Pith. sign in

REVIEW 4 major objections 5 minor 23 references

Beyond Templates: Dynamic Adaptation of Reasoning Demonstrations via Feasibility-Aware Exploration

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Small models beat full imitation by adapting expert reasoning paths

desk verdict A genuinely useful data-adaptation idea with a real novelty in step-level feasibility scoring, but the experiments as reported don't yet establish the data-efficiency claim — still worth refereeing. read the letter →

arxiv 2505.20700 v1 pith:QNBH2RNT submitted 2025-05-27 cs.CL

Yong Wu , Weihang Pan , Ke Li , Chen Binhui , Ping Li , Binbin Lin This is my paper · ORCID
classification cs.CL
keywords dynamicdataadaptationsmalllanguagemodelsmathematicalreasoningselectiveimitationgapoutcome-consistentexplorationdistillationadaptabilityestimation
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

Supervised fine-tuning on expert reasoning demonstrations often makes small language models worse, because many expert steps lie beyond what a small model can absorb. This paper claims that the fix is to measure, step by step, whether the student can realistically continue to the correct answer while imitating a given expert step, and to stop imitating at the first steep drop in that score. Beyond that 'imitation gap,' the student explores its own continuations and keeps only those that reach the expert's final answer. The resulting DART framework outperforms static fine-tuning on seven math benchmarks at 0.5B, 1.5B, and 3B scale, using fewer training examples. If the claim holds, it gives a model-aware alternative to static expert-trajectory distillation for reasoning alignment in resource-constrained models.

What carries the argument

The central object is the step-wise adaptability score $f_t = Q(s_{<t}, s_t)$, the fraction of $N_{sim}$ simulated continuations, sampled from the student conditioned on the expert prefix plus candidate step, that reach the expert's final answer. A drop of more than $\epsilon$ below the running maximum marks the 'imitation gap' and the truncation point; past that point the student explores with temperature 0.7 and retains trajectories satisfying the outcome-consistency check $O(\tau_{adapt}) = O(\tau_{expert})$, trained by the cross-entropy distillation loss of Eq. (6). The machinery turns a capacity mismatch into a per-example data-adaptation decision: imitate where adaptable, explore where not.

What would settle it

Train the student on expert trajectories with each expert step individually omitted (leave-one-step-out fine-tuning), then check whether the steps DART flags as low-adaptability are exactly the ones whose removal improves downstream accuracy; if the omitted-step ablation does not correlate with the adaptability scores, the imitation-gap mechanism collapses.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that outcome-consistent autonomous exploration, triggered by a per-step adaptability estimate, turns expert reasoning data into supervision that small models can actually use. The adaptability of an expert step is measured by Monte Carlo rollouts: given the expert prefix including that step, the student samples multiple continuations and the score is the fraction that land on the expert's final answer. Early expert steps typically raise this score, but it collapses mid-trajectory — the 'imitation gap' — and static fine-tuning that forces imitation through this region degrades accuracy. DART truncates the expert path at the gap, has the student sample continuations at higher temperature, and distills only trajectories whose final answer matches the expert's. The empirical pattern across 0.5B–3B Qwen models, on LIMO and Math-QwQ-32B data, is that this adaptation beats both no tuning and full imitation on the average of GSM8K, MATH, Minerva Math, GaoKao 2023 En, OlympiadBench, College Math, and MMLU-STEM.

Load-bearing premise

The adaptability score computed from the pre-fine-tuning student, with four rollouts at low temperature and exact final-answer matching, truly predicts which expert steps will hurt after fine-tuning.

