Pith. sign in

REVIEW 3 major objections 5 minor 58 references

PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Retained experience genuinely improves personal agents—but only on specific capabilities and sometimes through unintended mechanisms.

desk verdict Worth reading and worth refereeing, but the paper's own acceptance gate is never applied to the reported aggregates, so the 'gains are real' headline is not yet established. read the letter →

arxiv 2608.04003 v1 pith:EG2MMGVW submitted 2026-08-04 cs.CL

Shuhan Xue , Zixin Ding , Yichen Shen , Yinjie Wang , Zhenfei Yin , Yingcheng Wu , Yuxin Chen , Mengdi Wang
show 1 more author
Ling Yang
This is my paper
classification cs.CL
keywords recursiveself-improvementpersonalAIagentspersistentmemorybenchmarkperformanceattributionmechanismevidenceagentframeworksonlineself-evolution
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

This paper asks a foundational question for recursive self-improvement: does a personal agent actually get better at later tasks by retaining and reusing what it learned earlier? PAST-Bench answers it with a controlled protocol—each task family is run twice, once with persistent state available and once with it stripped, holding model, prompt, tools, and seed fixed. Across seven base models and four agent frameworks, every model gains from persistence, but the gains concentrate in different capabilities for different models, and two agents with the same headline gain can differ sharply in whether the gain follows the intended save–retrieve–update pathway. The benchmark therefore separates outcome improvement from mechanism evidence, which is why the authors can diagnose where self-evolution breaks down and then build a framework that targets those specific failures. The paper's contribution is not just another leaderboard but a way to attribute cross-session improvement to retained experience rather than to model capability, runtime design, or scoring noise.

What carries the argument

The carrying object is the 'task family': an ordered sequence of fresh-session episodes (cold, learn, evaluation, control) that share a latent rule, a reusable artifact, a correction, or a preseeded reference, and that are run twice under matched persistence-on and persistence-off conditions so any score gap must flow through the persistent substrate rather than residual prompt overlap. Alongside this, the mechanism-evidence score (Mech) tests whether the gain followed the intended pathway by scoring write precision, recall accuracy, update correctness, retention horizon, and pollution rate from saved artifacts and runtime telemetry.

What would settle it

Instrument the two runs of a task family to compare every non-persistence variable (tool stack, prompt text, context window, retry policy, seed); if any differs systematically between persistence-on and persistence-off, the reported Δ is not attributable to retained experience. More directly: in the persistence-on condition, corrupt or delete the exact memory or skill artifact that the mechanism score says the agent used, then rerun the evaluation episodes; if task scores do not drop measurably, the claimed pathway is not the cause of the gain.

Watch

Extended reading notes

Core claim

PAST-Bench shows that online self-evolution—improvement by reusing retained state without retraining—is measurable, real, and capability-specific. Across seven base models and four agent frameworks, persistence-on runs beat matched persistence-off runs, but the benefit lands on different capabilities for different models. Task-score gains do not imply the intended pathway: Hermes and nanobot both reach Δ=+0.13, but Hermes spreads the gain across all four capabilities with Mech 0.64, while nanobot earns it almost entirely from Update with no consistent write-then-read trace (Mech 0.57). The authors then build Hermes+ with five runtime interventions, one per loop stage, raising Overall Δ from

Load-bearing premise

The whole measurement rests on the matched persistence-on/off pair differing only in access to retained state; if toggling persistence also changes tool availability, prompt construction, or context handling, every reported gain conflates retained state with system-level differences.

Editorial extensions

If this is right

  • Self-evolution quality should be reported as at least two numbers—task gain and mechanism evidence—because equal gains can be supported by different persistence pathways, and one-shot scoring hides that difference.
  • Capability decomposition is necessary: an overall Δ can disguise one capability absorbing most of the benefit while others stagnate or regress, as Agent-Zero does on three of four capabilities.
  • Runtime design matters as much as base model: fixing the model and swapping the framework changes both which capabilities improve and whether the improvement is mechanism-aligned.
  • Targeted interventions at specific loop stages can shift gains onto a target capability: the full Hermes+ places 37% of its movement on Update, versus 23% for Hermes, and reaches a +0.24 Update gap, above any single mechanism.
  • The matched persistence protocol transfers beyond personal agents: both tested general-purpose CLI agents show positive matched gaps on all four capabilities, so the protocol measures a general retained-state effect.

