Pith. sign in

REVIEW 3 major objections 4 minor 43 references

LLM agents acting as data-centric researchers can make useful discoveries in most settings, but systematically fail to translate later feedback into consistent improvements.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 01:11 UTC pith:XRSB62GX

load-bearing objection A well-designed benchmark for data-centric research loops with a credible but statistically fragile central finding. the 3 major comments →

arxiv 2607.25886 v1 pith:XRSB62GX submitted 2026-07-28 cs.SE cs.CL

RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement

Fanqing Meng , Lingxiao Du , Qiguang Chen , Ziqi Zhao , Haocheng Lu , Mengkang Hu , Michael Qizhe Shieh This is my paper
classification cs.SE cs.CL
keywords recursive self-improvementdata-centric post-trainingLLM agentsbenchmark designtraining-data synthesisfeedback-driven revisiondiscovery-reliability gapclosed-loop evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether LLM agents can automate the data-centric research part of recursive self-improvement: diagnosing a model's capability gaps, designing training-data strategies, and revising those strategies from checkpoint feedback. To isolate that skill, it fixes the entire post-training stack—base model, LoRA training, serving, evaluation, budgets—and lets four frontier agents iterate on training-data synthesis across six benchmarks. The central empirical finding is a discovery–reliability gap: agents improve on their first valid attempt in 58.33% of settings, but among searches that continue after reaching their best observed score, 78.26% end with a lower-scoring final attempt and the rest only recover the same peak. The paper argues this shows current agents already have core data-centric research capabilities, yet cannot yet turn feedback into dependable improvement.

Core claim

Under a fixed post-training stack, four frontier LLM agents ran iterative data-synthesis loops across six benchmarks spanning software engineering, terminal use, science question answering, and mathematics. The paper's central claim is a discovery–reliability gap: iteration improved on the first valid training-data attempt in 14 of 24 agent–benchmark settings, but among the 23 settings that continued after reaching a peak selection score, 18 ended with a lower final candidate and 5 only matched the peak. The paper interprets this as evidence that agents possess core data-centric research capabilities—hypothesis formation, strategy revision, validation, checkpoint selection—but cannot yet con

What carries the argument

The controlled iterative data-synthesis loop: an agent proposes training-data artifacts and a whitelisted configuration, a shared LoRA SFT backend trains a checkpoint from a fixed base model, a sandboxed evaluator returns permitted selection signals (scores, trajectories, verifier outcomes, costs), and the agent revises its strategy until a budget is exhausted, then selects a checkpoint for official evaluation. This machinery isolates the research component by fixing training, serving, evaluation, and budgets while leaving data strategy and revision decisions free, making each attempt auditable and linking a strategy to its checkpoint's behavior.

Load-bearing premise

The headline statistics come from a single representative run per agent–benchmark setting, with checkpoint selection and official evaluation sharing the same task subset; if that one run is unrepresentative or the shared subset rewards selection over durable skill, the discovery–reliability gap is not established.

What would settle it

Run each of the 24 settings several times with fresh evaluator-private task splits and recompute the two headline rates; if across repeated trials the final-after-peak regression rate does not stay near 78.26%, or if official scores on held-out tasks diverge strongly from selection scores, the gap is an artifact of evaluation noise rather than a stable property of current agents.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Autonomous data-centric research is measurable: with a fixed stack, differences in outcome can be attributed to the agent's research policy rather than infrastructure changes.
  • Historical-best checkpoint selection materially protects final performance: across GPQA, AIME, and Terminal-Bench, every run selected a checkpoint matching its best observed selection score, even when later attempts regressed.
  • Reasoning effort changes the research policy itself, not just compute: a max-effort diagnostic traded fewer candidates for larger, more expensive datasets and a stronger first candidate.
  • Current agents are not yet reliable enough to close the recursive self-improvement loop: several trained checkpoints fail to beat the unadapted base model, and post-peak revisions frequently degrade the candidate frontier.
  • Benchmark family labels do not determine search dynamics; the agent–benchmark interaction is the relevant unit, so progress should be measured per setting rather than across broad task families.

