Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Post-Completion Learning for Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Training a language model to keep writing after its final answer—generating a self-evaluation and predicted reward scores in a <post-completion> region—improves reasoning accuracy on math benchmarks while adding zero inference tokens…

desk verdict A genuinely new training-time idea with plausible gains, but the exact-match evaluation protocol and hand-wavy theory keep the headline numbers from being fully convincing. read the letter →

arxiv 2507.20252 v3 pith:2MGVWQN3 submitted 2025-07-27 cs.CL cs.AI

Xiang Fei , Siqi Wang , Shu Wei , Yuxiang Nie , Wei Shi , Hao Feng , Chao Feng , Can Huang This is my paper · ORCID
classification cs.CLcs.AI
keywords post-completionlearningself-evaluationwhite-boxreinforcementconsistencyrewardGRPOchain-of-thoughtreasoninglanguagemodeltraininginferenceefficiency
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

Post-Completion Learning (PCL) is a training scheme that lets a language model keep generating after it has produced its final answer, writing a self-assessment and predicted reward values in a new region. At deployment the model stops at that marker, so the reflection costs nothing at inference time. The paper reports that models trained this way score higher on GSM8K, StrategyQA, and MathQA than models trained with supervised fine-tuning, sequential SFT-then-RL, or joint SFT+RL, with average gains over SFT of 7.81, 6.11, and 6.27 percentage points respectively. The central bet is that forcing a model to compute and verbalize reward-function scores during training makes it internalize self-evaluation, and that this internalized evaluation improves the answers that are actually served.

What carries the argument

The load-bearing object is the <post-completion> token, a separator that splits the sequence into a reasoning region (<think> + <answer>) and a reflection region (<evaluation> + <reward>). Around it the paper builds a white-box reinforcement learning loop: the model is given the reward rules in its prompt, works through accuracy and format scoring inside the evaluation passage, outputs predicted values in the reward block, and is supervised by a consistency reward that compares those predictions with the true reward-function outputs. Dual-track SFT trains the two regions separately, and GRPO samples eight responses per group and optimizes the sum of accuracy, format, and consistency rewards. The token is what makes the scheme zero-cost at inference: because it is a stop word for generation, the reflection region is trained but never decoded.

What would settle it

Re-run the comparison with a lenient answer extractor that searches the full response for the final number, as traditional math-verify scoring does, and see whether PCL's accuracy advantage over SFT and SFT+RL survives; if the gap collapses, the gains are format compliance rather than reasoning.

Watch

Extended reading notes

Core claim

PCL's central claim is that the sequence space after the end-of-sequence token is a usable training resource. Instead of ending learning at <eos>, the model is trained to emit <post-completion>, then an <evaluation> passage that recomputes the accuracy and format rewards, then a <reward> block containing predicted scores. A consistency reward $R_c = 1 - |R_{pred} - R_{true}|_1$ aligns those predicted scores with true reward-function outputs, GRPO optimizes the full sequence, and dual-track SFT keeps the reasoning loss (think + answer) and evaluation loss (evaluation + reward) separate. At inference <post-completion> is registered as a stop word, so the reflection content is never generated. On GSM8K, StrategyQA, and MathQA, PCL outperforms SFT, SFT+RL, and joint SFT+RL baselines on exact-match answer accuracy, with average improvements over SFT of +7.81, +6.11, and +6.27 percentage points and over SFT+RL of +3.45, +4.56, and +1.14 percentage points respectively.

Load-bearing premise

The reported gains assume exact matching on the <answer> region is equally fair to PCL and to baseline models, even though PCL is the only method explicitly trained with a format reward to produce those tags.