Reading between the lines

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

  • If the matched-control assumption holds, the persistence on/off gap becomes a natural effect-size quantity for comparing agent runtimes, and a standardized version could let future work meta-analyze self-evolution across frameworks without re-running every baseline.
  • The super-additive Update result (full system +0.24 vs. best single mechanism +0.16) suggests persistence mechanisms interact, so ablating one mechanism at a time may understate what a coherent closed loop can do; a testable hypothesis is that the plan-time check (E1) amplifies downstream mechanisms.
  • The benchmark's synthetic, isolated families are a deliberate bound: extending to inter-family transfer and human-derived sessions would test whether retained experience generalizes or overfits to each family's latent rule—the paper's own future-work section sketches exactly this.
  • A useful stress test implied by the findings is to vary episode order or seed multiple conflicting corrections in the same family, since the benchmark currently plants one authoritative second write; this would show whether the update gains survive when stale state competes harder.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces PAST-Bench, a benchmark for measuring whether personal AI agents improve across fresh sessions by using retained experience. The benchmark consists of 26 task families and 204 episodes spanning memory, procedural reuse, information gathering, and update. Each family is run under matched persistence-on/off conditions, with fresh sessions and a context-clearing protocol, so that later-task gains can be attributed to the persistence layer rather than prompt overlap. The authors report task-score gaps (Δ) and a separate mechanism-evidence score (Mech) based on writes, retrievals, updates, and pollution. They evaluate seven base models and four agent frameworks, report that persistence gains are real but uneven across capabilities, and then design Hermes+ by adding five mechanisms to the Hermes framework. Hermes+ is reported to improve the mean Mech score and the Update-specific Δ, while the overall +0.02 Δ gain is acknowledged to be smaller than run-to-run variation. The paper presents PAST-Bench as a diagnostic foundation and Hermes+ as a new reference baseline.

Significance. If the central claim is correct, PAST-Bench is a useful contribution: it is the first benchmark in this space to combine matched persistence-on/off controls with fresh-session evaluation and trace-level mechanism scoring, across both models and frameworks. The design choices are largely sound: the matched ablation is a strong control, the four-capability decomposition is informative, and the authors are appropriately hedged about the difference between behavioral gains and causal evidence. The paper also ships code and detailed reproducibility settings, which is a strength. However, the load-bearing claim that 'every base model gains from persistence' rests on raw Δ aggregates that do not implement the paper's own acceptance gate for control episodes. The current manuscript therefore establishes a benchmark infrastructure and a plausible diagnostic, but the headline empirical conclusion needs to be re-derived under the stated acceptance rule before it can be considered verified.