Where Pith is reading between the lines

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

  • If the discovery–reliability gap holds up under repeated trials, then the field should benchmark stop/rollback policies and 'knowing when to stop' as first-class research capabilities, not just as side effects of final-score evaluation.
  • Selection and official scores diverge by about 3.9 points on average, so optimizing a development-set selection metric could actively mislead; a natural extension is to add a held-out selection set and test whether agents that do well on it also do well officially.
  • Several trajectories showed late discrete jumps after many weak rounds, suggesting that adaptive budget allocation across candidates might convert more of the ten settings where the first candidate remained best into actual improvements.
  • The 78.26% regression rate implies that simple 'keep trying harder' loops can be wasteful; a testable fix is to give agents explicit regression-triggered reflection or rollback mechanisms and measure whether the post-peak failure rate drops.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces RSIBench-Data, a benchmark that isolates the data-centric research component of recursive self-improvement. Four frontier LLM agents (Claude Code Opus-4.8/Sonnet-5 and Codex gpt-5.6-sol/terra) act as researcher agents that iteratively propose training-data strategies for a fixed Qwen3.5-35B-A3B base model through a shared LoRA SFT service; evaluation is run in E2B sandboxes through Harbor. Six benchmarks are used: three SWE-bench variants, Terminal-Bench 2.0, GPQA Diamond, and AIME 2026. The central empirical finding is a 'discovery–reliability gap': iteration improves on the first valid candidate in 14/24 settings (58.33%), but among 23 searches that continue after reaching a peak, 18 (78.26%) end with a lower-scoring final attempt and the remaining 5 only recover the peak. The paper also reports trajectory analyses identifying mechanisms such as accurate hypotheses, validation-grounded supervision, behavior-aligned data, and checkpoint preservation, plus an exploratory same-family RSI experiment.

Significance. If the empirical claims hold, RSIBench-Data would be a valuable, auditable testbed: it fixes the training/serving/evaluation stack, makes the loop recordable, and provides a concrete measurement of non-monotonic feedback-driven research. The authors are also appropriately careful in places: they distinguish selection scores from official scores, report resource usage, and explicitly list limitations in Section 6. The benchmark artifact itself, with open-source code, is a useful contribution. However, the headline quantitative claims rest on a single run per cell and on a fixed evaluation subset shared between selection and official evaluation; the current data do not yet establish the broad conclusion about 'current agents' that the abstract and conclusion assert. The paper's own Section 6 concession should govern the interpretation.

major comments (3)
  1. [Section 5.1 and Table 2] Each agent–benchmark cell is a single 'pre-specified representative run'; no repeated trials or error bars are reported. The headline rates 14/24 (58.33%) and 18/23 (78.26%) therefore have wide binomial uncertainty (approximate 95% intervals of roughly 37–78% and 56–93%, respectively), and the trajectory-level statements in §5.2 are single paths. Since the conclusion describes a property of 'current agents' generally, the paper needs repeated runs at least for a subset of cells and a reporting convention that separates stochastic variation from policy-level effects. Otherwise the 'discovery–reliability gap' is a suggestive pattern, not an established result.
  2. [Sections 3.4, 4.3, and 6] Selection feedback includes task trajectories, verifier outcomes, and execution diagnostics from the same fixed evaluation subset that is later used for official evaluation. The protocol therefore permits the agent's revisions to be conditioned on the exact official task distribution, and the official score measures fit to a fixed sample rather than held-out generalization. This is not a minor caveat: it is load-bearing for the claim that agents 'can make useful data-centric discoveries,' because a candidate that scores well on the subset may be overfitting to the selection signal. The paper's own Section 6 concession ('checkpoint selection and official evaluation use the same task subset... not generalization to statistically held-out tasks') should be reflected in the abstract and conclusion. An evaluator-private split or refreshed tasks is needed before the broad conclusion can be dra
  3. [Section 5.2 and abstract] The 'improve upon the first valid attempt' statistic is computed from selection scores (first vs best valid candidate under the same selection protocol), not from official evaluation of the two checkpoints. The abstract and conclusion state it without this caveat. Since selection and official scores differ substantially in places (e.g., §5.3: Codex gpt-5.6-sol on AIME rises from 48.33% selection to 53.33% official; Claude Code Sonnet-5 on GPQA falls from 64.00% to 52.00%), the 58.33% figure cannot be read as an official-evaluation improvement rate. Please state explicitly which protocol supports each headline number.
minor comments (4)
  1. [Typos and presentation] Figure 1(a) caption has 'post-taining'; Figure 3 axis shows 'T erminal-Bench 2.0'; §5.5 has 'In conclusion, Although' with a capitalization/grammar error. These should be cleaned up.
  2. [Section 4.1] Clarify that 'Claude Code' and 'Codex' are agent scaffolds; the model version names (Opus-4.8, Sonnet-5, gpt-5.6-sol, gpt-5.6-terra) do not uniquely determine agent behavior. Currently the reader must infer the scaffold/model relationship.
  3. [Section 4.3] The fixed 100-instance subsets for SWE-bench variants and GPQA Diamond should be defined by task IDs or an explicit deterministic selection rule. Without this, readers cannot assess overlap with the public seed repositories or contamination risk, which is central to the benchmark's validity.
  4. [References] The DataComp-LM reference appears to have a corrupted author name ('Saurabh Garg Jiang'); verify all references for formatting and completeness.