Editorial extensions

If this is right

  • Static fine-tuning can be actively harmful for small models: on Math-QwQ-32B it underperforms no tuning by 8.7, 10.7, and 6.7 points at 0.5B, 1.5B, and 3B scales, while DART's Adaptation-Full exceeds no-tuning accuracy with gains of +8.5, +6.7, and +7.8 over static.
  • Data efficiency follows from the same mechanism: on LIMO, Adaptation-Full uses only 202 (0.5B), 546 (1.5B), and 621 (3B) of the 817 examples, yet achieves the best average accuracy among the three strategies.
  • The imitation gap is a predictable, model-dependent phenomenon: adaptability rises early, collapses mid-trajectory, and recovers near the final answer across all three scales and both datasets.
  • Truncating at the gap (Adaptation-Gap) beats stopping at the first feasible state (Adaptation-First) by 5.2 and 12.1 points on Math-QwQ-32B at 1.5B and 3B, showing that detecting the adaptability decline, not just any feasible prefix, is what carries the benefit.
  • Exploration is necessary, not decorative: without the autonomous search (Adaptation-Gap), gains shrink substantially; with it, the model learns goal-directed language, with fewer 'but' and 'wait' tokens and more 'find' and 'need' tokens.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the adaptability score is a stable predictor of post-fine-tuning usefulness, DART's per-step scores could be reused as a data-selection filter for any expert dataset, without re-running fine-tuning; one test would be to sort examples by their average adaptability and train on increasing quantiles.
  • The outcome-consistency constraint generalizes naturally to code generation (unit tests as outcome) and structured tasks with verifiable outputs; the paper's own limitation section stops short of open-ended generation, where a learned verifier would be needed.
  • The observed keyword shifts ('but'/'wait' down, 'find'/'need' up) suggest adapted trajectories are more decisive; a direct ablation could test whether suppressing exploratory tokens in the expert trajectories, without exploration, reproduces the gains, isolating the supervision content from the exploration mechanism.
  • One could view the imitation gap as an automatic curriculum signal: steps below the adaptability threshold are precisely the subskills the model has not mastered, so the truncation point doubles as a per-example difficulty estimate for ordering training data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DART, a framework for adapting expert reasoning demonstrations to small student models. DART computes a step-wise adaptability score f_t by simulating N_sim rollouts from each expert prefix and checking whether the rollout reaches the ground-truth answer. It then truncates the expert trajectory at the first step where f_t falls more than epsilon below its running peak, generates a continuation from the student's own policy, and keeps the continuation only if its final answer matches the expert's. Training uses cross-entropy on the resulting adapted trajectories. The authors evaluate on LIMO and Math-QwQ-32B with Qwen2.5-Instruct 0.5B/1.5B/3B across seven math/STEM benchmarks, reporting average gains over static fine-tuning and claiming improved data efficiency.

Significance. If the claims are established, DART is a useful step toward model-aware data selection for reasoning: it replaces uniform imitation with a feasibility check and lets the student generate outcome-consistent alternatives, which is a sensible way to mitigate distributional mismatch between expert demonstrations and small models. The paper's empirical pattern of non-monotonic adaptability (Figure 2) is interesting and testable, and the ablation of truncation strategies (Table 2) addresses search restriction; the repetition-cleaning analysis in Appendix B is a useful sanity check. The method is also presented with a clear algorithm and an explicit limitation to verifiable tasks. However, the central data-efficiency and mechanism claims currently rest on confounded comparisons, an unresolved internal inconsistency, and an unassigned threshold, so the empirical contribution is not yet established at the level claimed.

