Pith.

Computation and Language · arXiv:2611.26510 v1 · submitted 19 Nov 2026 · 38 references

AutoMaster: Autonomous Post-Training by Making the Same Model a Better Teacher

Guoxin Chen, Shuang Sun, Fanzhe Meng, Xinyu Tang, Ruihua Song, Wayne Xin Zhao

The paper says

“Task success alone does not specify which behaviors post-training should reinforce. Even successful teacher trajectories can contain unsupported assumptions, unproductive repetition, or superficial verification.”

more of the abstract

We propose Skill-FT (Skill-Guided Fine-Tuning), which uses natural-language skills to suppress undesirable habits and elicit underrepresented effective behaviors during teacher rollouts, then fine-tunes a target model on the resulting trajectories. We introduce AutoMaster, an autonomous post-training system in which an LLM agent develops these interventions through trajectory analysis, skill design, data selection, and feedback from rollouts and target-model evaluation. Across Terminal, Search, and software engineering tasks with 4B and 9B target models, AutoMaster improves mean performance over Vanilla SFT on successful unguided teacher trajectories by 1.42-8.99 percentage points, with the teacher, task pool, training set size, and SFT recipe held fixed. Together with the ablations and trajectory analyses, these results suggest that behavioral shaping can make a fixed model a more effective teacher. We further find that retaining skill instructions during SFT yields higher mean performance than removing them across all six settings, despite their absence at evaluation and identical supervised outputs. These findings suggest that preserving skill context as part of a training demonstration can support better behavioral internalization.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 20 Nov 2026 · read the full text (88,019 characters) · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Skill-guided teacher rollouts plus retaining skill context during SFT give consistent student gains over Vanilla SFT in 4 of 6 settings, with the headline caveat that the agent picks the final checkpoint from evaluation feedback across up to five rounds. This line is wrong

10 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

The paper, first character to last

The reviewer was given the whole text, 88,019 characters. Not checked: figures, code, what the 38 references say.

10 minor objections

The paper says, at §4.1 (Main Results), Table 1

Pith says: minor objection 1 of 10

The table reports mean ± SD over three evaluation runs per checkpoint. For Vanilla SFT, the SD is sometimes larger than for AutoMaster (e.g., Doc2Repo 4B: 4.00 vs 3.58). The paper should clarify whether the three runs are independent (different random seeds) or the same checkpoint evaluated three times. If the latter, the SD reflects evaluation noise, not training variability, and the comparison may be affected.

The paper says, at §4.2 (Model Evaluation), Appendix A.4

Pith says: minor objection 2 of 10

The paper states that at the end of autonomous experimentation, AutoMaster selects a final checkpoint based on accumulated evaluation feedback and records its choice. The selection criteria are not specified. The paper should describe how the agent chooses the final checkpoint (e.g., highest mean, highest single run, or some other heuristic) to allow reproducibility.

show 8 more

The paper says, at §4.3 (Ablation Studies), Figure 3

Pith says: minor objection 3 of 10

The Task-Matched SFT baseline uses unguided rollouts matched to AutoMaster's final training tasks, but the paper does not specify whether the unguided rollouts are from the same teacher or from a different run. The caption says 'unguided rollouts from the same teacher', which is clear, but the text should also state that the unguided rollouts are from the initial unguided teacher rollouts (not from a later round).

The paper says, at §5.2 (Preserving Skill Context), Figure 6

Pith says: minor objection 4 of 10

The comparison between Keep and Remove is a within-trajectory ablation, but the paper does not report whether the difference is statistically significant. Given the small number of evaluation runs (three), the paper should provide confidence intervals or a significance test for the Keep vs Remove comparison.

The paper says, at §5.3 (Autonomous Experimentation), Figure 7

Pith says: minor objection 5 of 10

The paper shows performance across five rounds, but does not report how many rounds the agent actually ran in each setting (the budget is five, but the agent may stop earlier). The paper should clarify whether all five rounds were completed for each setting and whether the final checkpoint is always from the round with the highest mean.