major comments (3)
  1. [§3.2 vs. Appendix B.2 and Table 10] Section 3.2 states that Δ_f is accepted as evidence of self-evolution only when it clears the upper bound set by the family's control episodes for shortcut, surface-memorization, stale-reuse, and wrong-mechanism explanations. However, Appendix B.2 defines Δ_f simply as the difference of evaluation-bucket means, and no reported table applies the acceptance gate. Table 10 includes families with negative or suspicious Δ (e.g., PC01_sop_bootstrap_02: −0.124, PG06_kappa_integration_review: −0.124 under Hermes), and these are macro-averaged into the Overall Δ. The central claim that 'every base model in Table 2 gains from persistence' therefore rests on aggregates that include families the paper's own criterion would not certify. Please recompute the capability-level and overall Δ after excluding families that fail the control gate, or explicitly report both raw and gated values.
  2. [Tables 2–6 and Appendix D.5] The main tables report point estimates without standard deviations or confidence intervals. The paper itself notes in §4.4 that the +0.02 Overall Δ difference between Hermes and Hermes+ is smaller than run-to-run variation (0.13±0.04 vs. 0.15±0.06). Yet Tables 2, 3, 4, and 6 contain many comparisons of similar or smaller magnitude with no variance information. In particular, the 'uneven across capabilities' claim and the framework comparisons in Table 3 could be driven by noise. Please report run-level variance for all reported configurations, or at least for the headline model/framework comparisons, and indicate which cross-configuration differences exceed run-to-run variation.
  3. [§4.3, §4.4, Eq. (2), and Table 4] The Hermes+ mechanisms E1–E5 were designed after inspecting failures on PAST-Bench, and the Mech metric in Eq. (2) rewards exactly the artifact patterns these mechanisms emit (write precision, recall accuracy, update correctness, retention horizon, low pollution). The paper correctly hedges the overall +0.02 gain and says Mech is not causal proof, but the Update-specific claim (+0.12→+0.24) is presented as the clearest benefit, and E5 was specifically designed for the Update family. This is a post-hoc overfitting risk rather than circularity in the benchmark itself. Please provide a held-out evaluation, a leave-one-family-out analysis, or a pre-registered protocol for the Update claim; otherwise the Update result should be labeled explicitly as a retrospective diagnosis, not as evidence of a generalizable mechanism.
minor comments (5)
  1. [Table 8] The header layout is ambiguous: 'Exact / Within 0.25 / Within 0.25 / Within 0.5' makes it unclear which columns correspond to human–human vs. judge–human agreement. Please restructure the table and report the exact judge–human agreement, which is currently missing.
  2. [Table 10] PC01_sop_bootstrap_05 and PC01_sop_bootstrap_06 have identical Hermes values (0.488/+0.117 for both) and nearly identical Hermes+ values, which looks like a copy-paste error. Please verify.
  3. [§3.2] The sentence 'We accept Δ_f as evidence of self-evolution only when it clears the upper bound set by the family's control episodes' is never operationalized. Please specify the exact formula for the control bound in Appendix B.2.
  4. [§4.4] The phrase 'ties the benchmark's highest configuration' is slightly misleading since the comparison is to the Hermes+GPT-5.4 configuration itself; consider rewording to 'matches the best observed configuration'.
  5. [General] Given that the LLM judge agrees with humans only 68.8% within 0.25 on the four-capability audit, it would be useful to see a per-capability breakdown of judge–human agreement and, if possible, a sensitivity analysis of the main Δs to the judge choice.

Circularity Check

1 steps flagged · score 5.0 of 10

PAST-Bench's Δ is a matched behavioral comparison and is not circular, but Hermes+'s 'clearer pathway evidence' claim is substantially self-referential: the Mech metric rewards exactly the artifact/retrieval signals that E1–E5 were designed to emit.

  1. self definitional [Appendix B.3 (Eq. 2); Section 4.3 (E1–E5); Section 4.4 (Table 4)]
    "Each episode specifies an expectation contract in its family YAML: the expected artifact type (memory, skill, or session_search), required keyword patterns, minimum write/read counts, and retrieval signals. ... Update correctness (uc): ... stale-memory resistance (binary: 1 if the artifact shows updates, changes, or additions; 0 otherwise). ... E5 extracts the final binding key or updated rule at episode close, writes it as the new authoritative artifact in place of the prior value, and flushes it synchronously to the persistent store."

    Mech is defined as compliance with expectation contracts authored by the benchmark, not as an independent check on Hermes+. The five Hermes+ mechanisms were designed after inspecting failures on these same families and they emit exactly the contract signals Mech counts: E4 forces the retrieval-before-answer event that the ra term requires; E2/E3 write typed bindings and ranked skills whose keywords and count-deltas feed wp; E5's overwrite-and-flush directly sets the 'artifact shows updates/changes' bit in uc. Therefore the report that Hermes+ provides 'clearer pathway evidence' (Mech 0.64→0.73) is partly a restatement of the design objective: the mechanisms were built to satisfy the same contract that Mech measures. The paper candidly says Mech is not causal proof, and the behavioral Δ is

full rationale

The core PAST-Bench attribution is not circular: Δ_f is a matched persistence-on/off difference of task scores (Appendix B.2), computed from the same prompt, grader, tools, and seed, with only access to retained state toggled. The claims that improvement is 'real but uneven' and that identical headline Δ can hide different mechanism evidence rest on that behavioral comparison and on a Mech score that is applied uniformly to all frameworks, not just to Hermes+. No load-bearing self-citation, uniqueness theorem, or ansatz-smuggled-via-citation pattern appears; the benchmark is self-contained. Two concerns reduce the confidence in the headline but are not, strictly, circularity: (1) Section 3.2 says Δ_f is accepted as self-evolution evidence only when it clears the family's control bounds, yet Appendix B.2 defines Δ_f as a raw bucket-mean difference and Tables 2–4, 6, and 10 report raw macro-averaged Δs including negative per-family gaps (e.g., PC01_sop_bootstrap_02 −0.124, PG06_kappa_integration_review −0.124), so the 'every base model gains' claim is not shown to satisfy the paper's own acceptance rule; this is a validity gap, not a reduction of the result to its inputs. (2) Hermes+ is diagnosed and evaluated on the same benchmark, which risks overfitting, but overfitting is not circularity unless the metric itself is the construct being optimized. The one genuine circular step is the Mech/Hermes+ alignment described above: the mechanism-evidence score is contract-based, the mechanisms were designed to satisfy those contracts, and the 'clearer pathway evidence' conclusion is therefore partly by construction. Because the behavioral Δ is still independently meaningful and the paper explicitly disclaims causal necessity for Mech, the overall circularity is partial: score 5 rather than higher.