major comments (4)
  1. [§4.2, Table 1] The central claim that DART 'significantly improves generalization and data efficiency over static fine-tuning' is not isolated: Adaptation-Full is trained on a subset (202/546/621 of the 817 LIMO examples and 1829/3922/4100 of the 5383 Math-QwQ examples), while Static is trained on the full set. Accuracy could improve simply because the retained examples are easier or less noisy, not because of feasibility-aware adaptation. To support the data-efficiency claim, the paper needs a same-size Static baseline (e.g., random subsets of equal size, or a difficulty-matched selection), and ideally variance across seeds, given that exploration uses stochastic sampling. Without these controls, the comparison conflates data selection with the adaptation mechanism.
  2. [§5.2/Table 2 vs §4.1/Table 1] The same condition, Adaptation-Full with Qwen2.5-1.5B on Math-QwQ-32B, is reported with average accuracy 44.6 in Table 1 (GSM8K 74.2, MATH 55.1, MMLU-STEM 57.7) but 40.4 in Table 2 (GSM8K 72.1, MATH 49.9, MMLU-STEM 44.1). The 3B row is identical in both tables, so this is not a different experiment configuration. Because Table 2 is the evidence that Adaptation-Gap and Adaptation-First underperform Adaptation-Full, the discrepancy changes the quantitative conclusions. Please reconcile the two tables or explain what differs between the runs.
  3. [Algorithm 1, §3.2] The imitation-gap detection in Algorithm 1 requires the threshold epsilon (line 5), but epsilon is never given a numerical value in the main text or Appendix A, and no sensitivity analysis is reported. The definition of the 'local maximum' t_peak and the behavior when no step exceeds the threshold are also unspecified. Since the entire data-adaptation procedure is defined by the truncation point, these omissions make the method non-reproducible. Please report the exact epsilon values used for each table, the smoothing/peak-finding procedure behind Figure 2, and the fallback behavior when no imitation gap is found.
  4. [§5.2, Table 2; Eq. (3)] The paper's mechanistic claim is that step-level adaptability f_t identifies expert steps that are harmful to imitate, but the experiments do not separate the feasibility signal from the generic exploration-and-filtering machinery. Adaptation-Gap and Adaptation-First differ only in where they stop imitating; both use the same autonomous rollouts and outcome filtering. Add a STaR-style control (sample k completions from the student without conditioning on the expert prefix, keep answer-correct ones, and fine-tune on them) and, if possible, a per-step analysis comparing f_t with post-training outcomes. These controls are necessary to support the conclusion that feasibility-aware selection, rather than self-training with answer filtering, drives the gains.
minor comments (5)
  1. [Abstract/§3.1] There are repeated spacing errors ('as aselective', 'animitation gap'); the manuscript needs a copyedit pass.
  2. [Appendix A.2] The text refers to 'Algorithm 3.1', but the paper contains only Algorithm 1; correct the cross-reference.
  3. [Figure 2/Section 5.1] The caption and text do not describe how the curves are smoothed or how standard error is computed; add this information so the reader can interpret the non-monotonic pattern.
  4. [Table 3] The table is described as top-20 keyword shifts, but for 1.5B and 3B many rows contain dashes; clarify whether the keyword set is the union across models and what the dashes mean.
  5. [§3.2] The citation to Lightman et al. and Zhang et al. contains a malformed comma ('Zhang et al. [2025],)'); fix the citation formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the evaluation is held-out and the filtering is outcome-based, so the central empirical claim does not reduce to its inputs by construction.

full rationale

The paper's derivation chain is not circular. The core mechanism estimates step adaptability f_t = Q(s_<t, s_t) by rolling out the pre-finetuning student and comparing final answers against the ground truth; the imitation gap is then read off these empirical scores. The adapted trajectories are generated by the same student and filtered by exact outcome match before being distilled with cross-entropy. This is self-training/self-distillation rather than circular derivation: no parameter is fitted to the evaluation benchmarks, and the evaluation metrics (GSM8K, MATH, Minerva, GaoKao, OlympiadBench, College Math, MMLU-STEM) are external held-out sets. The pre-training adaptability signal's predictive validity is a potential correctness risk because no STaR-style same-size baseline or same-size Static fine-tuning baseline is reported, and the duplicate Adaptation-Full rows in Table 1 and Table 2 disagree (44.6 vs 40.4 avg for Math-QwQ-32B at 1.5B), but these are empirical-validity and reproducibility concerns, not instances where a prediction equals an input by construction. Section 8's limitation to structured reasoning tasks with verifiable outcomes is an honest scope restriction and does not hide a circular step. Therefore no step meets the standard of exhibiting Eq. X = Eq. Y by construction, a fitted parameter renamed as a prediction, or a load-bearing self-citation chain.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

DART is an empirical data-processing method, not a first-principles derivation. It relies on assumptions about answer verification, the predictive validity of pre-training rollout scores for post-training benefit, and the existence of a detectable imitation gap in every trajectory. No new physical or theoretical entities are introduced.