The paper says, at §2 (Related Work)

Pith says: minor objection 6 of 10

The related work section cites several papers but does not discuss the key difference between AutoMaster and prior work on prompt optimization (e.g., GEPA) or skill libraries (e.g., SkillRL). The paper should clarify that AutoMaster uses skills to shape teacher rollouts for SFT data generation, rather than optimizing prompts for the student's own performance.

The paper says, at §4.1 (Main Results), Table 1

Pith says: minor objection 7 of 10

The paper reports gains of 1.42-8.99 percentage points, but the baseline Vanilla SFT uses 512 trajectories randomly sampled from successful unguided teacher rollouts. The paper should discuss whether the random sampling of 512 from a larger pool introduces variance that could affect the comparison, and whether AutoMaster's selection of 512 trajectories from a potentially different distribution (including some unsuccessful ones?) is a confounding factor.

The paper says, at §4.3 (Ablation Studies), Table A5

Pith says: minor objection 8 of 10

Table A5 shows that for Terminal and Search, 0% of pairs have 'Success → failure' (i.e., the unguided rollout succeeds but the AutoMaster demonstration fails). This suggests that skill guidance rarely hurts task success. However, the paper should note that this is on the training tasks, not the evaluation set, and that the agent may avoid applying skills to tasks where they would hurt.

The paper says, at §5.1 (Behavioral Shaping), Figure 5

Pith says: minor objection 9 of 10

The behavioral shaping examples are illustrative but the paper does not quantify how often such improvements occur across the training set. The paper should provide a systematic analysis (e.g., human evaluation or automated metrics) of how often skill guidance changes teacher behavior in the intended direction.

The paper says, at §4.2 (Model Evaluation), Appendix A.4

Pith says: minor objection 10 of 10

The paper states that evaluation uses a temperature of 1.0 and top-p 0.95. The paper should clarify whether these settings are the same as those used during teacher rollouts (temperature 1.0, top-p 1.0) and whether the difference in top-p could affect the comparison.

What would settle it

Run AutoMaster with the same teacher, task pool, and SFT recipe but with the agent's skill design and data selection replaced by random selection of 512 trajectories from the same pool of unguided successful rollouts, and compare to Vanilla SFT on a single fixed run. If the gains disappear, the headroom is due to selection rather than behavioral shaping.

from Pith's summary of the paper

How the review went

5 passes by deepseek-v4-flash · 20 Nov 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic concern withdrawn: Pith's reviewer misread the paper 8 s · 26,351 tokens read, 803 written
  3. Referee 10 minor 17 s · 26,632 tokens read, 2,377 written
  4. Check against the paper kept 0 of 4 objections 24 s · 55,101 tokens read, 4,032 written
  5. Editor wrote the verdict above 7 s · 26,900 tokens read, 872 written

Not run on this paper: the simulated authors' defence.

Pith's passes raised 4 objections. A check against the paper's text kept 0 and withdrew 4 (4 Pith's reviewer misread the paper).

The review in full

The review in brief, machine-written

The paper's core claim is that natural-language skills appended to the teacher's system prompt can shape which behaviors appear in successful trajectories, and that these shaped demonstrations train better students than unguided ones, all with the teacher, task pool, training-set size, and SFT recipe held fixed. That claim is mostly supported by the evidence, and the authors report it honestly: paired bootstrap confidence intervals, a task-matched SFT control, and component ablations are all good practice, and they disclose that the 4B gains on BrowseComp and Doc2Repo have intervals including zero.

The cleanest result is the Keep-versus-Remove comparison: identical trajectories and SFT recipe, skill instructions either retained or dropped from the training input, no skills at evaluation, and Keep beats Remove by 2.15 to 4.01 points in all six settings. That is a controlled internalization finding and, as far as I can tell from the cited literature, genuinely new. It deserves to be the centerpiece.

