Artificial Intelligence · arXiv:2610.23599 v1 · submitted 16 Oct 2026 · 59 references
From Trajectories to Differentiated Skills: Cognitive Differentiation for Self-Improving Agents
The paper says
“Large language model (LLM) agents tackle complex tasks through reasoning and interaction with external environments. Agent skills provide persistent, transferable capabilities that guide task planning, inform contextual decisions, support tool use, and enable executable procedures.”
more of the abstract
Yet existing skill-evolution methods largely organize knowledge around skill artifacts, overlooking the functional heterogeneity within individual skills and its implications for knowledge maintenance and use. Inspired by cognitive science, we propose Cognitive Differentiation in Skill Evolution (CDSE), which evolves a unified skill through four functionally distinct yet coordinated components: Meta, Experience, Tool, and Script. CDSE extracts evidence-grounded semantic gradients from successful and failed trajectories, consolidates them by functional role, and applies component-specific updates and validation with cross-component coordination and adaptive invocation. An analysis of 500 top-ranked skill packages shows that 96.8% contain knowledge spanning at least two functional roles. Across six benchmarks, three agent models, and two execution harnesses, CDSE achieves the highest scores in all 36 settings, improving over the Base agent by 7.6 percentage points and the strongest competing method in each setting by 2.6 points on average. Ablation studies further support the importance of functional differentiation and adaptive invocation.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 19 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A well-motivated framework for functionally differentiated skill evolution with impressively broad experiments, but the central claim of universal superiority is not statistically supported due to missing variance estimates and very small margins in several settings. This line is wrong
1 major objection · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 59 references say.
The paper says, at §4.2, Table 1
The review did not quote this passage.
Pith says: Objection 1 of 1 · major
The central claim that CDSE achieves the highest score in all 36 settings rests entirely on single-run point estimates without any reported measure of variability (standard deviations, confidence intervals, or multiple independent trials). In several configurations the margin over the best baseline is very small; the narrowest is 1.0 percentage point on LiveMathBench for MiniMax-M3 under Codex (CDSE 74.5% vs.
more of objection 1
SkillOpt-lite 73.5%). Without variance estimates or significance testing, it is impossible to assess whether these differences are statistically reliable or could be reversed by stochasticity in model outputs, environment interactions, or the skill-evolution process. The universal claim of superiority across all 36 settings is therefore not formally supported in its current form. The authors should either (a) provide variability measures (e.g., 3–5 independent runs with different seeds), or (b) report confidence intervals via bootstrapping, or (c) soften the claim to reflect that not all individual comparisons may be statistically significant. This is the most critical issue for publication.
6 minor objections
The paper says, at §4.3, Figure 4
Pith says: minor objection 1 of 6
The y-axis label 'Removal delta' is ambiguous. Clarify in the caption that the bars represent the change in task score (in percentage points) when the indicated component is removed, e.g., 'Change in accuracy when component is removed (pp, negative = decrease)'.
The paper says, at Table 2, caption
Pith says: minor objection 2 of 6
The parenthesized values in Table 2 are described as 'explicit component-loading rates' in the body text, but the caption does not explain what these numbers represent or that descriptions remain visible before loading (which is important for interpreting zero loading). Include the explanation in the caption.
show 4 more
The paper says, at §4.4, Figure 5
Pith says: minor objection 3 of 6
The y-axis label 'Score change from R0 (pp)' is clear, but the caption should explicitly state that the same trajectory corpus is reused across rounds (this is mentioned later but should be in the caption or the main text description of the figure at first mention).
The paper says, at Appendix F.1, Table 13
Pith says: minor objection 4 of 6
The validation of GPT-6-Astra annotations against 100 expert-annotated packages shows macro-average F1=0.875. While acceptable, the 95% confidence intervals for these F1 scores would help gauge reliability. Consider adding them.
The paper says, at §3.2, Eq. (2)
Pith says: minor objection 5 of 6
The notation A_{s(y_i)} is used for the analyzer conditioned on the success/failure label. It would be helpful to define explicitly that s(y_i) ∈ {+, −} maps the outcome to a binary label, as is done later in Appendix G.1 but not in the main text.
The paper says, at Appendix H.4, OfficeQA template
Pith says: minor objection 6 of 6
The command line includes a JSON schema for final_answer and reasoning. If this schema is used for answer extraction, the scoring procedure (Appendix H.3) should confirm that the reasoning field does not influence the task score. This is presumably the case, but stating it explicitly avoids confusion.