Circularity Check

0 steps flagged

No significant circularity: benchmark scores are measured against external benchmarks; the same-subset evaluation concern is an acknowledged validity limitation rather than a definitional reduction.

full rationale

RSIBench-Data is an empirical benchmark paper, not a derivation. The headline statistics (58.33%, 78.26%) are descriptive summaries of measured agent trajectories across external benchmarks (SWE-bench variants, Terminal-Bench 2.0, GPQA Diamond, AIME 2026), and the reported official scores are outputs of a fixed evaluation protocol rather than quantities fitted from the paper's own assumptions. The paper's Section 3.1 explicitly separates selection feedback from official evaluation, and Section 6 acknowledges that 'checkpoint selection and official evaluation use the same task subset' and that 'repeated trials, evaluator-private splits, and refreshed tasks are needed'; this is a generalization-validity limitation, not circularity, because no fitted parameter is renamed as a prediction and no defining equation reduces the conclusion to its input. The only author-overlapping citations (e.g., Chen et al., 2026a in the introduction) support background remarks about agent trajectories and are not load-bearing for the benchmark's central claim. No self-definitional, fitted-input-as-prediction, self-citation-load-bearing, uniqueness-imported-from-authors, or ansatz-smuggling pattern is present, so the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claims rest on protocol choices that are not externally calibrated: fixed budgets, fixed subsets, a single representative run per setting, and the assumption that four proprietary agents represent the frontier. The paper is transparent about some of these (Section 6), but the headline percentages should be read as conditional on these choices.

free parameters (5)
  • Per-run budget (wall-clock and Tinker cost) = 16 hours / $500 for main matrix; 20 hours / $2000 for early RSI
    Hand-chosen budget constraint; it shapes how many attempts an agent can make and may explain part of the discovery–reliability gap.
  • Evaluation subset sizes = 100 (SWE-bench Verified/Multilingual/Pro and GPQA), 89 (Terminal-Bench), 30 (AIME)
    Hand-chosen fixed subsets; scores are not comparable to full-benchmark leaderboards and may not generalize, as the paper acknowledges.
  • Reasoning effort settings = high (Claude Code), max (Codex) in main matrix
    Hand-chosen per agent family; the paper runs one diagnostic comparing high vs max effort on SWE-bench Verified and reports a 52% vs 35% official score difference, showing sensitivity.
  • Training pipeline hyperparameters (LoRA rank, learning rate, steps)
    Shared across all agents but chosen by the benchmark authors; the exact values are not in the text. These interact with data strategy and could affect the measured capabilities.
  • Number of representative runs = 1 per setting
    The headline percentages are computed from one run per agent–benchmark cell; no repeated trials or error bars.
axioms (4)
  • domain assumption Selection scores from the same task subset are a valid basis for checkpoint selection and iteration during search.
    The entire closed-loop protocol relies on the assumption that feedback from the fixed evaluation subset is informative for improving the training-data strategy. The paper uses this for all search decisions (Section 4.2, Table 4).
  • domain assumption The four proprietary agent systems (Claude Code with Opus-4.8/Sonnet-5, Codex with gpt-5.6-sol/terra) are representative of 'current agents' at frontier level.
    The paper generalizes from these four systems to 'current agents' in the abstract and conclusion. No evidence is provided that other agents, or different reasoning-effort settings, would behave similarly.
  • domain assumption LoRA SFT on Qwen3.5-35B-A3B-Base via the shared Tinker service is an adequate optimization interface for eliciting data-centric research capability.
    The benchmark fixes this training backend; if a different backend (e.g., full fine-tuning, RL) would allow better data strategies to shine, the measured capability ceiling is an artifact of the chosen stack.
  • domain assumption The public seed repositories and data-use constraints prevent evaluation-task contamination of the training data.
    The protocol forbids using evaluation-only tasks as supervision (Section 3.3), but contamination could still occur via the shared seed repos; the paper provides no automated contamination test.
invented entities (2)
  • Data-centric researcher agent no independent evidence
    purpose: The paper's central unit of analysis: an LLM agent that iteratively forms capability hypotheses and revises training-data strategies under a fixed training/evaluation stack.
    This is a conceptual role defined by the paper; there is no externally verified instrument or measurement associated with it beyond the benchmark itself.
  • Discovery–reliability gap no independent evidence
    purpose: A named pattern describing the observation that agents improve on their first attempt in 58.33% of settings but regress after peaking in 78.26% of continuing searches.
    The pattern is computed from the paper's own single-run trajectories; no independent dataset confirms it.