Two soft spots, in proportion. First, the Appendix's Model Evaluation section states that the agent selects the final AutoMaster checkpoint based on accumulated evaluation scores and trajectories over up to five rounds, then re-evaluates it three times, while Vanilla SFT is a single fixed run. The later rounds section shows per-round variation of several points, so part of the reported headroom may be best-of-rounds selection rather than skill-guided generation alone. This is the main thing a referee should push on. The component ablations do not fully resolve it, since those variants also get agent selection. Second, the two CIs already including zero mean the effect is not established for 4B BrowseComp or 4B Doc2Repo even before that confound.

Minor notes: the closest skill-distillation baselines, Skill-SD and Skill-to-LoRA, are cited but not run, and several baselines are 2026 arXiv preprints, which made related-work positioning harder for me to check. One claimed result I could not verify: the nine-point stencil claim in the Terminal case relies on a figure I could not see.

This paper deserves a serious referee. The internalization result is well controlled, the main comparison is carefully reported, and the selection confound is fixable. A revision that either holds out an evaluation-excluded selection criterion or analyzes the final round's checkpoint under the same protocol as Vanilla SFT would substantially firm up the central claim.

The stress test's concern

Withdrawn: Pith's reviewer misread the paper.

The paper says

“At the end of autonomous experimentation, \papername{} selects a final checkpoint based on accumulated evaluation feedback and records its choice in a final report.”

The referee's report

Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 3 of its 3 major objections did not survive that check.

Pith's reading of the paper (a machine's summary, not the authors')

The paper proposes Skill-FT (Skill-Guided Fine-Tuning), a method that uses natural-language skills to shape the behaviors a teacher model demonstrates during rollout, then fine-tunes a student on the resulting trajectories. The authors introduce AutoMaster, an autonomous system where an LLM agent designs, tests, and refines these skills through iterative rounds of trajectory analysis, skill design, data selection, and evaluation feedback. Across Terminal, Search, and software engineering tasks with 4B and 9B student models, AutoMaster improves mean performance over standard SFT on successful unguided teacher trajectories by 1.42–8.99 percentage points, with the teacher, task pool, training set size, and SFT recipe held fixed. The paper further reports that retaining skill instructions in the SFT input yields higher mean performance than removing them across all six settings, even though students are evaluated without skills. The authors argue that behavioral shaping can make a fixed model a more effective teacher, and that preserving skill context during training supports better behavioral internalization.

What carries the argument. Skill-FT (Skill-Guided Fine-Tuning): natural-language skills appended to the teacher's system prompt to suppress undesirable habits and elicit effective behaviors during rollout. AutoMaster: an autonomous system with an inner loop (skill design, pilot rollouts, refinement, data selection) and an outer loop (SFT from a fixed initial checkpoint, evaluation feedback, iterative refinement). The central mechanism is behavioral shaping of teacher demonstrations before student training, combined with retaining skill instructions in the training input.

Core claim. The paper claims that a fixed teacher model can be made more effective for post-training by shaping its behavioral patterns during rollout using natural-language skills, and that the resulting skill-guided demonstrations improve student performance beyond what is achievable by selecting from unguided successful trajectories alone. The authors further claim that retaining the skill instructions in the SFT input, even though they are absent at evaluation, improves student performance compared to removing them, suggesting that preserving the context behind demonstrated reasoning and actions aids behavioral internalization.

The premise it rests on. The reported AutoMaster checkpoint is selected by the agent using accumulated evaluation scores and trajectories over up to five rounds, then re-evaluated three times, while Vanilla SFT is a single fixed run; if the headroom comes from best-of-rounds selection rather than from skill-guided generation and data selection, the central claim collapses to selection bias.

If the paper is right (Pith's inferences, not the authors')

If correct, the approach suggests that post-training headroom from a fixed teacher is larger than what is captured by distilling successful unguided trajectories alone.

The finding that retaining skill instructions during SFT improves performance, despite a training-evaluation prompt mismatch, challenges the assumption that minimizing prompt mismatch is always beneficial.