What would settle it
Re-run the LiveMathBench evaluation for MiniMax-M3 under Codex at least five times with different random seeds (e.g., seeds 1–5), using the same pre-evolved skill state. Compute the mean and standard deviation of CDSE and the strongest baseline (SkillOpt-lite). If the 95% confidence intervals overlap, then the claim that CDSE achieves the highest score in this setting is not statistically supported. Repeat this test for the two or three other smallest-margin settings (e.g., SealQA for GLM-5.2/Claude Code, and SpreadsheetBench2 for MiniMax-M3/Codex). If any of these comparisons yield non-significant differences, the headline claim of 'highest scores in all 36 settings' must be qualified or the paper should report effect sizes with confidence intervals.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 19 Oct 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 1 major objection, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
This paper does something genuinely new: it organizes trajectory-derived agent skill knowledge into four functionally distinct components (Meta, Experience, Tool, Script) with component-specific update and adaptive invocation. The motivation from cognitive science and the analysis of 500 real-world skill packages make the case that functional heterogeneity matters. The experimental breadth is a real strength—36 settings across six benchmarks, three models, and two harnesses. Ablations and transfer studies give a useful picture of how the components behave. The consistent pattern of improvement is hard to dismiss as noise.
The soft spot is the statistical case for the central claim. The paper says CDSE 'achieves the highest scores in all 36 settings,' but every reported number is a point estimate with no measure of variability—no standard deviations, confidence intervals, or mention of multiple runs. The stress-test note is right: several margins are razor-thin. The narrowest is 1.0 percentage point on LiveMathBench for MiniMax-M3 under Codex (74.5 vs. 73.5). Another is 1.7 points on SealQA. Without knowing whether these differences could flip under stochasticity, the universal claim is not formally supported. The paper does not say whether it used fixed seeds or temperature 0, so we have to assume variability is plausible.
That said, the overall pattern across 36 settings is strong evidence that CDSE is genuinely better on average. The flaw is in how the claim is framed, not in the underlying engineering or experimental design. The paper also acknowledges its own limitations: fixed trajectory corpus for multi-round, no proof that exactly four components are optimal, no code or data released yet. These are honest and appropriate.
Who this is for: researchers in LLM self-improvement, agent skill architectures, and anyone thinking about principled knowledge organization in agents. It deserves a serious referee—the idea is worthwhile and the experiments are extensive enough to merit discussion. But the authors should be asked to report variance (multiple runs or bootstrapped intervals) and to tone down the universality claim. Minor revision.
Recommendation: send to peer review, but flag the statistical support issue prominently.
The strongest objection
The central claim that CDSE achieves the highest scores in all 36 evaluation settings rests on point estimates without any reported measure of variability (standard deviations, confidence intervals, or multiple independent trials). In several configurations the margin over the best baseline is very small. The narrowest such margin occurs on LiveMathBench for MiniMax-M3 under Codex: CDSE scores 74.5% vs. SkillOpt-lite at 73.5%, a difference of only 1.0 percentage point. Other small margins include SealQA for GLM-5.2/Claude Code (CDSE 54.1% vs. EvoSkill 52.4%, delta 1.7 pp) and AutomationBench for DeepSeek-V4-Flash/Claude Code (CDSE 28.9% vs. SkillOpt-lite 22.6%, delta 6.3 pp, though that is larger). Without variance estimates or multiple runs, it is impossible to assess whether these differences are statistically significant or could be reversed by stochasticity in model outputs, environment interactions, or the skill-evolution process itself. The paper does not specify whether the evaluation uses a fixed seed or temperature 0 for agent generations, so variability is plausible. If even a single configuration fails to show a reliable advantage, the universal claim of superiority across all 36 settings is undermined. While the overall pattern across many settings is strong evidence, the absence of any significance testing means the claim is not formally supported in its current form.