pith-pipeline@v1.3.0-alltime-deepseek · 16905 in / 14063 out tokens · 110246 ms · 2026-08-01T01:11:12.294751+00:00 · methodology

0 comments
read the original abstract

Recursive self-improvement requires turning evidence of model failures into better models. Data-centric post-training research entails diagnosing capability gaps, designing and validating training-data strategies, and learning from checkpoint feedback. Can LLM agents automate this loop? Existing benchmarks entangle research decisions with optimization, serving, evaluation, and systems implementation, obscuring agents' research capability. We introduce RSIBench-Data, a controlled benchmark of LLM agents as data-centric researchers with a fixed post-training stack. Agents iteratively revise training-data strategies for a fixed target model; training and serving use Tinker-backed services, official evaluation runs through Harbor and E2B sandboxes, and budgets are fixed across agents. We evaluate four frontier agents on six benchmarks across software engineering, terminal use, scientific question answering, and mathematics. Agents demonstrate core data-centric research capabilities: in 58.33\% of settings, they improve upon the first valid attempt by refining strategies from feedback. However, improvement is inconsistent. Among searches continuing after the best observed score, 78.26\% end with a lower-scoring final attempt, while the rest only recover the same peak. A strong candidate may therefore appear early or midway through a run even as later revisions fail. Trajectory analysis identifies four patterns in stronger runs: accurate hypotheses, validation-grounded supervision, behavior-aligned data, and preservation of strong checkpoints. These findings suggest that current agents can make useful data-centric discoveries but cannot yet translate feedback into consistent improvements. RSIBench-Data provides a measurable, auditable testbed for the research capabilities required for recursive self-improvement. We open-source our code at https://github.com/evolvent-ai/RSIBench-Data.

Figures

Figures reproduced from arXiv: 2607.25886 by Fanqing Meng, HaoCheng Lu, Lingxiao Du, Mengkang Hu, Michael Qizhe Shieh, Qiguang Chen, Ziqi Zhao.