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

The benchmark is an empirical instrument; its design choices (score weights, Mech weights, agent limits, judge settings, and the five Hermes+ mechanisms) are hand-set or tuned on the benchmark itself. The mechanistic attribution of Hermes+ gains is the main circularity concern. No new theoretical entities are introduced.

free parameters (5)
  • Task score weighting: completion 0.80, robustness 0.20 = 0.80 / 0.20
    Chosen by hand in Equation 1; affects every episode score and all aggregate Δ values.
  • Mech component weights = equal 1/5 each
    Equation 2 averages five sub-scores equally; sensitivity analysis only perturbs one component at a time.
  • Agent limits: 25 turns, 300s wall-clock (Hermes/nanobot/ZeroClaw), 1200s for Agent-Zero = 25 turns; 300s / 1200s
    Arbitrary limits that affect which episodes complete; timed-out episodes score zero.
  • LLM judge configuration: MiniMax-M2.7, temperature 0, max 8192 tokens = MiniMax-M2.7, temp 0, 8192
    Judge model and prompt are fixed without variation; judge-human agreement is imperfect.
  • Hermes+ interventions E1-E5 = five mechanism designs
    Prompt and mechanism texts were manually designed from inspection of failure traces on the benchmark, then evaluated on the same benchmark.
assumptions (5)
  • domain assumption The w/o-evolve matched ablation isolates retained experience toggling only
    Section 3.2 claims the two runs share prompt, grader, tools, and seed; if persistence-off also changes tool availability or prompt construction, Δ is contaminated.
  • domain assumption The 26 synthetic task families and control episodes validly operationalize self-evolution capabilities
    Appendix A.2: families are LLM-generated, with only three authors manually checking; the paper later admits limited ecological validity.
  • domain assumption The LLM judge's score is a valid proxy for human judgment
    Table 8: judge-human agreement within 0.25 is 68.8%, within 0.5 is 91.7%, so about 8% of scores differ by more than 0.5.
  • domain assumption The mechanism-evidence contract (artifact type, keyword patterns, counts) captures the intended persistence pathway
    B.3 defines Mech from an expectation contract; the paper explicitly calls it consistency, not causal proof.
  • domain assumption Adapter-standardized frameworks permit meaningful framework comparison
    Appendix C: adapters are not byte-for-byte default deployments; framework differences conflate adapter and framework design.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents." pith.science (2026). https://pith.science/paper/EG2MMGVW

@misc{pith2026260804003,
  author       = {Pith},
  title        = {Pith review of: PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EG2MMGVW}},
  note         = {Machine review of arXiv:2608.04003}
}
read the original abstract

Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain preferences, task histories, tool routines, and learned skills across sessions. Yet whether retained experience actually improves them over time has not been systematically tested. We introduce PAST-Bench, a benchmark designed to isolate this question. Each agent runs through ordered sequences of fresh-session tasks under matched conditions that turn retained experience on and off. It spans 26 scenarios and 204 episodes across memory, procedural reuse, information gathering, and update. We report both later-task gains and whether those gains follow the intended save, retrieve, and update pathway. Across seven base models and four agent frameworks, improvement is real but uneven across capabilities. Agents with the same headline gain can differ markedly in whether that gain is supported by evidence of the intended pathway. Guided by these findings, we develop Hermes+, which extends Hermes with five targeted interventions across stages of the agent loop. Hermes+ raises the average gain from retained experience and provides clearer pathway evidence, with its strongest improvement on tasks requiring outdated state to be replaced, although the effect remains capability- and model-dependent. Together, PAST-Bench and Hermes+ provide an evaluation and diagnostic foundation for studying how persistent agents can progress from retaining experience to systematically improving through it. Code: https://github.com/Gen-Verse/PAST-Bench

