Pith. sign in

REVIEW 3 major objections 5 minor 20 references

SHARP: Synthesizing High-quality Aligned Reasoning Problems for Large Reasoning Models Reinforcement Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A self-alignment prompt pipeline (SHARP) generated roughly 190,000 verifiable STEM problems that, when used to distill or RL-train 7B models, improved GPQA Diamond accuracy in the reported runs.

desk verdict A plausible and useful synthetic-data pipeline for STEM reasoning, undercut by an unverified 'ground-truth' premise and thin statistics; the chemistry example in Table 4 is genuinely inconsistent. read the letter →

arxiv 2505.14147 v3 pith:5CA5JGEL submitted 2025-05-20 cs.AI

Xiong Jun Wu , Zhenduo Zhang , ZuJie Wen , Zhiqiang Zhang , Wang Ren , Lei Shi , Cai Chen , Deng Zhao
show 6 more authors
Qing Wang Xudong Han Chengfu Tang Dingnan Jin Qing Cui Jun Zhou
This is my paper · ORCID
classification cs.AI
keywords reasoningsharplearningreinforcementverifiablehigh-qualityalignedcomplex
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

SHARP is a recipe for creating hard STEM training questions without human authors. A seed topic list is organized into a three-level hierarchy, for example Chemistry, Organic Chemistry, Elimination reactions. A large reasoning model such as DeepSeek R1 is prompted with a strict template requiring graduate or Olympiad difficulty, step-by-step reasoning, a single unambiguous numeric answer, and no copied text. The generated questions and answers are then filtered with a verifier (Math-Verify) and model self-checks. The surviving samples, about 190,000 after cleaning, are used either to fine-tune a 7B model or as problems in reinforcement learning with a binary correctness reward (GRPO).
Extended reading notes

Core claim

The paper's central claim is that training on SHARP-generated problems improves complex STEM reasoning: the distilled model reaches 54.7 GPQA Diamond versus 46.4 baseline and 49.9 DeepSeek-R1-Distill-Qwen-7B, and the RL Zero model reaches 37.0 versus 35.5 baseline. The abstract states SHARP-augmented training 'substantially outperforms existing methods' and pushes performance 'closer to expert-level proficiency.' If correct, it means a self-aligned synthesis pipeline can produce large volumes of verifiable hard STEM problems that transfer to benchmark reasoning performance.

Load-bearing premise

SHARP trusts model-generated reference answers as ground truth. The Verifying stage combines a model self-check and a rule-based verifier (Math-Verify) but includes no human expert validation; Appendix B.2 says the answer is 'similar to the ground truth solution a_i prime' where the ground truth is itself generated by the prompting pipeline. If a meaningful fraction of these reference answers are wrong, RL rewards and distillation labels will reinforce incorrect reasoning. This is structurally distinct from the claim, because the claim depends on the correctness of the training signal.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces SHARP, a three-phase pipeline (Alignment, Instantiation, Inference) plus a Verifying stage for synthesizing graduate- and Olympiad-level STEM reasoning problems with a state-of-the-art LRM (DeepSeek R1). The pipeline is governed by explicit self-alignment principles, uses a three-tier subject taxonomy seeded by a topic library that is partly derived from GPQA knowledge-point distributions, and filters outputs through a model self-check and Math-Verify. The resulting problems are used for SFT distillation and for GRPO-based RLVR training. Experiments report GPQA Diamond accuracy of 54.7 for a SHARP-distilled Qwen2.5-7B model versus 46.4 for the baseline and 49.9 for DeepSeek-R1-Distill-Qwen-7B, and 37.0 versus 35.5 for RL-Zero training. The paper claims that SHARP-augmented training substantially improves complex STEM reasoning.