Figure 1
Figure 1. Figure 1: Existing evaluations entangle multiple post-training choices; RSIBench-Data isolates the LLM agent’s ability to conduct data-centric post-training research. evidence-to-training pipeline [Gadre et al., 2023, Pan et al., 2024, Rank et al., 2026]. SWE-Gym[Pan et al., 2024] and R2E-Gym [Jain et al., 2025] provide executable environments and training resources for software-engineering agents. DataComp [Gadre e… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of RSIBench-Data, which evaluates how an LLM researcher agent evolves a training-data research policy under shared training, serving, evaluation, and budget controls. 3.2. Fixed Infrastructure RSIBench-Data fixes the infrastructure that turns a training-data strategy into an evaluated checkpoint, so that a researcher agent works inside a fixed boundary rather than over the whole post-training stac… view at source ↗
Figure 3
Figure 3. Figure 3: Selection-score trajectories across valid attempts in their original order. Stars mark the first candidate that reaches the historical-best selection score. hours with $4.80. Other settings consume far more budget for modest or weak outcomes: Claude Code Sonnet-5 on Terminal-Bench 2.0 uses 8.87 hours and $156.93 but scores 5.62%, while Codex gpt-5.6-sol on SWE-bench Pro spends $300.49 to reach 9.00%. These… view at source ↗
Figure 4
Figure 4. Figure 4: Selection score versus cumulative Tinker cost across benchmarks. Filled markers indicate candidates that establish a new within-run best, hollow markers indicate the remaining candidates, and stars mark selected checkpoints. The value and speed of iteration are strongly benchmark dependent. To distinguish where iteration helps from how quickly it helps, we compare first-to-best gains and time to peak. AIME… view at source ↗
Figure 5
Figure 5. Figure 5: Reasoning-effort diagnostic on SWE￾bench Verified with the Claude Code Sonnet-5 researcher agent. The figure compares within-run selection scores under high and max effort. Stars mark the first historical-best candidate, and dia￾monds show the selected checkpoint’s official eval￾uation score. Higher reasoning effort improves this diagnostic setting by changing the training-data research policy. We compare … view at source ↗
Figure 6
Figure 6. Figure 6: Early RSI trajectory on SWE-bench Pro. The Kimi K2.6 researcher evolves its synthetic-data strategy across seven Tinker LoRA attempts, improving the candidate score but remaining below the 33% unadapted-model reference. Results. The trajectory shows clear strategy evolution but no effective model improvement. The agent first converts 1,500 resolved SWE-smith trajectories into full assistant–tool conversati… view at source ↗
Figure 7
Figure 7. Figure 7: Representative evolution trajectories on AIME 2026, SWE-bench Pro, and Terminal-Bench 2.0. Circles show valid-attempt selection scores, stars mark selected historical-best candidates, and open diamonds show official evaluations. B.5. Instantiation Across Benchmark Families For software-engineering benchmarks, the loop may connect repository state, issue interpretation, code localization, patch generation, … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 19 linked inside Pith

  1. [1]

    and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik , year =

    Jimenez, Carlos E. and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik , year =. 2310.06770 , archivePrefix =

    Pith/arXiv arXiv

  2. [2]

    Yao, Shunyu and Shinn, Noah and Razavi, Pedram and Narasimhan, Karthik , booktitle =

  3. [3]

    2304.14108 , archivePrefix =

    Gadre, Samir Yitzhak and Ilharco, Gabriel and Fang, Alex and Hayase, Jonathan and Smyrnis, Georgios and Nguyen, Thao and Marten, Robert and Wortsman, Mitchell and Ghosh, Dhruba and Zhang, Jieyu and Orgad, Eyal and Entezari, Rahim and Daras, Giannis and Pratt, Sarah and Ramanujan, Vivek and Bitton, Yonatan and Marathe, Jitesh and Mussmann, Stephen and Venc...

    Pith/arXiv arXiv

  4. [4]

    2406.11794 , archivePrefix =

    Li, Jeffrey and Fang, Alex and Smyrnis, Georgios and Ivgi, Maor and Jordan, Matt and Gadre, Samir Yitzhak and Bansal, Hritik and Guha, Etash and Keh, Sedrick and Arora, Kushal and Jiang, Saurabh Garg and Dziugaite, Gintare Karolina and Wortsman, Mitchell and Ilharco, Gabriel and Schmidt, Ludwig and Hajishirzi, Hannaneh , year =. 2406.11794 , archivePrefix =

    Pith/arXiv arXiv

  5. [5]

    2603.08640 , archivePrefix =

    Rank, Ben and Bhatnagar, Hardik and Prabhu, Ameya and Eisenberg, Shira and Nguyen, Karina and Bethge, Matthias and Andriushchenko, Maksym , year =. 2603.08640 , archivePrefix =

    arXiv

  6. [6]

    2604.10547 , archivePrefix =

    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 =. 2604.10547 , archivePrefix =

    Pith/arXiv arXiv

  7. [7]

    Advances in Neural Information Processing Systems , year =

    The Reusable Holdout: Preserving Validity in Adaptive Data Analysis , author =. Advances in Neural Information Processing Systems , year =

  8. [8]

    2412.21139 , archivePrefix =

    Pan, Jiayi and Wang, Xingyao and Neubig, Graham and Suhr, Alane and Fried, Daniel , year =. 2412.21139 , archivePrefix =

    Pith/arXiv arXiv

  9. [9]