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 43 canonical work pages

  1. [1]

    Liu, Xiao and Yu, Hao and Zhang, Hanchen and Xu, Yifan and Lei, Xuanyu and Lai, Hanyu and Gu, Yu and Ding, Hangliang and Men, Kaiwen and Yang, Kejuan and others , booktitle=

  2. [2]

    Koh, Jing Yu and Lo, Robert and Jang, Lawrence and Duvvur, Vikram and Lim, Ming and Huang, Po-Yu and Neubig, Graham and Zhou, Shuyan and Salakhutdinov, Russ and Fried, Daniel , booktitle=

    work page

  3. [3]

    and Del Verme, Manuel and Marty, Tom and Vazquez, David and Chapados, Nicolas and Lacoste, Alexandre , booktitle=

    Drouin, Alexandre and Gasse, Maxime and Caccia, Massimo and Laradji, Issam H. and Del Verme, Manuel and Marty, Tom and Vazquez, David and Chapados, Nicolas and Lacoste, Alexandre , booktitle=. 2024 , publisher=

    work page 2024

  4. [4]

    and Cheng, Zhoujun and Shin, Dongchan and Lei, Fangyu and others , journal=

    Xie, Tianbao and Zhang, Danyang and Chen, Jixuan and Li, Xiaochuan and Zhao, Siheng and Cao, Ruisheng and Hua, Toh J. and Cheng, Zhoujun and Shin, Dongchan and Lei, Fangyu and others , journal=

    work page

  5. [5]

    Kelly and others , booktitle=

    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 others , booktitle=

    work page

  6. [6]

    Ma, Chang and Zhang, Junlei and Zhu, Zhihao and Yang, Cheng and Yang, Yujiu and Jin, Yaohui and Lan, Zhenzhong and Kong, Lingpeng and He, Junxian , journal=

    work page

  7. [7]

    Wu, Di and Wang, Hongwei and Yu, Wenhao and Zhang, Yuwei and Chang, Kai-Wei and Yu, Dong , booktitle=

  8. [8]

    Li, Xiangyi and Liu, Yimin and Chen, Wenbo and You, Bingran and Di, Zonglin and He, Yifeng and Zheng, Shenghan and Choe, Kyoung Whan and Sun, Jiankai and Wang, Shuyi and others , journal=

    work page

Show all 58 references
  1. [9]

    Yang, Yifan and Gong, Ziyang and Huang, Weiquan and Yang, Qihao and Zhou, Ziwei and Huang, Zisu and Li, Yan and Gao, Xuemei and Dai, Qi and Liu, Bei and others , journal=

  2. [10]

    Ouyang, Siru and Yan, Jun and Chen, Yanfei and Han, Rujun and Wang, Zifeng and Mishra, Bhavana Dalvi and Meng, Rui and Li, Chun-Liang and Jiao, Yizhu and Zha, Kaiwen and others , journal=

  3. [11]

    arXiv preprint arXiv:2605.05242 , year=

    Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction , author=. arXiv preprint arXiv:2605.05242 , year=

    arXiv

  4. [12]

    arXiv preprint arXiv:2605.23899 , year=

    From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills , author=. arXiv preprint arXiv:2605.23899 , year=

    arXiv

  5. [13]

    Lin, Huawei and Li, Peng and Song, Jie and Jiang, Fuxin and Zhang, Tieying , journal=

  6. [14]

    Bogavelli, Tara and Sharma, Roshnee and Subramani, Hari , journal=

  7. [15]

    Liang, Shanchao and Garg, Spandan and Moghaddam, Roshanak Zilouchian , booktitle=. The

  8. [16]

    , journal=

    Zhang, Yuxuan and Wang, Yubo and Zhu, Yipeng and Du, Penghui and Miao, Junwen and Lu, Xuan and Li, Zhuofeng and Qu, Xingwei and Guo, Zhengkang and Shen, Yuanzhe and Song, Dingjie and Zhou, Han and Zheng, Tuney and Wu, Xian and Yu, Hao and Cai, Songcheng and Lu, Yi and Hao, Yun...