Editorial extensions

If this is right

  • PCL models serve answers at the same latency as baselines: the <post-completion> token stops generation before any self-evaluation text is produced, so the training-time reflection is free at inference.
  • Smaller models get the largest bumps on structured reasoning, with LLaMA3.2-3B up 10.47 points and Qwen2.5-1.5B up 7.20 points over SFT on GSM8K, suggesting the self-evaluation signal helps most when raw capacity is limited.
  • Both evaluation-track SFT and the consistency reward carry weight: dropping the evaluation SFT lowers GSM8K accuracy to 67.02 and dropping the consistency reward lowers it to 66.08, compared with 67.93 for the full method.
  • Training reasoning and evaluation objectives jointly in one stage beats sequential SFT-then-RL in the reported experiments, matching the paper's claim that avoiding the forgetting penalty improves sample efficiency.
  • The gains appear across three model families (Qwen2.5 and LLaMA3.x) and three reasoning benchmarks, which supports treating PCL as a general training recipe rather than a dataset-specific adjustment.

Reading between the lines

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

  • If PCL's benefit comes from internalized self-evaluation rather than from the extra output tokens, the <post-completion> recipe should transfer to code generation and multi-turn dialogue, where the model could reflect on compiler errors or conversation goals during training without paying for that reflection at serving time.
  • Because the consistency reward penalizes mismatches between predicted and true reward scores, PCL is implicitly a calibration training method; a testable prediction is that PCL models would show better confidence calibration on out-of-domain questions, not merely higher accuracy.
  • The white-box format also creates an audit trail: since the model writes out its reward calculations, a deployed system could read the evaluation text to detect fabricated self-assessments, turning reward hacking into an inspectable failure mode instead of a hidden one.
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

4 major / 5 minor

Summary. The paper proposes Post-Completion Learning (PCL), a training framework in which the model continues generating a self-evaluation and reward-prediction segment after the answer, separated by a dedicated <post-completion> token. Training combines dual-track SFT (one track for reasoning, one for evaluation) with GRPO-style RL using accuracy, format, and consistency rewards. At inference the model stops at the <post-completion> token, so no extra tokens are generated. Experiments on GSM8K, StrategyQA, and MathQA across several small models report consistent gains over SFT and SFT+RL baselines, with ablations on GSM8K attributing parts of the gain to the evaluation SFT and the consistency reward. The paper also presents an information-theoretic and convergence analysis intended to justify the method.

Significance. If the empirical claims hold, PCL is a practically appealing idea: it internalizes self-evaluation during training while adding zero inference-time tokens, and the GSM8K ablations give a credible decomposition of the gains into reasoning SFT, evaluation SFT, format rewards, and consistency rewards. The teacher-data filtering step (retaining only samples where the teacher's self-evaluation matches the recalculated reward) is a sensible anti-circularity measure, and the consistency reward is computed against fixed ground-truth-based reward functions rather than against the model's own outputs. The main value is the training recipe and the empirical comparison, not the theoretical section as written. The central weakness is that all methods are scored by exact matching on the <answer> region while only PCL and one control are trained to emit the required tags, so the reported margins may partly measure format compliance rather than reasoning quality. The paper would be strengthened by a format-agnostic rescoring, by significance information, and by either proving or explicitly reframing the theoretical claims.