    2504.07164 , archivePrefix =

    Jain, Naman and Han, King and Li, Alex Gu and Zhang, Wen-Ding and Albalak, Alon and Wang, Yuxiang and Kim, Wooseok and Li, Chen and Abu-Salih, Bashar and Hu, Yebowen and Habib, Andrew and Raman, Karthik and Neel, Susmit and Dziri, Nouha and Keutzer, Kurt and Stoica, Ion , year =. 2504.07164 , archivePrefix =

    Pith/arXiv arXiv

  10. [10]

    2505.20411 , archivePrefix =

    Badertdinov, Izzat and Golubev, Yaroslav and Mikhrin, Dmitry and Afanasiev, Mikhail and Kovrigin, Danil and Abramov, Mikhail and Vatsurenko, Alexey and Timofeev, Vladimir and Khabibullin, Marat and Khachaturov, David and Sadykov, Timur and Khoroshilov, Alexey , year =. 2505.20411 , archivePrefix =

    arXiv

  11. [11]

    2505.23419 , archivePrefix =

    Zhang, Linghao and He, Shilin and Zhang, Chaoyun and Kang, Yu and Li, Bowen and Xie, Chengxing and Wang, Junhao and Wang, Maoquan and Huang, Yufan and Fu, Shengyu and Nallipogu, Elsie and Lin, Qingwei and Dang, Yingnong and Rajmohan, Saravan and Zhang, Dongmei , year =. 2505.23419 , archivePrefix =

    Pith/arXiv arXiv

  12. [12]

    2025 , eprint=

    Multi-SWE-bench: A Multilingual Benchmark for Issue Resolving , author=. 2025 , eprint=

    2025

  13. [13]

    2025 , eprint=

    SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks? , author=. 2025 , eprint=

    2025

  14. [14]

    2026 , eprint=

    Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces , author=. 2026 , eprint=

    2026

  15. [15]

    2023 , eprint=

    GPQA: A Graduate-Level Google-Proof Q&A Benchmark , author=. 2023 , eprint=

    2023

  16. [16]

    2024 , eprint =

    Large Language Model based Multi-Agents: A Survey of Progress and Challenges , author =. 2024 , eprint =

    2024

  17. [17]

    Chen, Jianwen and Wang, Tonghui and Mo, Shentong and Mao, Xiaofeng and Wang, Ming and Guo, Changran and Zhang, Alan Y. H. and Shen, Tieliang and Wang, Qidong and Yasseen, Amr and Du, Mengnan and Shi, Weishi and Ji, Shuiwang and Cho, Kyunghyun and Zeng, Minghua and Shen, Dinggang and Wang, Zhangyang and Wei, Hongxin , year =. 2507.01903 , archivePrefix =

    Pith/arXiv arXiv

  18. [18]

    2603.08806 , archivePrefix =

    Barke, Shraddha and Kuo, Jialin and Polikarpova, Nadia , year =. 2603.08806 , archivePrefix =

    arXiv

  19. [19]

    Science China Information Sciences , volume =

    Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models , author =. Science China Information Sciences , volume =. 2026 , publisher =

    2026

  20. [20]

    and Ma, Tengyu and Yu, Adams Wei , year =

    Xie, Sang Michael and Pham, Hieu and Dong, Xuanyi and Du, Nan and Liu, Hanxiao and Lu, Yifeng and Liang, Percy and Le, Quoc V. and Ma, Tengyu and Yu, Adams Wei , year =. 2305.10429 , archivePrefix =

    Pith/arXiv arXiv

  21. [21]

    2407.01492 , archivePrefix =

    Liu, Qian and Zheng, Xiaosen and Muennighoff, Niklas and Zeng, Guangtao and Dou, Longxu and Pang, Tianyu and Jiang, Jing and Lin, Min , year =. 2407.01492 , archivePrefix =

    Pith/arXiv arXiv

  22. [22]

    2402.04333 , archivePrefix =

    Xia, Mengzhou and Malladi, Sadhika and Gururangan, Suchin and Arora, Sanjeev and Chen, Danqi , year =. 2402.04333 , archivePrefix =

    Pith/arXiv arXiv

  23. [23]

    2023 , eprint =

    Self-Instruct: Aligning Language Models with Self-Generated Instructions , author =. 2023 , eprint =

    2023

  24. [24]

    2304.12244 , archivePrefix =

    Xu, Can and Sun, Qingfeng and Zheng, Kai and Geng, Xiubo and Zhao, Pu and Feng, Jiazhan and Tao, Chongyang and Jiang, Daxin , year =. 2304.12244 , archivePrefix =

    Pith/arXiv arXiv

  25. [25]

    2024 , eprint =

    Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing , author =. 2024 , eprint =

    2024

  26. [26]

    2410.07706 , archivePrefix =

    Song, Yifan and Xiong, Weimin and Zhao, Xiutian and Zhu, Dawei and Wu, Wenhao and Wang, Ke and Li, Cheng and Peng, Wei and Li, Sujian , year =. 2410.07706 , archivePrefix =

    Pith/arXiv arXiv

  27. [27]

    2412.09605 , archivePrefix =

    Xu, Yiheng and Lu, Dunjie and Shen, Zhennan and Wang, Junli and Wang, Zekun and Mao, Yuchen and Xiong, Caiming and Yu, Tao , year =. 2412.09605 , archivePrefix =

    Pith/arXiv arXiv

  28. [28]

    2023 , eprint =

    Reflexion: Language Agents with Verbal Reinforcement Learning , author =. 2023 , eprint =

    2023

  29. [29]

    2604.06008 , archivePrefix =

    Gao, Kuofeng and Hong, Sirui and He, Bryan and Liu, Tengxiao and Shu, Yiheng and Zheng, Huaixiu Steven , year =. 2604.06008 , archivePrefix =

    Pith/arXiv arXiv

  30. [30]

    2026 , eprint =

    Training Language Agents to Learn from Experience , author =. 2026 , eprint =

    2026

  31. [31]

    2603.02007 , archivePrefix =

    Hu, Jiani and Xie, Jianwen and Su, Yixuan and Yin, Zhenfei and Li, Qian and An, Bo and Li, Xiu , year =. 2603.02007 , archivePrefix =

    arXiv

  32. [32]

    2606.07412 , archivePrefix =

    Xiao, Bin and Liu, Pengfei and Liu, Qian , year =. 2606.07412 , archivePrefix =

    Pith/arXiv arXiv

  33. [33]

    2606.00458 , archivePrefix =

    Zhang, Shenao and Zhang, Zixuan and Song, Chiyu and Zhou, Wuhang and Liu, Jiawei and He, Junxian and Koh, Pang Wei and Wang, Liwei , year =. 2606.00458 , archivePrefix =

    Pith/arXiv arXiv

  34. [34]

    2604.01658 , archivePrefix =

    Qu, Ao and Zheng, Han and Zhou, Zijian and Yan, Yihao and Tang, Yihong and Ong, Shao Yong and Hong, Fenglu and Zhou, Kaichen and Jiang, Chonghe and Kong, Minwei and Zhu, Jiacheng and Jiang, Xuan and Li, Sirui and Wu, Cathy and Low, Bryan Kian Hsiang and Zhao, Jinhua and Liang, Paul Pu , year =. 2604.01658 , archivePrefix =

    Pith/arXiv arXiv

  35. [35]

    2025 , eprint =

    Evaluating Language Models as Synthetic Data Generators , author =. 2025 , eprint =

    2025

  36. [36]

    2406.07275 , archivePrefix =

    Huang, Benhao and Yu, Yingzhuo and Huang, Jin and Zhang, Xingjian and Ma, Jiaqi , year =. 2406.07275 , archivePrefix =

    Pith/arXiv arXiv

  37. [37]

    2025 , eprint =

    Khan, Zaid and Stengel-Eskin, Elias and Cho, Jaemin and Bansal, Mohit , booktitle =. 2025 , eprint =

    2025

  38. [38]

    2026 , eprint =

    Can Generalist Agents Automate Data Curation? , author =. 2026 , eprint =

    2026

  39. [39]

    2025 , eprint =

    Qwen3 Technical Report , author =. 2025 , eprint =

    2025

  40. [40]

    and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik R

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

    2024

  41. [41]

    2025 , howpublished =

    Tinker: A Training. 2025 , howpublished =

    2025

  42. [42]

    2025 , howpublished =

    Claude Code , author =. 2025 , howpublished =

    2025

  43. [43]

    2025 , howpublished =

    Codex. 2025 , howpublished =

    2025