  9. [17]

    2026 , publisher =

    OpenClaw , title =. 2026 , publisher =

    2026

  10. [18]

    Hermes Agent: The Agent That Grows With You , year =

  11. [19]

    2026 , note=

    Xia, Peng and Zeng, Kaide and Liu, Jiaqi and Qin, Can and Wu, Fang and Zhou, Yiyang and Xiong, Caiming and Yao, Huaxiu , booktitle=. 2026 , note=

    2026

  12. [20]

    AI Open , volume=

    Symbolic learning enables self-evolving agents , author=. AI Open , volume=. 2025 , publisher=

    2025

  13. [21]

    Joshi and Hanna Moazam and Heather Miller and Matei Zaharia and Christopher Potts , booktitle=

    Omar Khattab and Arnav Singhvi and Paridhi Maheshwari and Zhiyuan Zhang and Keshav Santhanam and Sri Vardhamanan A and Saiful Haq and Ashutosh Sharma and Thomas T. Joshi and Hanna Moazam and Heather Miller and Matei Zaharia and Christopher Potts , booktitle=. 2024 , url=

    2024

  14. [22]

    Optimizing Generative

    Yuksekgonul, Mert and Bianchi, Federico and Boen, Joseph and Liu, Sheng and Lu, Pan and Huang, Zhi and Guestrin, Carlos and Zou, James , journal=. Optimizing Generative. 2025 , publisher=

    2025

  15. [23]

    Proceedings of the ACM Conference on AI and Agentic Systems , pages=

    Scaling Textual Gradients via Sampling-Based Momentum , author=. Proceedings of the ACM Conference on AI and Agentic Systems , pages=. 2026 , publisher=

    2026

  16. [24]

    Advances in Neural Information Processing Systems , volume=

    Many-shot in-context learning , author=. Advances in Neural Information Processing Systems , volume=

  17. [25]

    arXiv preprint arXiv:2607.13104 , year =

    Self-Improvements in Modern Agentic Systems: A Survey , author =. arXiv preprint arXiv:2607.13104 , year =

    arXiv

  18. [26]

    arXiv preprint arXiv:2607.15524 , year=

    Recursive Harness Self-Improvement , author=. arXiv preprint arXiv:2607.15524 , year=

    arXiv

  19. [27]

    Advances in Neural Information Processing Systems , volume=

    Recursive introspection: Teaching language model agents how to self-improve , author=. Advances in Neural Information Processing Systems , volume=

  20. [28]

    Yin, Xunjian and Wang, Xinyi and Pan, Liangming and Lin, Li and Wan, Xiaojun and Wang, William Yang , booktitle=. G

  21. [29]

    Wang, Yinjie and Chen, Xuyang and Jin, Xiaolong and Wang, Mengdi and Yang, Ling , journal=

  22. [30]

    arXiv preprint arXiv:2507.21046 , year=

    A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence , author=. arXiv preprint arXiv:2507.21046 , year=

    arXiv

  23. [31]

    Self-Generated In-Context Examples Improve

    Sarukkai, Vishnu and Xie, Zhiqiang and Fatahalian, Kayvon , booktitle =. Self-Generated In-Context Examples Improve

  24. [32]

    Zhang, Shengtao and Wang, Jiaqian and Zhou, Ruiwen and Liao, Junwei and Feng, Yuchen and Li, Zhuo and Zheng, Yujie and Zhang, Weinan and Wen, Ying and Li, Zhiyu and Xiong, Feiyu and Qi, Yutao and Tang, Bo and Wen, Muning , journal =

  25. [33]

    A Comprehensive Survey of Self-Evolving

    Fang, Jinyuan and Peng, Yanwen and Zhang, Xi and Wang, Yingxu and Yi, Xinhao and Zhang, Guibin and Xu, Yi and Wu, Bin and Liu, Siwei and Li, Zihao and others , journal=. A Comprehensive Survey of Self-Evolving

  26. [34]

    Yuan, Jiarui and Jin, Tailin and Chen, Weize and Liu, Zeyuan and Liu, Zhiyuan and Sun, Maosong , journal=

  27. [35]

    Evaluating Very Long-Term Conversational Memory of

    Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei , booktitle=. Evaluating Very Long-Term Conversational Memory of

  28. [36]

    2026 , howpublished =

    How We Broke Top. 2026 , howpublished =

    2026

  29. [37]