The results imply that autonomous post-training systems can improve student performance by jointly optimizing how demonstrations are generated and selected, rather than only selecting from a fixed pool.

The larger gains for the 9B model compared to the 4B model suggest that greater student capacity may better leverage skill-guided demonstrations.

The approach could be extended to other forms of behavioral guidance beyond natural-language skills, such as structured constraints or reward signals during rollout.

The finding that skill instructions aid internalization even when absent at evaluation raises the question of whether similar context-preserving strategies could benefit other forms of instruction-tuned or demonstration-based learning.

The reliance on a strong LLM agent (GPT-6 Astra) for skill design and refinement means the method's effectiveness may be sensitive to the agent's capabilities; testing with weaker agents would clarify the generality.

The paired confidence intervals for 4B BrowseComp and Doc2Repo include zero, so the claim of improvement in those settings is not statistically established even before considering the selection-bias confound.

Check this review

What was read

The passes were given all 88,019 characters of the paper's LaTeX source.

Fingerprint of that text (SHA-256): f87502c5e4e6145e4d82eaf4f9558af90348ef355b94105a49c1ff38947ed5b4

Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:

PYTHONPATH=. python scripts/review_input_check.py 2611.26510

The passes, in the order they ran

  1. Reader deepseek-v4-flash · medium effort · $0.0077 · instructions of the Reader, fingerprint f9ead7300382…
  2. Skeptic deepseek-v4-flash · medium effort · $0.0025 · instructions of the Skeptic, fingerprint 8f88c0f61488…
  3. Referee deepseek-v4-flash · medium effort · $0.0029 · instructions of the Referee, fingerprint a852d99aea0d…
  4. Check against the paper deepseek-v4-flash · medium effort · $0.0039 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
  5. Editor deepseek-v4-flash · medium effort · $0.0039 · instructions of the Editor, fingerprint e7211f6a3d27…
  6. Circularity check deepseek-v4-flash · medium effort · $0.0040 · instructions of the Circularity check, fingerprint e9946e39d67b…
  7. Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0028 · instructions of the Summary for the page, fingerprint e4f56bf568e7…

The passes cost $0.028 in model calls together.

Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00

Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.

curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum

prints the fingerprint of the reader's instructions. The same works for every pass.

The receipt

The review's fingerprint was taken on 20 Nov 2026. The review you are reading still matches it.

Fingerprint of the review (SHA-256): 0bffabe5b4261f9d3df8a10a7be5ad68f6113cd101afe0ed244ca6487c48e2d6

curl -s 'https://pith.science/api/v1/verdicts/2611.26510?form=canonical' | sha256sum

prints the fingerprint of the review as it stands now. The rule, in four steps.

A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Task success alone does not specify which behaviors post-training should reinforce. Even successful teacher trajectories can contain unsupported assumptions, unproductive repetition, or superficial verification. We propose Skill-FT (Skill-Guided Fine-Tuning), which uses natural-language skills to suppress undesirable habits and elicit underrepresented effective behaviors during teacher rollouts, then fine-tunes a target model on the resulting trajectories. We introduce AutoMaster, an autonomous post-training system in which an LLM agent develops these interventions through trajectory analysis, skill design, data selection, and feedback from rollouts and target-model evaluation. Across Terminal, Search, and software engineering tasks with 4B and 9B target models, AutoMaster improves mean performance over Vanilla SFT on successful unguided teacher trajectories by 1.42-8.99 percentage points, with the teacher, task pool, training set size, and SFT recipe held fixed. Together with the ablations and trajectory analyses, these results suggest that behavioral shaping can make a fixed model a more effective teacher. We further find that retaining skill instructions during SFT yields higher mean performance than removing them across all six settings, despite their absence at evaluation and identical supervised outputs. These findings suggest that preserving skill context as part of a training demonstration can support better behavioral internalization.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-20

The authors' main claim, as the classifier picked it out 0.99

“Together with the ablations and trajectory analyses, these results suggest that behavioral shaping can make a fixed model a more effective teacher.”