free parameters (5)
  • epsilon (imitation gap threshold) = not reported
    Defines the drop in adaptability score that triggers truncation in Algorithm 1. Its value is never given, so the central mechanism is not reproducible.
  • N_sim (rollouts per step) = 4
    Number of Monte Carlo rollouts per candidate step in Equation (3), chosen by hand in Appendix A.2 with no sensitivity analysis.
  • Simulation temperature = 0.1
    Sampling temperature used for adaptability estimation; affects the score and is chosen in Appendix A.2.
  • NUM_SAMPLES (exploration candidates) = 8
    Beam size for adaptive path exploration in Appendix A.3; chosen by hand and not varied.
  • Exploration temperature = 0.7
    Sampling temperature for the autonomous exploration phase in Appendix A.3; affects path diversity and downstream training data.
assumptions (4)
  • domain assumption Ground-truth final answer a* exists and exact-match comparison O(τ_adapt)=O(τ_expert) is a sufficient correctness criterion for retaining adapted trajectories.
    Used in Equation (3) and Equation (5). It is plausible for boxed math answers but brittle to formatting variants and, as the paper's limitations section notes, fails for open-ended outputs.
  • domain assumption The pre-training rollout success rate predicts post-training learning utility of an expert step.
    The imitation-gap truncation in Algorithm 1 is based on f_t measured with the untrained student. No experiment validates the transfer of this score to the fine-tuned model.
  • domain assumption Every expert trajectory contains a detectable local maximum and a later significant drop in adaptability, the imitation gap.
    Algorithm 1 assumes t_peak and t_gap exist. Figure 2 shows smoothed averages, but the paper does not report the fraction of individual trajectories where such a gap is detected.
  • domain assumption Expert reasoning can be segmented into discrete steps and conditioned prefixes s_<t.
    The method treats steps as atomic units; the paper does not specify how steps are tokenized or sentence-segmented for the rollouts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Templates: Dynamic Adaptation of Reasoning Demonstrations via Feasibility-Aware Exploration." pith.science (2026). https://pith.science/paper/QNBH2RNT

@misc{pith2026250520700,
  author       = {Pith},
  title        = {Pith review of: Beyond Templates: Dynamic Adaptation of Reasoning Demonstrations via Feasibility-Aware Exploration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QNBH2RNT}},
  note         = {Machine review of arXiv:2505.20700}
}
read the original abstract

Large language models (LLMs) have shown remarkable reasoning capabilities, yet aligning such abilities to small language models (SLMs) remains a challenge due to distributional mismatches and limited model capacity. Existing reasoning datasets, typically designed for powerful LLMs, often lead to degraded performance when directly applied to weaker models. In this work, we introduce Dynamic Adaptation of Reasoning Trajectories (DART), a novel data adaptation framework that bridges the capability gap between expert reasoning trajectories and diverse SLMs. Instead of uniformly imitating expert steps, DART employs a selective imitation strategy guided by step-wise adaptability estimation via solution simulation. When expert steps surpass the student's capacity -- signaled by an Imitation Gap -- the student autonomously explores alternative reasoning paths, constrained by outcome consistency. We validate DART across multiple reasoning benchmarks and model scales, demonstrating that it significantly improves generalization and data efficiency over static fine-tuning. Our method enhances supervision quality by aligning training signals with the student's reasoning capabilities, offering a scalable solution for reasoning alignment in resource-constrained models.

Figures

Figures reproduced from arXiv: 2505.20700 by the authors.

Figure 1
Figure 1. Overview of the DART framework. where Nsim denotes the total number of rollouts performed for each candidate step st, with each rollout simulating a complete reasoning trajectory conditioned on the prefix s<t and the adoption of step st. Empirically observed patterns (see Section 5.1) suggest that adaptability tends to rise in the early stages of expert trajectories, but drops sharply beyond a certain point. This no… view at source ↗
Figure 3
Figure 3. Repetition ratios(%) in search paths across different model sizes and datasets. Smaller [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 7 canonical work pages

  1. [5]

    Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. In Jian Su, Kevin Duh, and Xavier Carreras, editors,Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1317–1327, Austin, Texas, November

    work page 2016

  2. [9]

    MARIO: math reasoning with code interpreter output - a reproducible pipeline

    Minpeng Liao, Chengxi Li, Wei Luo, Jing Wu, and Kai Fan. MARIO: math reasoning with code interpreter output - a reproducible pipeline. In Lun-Wei Ku, André Martins, and Vivek Srikumar, editors,Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thai- land and virtual meeting, August 11-16, 2024, pages 905–924. Association for Com...

    work page 2024

  3. [10]

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto

    URLhttps://arxiv.org/abs/2503.16385. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling,

    arXiv

  4. [11]

    URLhttps://arxiv.org/abs/2501.19393. OpenAI. Openai o1 system card,

    arXiv

  5. [12]

    URLhttps://arxiv.org/abs/2412.16720. Dean A. Pomerleau. Efficient training of artificial neural networks for autonomous navigation.Neural Computation, 3(1):88–97,

    arXiv

  6. [14]

    David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al

    URLhttps://arxiv.org/abs/2402.03300. David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search.nature, 529(7587):484–489,

    arXiv

  7. [15]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou

    URLhttps://arxiv.org/abs/2403.02884. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume...

    arXiv

  8. [16]

    11 Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, and Radha Poovendran

    URL https://proceedings.neurips.cc/paper_files/paper/2022/file/ 9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf. 11 Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, and Radha Poovendran. Stronger models are not stronger teachers for instruction tuning,

    work page 2022

Show all 23 references
  1. [17]

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu

    URL https://arxiv.org/abs/ 2411.07133. Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning,

    arXiv

  2. [18]

    Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue

    URLhttps://arxiv.org/abs/2502.03387. Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. Demystifying long chain-of-thought reasoning in llms,

    arXiv

  3. [19]

    Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T

    URLhttps://arxiv.org/abs/2502.03373. Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T. Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models,

    arXiv

  4. [20]

    Xiang Yue, Tuney Zheng, Ge Zhang, and Wenhu Chen

    URLhttps://arxiv.org/abs/2309.12284. Xiang Yue, Tuney Zheng, Ge Zhang, and Wenhu Chen. Mammoth2: Scaling instructions from the web,

    arXiv

  5. [21]

    Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin

    URLhttps://arxiv.org/abs/2405.03548. Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning,

    arXiv

  6. [22]

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al

    URLhttps://arxiv.org/abs/2501.07301. Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. Lima: Less is more for alignment.Advances in Neural Information Processing Systems, 36,

    arXiv

  7. [23]

    The primary simulation prompt used in our study is defined as follows: Simulation Prompt Question:[Question] Rationale so far:s 1, s2,

    12 Appendix A Experimental Details A.1 Experiment Prompts In our simulation experiments, we employed a structured prompting approach to guide the language model through multi-step reasoning tasks. The primary simulation prompt used in our study is defined as follows: Simulatio...

    2000

  8. [1991]

    Stephane Ross, Geoffrey Gordon, and Drew Bagnell

    doi: 10.1162/neco.1991.3.1.88. Stephane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and struc- tured prediction to no-regret online learning. In Geoffrey Gordon, David Dunson, and Miroslav Dudík, editors,Proceedings of the Fourteenth Internationa...

    1991 doi

  9. [2006]

    Solving quantitative reasoning prob- lems with language models

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, 10 Behnam Neyshabur, Guy Gur-Ari, and Vedant Misra. Solving quantitative reasoning prob- lems with language mod...

    2022

  10. [2015]

    Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang

    URLhttps://arxiv.org/abs/1503.02531. Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning.arXiv preprint arXiv:2504.01296,

    arXiv

  11. [2016]

    doi: 10.18653/v1/D16-1139

    Association for Computational Linguistics. doi: 10.18653/v1/D16-1139. URL https://aclanthology.org/ D16-1139/. Levente Kocsis and Csaba Szepesvári. Bandit based monte-carlo planning. InEuropean conference on machine learning, pages 282–293. Springer,

    doi

  12. [2021]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z

    CoRR, abs/2110.14168. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, et al. Deepseek-r1: Incentivizing reasoning capability in llms via...

    arXiv

  13. [2022]

    Jia Li, Edward Beeching, Lewis Tunstall, Ben Lipkin, Roman Soletskyi, Shengyi Huang, Kashif Rasul, Longhui Yu, Albert Q Jiang, Ziju Shen, et al

    URL https://papers.nips.cc/paper_files/paper/2022/hash/ 18abbeef8cfe9203fdf9053c9c4fe191-Abstract-Conference.html. Jia Li, Edward Beeching, Lewis Tunstall, Ben Lipkin, Roman Soletskyi, Shengyi Huang, Kashif Rasul, Longhui Yu, Albert Q Jiang, Ziju Shen, et al. Numinamath: The l...

    2022

  14. [2024]

    doi: 10.18653/v1/2024.acl-long.211

    Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.211. URL https://aclanthology.org/2024.acl-long.211. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understand...

    2024 doi

  15. [2025]

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, Jie Liu, Lei Qi, Zhiyuan Liu, and Maosong Sun

    URLhttps://arxiv.org/abs/2501.12948. Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, Jie Liu, Lei Qi, Zhiyuan Liu, and Maosong Sun. Olympiad- bench: A challenging benchmark for promoting agi with olym...

    arXiv

Pith tools

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