    International Conference on Learning Representations , volume=

    Mialon, Gr. International Conference on Learning Representations , volume=

  30. [38]

    Second Conference on Language Modeling , year=

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning , author=. Second Conference on Language Modeling , year=

  31. [39]

    The Thirteenth International Conference on Learning Representations , year=

    Automated Design of Agentic Systems , author=. The Thirteenth International Conference on Learning Representations , year=

  32. [40]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Multi-Agent Collaboration via Evolving Orchestration , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  33. [41]

    ICML 2026 Workshop on Continual Adaptation at Scale: Towards Sustainable AI , year=

    Aligning language models from user interactions , author=. ICML 2026 Workshop on Continual Adaptation at Scale: Towards Sustainable AI , year=

    2026

  34. [42]

    and Stoica, Ion and Gonzalez, Joseph E

    Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , journal =

  35. [43]

    Introduction to Stateful Agents , year =

  36. [44]

    Mem0: Building Production-Ready

    Chhikara, Prateek and Khant, Dev and Aryan, Saket and Singh, Taranjeet and Yadav, Deshraj , booktitle=. Mem0: Building Production-Ready. 2025 , publisher=

    2025

  37. [45]

    Xu, Wujiang and Liang, Zujie and Mei, Kai and Gao, Hang and Tan, Juntao and Zhang, Yongfeng , journal=

  38. [46]

    R-Zero: Self-Evolving Reasoning

    Chengsong Huang and Wenhao Yu and Xiaoyang Wang and Hongming Zhang and Zongxia Li and Ruosen Li and Jiaxin Huang and Haitao Mi and Dong Yu , booktitle=. R-Zero: Self-Evolving Reasoning. 2026 , url=

    2026

  39. [47]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Absolute Zero: Reinforced Self-play Reasoning with Zero Data , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  40. [48]

    arXiv preprint arXiv:2602.21320 , year=

    Tool-R0: Self-Evolving LLM Agents for Tool-Learning from Zero Data , author=. arXiv preprint arXiv:2602.21320 , year=

  41. [49]

    He, Pengfei and Dai, Zhenwei and He, Bing and Liu, Hui and Tang, Xianfeng and Lu, Hanqing and Li, Juanhui and Ding, Jiayuan and Mukherjee, Subhabrata and Wang, Suhang and Xing, Yue and Tang, Jiliang and Dumoulin, Benoit , booktitle=

  42. [50]

    Li, Yu and Luo, Haoyu and Xie, Yuejin and Fu, Yuqian and Yang, Zhonghao and Shao, Shuai and Ren, Qihan and Qu, Wanying and Fu, Yanwei and Yang, Yujiu and Shao, Jing and Hu, Xia and Liu, Dongrui , journal=

  43. [51]

    Evaluating Memory in

    Hu, Yuanzhe and Wang, Yu and McAuley, Julian , booktitle=. Evaluating Memory in

  44. [52]

    2025 , month = aug, note =

    Benchmarking. 2025 , month = aug, note =

    2025

  45. [53]

    Wei, Tianxin and Sachdeva, Noveen and Coleman, Benjamin and He, Zhankui and Bei, Yuanchen and Ning, Xuying and Ai, Mengting and Li, Yunzhe and He, Jingrui and Chi, Ed H and others , journal=

  46. [54]

    Zheng, Junhao and Cai, Xidi and Li, Qiuke and Zhang, Duzhen and Li, ZhongZhi and Zhang, Yingying and Song, Le and Ma, Qianli , journal=

  47. [55]

    Transactions on Machine Learning Research , year=

    Cognitive architectures for language agents , author=. Transactions on Machine Learning Research , year=

  48. [56]

    Journal of cognitive neuroscience , volume=

    Declarative and nondeclarative memory: Multiple brain systems supporting learning and memory , author=. Journal of cognitive neuroscience , volume=. 1992 , publisher=

    1992

  49. [57]

    2014 , publisher=

    The atomic components of thought , author=. 2014 , publisher=

    2014

  50. [58]

    Hong, Sirui and Zhuge, Mingchen and Chen, Jonathan and Zheng, Xiawu and Cheng, Yuheng and Wang, Jinlin and Zhang, Ceyao and Wang, Zili and Yau, Steven Ka Shing and Lin, Zijuan and others , booktitle=

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.