major comments (4)
  1. [Experiments (Main Results), Tables 2–4] The evaluation protocol is not symmetric across methods. Accuracy is computed by exact matching on the <answer> region for every method, while PCL is trained with format rewards and evaluation SFT to emit the required tags; SFT and SFT+RL are not. The paper itself states that the criteria are "relatively more stringent, which may result in lower baseline performance metrics compared to other studies." A baseline that produces a correct answer outside the tags is scored as wrong, whereas PCL is rewarded for tag compliance. The reported gaps (e.g., +3.45 on GSM8K, +4.56 on StrategyQA) therefore conflate format compliance with reasoning gains. Please re-score all methods with a format-agnostic answer extractor (e.g., regular-expression extraction from the full generation) and report both exact-match and format-agnostic accuracies. The Table 2 "w/ eval output" row, which receives format rewards and post-completion outputs, narrows the GSM8K gap to about +2.00, suggesting a residual effect, but that control is absent from Tables 3 and 4, so the magnitude of the claimed "enhanced output quality" is not yet established.
  2. [Theoretical Analysis, Theorems 1–3, Eqs. (7)–(13)] The main theoretical claims are asserted rather than proved. Theorem 1 needs the strict inequality I(θ;e,r|x,y)>0 in Eq. (8), but the chain rule alone gives only non-negativity; the "not redundant" assumption is exactly what must be shown, especially because in the data-generation process e and r are constructed from y and fixed reward rules. Theorem 2's Eq. (11) states O(log n/n) versus O(sqrt(log n)/sqrt n) sample complexity without defining the error classes, the learning algorithm, or the "parameter separation property," and Eqs. (9)–(10) are not derivations. Theorem 3 likewise relies on an undefined claim that |Θ_R ∩ Θ_E| ≪ |Θ_R ∪ Θ_E| and on unspecified "standard regularity conditions." If these results are intended as a theoretical contribution, they need real assumptions and proofs; otherwise the section should be reframed as intuition, because the current "Proof" labels overclaim.
  3. [Tables 3–4 and Limitations] The cross-dataset evidence is weaker than the text implies. StrategyQA and MathQA report only three settings (SFT, SFT+RL, PCL), so neither the ablations nor the format-trained "w/ eval output" control is available on those datasets. MathQA results for the two smaller models are -0.04 and +0.24 points, which the paper itself calls "mixed results" in the Limitations section, yet the conclusion claims "consistent improvements" across datasets and scales. With no repeated runs, confidence intervals, or significance tests, these near-zero differences cannot support the cross-dataset claim. Please add variance information, include ablation or control rows for at least StrategyQA, and analyze the MathQA non-improvement quantitatively instead of attributing it to an unplotted reward oscillation.
  4. [Method, Unified Training Framework, Eq. (5)] Equation (5) presents the RL loss as REINFORCE with the raw sum of rewards as the weighting term, but the text says the method uses GRPO with 8 sampled responses, group-relative contrast, and a KL constraint of β=0.04. None of these ingredients appears in Eq. (5). The exact objective matters for reproducibility and for the claimed connection to GRPO. Please write the actual loss, including the advantage normalization and KL penalty, or explicitly state that Eq. (5) is schematic and give the precise training objective in an appendix.
minor comments (5)
  1. [Experiments, Main Results] The sentence "PCL demonstrates consistent and improvements over all baseline methods" is missing a noun; it should read "consistent and significant improvements" or similar.
  2. [Table 2 / Table 3] Model names are inconsistent: Table 2 uses "LlaMA3.2" while Table 3 uses "LlaMA3.1"; the standard spelling is "LLaMA."
  3. [Table 2 / Main Results] The text refers to "RFT (Joint SFT+RL) w/ eval output" but the table row is labeled "w/ eval output + post-completion output"; please align the nomenclature.
  4. [Method, Evaluation Data Preparation] The description of the teacher-distillation baseline is confusing: the table says "Teacher distillation only" is SFT on reasoning+evaluation data, while the text calls it "corrected responses from teacher models"; clarify what is distilled and what is filtered.
  5. [Figure 2] The example text "Then, 7 + 5 = 14 is wrong" is confusing because 7+5=12; the intended point is that the earlier step was wrong. Rephrase the self-evaluation to say explicitly that 7+5=12, so the stated final answer of 14 is incorrect.

Circularity Check

1 steps flagged · score 3.0 of 10

Theorem 1 assumes the information gain it claims to prove; the empirical comparisons themselves are not circular.

  1. self definitional [Theoretical Analysis, 'Information-Theoretic Foundation and Optimality', Theorem 1, Eqs. (7)–(8)]
    "By the chain rule of mutual information, the mutual information between model parameters θ and the extended target distribution can be decomposed as: I(θ;y+|x) = I(θ;y|x) + I(θ;e,r|x,y) (7) Since evaluatione and reward prediction r provide additional supervisory signals about the quality and characteristics of reasoningy, and these signals are not redundant with the reasoning content itself, we have: I(θ;e,r|x,y )> 0 (8) Therefore:I(θ;y+|x)>I (θ;y|x)"

    The theorem's conclusion, that extending the sequence with e and r increases learnable information, is identical to the asserted inequality in Eq. (8). Eq. (7) is only the chain-rule decomposition and contributes nothing substantive; the result I(θ;y+|x)>I(θ;y|x) is equivalent, through Eq. (7), to I(θ;e,r|x,y)>0. That latter inequality is exactly what the theorem is supposed to establish, but the 'proof' merely re-asserts it with the phrase 'provide additional supervisory signals ... not redundant.' No independent argument is given for non-redundancy, so the theoretical foundation is circular at this step.