The reader's verdict is CONDITIONAL, with the weakest assumption being the necessity and sufficiency of the four-component decomposition. While that is a reasonable conceptual concern, it is less directly load-bearing for the central empirical claim because the ablations already show that the full CDSE outperforms a unified skill. A more pressing threat to the central claim is the complete absence of any statistical significance assessment. The paper reports point estimates only, and several margins are small (e.g., 1.0 pp on one setting). Without variability measures, the reader cannot evaluate whether the claimed universal superiority is robust. The reader's rationale does mention 'lack of statistical significance testing' as part of the justification for CONDITIONAL, so there is partial agreement. However, the reader's designated weakest assumption differs, leading to 'partial' alignment. My analysis strengthens the condition: even if the decomposition were optimal, the claim that CDSE is best in all 36 settings remains unverified without variance estimates. The recommended verdict remains UNCHANGED (CONDITIONAL) because the paper's overall evidence is strong and the concern can be addressed through additional runs or confidence interval reporting. The concrete test would settle the concern for the tightest margins and could lead to a more precise claim.
The referee's report
The paper tackles a well-motivated problem—how to organize and maintain functionally heterogeneous knowledge from agent execution trajectories. The four-component decomposition is grounded in cognitive science and supported by an empirical analysis of existing skill packages. The experimental evaluation is broad (36 settings across diverse benchmarks, models, and harnesses) and includes ablation, transfer, and token-efficiency analyses that strengthen the empirical case. If the claimed results are robust, CDSE offers a principled and effective approach to skill evolution that could serve as a strong baseline for future work. The detailed implementation in the appendices supports reproducibility.
Pith's reading of the paper (a machine's summary, not the authors')
The paper argues that skills derived from agent execution trajectories contain functionally heterogeneous knowledge—task-level control, contextual heuristics, tool guidance, and executable procedures—and that organizing this knowledge by functional role rather than as a monolithic artifact improves how it is maintained and used. It proposes CDSE, which extracts semantic gradients from successful and failed trajectories, assigns them to four distinct components (Meta, Experience, Tool, Script), and evolves each through specialized updates, cross-component coordination, and adaptive invocation. Across 36 settings (six benchmarks, three LLMs, two execution harnesses), CDSE achieves the highest scores in all settings, averaging 7.6 percentage points over a no-skill baseline and 2.6 points over the best prior skill-evolution method. An analysis of 500 real-world skill packages found that 96.8% contain content spanning at least two functional roles, supporting the premise that functional heterogeneity is prevalent.
What carries the argument. The key machinery is a three-stage pipeline: (1) Outcome-Contrastive Semantic Gradient Discovery extracts reusable behavioral lessons from successful and failed trajectories, keeping supporting evidence attached; (2) Cognition-Guided Gradient Consolidation retains evidence-supported gradients via a retention gate and assigns each to one of four functional components (Meta, Experience, Tool, Script) according to its behavioral role; (3) Adaptive Self-Improvement Update applies component-specific maintenance (rule alignment, revision, induction for Meta and Experience; category-level capacity control for Tool; executable implementation and independent verification for Script), followed by cross-component coordination for conflict resolution and description rewriting that guides adaptive invocation during subsequent tasks. The architecture instantiates cognitive-science distinctions (control vs. memory, knowing-how vs. knowing-that, unity and diversity of executive functions) in agent skill evolution.
Core claim. CDSE organizes trajectory-derived agent skill knowledge into four functionally distinct components—Meta (task-level control and planning), Experience (context-conditioned heuristics and recovery strategies), Tool (tool-specific operational guidance), and Script (reusable executable procedures with verification)—and evolves them through specialized update mechanisms, cross-component conflict resolution, and an adaptive invocation interface that rewrites component descriptions to guide selection during execution. This functional differentiation enables the agent to maintain and deploy knowledge according to its behavioral role, rather than treating all skill knowledge as a homogeneous artifact. The result is a 7.6 percentage point average improvement over a no-skill baseline and a 2.6 point improvement over the strongest competing skill-evolution method in each of 36 evaluated settings, covering six benchmarks, three LLMs (GLM-5.2, DeepSeek-V4-Flash, MiniMax-M3), and two harnesses (Claude Code, Codex).
The premise it rests on. The functional decomposition into exactly four components (Meta, Experience, Tool, Script) is assumed to be the correct and sufficient organization for all trajectory-derived agent skill knowledge, but the paper acknowledges that alternative decompositions remain to be studied.
If the paper is right (Pith's inferences, not the authors')
Agent skill systems structured as monolithic artifacts can be improved by adopting functional decomposition, enabling specialized update rules and reducing maintenance conflicts.
Component-specific validation (e.g., independent verification for Script, rule alignment for Meta and Experience) raises the reliability of skill updates compared to generic editing.
Content-aware invocation descriptions allow an agent to select relevant components adaptively as task demands change, rather than loading all knowledge into every execution.
Cross-model and cross-harness transfer of skill states retains substantial gains (4.2 pp across models, 3.3–3.4 pp across harnesses), suggesting that learned behavioral knowledge is partially independent of the underlying LLM or harness.
The first evolution round accounts for most of the improvement, with diminishing returns on a fixed trajectory corpus, indicating that sustained self-improvement would benefit from new execution experience.
The four-component decomposition may not be exhaustive for all domains; future tasks requiring ethical constraints, multi-agent coordination, or real-time sensor integration could expose missing functional roles, and the framework's modularity would need to accommodate them.
The strong benchmark performance could depend on the specific LLM and harness used for training; experiments with three models and two harnesses provide breadth, but the method's robustness to very different skill-representation formats (e.g., binary or structured APIs vs. natural language) remains unexplored.
The cognitive-science analogy (Atkinson-Shiffrin, Cohen-Squire, Miyake) is used as design inspiration rather than a direct mapping; a natural extension would be to test whether alternative decompositions (e.g., more or fewer components) yield further gains, which the paper itself acknowledges as open.
The fixed-corpus multi-round results show non-monotonic fluctuations, suggesting that repeated consolidation on the same examples can produce minor regressions; online or self-curricula settings might avoid this plateau and achieve larger cumulative improvements.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
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.
The paper
Abstract
Large language model (LLM) agents tackle complex tasks through reasoning and interaction with external environments. Agent skills provide persistent, transferable capabilities that guide task planning, inform contextual decisions, support tool use, and enable executable procedures. Yet existing skill-evolution methods largely organize knowledge around skill artifacts, overlooking the functional heterogeneity within individual skills and its implications for knowledge maintenance and use. Inspired by cognitive science, we propose Cognitive Differentiation in Skill Evolution (CDSE), which evolves a unified skill through four functionally distinct yet coordinated components: Meta, Experience, Tool, and Script. CDSE extracts evidence-grounded semantic gradients from successful and failed trajectories, consolidates them by functional role, and applies component-specific updates and validation with cross-component coordination and adaptive invocation. An analysis of 500 top-ranked skill packages shows that 96.8% contain knowledge spanning at least two functional roles. Across six benchmarks, three agent models, and two execution harnesses, CDSE achieves the highest scores in all 36 settings, improving over the Base agent by 7.6 percentage points and the strongest competing method in each setting by 2.6 points on average. Ablation studies further support the importance of functional differentiation and adaptive invocation.
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 · 2026-10-20
The authors' main claim, as the classifier picked it out 0.95
“Across six benchmarks, three agent models, and two execution harnesses, CDSE achieves the highest scores in all 36 settings, improving over the Base agent by 7.6 percentage points and the strongest competing method in each setting by 2.6 points on average.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.08 · human participants 0.04 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · 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.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Large language model (LLM) agents tackle complex tasks through reasoning and interaction with external environments. | background 1.00 | 0.05 | 0.02 | asserted 1.00 |
| Agent skills provide persistent, transferable capabilities that guide task planning, inform contextual decisions, support tool use, and enable executable procedures. | background 1.00 | 0.06 | 0.04 | asserted 1.00 |
| Yet existing skill-evolution methods largely organize knowledge around skill artifacts, overlooking the functional heterogeneity within individual skills and its implications for knowledge maintenance and use. | gap 1.00 | 0.11 | 0.05 | asserted 1.00 |
| Inspired by cognitive science, we propose Cognitive Differentiation in Skill Evolution (CDSE), which evolves a unified skill through four functionally distinct yet coordinated components: Meta, Experience, Tool, and Script. | contribution 0.78 | 0.83 | 0.14 | asserted 0.99 |
| CDSE extracts evidence-grounded semantic gradients from successful and failed trajectories, consolidates them by functional role, and applies component-specific updates and validation with cross-component coordination and adaptive invocation. | method 1.00 | 0.90 | 0.11 | asserted 0.99 |
| An analysis of 500 top-ranked skill packages shows that 96.8% contain knowledge spanning at least two functional roles. | result 0.94 | 0.84 | 0.06 | asserted 1.00 |
| Across six benchmarks, three agent models, and two execution harnesses, CDSE achieves the highest scores in all 36 settings, improving over the Base agent by 7.6 percentage points and the strongest competing method in each setting by 2.6 points on average. | result 1.00 | 0.96 | 0.50 | asserted 1.00 |
| Ablation studies further support the importance of functional differentiation and adaptive invocation. | result 0.79 | 0.92 | 0.13 | asserted 0.98 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.91
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, 2026-10-20: 1 request, 6,319 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 (7)
Figures are the authors', reproduced from the paper.
References (59)
- [1] Reflexion: Language Agents with Verbal Reinforcement Learning , author =. Advances in Neural Information Processing Systems , volume =. 2023 , doi = on Pith
- [2] Zhao, Andrew and Huang, Daniel and Xu, Quentin and Lin, Matthieu and Liu, Yong-Jin and Huang, Gao , booktitle =. 2024 , doi = on Pith
- [3] Fu, Yao and Kim, Dong-Ki and Kim, Jaekyeom and Sohn, Sungryull and Logeswaran, Lajanugen and Bae, Kyunghoon and Lee, Honglak , booktitle =. 2024 , url = on Pith
- [4] Agent Workflow Memory , author =. Proceedings of the 42nd International Conference on Machine Learning , series =. 2025 , publisher = on Pith
- [5] Ouyang, Siru and Yan, Jun and Hsu, I-Hung and Chen, Yanfei and Jiang, Ke and Wang, Zifeng and Han, Rujun and Le, Long T. and Daruki, Samira and Tang, Xiangru and Tirumalashetty, Vishy and Lee, George and Rofouei, Mahsan and Lin, Hangfei and Han, Jiawei and Lee, Chen-Yu and Pfi... on Pith
- [6] Fang, Runnan and Liang, Yuan and Wang, Xiaobin and Wu, Jialong and Qiao, Shuofei and Xie, Pengjun and Huang, Fei and Chen, Huajun and Zhang, Ningyu , booktitle =. 2026 , url = on Pith
- [7] Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory , author =. 2025 , eprint = on Pith
- [8] Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models , author =. International Conference on Learning Representations , year = on Pith
- [9] Voyager: An Open-Ended Embodied Agent with Large Language Models , author =. 2024 , journal = on Pith
- [10] Zheng, Boyuan and Fatemi, Michael Y. and Jin, Xiaolong and Wang, Zora Zhiruo and Gandhi, Apurva and Song, Yueqi and Gu, Yu and Srinivasa, Jayanth and Liu, Gaowen and Neubig, Graham and Su, Yu , year =. 2504.07079 , archivePrefix = on PithPith's reviewarXiv
- [11] Li, Xiangyi and Liu, Yimin and Chen, Wenbo and others , year =. 2602.12670 , archivePrefix = on PithPith's reviewarXiv
- [12] Alzubi, Salaheddin and Provenzano, Noah and Bingham, Jaydon and Chen, Weiyuan and Vu, Tu , year =. 2603.02766 , archivePrefix = on PithPith's reviewarXiv
- [13] Ni, Jingwei and Liu, Yihao and Liu, Xinpeng and Sun, Yutao and Zhou, Mengyu and Cheng, Pengyu and Wang, Dexin and Zhao, Erchao and Jiang, Xiaoxi and Jiang, Guanjun , year =. 2603.25158 , archivePrefix = on PithPith's reviewarXiv
- [14] 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 Qiu, Kai and Yang, Yuqing and Chen, Dongdong and Yang, Xue and Luo, Chong , year =. 2605.23904 , archivePrefix = on PithPith's reviewarXiv
- [15] Gautam, Srishti and Radhakrishna, Arjun and Gulwani, Sumit , year =. 2606.10546 , archivePrefix = on PithPith's reviewarXiv
- [16] Moll, Johannes and Corbeil, Jean-Philippe and Pan, Jiazhen and Hadamitzky, Martin and Rueckert, Daniel and Adams, Lisa and Bressem, Keno , year =. 2605.29668 , archivePrefix = on PithPith's reviewarXiv
- [17] He, Yu and Yang, Weikai , year =. 2608.15165 , archivePrefix = on PithPith's reviewarXiv
- [18] Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries , author =. 2026 , eprint = on Pith
- [19] Zhang, Hanrong and Fan, Shicheng and Zou, Henry Peng and Chen, Yankai and Wang, Zhenting and Zhou, Jiayu and Li, Chengze and Huang, Wei-Chieh and Yao, Yifei and Zheng, Kening and Liu, Xue and Li, Xiaoxiao and Yu, Philip S. , year =. 2604.01687 , archivePrefix = on PithPith's reviewarXiv
- [20] Yang, Min and Piao, Jinghua and Xia, Xu and Lan, Xiaochong and Chen, Jiaju and Gong, Yongshun and Li, Yong , year =. 2605.08693 , archivePrefix = on PithPith's reviewarXiv
show 39 more
- [21] Lin, Huawei and Li, Peng and Song, Jie and Jiang, Fuxin and Zhang, Tieying , year =. 2605.27366 , archivePrefix = on PithPith's reviewarXiv
- [22] Progressive Agent Skill Generation via Reinforcement Learning , author =. 2026 , eprint = on Pith
- [23] Wang, Zefeng and Yan, Minxi and Bi, Jinhe and Yan, Sikuan and Tresp, Volker and Ma, Yunpu , year =. 2607.05297 , archivePrefix = on PithPith's reviewarXiv
- [24] Qian, Cheng and Han, Chi and Fung, Yi and Qin, Yujia and Liu, Zhiyuan and Ji, Heng , booktitle =. 2023 , publisher =. doi:10.18653/v1/2023.findings-emnlp.462 , url = on Pithdoi:10.18653/v1/2023.findings-emnlp.462
- [25] Large Language Models as Tool Makers , author =. 2023 , eprint = on Pith
- [26] Nguyen, Dang and Lai, Viet Dac and Yoon, Seunghyun and Rossi, Ryan A. and Zhao, Handong and Zhang, Ruiyi and Mathur, Puneet and Lipka, Nedim and Wang, Yu and Bui, Trung and Dernoncourt, Franck and Zhou, Tianyi , year =. 2411.01747 , archivePrefix = on PithPith's reviewarXiv
- [27] W. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics , pages =. 2025 , publisher =. doi:10.18653/v1/2025.acl-long.1266 , url = on Pithdoi:10.18653/v1/2025.acl-long.1266
- [28] Luo, Jianwen and Huang, Yiming and Meng, Jinxiang and Lei, Fangyu and He, Shizhu and Liu, Xiao and Jiang, Shanshan and Dong, Bin and Zhao, Jun and Liu, Kang , booktitle =. 2026 , publisher =. doi:10.18653/v1/2026.acl-long.87 , url = on Pithdoi:10.18653/v1/2026.acl-long.87
- [29] Zhang, Zhang and Lu, Shuqi and Qian, Hongjin and He, Di and Liu, Zheng , booktitle =. 2026 , publisher =. doi:10.18653/v1/2026.acl-demo.81 , url = on Pithdoi:10.18653/v1/2026.acl-demo.81
- [30] Harnessing LLM Agents with Skill Programs , author =. 2026 , eprint = on Pith
- [31] Structure and Function of Declarative and Nondeclarative Memory Systems , author =. Proceedings of the National Academy of Sciences , volume =. 1996 , doi = on Pith
- [32] Metamemory: A Theoretical Framework and New Findings , author =. Psychology of Learning and Motivation , volume =. 1990 , publisher = on Pith
- [33] Instance-Based Learning in Dynamic Decision Making , author =. Cognitive Science , volume =. 2003 , doi = on Pith
- [34] Acquisition of Cognitive Skill , author =. Psychological Review , volume =. 1982 , doi = on Pith
- [35] The Neural Bases of Complex Tool Use in Humans , author =. Trends in Cognitive Sciences , volume =. 2004 , doi = on Pith
- [36] A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics , author =. AI Magazine , volume =. 2017 , doi = on Pith
- [37] The Soar Cognitive Architecture , author =. 2012 , isbn = on Pith
- [38] Stocco, Andrea and Sibert, Catherine and Steine-Hanson, Zoe and Koh, Natalie and Laird, John E. and Lebiere, Christian J. and Rosenbloom, Paul , journal =. Analysis of the Human Connectome Data Supports the Notion of a. 2021 , doi = on Pith
- [39] A Proposal to Extend the Common Model of Cognition with Metacognition , author =. 2025 , eprint = on Pith
- [40] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =. 2023 , url = on Pith
- [41] Zhang, Barry and Lazuka, Keith and Murag, Mahesh , year =. Equipping Agents for the Real World with on Pith
- [42] Pryzant, Reid and Iter, Dan and Li, Jerry and Lee, Yin Tat and Zhu, Chenguang and Zeng, Michael , booktitle =. Automatic Prompt Optimization with. 2023 , publisher =. doi:10.18653/v1/2023.emnlp-main.494 , url = on Pithdoi:10.18653/v1/2023.emnlp-main.494
- [43] Zhang, Jenny and Hu, Shengran and Lu, Cong and Lange, Robert and Clune, Jeff , booktitle =. 2026 , url = on Pith
- [44] Human Memory: A Proposed System and Its Control Processes , author =. The Psychology of Learning and Motivation: Advances in Research and Theory , editor =. 1968 , publisher =. doi:10.1016/S0079-7421(08)60422-3 , url = on Pithdoi:10.1016/s0079-7421(08)60422-3
- [45] Preserved Learning and Retention of Pattern-Analyzing Skill in Amnesia: Dissociation of Knowing How and Knowing That , author =. Science , volume =. 1980 , doi = on Pith
- [46] Miyake, Akira and Friedman, Naomi P. and Emerson, Michael J. and Witzki, Alexander H. and Howerter, Amy and Wager, Tor D. , journal =. The Unity and Diversity of Executive Functions and Their Contributions to Complex. 2000 , doi = on Pith
- [47] Hyperagents , author =. 2026 , eprint = on Pith
- [48] A Survey on Large Language Model Based Autonomous Agents , author =. Frontiers of Computer Science , volume =. 2024 , doi = on Pith
- [49] Yang, John and Jimenez, Carlos E. and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , booktitle =. 2024 , doi = on Pith
- [50] Absolute Zero: Reinforced Self-play Reasoning with Zero Data , author =. 2025 , eprint = on Pith
- [51] Huang, Chengsong and Yu, Wenhao and Wang, Xiaoyang and Zhang, Hongming and Li, Zongxia and Li, Ruosen and Huang, Jiaxin and Mi, Haitao and Yu, Dong , booktitle =. 2026 , url = on Pith
- [52] 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 , url = on Pith
- [53] Opsahl-Ong, Krista and Singhvi, Arnav and Collins, Jasmine and Zhou, Ivan and Wang, Cindy and Baheti, Ashutosh and Oertell, Owen and Portes, Jacob and Havens, Sam and Elsen, Erich and Bendersky, Michael and Zaharia, Matei and Chen, Xing , year =. 2603.08655 , archivePrefix = on PitharXiv
- [54] Zhu, Jian and Zhang, Yuzheng and Ma, Zeyao and Zhang, Bohan and Schoepf, Armin and Woloch, Daniel and Wang, Peter Yiliu and Yang, Guangyu Robert and Jacob, Samuel and Nagisetty, Siddharth and Chundru, Abhiram and Lin, Jean and Mateega, Spencer and Zhang, Jing , year =. 2606.29... on PithPith's reviewarXiv
- [55] Pham, Thinh and Nguyen, Nguyen and Zunjare, Pratibha and Chen, Weiyuan and Tseng, Yu-Min and Vu, Tu , booktitle =. 2026 , url = on Pith
- [56] Liu, Junnan and Liu, Hongwei and Xiao, Linchen and Wang, Ziyi and Liu, Kuikun and Gao, Songyang and Zhang, Wenwei and Zhang, Songyang and Chen, Kai , booktitle =. Are Your. 2025 , pages =. doi:10.18653/v1/2025.findings-acl.905 , url = on Pithdoi:10.18653/v1/2025.findings-acl.905
- [57] Deng, Xiang and Da, Jeff and Pan, Edwin and He, Yannis Yiming and Ide, Charles and Garg, Kanak and Lauffer, Niklas and Park, Andrew and Pasari, Nitin and Rane, Chetan and Sampath, Karmini and Krishnan, Maya and Kundurthy, Srivatsa and Hendryx, Sean and Wang, Zifan and Zhang, C... on PithPith's reviewarXiv
- [58] Shepard, Daniel and Salimans, Robin , year =. 2604.18934 , archivePrefix = on PithPith's reviewarXiv
- [59] Lai, Xunhao and Xu, Weiqi and Yang, Yufeng and Chen, Qiaorui and Xu, Yang and Zeng, Lunbin and Li, Xiaolong and Sun, Haohai and Zhu, Haichao and Zhang, Vito and Hu, Jinkai and Li, Jiayao and Gao, Rui and Li, Zekun and Zhu, Songquan and Zhou, Jingkai and Zhao, Pengyu , year =. ... on PithPith's reviewarXiv
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.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment