Artificial Intelligence · arXiv:2610.20845 v1 · submitted 14 Oct 2026 · 43 references
Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents
The paper says
“Skill-policy co-evolution, in which an agent's policy model and its skill bank improve each other during training, is a promising paradigm for self-improving agents. Current evaluations score the co-evolved policy on the final bank it trained with (the delivered bank).”
more of the abstract
After deployment, however, skills may be rewritten, added or removed more frequently than the policy model is retrained, leaving the policy to operate on a drifted skill bank that neither the training objective nor existing evaluation protocols address. In this paper, we first establish, through theoretical and empirical analysis, that the fragility hidden by this blind spot is not incidental. The acceptance rule these methods share (cooperative acceptance) ties a policy's competence to the bank it co-evolved with, leaving it fragile once skills drift, regardless of the implementation; empirically, sensitivity to skill drift grows as co-evolution training proceeds. We therefore introduce the CO-evolution Drift Audit (CODA), a deployment-oriented evaluation protocol that simulates skill drift and, alongside success rate under worst-case drift, introduces two new metrics: a consistency rate under meaning-preserving edits and a resistance rate to corrupted skills. Across four co-evolution methods and five benchmarks (six settings), worst-case success is 5-13 points below the peak measured for each delivered pair, erasing 52-93% of the delivered bank's gain over the no-skill baseline. CODA thus reframes the target of co-evolution, from peak performance on the co-evolved bank to the ability to harness skills however they evolve, and provides a shared standard against which future methods can be audited.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 16 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: minor revision.
A clear and honest diagnosis of a real blind spot in co-evaluation, backed by theory and a well-designed audit protocol. The single-training-seed fragility trends are a real limitation, but the paper's contribution stands. This line is wrong
3 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 43 references say.
The paper says, at §3.3, Appendix G
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The core empirical claim—that fragility grows with co-evolution—is supported by fragility curves from a single training run per setting. The main text does not state this; it appears only in Appendix G ('Each setting comes from a single co-evolution run').
more of objection 1
Training randomness (policy initialization, data order, skill proposal stochasticity) is unmeasured, so it is unclear whether the increasing trend is a robust property of co-evolution or an artifact of a particular run. For example, in Table 10 (RESKILL-ScienceWorld T1) the 33.7 pp drop at s240 drives much of the trend. The authors should either (a) provide replication with additional seeds for at least one setting to demonstrate reproducibility, or (b) add an explicit statement of this limitation in Section 3.3 and discuss its implications for the generality of the conclusion.
The paper says, at §2.2, Assumptions 1–2
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
Proposition 1 relies on Assumption 2 (candidate coverage at delivery: B(S*) ⊆ C_T). The paper tests the prediction (no content-level rewrite improves performance) and finds it holds 18/18, which is good. However, the candidate set C_T is never empirically characterized for any method. The paper should clarify what constitutes the candidate set in each method and argue why the tested T1–T3 rewrites are plausibly covered. Without this, the link between the theory and the experiment is incomplete.
The paper says, at §4.3, Table 2
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The CODA verdict is an absolute pass/fail, and all six methods fail the consistency criterion. The paper states that 'CODA does not declare these methods useless,' but a protocol that no current method passes may be interpreted as setting an unrealistic bar. The authors should discuss whether the current pass/fail threshold is intended to be aspirational or achievable with existing techniques, and whether the verdict is meant to be a research challenge rather than a practical audit.
6 minor objections
The paper says, at Abstract
Pith says: minor objection 1 of 6
The phrase 'the ability to harness skills however they evolve' is slightly ambiguous; consider 'the ability to harness skills as they evolve' or 'however they may evolve'.
The paper says, at Figure 1, panel (2)
Pith says: minor objection 2 of 6
The label 'current co-evolved policy fails once skills change' could be more precise: 'the co-evolved policy fails when skills drift' to match the paper's terminology.
show 4 more
The paper says, at Table 1
Pith says: minor objection 3 of 6
For ReSkill, the 'Selection bias' column says 'per-step'. From Appendix C, the gate uses Thompson sampling; consider 'Thompson sampling (per-step)' for clarity.
The paper says, at §3.1
Pith says: minor objection 4 of 6
When introducing the five axes, note explicitly that T4 (ecosystem) lies outside B(S*) because it changes the skill set, not just content. This is stated later but a quick flag here would help the reader.
The paper says, at Table 2
Pith says: minor objection 5 of 6
Add a brief footnote or parenthetical referencing Appendix E.8 for the computation of the noise floor CR0, since the main text only cites the appendix.
The paper says, at §5
Pith says: minor objection 6 of 6
The related work section is thorough. Consider moving the discussion of adversarial co-evolution (Appendix D.5) to the main text, as it provides an important contrast that supports the causal role of the acceptance rule.
What would settle it
Reproduce the co-evolution fragility curve for RESKILL-ALFWorld (the setting with the clearest monotonic trend) using three independent training seeds. For each seed, evaluate all checkpoints on the same fixed perturbed banks and compute the OLS slope of Drop over training steps for each content-level axis. Require that all three slopes are positive and that the 95% confidence intervals of the pooled slope exclude zero. If any seed yields a non-positive slope, the fragility trend is not robust to training randomness.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 16 Oct 2026
- Reader holds up
- Skeptic partly agrees: says it holds up only with conditions Fragility trends rely on single training seeds per setting; training variability is unmeasured.
- Editor wrote the verdict above
- Referee minor revision: 3 major objections, 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 nails an overlooked problem. Current co-evolution evaluation reports only peak performance on the delivered bank, ignoring that skills drift after deployment. The authors show this isn't a minor oversight: cooperative acceptance ties the policy to its bank, so rewrites almost always hurt, and they build CODA to measure how much. The theory is clean—Propositions 1 and 2 are simple but exactly capture the selection bias that causes fragility—and the empirical sweep across four methods and six settings is convincing. The fragility curves, showing sensitivity growing with training, are the best evidence. The paper also does the right thing by releasing the perturbed banks, pinning the rewriting model, and being upfront about limitations (L1–L3).
The single-training-seed concern in the stress test is legitimate. Each setting comes from one co-evolution run (Appendix G), so we don't know if the increasing trend is robust to training randomness. The paper quantifies evaluation noise but not training variability, and the RESKILL-ScienceWorld T1 point (s240) shows how much one checkpoint can swing. That said, the trend is consistent across five of six settings, and the theory gives a reason to expect it, so I view this as a limitation to flag in review rather than a fatal flaw. The other soft spot is that CODA's content-level axes T1–T3, while well-designed, may not capture all realistic drift, and the paper acknowledges that rewriting is not bit-reproducible. These are reasonable compromises for a first audit protocol.
Overall, this is a serious, honest paper that reframes how co-evolution should be evaluated. It earns a careful referee. I'd bring it to a reading group, and I'd recommend acceptance after the authors address the training-seed variability (e.g., a two-seed replication on one setting).
The strongest objection
The paper’s core empirical claim—that sensitivity to skill drift grows with co-evolution training—is supported by fragility curves from one training run per setting (Appendix G: 'Each setting comes from a single co-evolution run'). Evaluation noise is quantified via multiple rollout seeds (Section 3.3), but training randomness (policy initialization, data order, skill proposal stochasticity) is not. A single trajectory could produce a spurious increasing trend (e.g., RESKILL-ScienceWorld T1 in Table 10: s240 has a 33.7 pp drop while s180 and s300 have drops of 2.0 and 10.2 pp; the trend is driven heavily by one anchor). Without replication across training seeds, it is unclear whether the fragility increase is a robust property of co-evolution or an artifact of a particular run. This directly affects the generalizability of the conclusion that 'co-evolution buys peak performance at the cost of fragility' (Section 3.2).
The reader's weakest_assumption focused on representativeness of perturbation axes and Assumption 2, which the paper explicitly qualifies. My concern targets a different gap: the absence of training-seed replication, which is not discussed as a limitation. Given the paper's otherwise thorough empirical design (multiple methods, benchmarks, repeated rollout evaluations, paired task setup), this is the most consequential unaddressed source of variance. A single seed per setting leaves the fragility trend statistically unverified across the training stochasticity that practitioners would face. If the trend holds across seeds, the paper's central claim is strongly supported; if not, the evidence weakens. Therefore, a conditional verdict requiring this replication is appropriate, even though the contribution and theory are otherwise sound.
The referee's report
If the results hold, the paper reframes the evaluation of co-evolution from peak performance on a fixed bank to robustness under bank drift. The theoretical analysis isolates cooperative acceptance as the source of fragility, which is actionable. The empirical evidence is broad across methods (four) and domains (six settings). The CODA protocol is a practical, released tool that can become a standard for auditing future methods. The paper is well structured, the theory is clean, the rewriting quality is validated via blinded audit, and the limitations are mostly acknowledged.
Pith's reading of the paper (a machine's summary, not the authors')
This paper argues that current skill-policy co-evolution methods train a policy and a skill bank that become specialized to each other, leaving the policy vulnerable when the bank evolves after deployment. The shared "cooperative acceptance" rule—retaining only skills that help the current policy—creates a peak on the delivered bank so that no content-level rewrite can improve performance, and usually degrades it. Empirically, across four methods and six settings, the worst-case success rate under drift falls 5–13 points below the reported peak, erasing 52–93% of the skill benefit. The paper introduces CODA, a deployment-oriented evaluation protocol that measures consistency under meaning-preserving rewrites and resistance to corrupted skills, reframing co-evaluation success from peak performance to robustness under realistic drift.
What carries the argument. The cooperative acceptance rule: a selection bias that retains a candidate bank only when it maximizes or non-decreasingly improves the current policy's return (idealized as S_t ∈ argmax_{S∈C_t} J(θ_t, S)). This rule, shared across all co-evolution methods despite differing implementations, ties the policy's competence to the specific bank it co-evolved with. The 'co-evolution peak' (Proposition 1) then states that for the delivered pair under candidate coverage of content-level perturbations, no covered rewrite can improve the policy, and the 'fragility' Φ(θ, S) measures the worst-case drop over perturbations. CODA operationalizes these concepts into three metrics—consistency rate, resistance rate, and worst-case success rate—to audit any delivered pair.
Core claim. The paper establishes that the cooperative acceptance rule—the gate common to all co-evolution methods that keeps only skill changes serving the current policy—places the delivered policy–bank pair at a local peak over the candidate banks considered during training. Consequently, under candidate coverage of content-level perturbations, no rewrite of the delivered skills can improve the policy's return; all covered perturbations either leave performance unchanged or degrade it. This theoretical prediction holds across 18 endpoint pairs (6 settings × 3 content axes), and the magnitude of fragility grows as co-evolution training proceeds in five of six settings. The resulting degradation is large: worst-case success under content-level drift trails the peak by 5–13 points, consuming 52–93% of the skill benefit, meaning the headline numbers that co-evolution methods report substantially overstate the competence a deployment can rely on.
The premise it rests on. The theoretical result (Proposition 1) depends on Assumption 2 that the candidate set at delivery covers all content-level perturbations considered, and the empirical protocol assumes that the three specific perturbation axes (synonym rewrite, incomplete, misleading) are representative of realistic post-deployment skill drift.
If the paper is right (Pith's inferences, not the authors')
Current co-evolution methods cannot be compared by peak success alone; future comparisons must include worst-case success under content-level drift, which CODA provides.
Developers of self-improving agents should retrain or adapt policies on a distribution of bank perturbations, not a single fixed bank, to avoid the point-mass evaluation blind spot (Eq. 6).
The finding that a meaning-preserving rewrite (T1) costs up to 12.6 points and flips 13–23% of task outcomes indicates that policies overfit to surface phrasing, not skill semantics, motivating learning algorithms that invariant to wording.
The CODA protocol sets a shared standard for auditing robustness, making it possible to fail a method that passes only the peak test—a gate that the paper shows is routinely failed by the consistency and resistance criteria.
If the acceptance sign is flipped from cooperative to adversarial (as in some adversarial co-evolution), robustness can be gained instead of lost, suggesting the bias in the rule is causal, not co-evolution per se.
The CODA metric set could be extended to include a 'skill-interpretation efficiency' measure: how much of the skill's benefit survives when the skill is compressed or summarised, which would test whether the policy understands the skill's purpose or merely memorises the exact string.
A natural next step is to test whether larger backbone models or more diverse training tasks reduce the phrasing overfit. The current paper leaves this as an open question; one could hypothesise that sufficiently large models trained on diverse data might naturally learn to ignore surface forms, but the ARISE-AMC23 result (6.5-point drop) suggests even large models overfit to the co-evolved bank.
The paper's diagnostic that dependence (drop when bank removed) predicts fragility could be used as an early stopping criterion: stop co-evolution when dependence begins to induce fragility, even if peak return is still rising, to trade off absolute performance for robustness.
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
Skill-policy co-evolution, in which an agent's policy model and its skill bank improve each other during training, is a promising paradigm for self-improving agents. Current evaluations score the co-evolved policy on the final bank it trained with (the delivered bank). After deployment, however, skills may be rewritten, added or removed more frequently than the policy model is retrained, leaving the policy to operate on a drifted skill bank that neither the training objective nor existing evaluation protocols address. In this paper, we first establish, through theoretical and empirical analysis, that the fragility hidden by this blind spot is not incidental. The acceptance rule these methods share (cooperative acceptance) ties a policy's competence to the bank it co-evolved with, leaving it fragile once skills drift, regardless of the implementation; empirically, sensitivity to skill drift grows as co-evolution training proceeds. We therefore introduce the CO-evolution Drift Audit (CODA), a deployment-oriented evaluation protocol that simulates skill drift and, alongside success rate under worst-case drift, introduces two new metrics: a consistency rate under meaning-preserving edits and a resistance rate to corrupted skills. Across four co-evolution methods and five benchmarks (six settings), worst-case success is 5-13 points below the peak measured for each delivered pair, erasing 52-93% of the delivered bank's gain over the no-skill baseline. CODA thus reframes the target of co-evolution, from peak performance on the co-evolved bank to the ability to harness skills however they evolve, and provides a shared standard against which future methods can be audited.
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 · 2026-10-16
The authors' main claim, as the classifier picked it out 0.68
“CODA thus reframes the target of co-evolution, from peak performance on the co-evolved bank to the ability to harness skills however they evolve, and provides a shared standard against which future methods can be audited.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.24 · states limitation 0.16 · human participants 0.03 · controlled comparison 0.93 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.94 · 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 |
|---|---|---|---|---|
| Skill-policy co-evolution, in which an agent's policy model and its skill bank improve each other during training, is a promising paradigm for self-improving agents. | background 1.00 | 0.06 | 0.03 | asserted 0.76 |
| Current evaluations score the co-evolved policy on the final bank it trained with (the delivered bank). | background 0.94 | 0.07 | 0.02 | asserted 1.00 |
| After deployment, however, skills may be rewritten, added or removed more frequently than the policy model is retrained, leaving the policy to operate on a drifted skill bank that neither the training objective nor existing evaluation protocols address. | gap 0.84 | 0.08 | 0.04 | hedged possibility 0.96 |
| In this paper, we first establish, through theoretical and empirical analysis, that the fragility hidden by this blind spot is not incidental. | contribution 0.79 | 0.85 | 0.11 | asserted 1.00 |
| The acceptance rule these methods share (cooperative acceptance) ties a policy's competence to the bank it co-evolved with, leaving it fragile once skills drift, regardless of the implementation; empirically, sensitivity to skill drift grows as co-evolution training proceeds. | interpretation 0.54 | 0.66 | 0.31 | asserted 1.00 |
| We therefore introduce the CO-evolution Drift Audit (CODA), a deployment-oriented evaluation protocol that simulates skill drift and, alongside success rate under worst-case drift, introduces two new metrics: a consistency rate under meaning-preserving edits and a resistance rate to corrupted skills. | contribution 0.60 | 0.93 | 0.13 | asserted 0.99 |
| Across four co-evolution methods and five benchmarks (six settings), worst-case success is 5-13 points below the peak measured for each delivered pair, erasing 52-93% of the delivered bank's gain over the no-skill baseline. | result 1.00 | 0.93 | 0.19 | asserted 1.00 |
| CODA thus reframes the target of co-evolution, from peak performance on the co-evolved bank to the ability to harness skills however they evolve, and provides a shared standard against which future methods can be audited. | interpretation 0.77 | 0.83 | 0.79 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 0.99 · evidence from observation or experiment 0.83
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, 2026-10-16: 1 request, 6,567 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 (17)
Figures are the authors', reproduced from the paper.
References (43)
- [1] Skillrl: Evolving agents via recursive skill-augmented reinforcement learning , author=. arXiv preprint arXiv:2602.08234 , year= on PithPith's reviewarXiv
- [2] ReSkill: Reconciling Skill Creation with Policy Optimization in Agentic RL , author=. arXiv preprint arXiv:2606.01619 , year= on PithPith's reviewarXiv
- [3] SkillForge: Evolving Verifiable Skills for Reinforcement Learning Agents , author=. arXiv preprint arXiv:2608.24747 , year= on PithPith's reviewarXiv
- [4] CoSkill: Joint Reinforcement Learning of Reasoning and Meta-Skill Agents for Hierarchical Skill Evolution , author=. arXiv preprint arXiv:2609.04865 , year= on PithPith's reviewarXiv
- [5] AUSO: Action-Level Unified Skill Optimization from Internalization to Utilization , author=. arXiv preprint arXiv:2608.21292 , year= on PithPith's reviewarXiv
- [6] Co-Evolving Skill Generation and Policy Optimization , author=. arXiv preprint arXiv:2606.08755 , year= on PithPith's reviewarXiv
- [7] Skill1: Unified evolution of skill-augmented agents via reinforcement learning , author=. arXiv preprint arXiv:2605.06130 , year= on PithPith's reviewarXiv
- [8] SkillMaster: Toward Autonomous Skill Mastery in LLM Agents , author=. arXiv preprint arXiv:2605.08693 , year= on PithPith's reviewarXiv
- [9] Arise: Agent reasoning with intrinsic skill evolution in hierarchical reinforcement learning , author=. arXiv preprint arXiv:2603.16060 , year= on PitharXiv
- [10] LLM-Based Human-Agent Collaboration and Interaction Systems: A Survey , author=. 2026 , eprint= on Pith
- [11] Deepscaler: Surpassing o1-preview with a 1.5 b model by scaling rl , author=. Notion Blog , volume= on Pith
- [12] Mind2web: Towards a generalist agent for the web , author=. Advances in Neural Information Processing Systems , volume= on Pith
- [13] Evolving-RL: End-to-End Optimization of Experience-Driven Self-Evolving Capability within Agents , author=. arXiv preprint arXiv:2605.10663 , year= on PithPith's reviewarXiv
- [14] The Path to Recursive Self-Improving Agents: Foundation, Framework, and Future Directions , author=. 2026 , publisher= on Pith
- [15] Co-evolving llm decision and skill bank agents for long-horizon tasks , author=. arXiv preprint arXiv:2604.20987 , year= on PithPith's reviewarXiv
- [16] Scienceworld: Is your agent smarter than a 5th grader? , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages= on Pith
- [17] Webshop: Towards scalable real-world web interaction with grounded language agents , author=. Advances in Neural Information Processing Systems , volume= on Pith
- [18] Alfworld: Aligning text and embodied environments for interactive learning , author=. arXiv preprint arXiv:2010.03768 , year= on PithPith's reviewarXiv
- [19] Large language model agents are not always faithful self-evolvers , author=. arXiv preprint arXiv:2601.22436 , year= on PitharXiv
- [20] Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use , author=. arXiv preprint arXiv:2607.01084 , year= on PithPith's reviewarXiv
show 23 more
- [21] Library drift: Diagnosing and fixing a silent failure mode in self-evolving LLM skill libraries , author=. arXiv preprint arXiv:2605.19576 , year= on PithPith's reviewarXiv
- [22] VCE-Skill: Enhancing Skill Self-Evolution with Version-Change Experience , author=. arXiv preprint arXiv:2608.16544 , year= on PithPith's reviewarXiv
- [23] SIA: Self Improving AI with Harness & Weight Updates , author=. arXiv preprint arXiv:2605.27276 , year= on PithPith's reviewarXiv
- [24] Rethinking Self-Evolving Agent Skills: Feedback Dynamics over Multiple Rounds , author=. arXiv preprint arXiv:2608.02636 , year= on PithPith's reviewarXiv
- [25] Learning Generalizable Skill Policy with Data-Efficient Unsupervised RL , author=. arXiv preprint arXiv:2607.00392 , year= on PithPith's reviewarXiv
- [26] Lifting traces to logic: Programmatic skill induction with neuro-symbolic learning for long-horizon agentic tasks , author=. arXiv preprint arXiv:2605.01293 , year= on PithPith's reviewarXiv
- [27] Game-Theoretic Co-Evolution for LLM-Based Heuristic Discovery , author=. arXiv preprint arXiv:2601.22896 , year= on PitharXiv
- [28] Advevo-marl: Shaping internalized safety through adversarial co-evolution in multi-agent reinforcement learning , author=. arXiv preprint arXiv:2510.01586 , year= on PitharXiv
- [29] The Regression Tax: Decomposing Why Skills Help and Hurt LLM Agents , author=. arXiv preprint arXiv:2607.22520 , year= on PithPith's reviewarXiv
- [30] SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision , author=. arXiv preprint arXiv:2606.01139 , year= on PithPith's reviewarXiv
- [31] Skill is Not One-Size-Fits-All: Model-Aware Skill Alignment for LLM Agents , author=. arXiv preprint arXiv:2605.30723 , year= on PithPith's reviewarXiv
- [32] How well do agentic skills work in the wild: Benchmarking llm skill usage in realistic settings , author=. arXiv preprint arXiv:2604.04323 , year= on PithPith's reviewarXiv
- [33] The Scaling Laws of Skills in LLM Agent Systems , author=. arXiv preprint arXiv:2605.16508 , year= on PithPith's reviewarXiv
- [34] A comprehensive survey on agent skills: Taxonomy, techniques, and applications , author=. arXiv preprint arXiv:2605.07358 , year= on PithPith's reviewarXiv
- [35] 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= on PithPith's reviewarXiv
- [36] Self-Improvements in Modern Agentic Systems: A Survey , author=. arXiv preprint arXiv:2607.13104 , year= on PithPith's reviewarXiv
- [37] MetaClaw: Just Talk--An Agent That Meta-Learns and Evolves in the Wild , author=. arXiv preprint arXiv:2603.17187 , year= on PitharXiv
- [38] Learning While Acting: A Skill-Enhanced Test-Time Co-Evolution Framework for Online Lifelong Learning Agents , author=. arXiv preprint arXiv:2606.04815 , year= on PithPith's reviewarXiv
- [39] MAGIC: A co-evolving attacker-defender adversarial game for robust LLM safety , author=. arXiv preprint arXiv:2602.01539 , year= on PitharXiv
- [40] “Other-play” for zero-shot coordination , author=. International conference on machine learning , pages=. 2020 , organization= on Pith
- [41] Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills , author=. arXiv preprint arXiv:2607.22529 , year= on PithPith's reviewarXiv
- [42] Benchmarking neural network robustness to common corruptions and perturbations , author=. arXiv preprint arXiv:1903.12261 , year= on PithPith's reviewarXiv
- [43] Beyond accuracy: Behavioral testing of NLP models with CheckList , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages= on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
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