Significance. If the claims hold, SHARP would provide a scalable route to generating verifiable hard STEM training data with reduced human annotation, which is a valuable contribution. The paper is unusually detailed about its pipeline, prompts, hyperparameters, and dataset-distribution analyses, and the distillation gains (+8.3 points on GPQA Diamond) are substantial and consistent across subjects. However, two load-bearing issues temper the significance. First, the training-data topic distribution is deliberately aligned to the evaluation benchmark, so the GPQA results are partly a benchmark-loop artifact. Second, the 'verified ground truth' is model-generated and filtered without human validation; the chemistry example in Table 4 is internally inconsistent, so the verification stage demonstrably passes domain-level errors. The RL-Zero gain of 1.5 points is also within plausible sampling noise. These issues must be addressed before the general claim of expert-level STEM reasoning gains is established.

major comments (3)
  1. [§3.2, §C.1, Tables 1–2] The seed-topic library is constructed in part by extracting topic keypoints from GPQA, and §C.1 states that the authors 'sampled according to the distribution of disciplines and the corresponding knowledge points' so that the synthetic data 'fully covers the relevant knowledge points of GPQA.' Evaluation is then reported on GPQA. The authors state that they do not directly rephrase validation queries, and this is a meaningful distinction, but aligning the training-data topic distribution to the test benchmark is still a benchmark-loop burden: the reported gains may reflect distribution matching rather than improved general reasoning. To support the paper's central claim, the authors should evaluate on held-out STEM benchmarks (e.g., MMLU-STEM, SciBench, or other graduate-level STEM sets) or, at minimum, ablate the evaluation with topic sampling that is independent of GPQA. The paper's decontamination step is described as exact-duplicate removal; it does not address this topical overlap.
  2. [§3.3, Appendix B.2, Table 4, Algorithm 1] The reference answers used for both distillation labels and RLVR rewards are generated by the same LRM pipeline and filtered only by a model self-check and Math-Verify. Appendix B.2 defines R_acc(a_i, a'_i) = 1 iff the model's answer 'is similar to the ground truth solution a'_i,' but a'_i is itself the pipeline's output. No human expert validation is reported. The chemistry example in Table 4 shows that this verification is not airtight: a mass loss of 2.64 g of CO2 implies 0.0600 mol of ZnCO3 and therefore 3.92 g of zinc, whereas the problem states that 5.89 g of zinc is produced, which requires 0.0901 mol of ZnO and a corresponding CO2 loss of 3.96 g. The stated reference answer of 58.9% is not derivable from the data as written without an unstated additional assumption. This example passed the SHARP verification pipeline, demonstrating that domain-level correctness is not reliably enforced. Since incorrect reference answers directly corrupt the reward signal and the distillation supervision, the paper should quantify reference-answer accuracy (e.g., a human expert audit of a random sample, or agreement with an independent solver) and either strengthen the verification or soften the 'verified ground truth' language.
  3. [Table 2, Checklist item 7] The RL-Zero result in Table 2 reports a GPQA Diamond gain of 1.5 points (37.0 vs. 35.5). GPQA Diamond contains roughly 198 questions, so this difference corresponds to about 3 questions and lies within one binomial standard error (approximately 3.4 points). The paper's own NeurIPS checklist item 7 answers 'No' to error bars or statistical significance tests. Appendix B.2 states that accuracy is averaged over '16 independent inference runs,' but no variance or confidence interval is reported, and it is unclear what varies between runs if decoding is greedy. The abstract's claim that SHARP-augmented training 'substantially outperforms existing methods' is not supported by the RL-Zero evidence alone. Please report confidence intervals, the number of evaluation questions, and the run-to-run variance; if the 16 runs are independent samples, report the standard error across them.
minor comments (5)
  1. [Table 4] The first example in Table 4 is labeled 'Particle Physics, High Energy Physics' but the problem text opens with 'Solve the following chemical problem'; this subject-label/content mismatch should be corrected.
  2. [Algorithm 1, §3.3 (Verifying)] The Verifying stage is described only as checking that 'all alignment checks' pass; the concrete checks performed by the model self-check and by Math-Verify, and any thresholds used (e.g., confidence score, pass-rate filter), are not specified. Please provide these details, including the pass-rate filtering threshold used to reduce the 229,452 generated pairs to the reported 190,000.
  3. [Appendix B.2] The reward function R_acc uses 'equal(a_i, a'_i)' but the equality relation is not defined; please specify whether it is exact string matching, numeric tolerance, or Math-Verify equivalence.
  4. [Appendix C.3] The pass-rate comparisons use different numbers of responses for the two models (10 vs. 5 for physics, 5 vs. 1 for biology, 10 vs. 5 for chemistry in the text and figures), which makes the cross-model comparisons difficult to interpret; please standardize the evaluation protocol or justify the differences.
  5. [Section 4, Appendix C.2] The dataset size is reported as 190,000 in Section 4 and as 229,452 generated in Appendix C; while the appendix clarifies that about 190,000 remain after filtering, the main text should state both the generated and post-filter counts for consistency. Also, the phrase 'disinfection' in Appendix C.2 should be 'detoxification' (as used elsewhere) or be defined.
Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a data quality assumption (generated ground truth is correct) and a benchmark alignment choice (seed topics follow GPQA's knowledge distribution). No mathematical derivation is involved; the main free parameters are data-generation design choices. The absence of human validation makes the correctness of the training signal the largest unverified input.

free parameters (3)
  • GPQA knowledge-point sampling distribution = aligned to GPQA benchmark categories
    Appendix C.1 says data is sampled according to the distribution of disciplines and knowledge points of GPQA, so the training distribution is shaped by the evaluation benchmark.
  • Pass-rate filtering threshold = not disclosed
    Appendix C.2 mentions 'pass ratio filtering' but does not give the cutoff used to retain questions; this affects dataset composition.
  • K in K-means clustering = about 1,000
    Appendix C says K is specified via elbow method on BGE-m3 embeddings; this is a design choice affecting topic sampling balance.
assumptions (4)
  • domain assumption Math-Verify correctly determines whether a generated answer matches the reference answer
    The whole verification stage relies on this; Appendix B.2 defines reward as equality or similarity between model response and generated ground truth.
  • domain assumption DeepSeek R1 generated reference answers are scientifically correct
    No human expert validation is reported; Section 3.3 says ground truth is obtained from generated candidate answers and template checks.
  • domain assumption The GPQA knowledge-point distribution can be used to seed training data without leaking GPQA problems
    The paper states only topic keypoints are extracted, not rephrased problems; no contamination test is provided to support this.
  • domain assumption Training on problems from the seed topics transfers to GPQA performance
    The evaluation assumes improvement on GPQA reflects general STEM reasoning; no held-out reasoning benchmark is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SHARP: Synthesizing High-quality Aligned Reasoning Problems for Large Reasoning Models Reinforcement Learning." pith.science (2026). https://pith.science/paper/5CA5JGEL

@misc{pith2026250514147,
  author       = {Pith},
  title        = {Pith review of: SHARP: Synthesizing High-quality Aligned Reasoning Problems for Large Reasoning Models Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5CA5JGEL}},
  note         = {Machine review of arXiv:2505.14147}
}
read the original abstract

Training large reasoning models (LRMs) with reinforcement learning in STEM domains is hindered by the scarcity of high-quality, diverse, and verifiable problem sets. Existing synthesis methods, such as Chain-of-Thought prompting, often generate oversimplified or uncheckable data, limiting model advancement on complex tasks. To address these challenges, we introduce SHARP, a unified approach to Synthesizing High-quality Aligned Reasoning Problems for LRMs reinforcement learning with verifiable rewards (RLVR). SHARP encompasses a strategic set of self-alignment principles -- targeting graduate and Olympiad-level difficulty, rigorous logical consistency, and unambiguous, verifiable answers -- and a structured three-phase framework (Alignment, Instantiation, Inference) that ensures thematic diversity and fine-grained control over problem generation. We implement SHARP by leveraging a state-of-the-art LRM to infer and verify challenging STEM questions, then employ a reinforcement learning loop to refine the model's reasoning through verifiable reward signals. Experiments on benchmarks such as GPQA demonstrate that SHARP-augmented training substantially outperforms existing methods, markedly improving complex reasoning accuracy and pushing LRM performance closer to expert-level proficiency. Our contributions include the SHARP strategy, framework design, end-to-end implementation, and experimental evaluation of its effectiveness in elevating LRM reasoning capabilities.

Figures

Figures reproduced from arXiv: 2505.14147 by the authors.

Figure 1
Figure 1. The SHARP Approach Algorithm 1 SHARP Self-Alignment Problem Synthesis Strategy Require: Seed topic set S = {s1, s2, . . . , sN }; alignment strategy constraints {xv}; base LRM model; reasoning spec Rspec; verifier V Ensure: Verified aligned question-answer pairs Q = {(q1, a1), . . . ,(qm, am)} 1: Initialize Q ← ∅ 2: for each seed topic si ∈ S do 3: ▷ Alignment Phase 4: Configure alignment constraints: Alignment cons… view at source ↗
Figure 2
Figure 2. The SHARP Framework coherent, complex reasoning problems that are carefully aligned with the nuanced demands of STEM disciplines. The primary objective of this framework is to generate high-quality, diverse training samples that drive the optimization of reinforcement learning (RL) models, especially in the context of high-difficulty STEM benchmarks. The SHARP framework is underpinned by its core modules, prominentl… view at source ↗
Figure 3
Figure 3. The SHARP Implementation for Large Reasoning Models Reinforcement Learning instantiations is introduced, distinct from the relatively free “Reasoning” step in traditional CoT. This instantiation phase integrates a “Three-Tier Category” knowledge framework to instantiate the strategy to different subjects’ characteristics and structures. The “Three-Tier Category” knowledge framework manages and organizes STEM knowled… view at source ↗
Figures from the paper (24 more)
Figure 4
Figure 4. Figure 4: GPQA score improvement of single STEM disciplines (physics, chemistry, and biology) of [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: GPQA score improvement of single STEM disciplines (physics, chemistry, and biology) of [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: GPQA score improvement of single STEM disciplines (physics, chemistry, and biology) of [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Open-Reasoner-Zero-7B performance in ablation of STEM data generated by the SHARP approach. (The x-axis represents the different running steps during the training of the reinforcement learning reasoning model, and the y-axis represents the GPQA score evaluation results…
Figure 8
Figure 8. Figure 8: The response length of RL Zero model in ablation of chemistry data from three different [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: The reward RL Zero model in ablation of chemistry data from three different problems [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: The overall dataflow process of the SHARP approach. the data flow and the difficulty of the STEM challenging problems generated by the SHARP method based on this category. The overall dataflow diagram of the construction of the Seed Topics Library and “Three-Tier Cate…
Figure 11
Figure 11. Figure 11: The K-means clustering results based on question embedding features extracted using [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: The overall subjects distribution of 229,452 question-answering problems generated by [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: The physics subject distribution of basic knowledge points word cloud of GPQA bench [PITH_FULL_IMAGE:figures/full_fig_p027_13.png]
Figure 14
Figure 14. Figure 14: The top 30 basic knowledge points of the physics subject of the GPQA benchmark. [PITH_FULL_IMAGE:figures/full_fig_p028_14.png]
Figure 15
Figure 15. Figure 15: The chemistry subject distribution of basic knowledge points word cloud of GPQA [PITH_FULL_IMAGE:figures/full_fig_p029_15.png]
Figure 16
Figure 16. Figure 16: The top 30 basic knowledge points of the chemistry subject of the GPQA benchmark. [PITH_FULL_IMAGE:figures/full_fig_p030_16.png]
Figure 17
Figure 17. Figure 17: The biology subject distribution of basic knowledge points word cloud of GPQA bench [PITH_FULL_IMAGE:figures/full_fig_p031_17.png]
Figure 18
Figure 18. Figure 18: The top 30 basic knowledge points of the biology subject of the GPQA benchmark. [PITH_FULL_IMAGE:figures/full_fig_p032_18.png]
Figure 19
Figure 19. Figure 19: The “Three-Tier Category” category distribution of physics subject for problems generated [PITH_FULL_IMAGE:figures/full_fig_p033_19.png]
Figure 20
Figure 20. Figure 20: The “Three-Tier Category” category distribution of biology subject for problems generated [PITH_FULL_IMAGE:figures/full_fig_p034_20.png]
Figure 21
Figure 21. Figure 21: The “Three-Tier Category” category distribution of chemistry subject for problems [PITH_FULL_IMAGE:figures/full_fig_p035_21.png]
Figure 22
Figure 22. Figure 22: The passrate distribution of the physics problems from open-source , the traditional CoT [PITH_FULL_IMAGE:figures/full_fig_p036_22.png]
Figure 23
Figure 23. Figure 23: The passrate distribution of the chemistry problems from open-source , the traditional [PITH_FULL_IMAGE:figures/full_fig_p036_23.png]
Figure 24
Figure 24. Figure 24: The passrate distribution of the biology problems from open-source , the traditional CoT [PITH_FULL_IMAGE:figures/full_fig_p037_24.png]
Figure 25
Figure 25. Figure 25: The passrate distribution of physics problems generated by the [PITH_FULL_IMAGE:figures/full_fig_p037_25.png]
Figure 26
Figure 26. Figure 26: The passrate distribution of biology problems generated by the [PITH_FULL_IMAGE:figures/full_fig_p038_26.png]
Figure 27
Figure 27. Figure 27: The passrate distribution of chemistry problems generated by the [PITH_FULL_IMAGE:figures/full_fig_p038_27.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

