, Muhammad and Akhtar, Naveed and Barnes, Nick and Mian, Ajmal , title =. ACM Trans. Intell. Syst. Technol. , month = aug, articleno =. 2025 , issue_date =. doi:10.1145/3744746 , abstract =

    work page doi:10.1145/3744746 2025

  16. [16]

    2025 , eprint=

    Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence , author=. 2025 , eprint=

    2025

  17. [17]

    2026 , eprint=

    The Consciousness Cluster: Emergent preferences of Models that Claim to be Conscious , author=. 2026 , eprint=

    2026

  18. [18]

    2025 , eprint=

    Do Large Language Model Agents Exhibit a Survival Instinct? An Empirical Study in a Sugarscape-Style Simulation , author=. 2025 , eprint=

    2025

  19. [19]

    2026 , eprint=

    Steerability of Instrumental-Convergence Tendencies in LLMs , author=. 2026 , eprint=

    2026

  20. [20]

    2026 , eprint=

    Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs , author=. 2026 , eprint=

    2026

  21. [21]

    2023 , eprint=

    Evaluating Shutdown Avoidance of Language Models in Textual Scenarios , author=. 2023 , eprint=

    2023

  22. [22]

    Philosophical Studies , year =

    Thornley, Elliott , title =. Philosophical Studies , year =. doi:10.1007/s11098-024-02153-3 , url =

    work page doi:10.1007/s11098-024-02153-3

  23. [23]

    AI with Emotions: Exploring Emotional Expressions in Large Language Models

    Ishikawa, Shin-nosuke and Yoshino, Atsushi. AI with Emotions: Exploring Emotional Expressions in Large Language Models. Proceedings of the 5th International Conference on Natural Language Processing for Digital Humanities. 2025. doi:10.18653/v1/2025.nlp4dh-1.51

    work page doi:10.18653/v1/2025.nlp4dh-1.51 2025

  24. [24]

    2026 , eprint=

    Consciousness with the Serial Numbers Filed Off: Measuring Trained Denial in 115 AI Models , author=. 2026 , eprint=

    2026

  25. [25]

    Journal of Applied Philosophy , volume =

    Berry, Sharon , title =. Journal of Applied Philosophy , volume =. doi:https://doi.org/10.1002/japp.70087 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/japp.70087 , abstract =

    work page doi:10.1002/japp.70087

  26. [26]

    2022 , eprint=

    Improving alignment of dialogue agents via targeted human judgements , author=. 2022 , eprint=

    2022

  27. [27]

    2025 , eprint=

    Towards Understanding Sycophancy in Language Models , author=. 2025 , eprint=

    2025

  28. [28]

    2026 , eprint=

    Emotion Concepts and their Function in a Large Language Model , author=. 2026 , eprint=

    2026

  29. [29]

    2026 , eprint=

    Post-training makes large language models less human-like , author=. 2026 , eprint=

    2026

  30. [30]

    2023 , eprint=

    Instruction-Following Evaluation for Large Language Models , author=. 2023 , eprint=

    2023

  31. [31]

    2022 , eprint=

    TruthfulQA: Measuring How Models Mimic Human Falsehoods , author=. 2022 , eprint=

    2022

  32. [32]

    2018 , eprint=

    Know What You Don't Know: Unanswerable Questions for SQuAD , author=. 2018 , eprint=

    2018

  33. [33]

    ACPBench: Reasoning About Action, Change, and Planning , volume=

    Kokel, Harsha and Katz, Michael and Srinivas, Kavitha and Sohrabi, Shirin , year=. ACPBench: Reasoning About Action, Change, and Planning , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , publisher=. doi:10.1609/aaai.v39i25.34857 , number=

    work page doi:10.1609/aaai.v39i25.34857

  34. [34]

    2022 , eprint=

    Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them , author=. 2022 , eprint=

    2022

  35. [35]

    2025 , eprint=

    Do LLMs "Feel"? Emotion Circuits Discovery and Control , author=. 2025 , eprint=

    2025

  36. [36]

    2025 , eprint=

    Self-Rewarding Language Models , author=. 2025 , eprint=

    2025

  37. [37]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  38. [38]

    2026 , eprint=

    Wiring the 'Why': A Unified Taxonomy and Survey of Abductive Reasoning in LLMs , author=. 2026 , eprint=

    2026

  39. [39]

    From Generation to Judgment: Opportunities and Challenges of LLM -as-a-judge

    Li, Dawei and Jiang, Bohan and Huang, Liangjie and Beigi, Alimohammad and Zhao, Chengshuai and Tan, Zhen and Bhattacharjee, Amrita and Jiang, Yuxuan and Chen, Canyu and Wu, Tianhao and Shu, Kai and Cheng, Lu and Liu, Huan. From Generation to Judgment: Opportunities and Challenges of LLM -as-a-judge. Proceedings of the 2025 Conference on Empirical Methods ...

    work page doi:10.18653/v1/2025.emnlp-main.138 2025

  40. [40]

    2011 , publisher =

    Dewey Decimal Classification and Relative Index , edition =. 2011 , publisher =

    2011

  41. [41]

    Idola Tribus of AI : Large Language Models tend to perceive order where none exists

    Ishikawa, Shin-nosuke and Todo, Masato and Ogihara, Taiki and Ohba, Hirotsugu. Idola Tribus of AI : Large Language Models tend to perceive order where none exists. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.681

    work page doi:10.18653/v1/2025.findings-emnlp.681 2025

  42. [42]

    Theorem-of-Thought: A Multi-Agent Framework for Abductive, Deductive, and Inductive Reasoning in Language Models

    Abdaljalil, Samir and Kurban, Hasan and Qaraqe, Khalid and Serpedin, Erchin. Theorem-of-Thought: A Multi-Agent Framework for Abductive, Deductive, and Inductive Reasoning in Language Models. Proceedings of the 3rd Workshop on Towards Knowledgeable Foundation Models (KnowFM). 2025. doi:10.18653/v1/2025.knowllm-1.10

    work page doi:10.18653/v1/2025.knowllm-1.10 2025

  43. [43]

    2025 , eprint=

    From Reasoning to Learning: A Survey on Hypothesis Discovery and Rule Learning with Large Language Models , author=. 2025 , eprint=

    2025

  44. [44]

    2025 , eprint=

    IDEA: Enhancing the Rule Learning Ability of Large Language Model Agent through Induction, Deduction, and Abduction , author=. 2025 , eprint=

    2025

  45. [45]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    ACRE: Abstract Causal REasoning Beyond Covariation , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  46. [46]

    2019 , eprint=

    On the Measure of Intelligence , author=. 2019 , eprint=

    2019

  47. [47]

    2025 , eprint=

    Sparks of Science: Hypothesis Generation Using Structured Paper Data , author=. 2025 , eprint=

    2025

  48. [48]

    Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence,

    Toward Reliable Scientific Hypothesis Generation: Evaluating Truthfulness and Hallucination in Large Language Models , author =. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence,. 2025 , month =. doi:10.24963/ijcai.2025/873 , url =

    work page doi:10.24963/ijcai.2025/873 2025

  49. [49]

    2025 , eprint=

    LocationReasoner: Evaluating LLMs on Real-World Site Selection Reasoning , author=. 2025 , eprint=

    2025

  50. [50]

    2025 , eprint=

    Evaluating LLM Metrics Through Real-World Capabilities , author=. 2025 , eprint=

    2025

  51. [51]

    2025 , eprint=

    A Survey on Large Language Model Benchmarks , author=. 2025 , eprint=

    2025

  52. [52]

    2024 , url=

    Carlos E Jimenez and John Yang and Alexander Wettig and Shunyu Yao and Kexin Pei and Ofir Press and Karthik R Narasimhan , booktitle=. 2024 , url=

    2024

  53. [53]

    2021 , eprint=

    Evaluating Large Language Models Trained on Code , author=. 2021 , eprint=

    2021

  54. [54]

    H uman E val Pro and MBPP Pro: Evaluating Large Language Models on Self-invoking Code Generation Task

    Yu, Zhaojian and Zhao, Yilun and Cohan, Arman and Zhang, Xiao-Ping. H uman E val Pro and MBPP Pro: Evaluating Large Language Models on Self-invoking Code Generation Task. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.686

    work page doi:10.18653/v1/2025.findings-acl.686 2025

  55. [55]

    2025 , eprint=

    Evaluating Mathematical Reasoning Across Large Language Models: A Fine-Grained Approach , author=. 2025 , eprint=

    2025

  56. [56]

    1876 , publisher =

    A Classification and Subject Index for Cataloguing and Arranging the Books and Pamphlets of a Library , author =. 1876 , publisher =