How the main claim is stated: hedged possibility 0.99

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.32 · human participants 0.06 · controlled comparison 0.96 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.95 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Task success alone does not specify which behaviors post-training should reinforce. background 0.60 0.10 0.03 asserted 1.00
Even successful teacher trajectories can contain unsupported assumptions, unproductive repetition, or superficial verification. background 0.74 0.18 0.03 asserted 0.89
We propose Skill-FT (Skill-Guided Fine-Tuning), which uses natural-language skills to suppress undesirable habits and elicit underrepresented effective behaviors during teacher rollouts, then fine-tunes a target model on the resulting trajectories. method 0.81 0.78 0.11 asserted 0.99
We introduce AutoMaster, an autonomous post-training system in which an LLM agent develops these interventions through trajectory analysis, skill design, data selection, and feedback from rollouts and target-model evaluation. contribution 0.67 0.92 0.08 asserted 0.99
Across Terminal, Search, and software engineering tasks with 4B and 9B target models, AutoMaster improves mean performance over Vanilla SFT on successful unguided teacher trajectories by 1.42-8.99 percentage points, with the teacher, task pool, training set size, and SFT recipe held fixed. result 1.00 0.97 0.24 asserted 1.00
Together with the ablations and trajectory analyses, these results suggest that behavioral shaping can make a fixed model a more effective teacher. interpretation 1.00 0.88 0.62 hedged possibility 0.99
We further find that retaining skill instructions during SFT yields higher mean performance than removing them across all six settings, despite their absence at evaluation and identical supervised outputs. result 1.00 0.97 0.11 asserted 1.00
These findings suggest that preserving skill context as part of a training demonstration can support better behavioral internalization. interpretation 1.00 0.86 0.17 hedged possibility 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-20: 1 request, 6,335 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (8)
Figure 1 of the paper
Figure 1. A correct result is only part of what a teacher demonstrates. AutoMaster designs natural-language skills to guide teacher rollouts and selects demonstrations for fine-tuning. (a) In this example, both pipelines process all 100 records with four workers. The skill-guided teacher also checks whether the fix holds with fewer or more workers. (b) With the same teacher, task pool, training set size, and SFT recipe, students trained with AutoMaster achieve higher performance than their Vanilla SFT counterparts on all three benchmarks at both model sizes (4B and 9B). Labels above the bars…
Figure 2 of the paper
Figure 2. Overview of AutoMaster. An LLM agent designs skills for subsets of a fixed training task pool and tests them in pilot teacher rollouts. It refines the skills as needed, runs promising interventions on larger batches, and selects demonstrations for SFT. In each outer-loop round, the student is trained from the same initial checkpoint, with any skill instructions used by the teacher retained in the training input. The student is evaluated without generated skills. 3.1 Skill-FT: Skill-Guided Fine-Tuning To shape the behavioral patterns represented in training trajectories, we guide the…
Figure 3 of the paper
Figure 3. Infrastructure of AutoMaster. The agent accesses data and submits rollout and SFT jobs through the MCP interface. After SFT, the execution controller automatically schedules evaluation on a separate GPU cluster, where SGLang workers load the checkpoint directly from storage shared with the training cluster. Linked experiment records let the agent trace each evaluation back to the training set and the skills that guided the teacher. weaknesses persist. These observations guide the skills it designs and the training tasks to which it applies them. As shown in
Figure 4 of the paper
Figure 4. Contributions of skill guidance and data selection. w/o Skill Guidance retains agent￾directed selection from unguided teacher rollouts. w/o Data Selection retains skill guidance on a fixed random subset of tasks but omits agent-directed trajectory selection. Removing either component lowers the mean score in each of the six settings. Task-Matched SFT uses unguided rollouts from the same teacher and matches the final training set of AutoMaster. Even with this match, AutoMaster scores 1.08–10.11 percentage points higher on average within each setting. Points show mean scores over…
Figure 5 of the paper
Figure 5. Skill-guided changes in teacher behavior. Each column compares unguided and skill-guided rollouts by the same teacher on the same task. (a) Terminal: Both solvers meet the error threshold. The unguided teacher claims fourth-order accuracy despite evidence of second-order convergence. The skill-guided teacher switches to a nine-point stencil and verifies fourth-order convergence using direct solves on successively finer grids. (b) Search: Both rollouts identify the same episode. The unguided teacher repeatedly guesses the award, while the skill-guided teacher identifies the director…
Figure 6 of the paper
Figure 6. Retaining skill instructions during SFT improves performance without test-time skills. Both conditions use the same teacher trajectories and SFT recipe. Keep retains the skill instructions in the training input, whereas Remove omits them. All students are evaluated without skills under the official benchmark protocols. Bars show mean scores over three evaluation runs per checkpoint, circles show individual runs, and error bars indicate ±1 SD. Labels give the gain of Keep over Remove in percentage points. 5.1 Behavioral Shaping Beyond Task Success Task success can conceal weaknesses…
Figure 7 of the paper
Figure 7. Explicit use of skill context in teacher demonstrations. Two examples from the SFT corpus illustrate how skill instructions inform the teacher’s reasoning and actions. (a) The teacher applies a conditional rule to distinguish interpreter invocation from direct execution, avoiding an unnecessary permission change. (b) The teacher references the supplied grammar to derive an expression’s grouping, then implements and tests the corresponding parser. Skill instructions and trajectory excerpts are condensed. Quotation marks indicate verbatim excerpts. 5.2 Preserving Skill Context for…
Figure 8 of the paper
Figure 8. Performance across five rounds of autonomous experimentation. For each model size, every round trains a student from the initial checkpoint. Points show means over three evaluation runs per checkpoint; error bars indicate ±1 SD. Dashed lines show the corresponding Vanilla SFT means. Filled markers indicate the highest observed mean for each model on each benchmark