full rationale

The empirical core of the paper is not circular. The consistency reward in Eq. (2) compares the model's predicted score with a score computed from ground-truth labels and fixed formatting rules, and the reported accuracy is measured on held-out test answers; no parameter is fitted to test labels and then renamed as a prediction. The ablation comparisons (PCL vs. SFT, SFT+RL, joint SFT+RL, w/ eval output) are therefore genuine held-out experimental comparisons. The exact-match evaluation protocol is more favorable to models trained with format rewards, and the paper explicitly concedes that the criteria are 'relatively more stringent,' but that is a benchmark-design concern, not a reduction of the claimed result to its inputs. The one clear circular step is Theorem 1, where the 'information expansion' is assumed through I(θ;e,r|x,y)>0 rather than derived. Because the central empirical claim does not depend on this theorem, the overall circularity score is modest rather than severe.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central method rests on a small set of training hyperparameters and on assumptions about information content and parameter separation of the post-completion tokens. The main invented element is a special token, which is testable. The teacher-data filtering rule is a free selection parameter that is not fully specified.

free parameters (6)
  • SFT track ratio = 1:1 reasoning:evaluation
    Chosen for simplicity in the unified training framework; no sweep or tuning reported.
  • RL to SFT ratio = 1:1
    Stated in Experimental Setup as a simplicity choice; affects the balance between supervised and reinforcement objectives.
  • GRPO group size = 8
    Eight sampled responses per group in the RL stage; taken from GRPO defaults rather than tuned for PCL.
  • KL divergence coefficient = 0.04
    Kept at the GRPO default to prevent policy drift; affects stability and final performance.
  • Training epochs = 2
    Chosen without reported ablation; the number of passes over the combined data.
  • Teacher data filtering criterion = Unspecified
    Teacher-generated samples are retained only when the model's self-evaluation matches recomputed reward scores, but the exact matching rule and threshold are not defined.
assumptions (5)
  • ad hoc to paper I(θ;e,r|x,y) > 0, i.e., the evaluation and reward tokens add information beyond the reasoning sequence.
    Theorem 1 requires this positivity; the proof asserts non-redundancy rather than deriving it.
  • ad hoc to paper The post-completion token induces near-disjoint parameter sets for reasoning and evaluation, |Θ_R ∩ Θ_E| ≪ |Θ_R ∪ Θ_E|.
    Theorem 3 relies on this separation, but transformer attention mixes all sequence positions, so parameters are not cleanly separable.
  • domain assumption GPT-4.1 in-context demonstrations provide high-quality PCL-format training data.
    Evaluation data quality is checked only by an internal self-consistency filter, not by an external benchmark.
  • standard math Standard PAC-learning and Lipschitz regularity assumptions apply to the transformer training objective.
    Invoked for the sample-complexity and convergence theorems; not verified for these models.
  • domain assumption Ground-truth labels are available during training for the accuracy reward.
    The accuracy reward compares the model answer with the label, which assumes labeled training data for each task.
invented entities (1)
  • <post-completion> token independent evidence
    purpose: Acts as the boundary between the reasoning region and the evaluation region during training, and as the stop token during inference to keep deployment cost unchanged.
    The token is a concrete vocabulary element whose behavior can be reproduced by training and inference runs; it is not an unobservable construct.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Post-Completion Learning for Language Models." pith.science (2026). https://pith.science/paper/2MGVWQN3

@misc{pith2026250720252,
  author       = {Pith},
  title        = {Pith review of: Post-Completion Learning for Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MGVWQN3}},
  note         = {Machine review of arXiv:2507.20252}
}
read the original abstract

Current language model training paradigms typically terminate learning upon reaching the end-of-sequence (<eos>) token, overlooking the potential learning opportunities in the post-completion space. We propose Post-Completion Learning (PCL), a novel training framework that systematically utilizes the sequence space after model output completion, to enhance both the reasoning and self-evaluation abilities. PCL enables models to continue generating self-assessments and reward predictions during training, while maintaining efficient inference by stopping at the completion point. To fully utilize this post-completion space, we design a white-box reinforcement learning method: let the model evaluate the output content according to the reward rules, then calculate and align the score with the reward functions for supervision. We implement dual-track SFT to optimize both reasoning and evaluation capabilities, and mixed it with RL training to achieve multi-objective hybrid optimization. Experimental results on different datasets and models demonstrate consistent improvements over traditional SFT and RL methods. Our method provides a new technical path for language model training that enhances output quality while preserving deployment efficiency.

Figures

Figures reproduced from arXiv: 2507.20252 by the authors.

Figure 1
Figure 1. The proposed post-completion learning method. (a). Traditional language model output its chain-of-thought (CoT) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustrative of the proposed white-box reinforce [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The unified training framework of our method. The model is trained with multi-objective SFT + RL optimization. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 4 canonical work pages

  1. [1]

    Amini, A.; Gabriel, S.; Lin, P.; Koncel-Kedziorski, R.; Choi, Y.; and Hajishirzi, H. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. arXiv preprint arXiv:1905.13319

    arXiv 2019

  2. [2]

    Bai, Y.; Kadavath, S.; Kundu, S.; Askell, A.; Kernion, J.; Jones, A.; Chen, A.; Goldie, A.; Mirhoseini, A.; McKinnon, C.; et al. 2022. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073

    arXiv 2022

  3. [3]

    D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

    Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877--1901

    2020

  4. [4]

    F.; Leike, J.; Brown, T.; Martic, M.; Legg, S.; and Amodei, D

    Christiano, P. F.; Leike, J.; Brown, T.; Martic, M.; Legg, S.; and Amodei, D. 2017. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30

    2017

  5. [5]

    W.; Hou, L.; Longpre, S.; Zoph, B.; Tay, Y.; Fedus, W.; Li, Y.; Wang, X.; Dehghani, M.; Brahma, S.; et al

    Chung, H. W.; Hou, L.; Longpre, S.; Zoph, B.; Tay, Y.; Fedus, W.; Li, Y.; Wang, X.; Dehghani, M.; Brahma, S.; et al. 2024. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70): 1--53

    2024

  6. [6]

    Cobbe, K.; Kosaraju, V.; Bavarian, M.; Chen, M.; Jun, H.; Kaiser, L.; Plappert, M.; Tworek, J.; Hilton, J.; Nakano, R.; et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    arXiv 2021

  7. [7]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), 4171--4186

    2019

  8. [8]

    Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Yang, A.; Fan, A.; et al. 2024. The llama 3 herd of models. arXiv e-prints, arXiv--2407

    2024

Show all 37 references
  1. [9]

    Finn, C.; Abbeel, P.; and Levine, S. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, 1126--1135. PMLR

    2017

  2. [10]

    Fu, Y.; Chen, T.; Chai, J.; Wang, X.; Tu, S.; Yin, G.; Lin, W.; Zhang, Q.; Zhu, Y.; and Zhao, D. 2025. SRFT: A Single-Stage Method with Supervised and Reinforcement Fine-Tuning for Reasoning. arXiv preprint arXiv:2506.19767

    2025 arXiv

  3. [11]

    Geva, M.; Khashabi, D.; Segal, E.; Khot, T.; Roth, D.; and Berant, J. 2021. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9: 346--361

    2021

  4. [12]

    Groot, T.; and Valdenegro-Toro, M. 2024. Overconfidence is key: Verbalized uncertainty evaluation in large language and vision-language models. arXiv preprint arXiv:2405.02917

    2024 arXiv

  5. [13]

    Hong, J.; Dragan, A.; and Levine, S. 2024. Q-sft: Q-learning for language models via supervised fine-tuning. arXiv preprint arXiv:2411.05193

    2024 arXiv

  6. [14]

    J.; Rohatgi, D.; Zhang, C.; Simchowitz, M.; Ash, J

    Huang, A.; Block, A.; Foster, D. J.; Rohatgi, D.; Zhang, C.; Simchowitz, M.; Ash, J. T.; and Krishnamurthy, A. 2024. Self-improvement in language models: The sharpening mechanism. arXiv preprint arXiv:2412.01951

    2024 arXiv

  7. [15]

    S.; Yu, A

    Huang, J.; Chen, X.; Mishra, S.; Zheng, H. S.; Yu, A. W.; Song, X.; and Zhou, D. 2023. Large language models cannot self-correct reasoning yet. arXiv preprint arXiv:2310.01798

    2023 arXiv

  8. [16]

    Hugging Face . 2025. Open R1: A fully open reproduction of DeepSeek-R1

    2025

  9. [17]

    Kamoi, R.; Zhang, Y.; Zhang, N.; Han, J.; and Zhang, R. 2024. When can llms actually correct their own mistakes? a critical survey of self-correction of llms. Transactions of the Association for Computational Linguistics, 12: 1417--1440

    2024

  10. [18]

    Lightman, H.; Kosaraju, V.; Burda, Y.; Edwards, H.; Baker, B.; Lee, T.; Leike, J.; Schulman, J.; Sutskever, I.; and Cobbe, K. 2023. Let's verify step by step. In The Twelfth International Conference on Learning Representations

    2023

  11. [19]

    Madaan, A.; Tandon, N.; Gupta, P.; Hallinan, S.; Gao, L.; Wiegreffe, S.; Alon, U.; Dziri, N.; Prabhumoye, S.; Yang, Y.; et al. 2023. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36: 46534--46594

    2023

  12. [20]

    Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 27730--27744

    2022

  13. [21]

    D.; Ermon, S.; and Finn, C

    Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C. D.; Ermon, S.; and Finn, C. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36: 53728--53741

    2023

  14. [22]

    A.; and DeSanctis, V

    Schon, D. A.; and DeSanctis, V. 1986. The reflective practitioner: How professionals think in action

    1986

  15. [23]

    Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

    2017 arXiv

  16. [24]

    S.; Xin, R.; Geng, S.; Wang, Y.; Oh, S.; Du, S

    Shao, R.; Li, S. S.; Xin, R.; Geng, S.; Wang, Y.; Oh, S.; Du, S. S.; Lambert, N.; Min, S.; Krishna, R.; et al. 2025. Spurious rewards: Rethinking training signals in rlvr. arXiv preprint arXiv:2506.10947

    2025 arXiv

  17. [25]

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  18. [26]

    Shinn, N.; Cassano, F.; Gopinath, A.; Narasimhan, K.; and Yao, S. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36: 8634--8652

    2023

  19. [27]

    Snell, J.; Swersky, K.; and Zemel, R. 2017. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30

    2017

  20. [28]

    Uesato, J.; Kushman, N.; Kumar, R.; Song, F.; Siegel, N.; Wang, L.; Creswell, A.; Irving, G.; and Higgins, I. 2022. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275

    2022 arXiv

  21. [29]

    Vinyals, O.; Blundell, C.; Lillicrap, T.; Wierstra, D.; et al. 2016. Matching networks for one shot learning. Advances in neural information processing systems, 29

    2016

  22. [30]

    Y.; Guu, K.; Yu, A

    Wei, J.; Bosma, M.; Zhao, V. Y.; Guu, K.; Yu, A. W.; Lester, B.; Du, N.; Dai, A. M.; and Le, Q. V. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652

    2021 arXiv

  23. [31]

    V.; Zhou, D.; et al

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824--24837

    2022

  24. [32]

    Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    2025 arXiv

  25. [33]

    Zelikman, E.; Wu, Y.; Mu, J.; and Goodman, N. 2022. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35: 15476--15488

    2022

  26. [34]

    Zhang, Z.; Zhang, A.; Li, M.; and Smola, A. 2022. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493

    2022 arXiv

  27. [35]

    Zhang, Z.; Zheng, C.; Wu, Y.; Zhang, B.; Lin, R.; Yu, B.; Liu, D.; Zhou, J.; and Lin, J. 2025. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301

    2025 arXiv

  28. [36]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  29. [37]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.