  1. [1]

    problem":

    problems synthesized through traditional COT, 2) problems augmented synthesized referencing to real challenging chemistry exercises and 3) problems synthesized SHARP approach. Through the experimental comparison of each stage, the difficulty of the sample problems generated by our SHARP approach has significantly increased the response length for the corr...

    work page 2024

  2. [2]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: The paper identifies several limitations in the end. Future work could explore applying this approach to other domains and more complex reasoning tasks, and further optimizing the SHARP approach on various larger-scale RL reasoni...

    work page

  3. [3]

    It does not appear to present new theoretical results in the form of theorems or mathematical proofs that would require a separate section for assumptions and proofs

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [NA] Justification: The paper introduces a new approach (SHARP) and a framework, supported by experimental results. It does not appear to present new theoretical results in the form of theorems...

    work page

  4. [4]

    Three-Tier Category

    and subject characteristics to ensure that the first-level sub-disciplines, second-level self- disciplines, and basic concepts of each discipline are covered. Then we built high-quality seed 22 Figure 11: The K-means clustering results based on question embedding features extracted using BGE-m3 from problems generated by the SHARP approach. documents to s...

    work page 2024

  5. [5]

    It claimsSHARP encompasses self-alignment principles and a three-phase framework

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The abstract states that this paper introduces SHARP, a unified approach for synthesizing high-quality reasoning problems for LRMs reinforcement learning with verifiable rewards (RLVR). It claimsSHA...

    work page

  6. [6]

    It names the models used for comparison

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: The paper specifies the training data used (baseline CoT samples and 190,000 SHARP-generated samples). It nam...

    work page

  7. [7]

    If needed, we will include them in the camera-ready version

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [No] Justification: The tables presenting the main experimental results (Table 1 and Table 2) show performance scores (e.g., GPQA Diamond scores, score...

    work page

  8. [8]

    It specifies the comparison models used for distillation and RL Zero training

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: The paper describes the tr...

    work page

Show all 20 references
  1. [9]

    Moreover, we will open-source all necessary codes and related data for industry use during the review period

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: Essential SHARP strategy pr...

  2. [10]

    It discusses the potential to push LRM performance closer to expert-level proficiency and superintelligence in STEM

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [No] Justification: The paper focuses on the technical contributions of the SHARP approach in enhancing LRM reasoning capabili...

  3. [11]

    The generated data consists of STEM problems

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? 43 Answer: [NA] Justification: The paper fo...

  4. [12]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The computational resou...

  5. [13]

    It does not involve human subjects or obviously ethically sensitive applications, and we assume it conforms to the NeurIPS Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: The research involves algorithmic development and evaluation on standard optimi...

  6. [14]

    The process involves using LLMs to generate and verify problems, and then training other LLMs

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  7. [15]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  8. [16]

    However, the specific licenses and terms of use for these assets are not explicitly mentioned in the paper text or the appendix

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [No] Justification: The paper properl...

  9. [17]

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: The main new asset introduced is the SHARP methodology and the dataset of 190,000 STEM problems generated using this...

  10. [20]

    We implement SHARP by leveraging a state-of-the-art LRM to infer and verify challenging STEM questions

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

    2025

  11. [757]

    calculate total moles of all possible products

    URL https://aclanthology.org/2024.emnlp-main.757/. Guanting Dong, Keming Lu, Chengpeng Li, Tingyu Xia, Bowen Yu, Chang Zhou, and Jingren Zhou. Self-play with execution feedback: Improving instruction-following capabilities of large language models. In The Thirteenth Internatio...

    2024 arXiv

  12. [2025]

    I don‘t know

    URL https://arxiv.org/abs/2501.12948. Yang Deng, Yong Zhao, Moxin Li, See-Kiong Ng, and Tat-Seng Chua. Don‘t just say “I don‘t know”! self-aligning large language models for responding to unknown questions with explanations. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.