Figures are the authors', reproduced from the paper.

References (38)
  1. [1] Kulikov, Ilia and Whitehouse, Chenxi and Wu, Tianhao and Nie, Yixin and Saha, Swarnadeep and Helenowski, Eryk and Yuan, Weizhe and Golovneva, Olga and Lanchantin, Jack and Bachrach, Yoram and Foerster, Jakob and Li, Xian and Fang, Han and Sukhbaatar, Sainbayar and Weston, Jaso... on PithPith's reviewarXivdoi:10.48550/arxiv.2606.25996
  2. [2] Chen, Wanyi and Yang, Xiao and Yang, Xu and Sha, Tianming and Li, Qizheng and Wang, Zhuo and Xian, Bowen and Kong, Fang and Liu, Weiqing and Bian, Jiang , year =. doi:10.48550/arXiv.2604.10547 , url =. 2604.10547 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2604.10547
  3. [3] Rank, Ben and Bhatnagar, Hardik and Prabhu, Ameya and Eisenberg, Shira and Nguyen, Karina and Bethge, Matthias and Andriushchenko, Maksym , booktitle =. 2026 , editor = on Pith
  4. [4] Ta, Anh and Zhu, Junjie and Shayandeh, Shahin , year =. doi:10.48550/arXiv.2608.23911 , url =. 2608.23911 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2608.23911
  5. [5] Chen, Baian and Shu, Chang and Shareghi, Ehsan and Collier, Nigel and Narasimhan, Karthik and Yao, Shunyu , year =. doi:10.48550/arXiv.2310.05915 , url =. 2310.05915 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2310.05915
  6. [6] Karpathy, Andrej , year = on Pith
  7. [7] Wei, Wei , year =. Autonomous on Pith
  8. [8] Li, Qizheng and Zhang, Yifei and Yang, Xiao and Yang, Xu and Wang, Zhuo and Xian, Bowen and Liu, Weiqing and Bian, Jiang , booktitle =. 2026 , volume = on Pith
  9. [9] Du, Yaxin and Yang, Xiyuan and Zhou, Zhifan and Ge, Yujie and Wang, Cheng and Wang, Jiajun and Chen, Sijie and Liu, Zehui and Zhang, Yuxin and Gu, Weicheng and Zhang, Julian and Lei, Zixing and Chen, Siheng , year =. doi:10.48550/arXiv.2609.35561 , url =. 2609.35561 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2609.35561
  10. [10] Toward autonomous long-horizon engineering for ml research , author=. arXiv preprint arXiv:2604.13018 , year= on PithPith's reviewarXiv
  11. [11] Agrawal, Lakshya A. and Tan, Shangyin and Soylu, Dilara and Ziems, Noah and Khare, Rishi and Opsahl-Ong, Krista and Singhvi, Arnav and Shandilya, Herumb and Ryan, Michael J. and Jiang, Meng and Potts, Christopher and Sen, Koushik and Dimakis, Alexandros G. and Stoica, Ion and ... on PithPith's reviewarXivdoi:10.48550/arxiv.2507.19457
  12. [12] Shi, Qiming and Dou, Yibo and Zhu, Jiawen and Tao, Yulong and Jin, Linbo and Kang, Zhaolu and Zhou, Yunfan and Weng, Di , year =. doi:10.48550/arXiv.2607.28048 , url =. 2607.28048 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2607.28048