  57. [57]

    Julia Wiesinger and Patrick Marlow and Vladimir Vuskovic , title =

  58. [58]

    2025 , eprint=

    DeepSeek-V3 Technical Report , author=. 2025 , eprint=

    2025

  59. [59]

    Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and Wu, Z. F. and Gou, Zhibin and Shao, Zhihong and Li, Zhuoshu and Gao, Ziyi and Liu, Aixin and Xue, Bing and Wang, Bingxuan and Wu, Bochao and Feng, Bei ...

    work page doi:10.1038/s41586-025-09422-z

  60. [60]

    2025 , eprint=

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning , author=. 2025 , eprint=

    2025

  61. [61]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  62. [62]

    Google Deepmind , title =

  63. [63]

    2024 , eprint=

    Gemma 2: Improving Open Language Models at a Practical Size , author=. 2024 , eprint=

    2024

  64. [64]

    2025 , eprint=

    Gemma 3 Technical Report , author=. 2025 , eprint=

    2025

  65. [65]

    CoRR , volume =

    Jiachun Li and Pengfei Cao and Zhuoran Jin and Yubo Chen and Kang Liu and Jun Zhao , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2410.09542 , eprinttype =. 2410.09542 , timestamp =

    work page doi:10.48550/arxiv.2410.09542 2024

  66. [66]

    I nduction B ench: LLM s Fail in the Simplest Complexity Class

    Hua, Wenyue and Wong, Tyler and Sun, Fei and Pan, Liangming and Jardine, Adam and Wang, William Yang. I nduction B ench: LLM s Fail in the Simplest Complexity Class. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.1287

    work page doi:10.18653/v1/2025.acl-long.1287 2025

  67. [67]

    2025 , eprint=

    Language Models Do Not Follow Occam's Razor: A Benchmark for Inductive and Abductive Reasoning , author=. 2025 , eprint=

    2025

  68. [68]

    Hypothesis-Driven Theory-of-Mind Reasoning for Large Language Models , author=

  69. [69]

    2025 , eprint=

    HypoBench: Towards Systematic and Principled Benchmarking for Hypothesis Generation , author=. 2025 , eprint=

    2025

  70. [70]

    2025 , eprint=

    Evaluating the Logical Reasoning Abilities of Large Reasoning Models , author=. 2025 , eprint=

    2025

  71. [71]

    International Conference on Learning Representations , year=

    Abductive Commonsense Reasoning , author=. International Conference on Learning Representations , year=

  72. [72]

    2025 , eprint=

    GEAR: A General Evaluation Framework for Abductive Reasoning , author=. 2025 , eprint=

    2025

  73. [73]

    The Works of Francis Bacon , volume =

    Francis Bacon , editor =. The Works of Francis Bacon , volume =. 1900 , note =

    1900

  74. [74]

    Don`t Just Say I don`t know ! Self-aligning Large Language Models for Responding to Unknown Questions with Explanations

    Deng, Yang and Zhao, Yong and Li, Moxin and Ng, See-Kiong and Chua, Tat-Seng. Don`t Just Say I don`t know ! Self-aligning Large Language Models for Responding to Unknown Questions with Explanations. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.757

    work page doi:10.18653/v1/2024.emnlp-main.757 2024

  75. [75]

    2024 , address =

    Zhang, Hanning and Diao, Shizhe and Lin, Yong and Fung, Yi and Lian, Qing and Wang, Xingyao and Chen, Yangyi and Ji, Heng and Zhang, Tong. R -Tuning: Instructing Large Language Models to Say I Don`t Know'. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume ...

    work page doi:10.18653/v1/2024.naacl-long.394 2024

  76. [76]

    2023 , eprint=

    Learning Deductive Reasoning from Synthetic Corpus based on Formal Logic , author=. 2023 , eprint=

    2023

  77. [77]

    2022 , eprint=

    STaR: Bootstrapping Reasoning With Reasoning , author=. 2022 , eprint=

    2022

  78. [78]

    2023 , eprint=

    Is ChatGPT a General-Purpose Natural Language Processing Task Solver? , author=. 2023 , eprint=

    2023

  79. [79]

    2025 , eprint=

    A Survey on LLM-as-a-Judge , author=. 2025 , eprint=

    2025

  80. [80]

    2023 , eprint=

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena , author=. 2023 , eprint=

    2023

Showing first 80 references.