  13. [13] Xia, Peng and Chen, Jianwen and Wang, Hanyang and Liu, Jiaqi and Zeng, Kaide and Wang, Yu and Han, Siwei and Zhou, Yiyang and Zhao, Xujiang and Chen, Haifeng and Zheng, Zeyu and Xie, Cihang and Yao, Huaxiu , year =. doi:10.48550/arXiv.2602.08234 , url =. 2602.08234 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2602.08234
  14. [14] Scaffold-Mediated Post-Training: Co-Evolving Model Parameters and Procedural Scaffold Graphs , author =. 2026 , eprint =. doi:10.48550/arXiv.2608.05156 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2608.05156
  15. [15] Zhang, Tianyi and Qi, Zhonghao , year =. doi:10.48550/arXiv.2606.16769 , url =. 2606.16769 , archivePrefix = on Pithdoi:10.48550/arxiv.2606.16769
  16. [16] Ge, Zhilong and Shao, Yuting and Yang, Yutao and Cai, Yuxuan and Zhou, Jie and Chen, Kai and Zhang, Bo and Chen, Qin and He, Liang , year =. doi:10.48550/arXiv.2609.27717 , url =. 2609.27717 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2609.27717
  17. [17] Zeng, Aohan and Liu, Mingdao and Lu, Rui and Wang, Bowen and Liu, Xiao and Dong, Yuxiao and Tang, Jie , booktitle =. 2024 , month = aug, address =. doi:10.18653/v1/2024.findings-acl.181 , url = on Pithdoi:10.18653/v1/2024.findings-acl.181
  18. [18] Fu, Dayuan and He, Keqing and Wang, Yejie and Hong, Wentao and GongQue, Zhuoma and Zeng, Weihao and Wang, Wei and Wang, Jingang and Cai, Xunliang and Xu, Weiran , booktitle =. 2025 , url = on Pith
  19. [19] Yuan, Siyu and Chen, Zehui and Xi, Zhiheng and Ye, Junjie and Du, Zhengyin and Chen, Jiecao , year =. doi:10.48550/arXiv.2501.11425 , url =. 2501.11425 , archivePrefix = on PithPith's reviewarXivdoi:10.48550/arxiv.2501.11425
  20. [20] Beneficial Reasoning Behaviors in Agentic Search and Effective Training Methods to Obtain Them , author =. Findings of the Association for Computational Linguistics:. 2026 , month = jul, address =. doi:10.18653/v1/2026.findings-acl.1400 , url = on Pithdoi:10.18653/v1/2026.findings-acl.1400
show 18 more
  1. [21] AlphaMath Almost Zero: Process Supervision without Process , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year= on Pith
  2. [22] Chen, Guoxin and Liao, Minpeng and Li, Chengxi and Fan, Kai. Step-level Value Preference Optimization for Mathematical Reasoning. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.463 on Pithdoi:10.18653/v1/2024.findings-emnlp.463
  3. [23] Merrill, Mike and Shaw, Alexander and Carlini, Nicholas and Li, Boxuan and Raj, Harsh and Bercovich, Ivan and Shi, Lin and Shin, Jeong and Walshe, Thomas and Buchanan, E. Kelly and Shen, Junhong and Ye, Guanghao and Lin, Haowei and Poulos, Jason and Wang, Maoyu and Nezhurina, ... on Pith
  4. [24] Jason Wei and Zhiqing Sun and Spencer Papay and Scott McKinney and Jeffrey Han and Isa Fulford and Hyung Won Chung and Alex Tachard Passos and William Fedus and Amelia Glaese , journal =. 2025 , url = on Pith
  5. [25] Jingyuan Ma and Lynx Aster and He Zhang and Siyao Song and Weijie Yuan and Zhe Zhang and Kai Jia and Zhifang Sui , journal =. 2026 , url = on Pith
  6. [26] Guoxin Chen and Fanzhe Meng and Jiale Zhao and Minghao Li and Daixuan Cheng and Huatong Song and Jie Chen and Yuzhi Lin and Hui Chen and Xin Zhao and Ruihua Song and Chang Liu and Cheng Chen and Kai Jia and Ji-Rong Wen , journal =. 2026 , url = on Pith
  7. [27] Fanzhe Meng and Guoxin Chen and Jiale Zhao and Shuang Sun and Yuzhi Lin and Wayne Xin Zhao and Ruihua Song and Ji-Rong Wen and Kai Jia , journal =. 2026 , url = on Pith
  8. [28] Jia Deng and Yimeng Chen and Xiaoqing Xiang and Ziyang Zeng and Shuo Tang and Zhengbo Zhang and Wayne Xin Zhao and Feng Chang and Chuan Hao and Yuan Wei and Ran Tao and Bryan Dai and Ji-Rong Wen , journal =. 2026 , url = on Pith
  9. [29] Jiale Zhao and Guoxin Chen and Fanzhe Meng and Wayne Xin Zhao and Ruihua Song and Ji-Rong Wen and Kai Jia , journal =. 2026 , url = on Pith
  10. [30] 2026 , month = feb, url = on Pith
  11. [31] arXiv preprint arXiv:2606.19348 , year = on PitharXiv
  12. [32] Sun, Shuang and Chen, Guoxin and Meng, Fanzhe and Deng, Jia and Song, Huatong and Jiang, Jinhao and Zhao, Wayne Xin and Xu, Hongteng and Wen, Ji-Rong , journal =. 2026 , url = on Pith
  13. [33] 2026 , howpublished = on Pith
  14. [34] Statistical Significance Tests for Machine Translation Evaluation , author =. Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing , year = on Pith
  15. [35] The Hitchhiker's Guide to Testing Statistical Significance in Natural Language Processing , author =. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =. doi:10.18653/v1/P18-1128 , url = on Pithdoi:10.18653/v1/p18-1128
  16. [36] Introducing the. 2024 , month = nov, url = on Pith
  17. [37] Zheng, Lianmin and Yin, Liangsheng and Xie, Zhiqiang and Sun, Chuyue and Huang, Jeff and Yu, Cody Hao and Cao, Shiyi and Kozyrakis, Christos and Stoica, Ion and Gonzalez, Joseph E. and Barrett, Clark and Sheng, Ying , title =. Advances in Neural Information Processing Systems ... on Pith
  18. [38] Wang, Hao and Wang, Guozhi and Xiao, Han and Zhou, Yufeng and Pan, Yue and Wang, Jichao and Xu, Ke and Wen, Yafei and Ruan, Xiaohu and Chen, Xiaoxin and Qi, Honggang , journal =. 2026 , url = on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment