Pith.

Artificial Intelligence · arXiv:2608.07449 v1 · submitted 7 Aug 2026 · 12 references

SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent

Mingxuan Zheng, Yujin Zhou, Chuxue Cao, Boqin Yin, Yuyao Zhang, Jiapeng Sun, Shuaishuai Gong, Sirui Han

show 1 more author

Yike Guo

The paper says

“LLM agents increasingly adapt to recurring tasks by accumulating procedural knowledge in skills. These skills are lightweight, reusable textual artifacts that are loaded into the agent's context without weight updates.”

more of the abstract

Recent methods refine skills through iterative task execution, failure diagnosis, and trajectory-guided text-space updates. However, existing frameworks lack explicit diagnosis–outcome feedback and treat deletion as a generic edit operation rather than a dedicated mechanism for consolidating accumulated knowledge. We introduce SkillProx, a proximal-gradient-inspired forward–backward framework that couples closed-loop diagnostic evolution with utility-aware proximal refinement. Motivated by a composite objective balancing task loss and skill complexity, the forward stage re-executes diagnosis-driven edits on the same task batch, rolls back regressions, and feeds measured outcomes into subsequent diagnoses. The backward stage decomposes the resulting skill into auditable knowledge units, estimates their contributions using a frozen leave-one-out utility audit, and applies validation-gated consolidation, demotion, or removal. Experiments on in-distribution and out-of-distribution benchmarks across multiple backbone LLMs show that SkillProx improves average accuracy by 3.0 percentage points over the strongest gradient-based baseline. Component ablations demonstrate the complementary effects of closed-loop diagnosis and proximal refinement.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 10 Aug 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. SkillProx has two genuinely useful mechanisms, but the 3.0 pp headline gain is overstated and the tau threshold may have been chosen on the test set. This line is wrong

3 major objections · 6 minor

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

Are you an author? What you can do.

Not checked: figures, code, what the 12 references say.

The paper says, at §5.4, Fig. 2, Table 7, Appendix C

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The candidate threshold τ appears to be selected on the OJ set, which is the same SpreadsheetBench test set reported in Table 2. Figure 2 reports OJ hard accuracy as a function of τ and identifies τ=−0.001 as the maximum (52.3%), and Table 7 then fixes τ=−0.001 for the main experiments.

more of objection 1

Unless a validation-only selection procedure exists and is disclosed, the headline gain for SkillProx on SpreadsheetBench is the result of test-set selection on the very metric used to claim improvement. This is load-bearing because Table 3 attributes 2.5 pp to the Prox stage, whose behavior depends on τ. Please re-select τ using only the validation split (or a nested procedure) and report the resulting Table 2, or clearly disclose the selection order and re-run the main experiments with a pre-registered τ.

The paper says, at Abstract, Table 2

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The abstract's claim of a 3.0 percentage-point gain over the strongest gradient-based baseline is not supported by Table 2. Recomputing the average margin over the per-cell best of SkillGrad and SkillOpt gives (1.7+8.8+3.8+0+1.6+1.0+1.2+1.4+0.3)/9 ≈ 2.2 pp, and several cells are ties or within one standard deviation (e.g., Qwen3.5-27B Spreadsheet: 51.3±1.2 vs 51.3±0.9; Qwen3.6-27B HiTab: 80.0±1.3 vs 79.7±0.8). Please report the exact aggregation used and correct the abstract, or justify why particular settings are excluded from the average.

The paper says, at §5.3, Table 3

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The component ablation that establishes the backward Prox stage's contribution is based on a single backbone (Qwen3.6-27B) with a fixed training seed and no paired significance test. The reported 'w/o Prox' mean of 52.0±1.0 overlaps with the full method's 54.5±0.5, so the −2.5 pp effect is not established beyond sampling noise.

more of objection 3

Because this ablation is the main evidence that proximal refinement—the paper's key novelty—matters, please add multiple training seeds and a paired comparison (or otherwise report confidence intervals) before drawing the complementarity conclusion.

6 minor objections

The paper says, at §5.2

Pith says: minor objection 1 of 6

The sentence 'SkillProx delivers consistent gains on every backbone' is not accurate for Qwen3.5-27B Spreadsheet, where SkillProx ties SkillGrad and SkillOpt at 51.3; please soften or clarify.

The paper says, at Table 3

Pith says: minor objection 2 of 6

The table reports mean±std but the text says a 'fixed training-set seed'; please specify whether the standard deviation is over evaluation reruns, random inference seeds, or some other source of variation.

show 4 more

The paper says, at Appendix D.4.0.0.1, Fig. 4

Pith says: minor objection 3 of 6

The text reports r=−0.628 while the figure caption says 'Pearson r = 0.628'; please use the signed value consistently.

The paper says, at Fig. 2

Pith says: minor objection 4 of 6

The no-Prox anchor point is not labeled in the rendered figure; add the τ=−∞ label or state it explicitly in the caption.

The paper says, at Eq. (8)

Pith says: minor objection 5 of 6

The expression 'i≺j⇐ ⇒' is a rendering artifact; it should read 'i ≺ j ⇔' or be otherwise formatted correctly.

The paper says, at §5.4, Table 4

Pith says: minor objection 6 of 6

The abbreviation G3D is used in the main text before being defined in Appendix C; please define it at first use.

What would settle it

Re-run the τ-sweep from Figure 2 on a fresh held-out split (e.g., a new OJ split for SpreadsheetBench or the WikiTQ/HiTab OOD sets) without consulting the original OJ results, and check whether τ=−0.001 remains the best threshold. If a different τ is optimal on the fresh split, the fixed τ in the main experiments was likely selected using the reported test set, and the main results should be re-estimated with τ chosen from a validation split only.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 10 Aug 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 51 s · 23,679 tokens read, 11,548 written
  3. Editor wrote the verdict above 42 s · 24,576 tokens read, 4,615 written
  4. Referee major revision: 3 major objections, 6 minor 1 min 37 s · 24,039 tokens read, 10,496 written

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

Read SkillProx. The two mechanisms—closed-loop forward updates that re-execute edits on the same batch and feed rejection outcomes into later diagnoses, and the backward utility-audit stage with validation-gated consolidation/demotion/removal—are real and fairly clean. The paper is better than the abstract suggests: it is explicit that Eq. (2) is conceptual, it runs three backbones and three benchmarks (one in-distribution, two OOD), and the appendix contains honest limitations, including the caveat that the utility audit is 'candidate evidence' and that the case studies reuse seed 8.

The main problem is the headline. I recomputed the margin from Table 2. Against the strongest gradient-based baseline per cell, SkillProx is about 2.2 pp better on average, not 3.0. Against SkillGrad averaged over all cells it is ~2.7. So the abstract is overstated, independent of any design issue.

Second, and more load-bearing: the tau sweep. Section 5.4 and Figure 2 evaluate a range of candidate thresholds on the OJ test set and report the maximum at tau=-0.001, which is then fixed for all main experiments (Appendix C). If the authors looked at that OJ sweep before fixing tau, the main results are selected on the test set. The paper never says when the sweep was run relative to the main table, and given that Prox contributes 2.5 pp in the ablation, this is not a minor detail. I would ask the authors to either show a validation-only tau selection or disclose the ordering and re-evaluate.

The 20-task leave-one-out audit is noisy, but the validation gate mitigates that; it is a minor concern. More minor: three seeds, ablation on one seed with no paired test, no code yet.

The paper is worth a serious referee. It introduces a genuinely useful mechanism and the authors are transparent about many limitations. The referee should push hard on the tau-selection question and the arithmetic of the headline claim, and ask for code or at least a detailed reproducibility appendix. If those get fixed, this is a solid methods paper.

The strongest objection

The central claim is that SkillProx improves average accuracy by 3.0 percentage points over the strongest gradient-based baseline. The backward Prox stage contributes substantially to this gain (removing it costs 2.5 pp in Table 3). However, the Prox configuration is not validated independently of the reported test set. In Section 5.4 and Figure 2, the authors sweep the candidate threshold τ from −0.05 to +0.10 and evaluate every configuration on the OJ (held-out) benchmark, reporting that τ=−0.001 attains the maximum OJ hard accuracy (52.3%). Appendix C then fixes τ=−0.001 for the main experiments. If the authors inspected the OJ sweep before fixing τ, the main results are selected on the test set: the hyperparameter was chosen to maximize the metric that is later reported as the headline gain. The paper does not disclose the ordering of these steps or provide a validation-only selection. This is load-bearing because the Prox stage—the main differentiator of the method—depends on this threshold, and the sweep's optimal value exactly coincides with the fixed value. A secondary arithmetic issue: recomputing the margin from Table 2 over the strongest per-row gradient baseline (SkillGrad or SkillOpt) yields roughly 2.2–2.7 pp, not 3.0, so the abstract already appears overstated independent of selection bias. The reader's concern about 20-task utility noise is related but secondary: even a noisy audit could be rescued by the re-evaluation gate, whereas test-set selection of τ directly biases the reported numbers.

The reader's weakest_assumption concerns the reliability of the frozen leave-one-out utility audit on 20 validation tasks. That is a legitimate risk, but the more load-bearing issue is upstream of it: the threshold τ that determines which audited units become candidates appears to have been chosen using the same OJ set that is later reported as the main evaluation. Figure 2 explicitly sweeps τ and evaluates each setting on the OJ benchmark; τ=−0.001 is both the argmax of that sweep and the value fixed for all main runs. Without a statement that τ was fixed before any OJ evaluation, the main results are vulnerable to test-set selection bias. The paper's own ablations show Prox is worth 2.5 pp, so the credibility of the headline 3.0 pp claim hinges on this threshold choice. In addition, recomputing the advertised margin from Table 2 against the strongest per-row gradient baseline gives approximately 2.2–2.7 pp, so the abstract's 3.0 pp figure is not directly supported by the table even before considering selection effects. I agree with the reader that the 20-task audit is a weak point, but the threshold-selection issue is sharper and more directly tied to the central claim. The concrete test—re-running the τ-sweep on a fresh split—would settle whether the fixed τ is genuinely well-chosen or an artifact of peeking at the test set.

The referee's report

If the reported gains are robust, SkillProx is a useful practical contribution: it couples outcome-gated forward edits with structure-aware backward consolidation, and the paper is unusually transparent about its limitations (e.g., Appendix D.3). Concrete strengths include three backbone LLMs, six baselines, OOD transfer checks, component ablations, detailed gate traces in Appendix D.1, and explicit acknowledgement that the proximal-gradient framing is an analogy. However, the headline empirical claim is currently not well supported because the key threshold appears to be selected on the test set and the abstract's 3.0 pp figure is not reproducible from Table 2. The contribution's magnitude is therefore uncertain until these issues are addressed.

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

The paper is trying to establish that skill evolution for LLM agents can be made reliable by pairing outcome-verified forward edits with utility-aware backward pruning, rather than trusting the semantic plausibility of a diagnosis or letting skills grow without bound. If this is right, agent skills can be refined as lightweight, readable artifacts: every proposed patch is re-executed on the same task batch and kept only if it measurably improves or preserves accuracy, and accumulated knowledge units are periodically audited and consolidated or removed when they hurt. The evidence is an average 3.0-percentage-point accuracy gain over the strongest gradient-based baseline across three backbone LLMs, with gains holding on out-of-distribution benchmarks. The practical upshot is that self-evolving procedural memory can be both effective and interpretable, without any weight updates.

What carries the argument. The load-bearing mechanism is a forward-backward loop borrowed conceptually from proximal gradient descent. The forward operator is the closed-loop gate of Eq. (5): a candidate skill produced by diagnose-and-patch is re-executed on the same batch and accepted only if hard accuracy and mean cell accuracy do not drop, with rejected edits recorded as feedback for the next diagnosis. The backward operator is the Prox stage: the skill is parsed into L2 sections and L3 reference files as auditable knowledge units, each unit's marginal utility is estimated once by leave-one-out evaluation on a fixed validation split (Eq. (7)), candidates with cell utility below $\tau=-0.001$ are processed in lexicographic order (Eq. (8)), and a Shrinker edits one unit at a time subject to structural validity, strict complexity reduction, a hard-accuracy tolerance $\delta_h=0$, a cell-accuracy tolerance $\delta_c=0.02$, and a soft compression cap $\rho=0.10$. This realizes the composite objective $J_\lambda(X)=L_T(X)+\lambda G(X)$ as two discrete, non-differentiable operators: forward diagnostic evolution as an inexact performance-gradient step and Prox as a complexity-regularizing proximal step.

Core claim. SkillProx claims that skill evolution should be treated as a forward-backward optimization over a composite objective balancing task loss and skill complexity, and that the two missing pieces in existing open-loop, growth-oriented methods are outcome-grounded verification of forward edits and utility-guided regularization of accumulated knowledge. Its forward stage re-executes each candidate edit on the same task batch, accepts only updates that pass a hard-and-cell accuracy gate, rolls back regressions, and feeds the measured effects into later diagnoses. Its backward stage parses the skill into auditable knowledge units, estimates each unit's marginal utility through a frozen leave-one-out evaluation on a 20-task validation split, and applies validation-gated consolidation, demotion, or removal. Across SpreadsheetBench in-distribution and WikiTQ and HiTab out-of-distribution, with three backbone LLMs, SkillProx reports an average accuracy improvement of 3.0 percentage points over the strongest gradient-based baseline, and component ablations show both stages contribute.

The premise it rests on. The entire backward stage rests on the assumption that deleting one rule at a time, measured on only 20 frozen validation tasks, gives a trustworthy ranking of which rules deserve consolidation or removal; if those scores are noisy or rules interact, the shrinkage may delete or reorder useful content.

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

If SkillProx is right, agents can evolve skills online without trusting the face plausibility of a diagnosis: a promising-sounding patch that lowers same-batch accuracy is rolled back, and its measured failure is given to the next diagnostic attempt.

Accumulated skill length is not a reliable proxy for quality; removing negative-utility units can simultaneously shrink the skill and raise accuracy, as in the reported lift from 46% to 54% hard accuracy with 3.12% compression.

Skill evolution is not inherently overfitting-prone: skills optimized on SpreadsheetBench generalize to WikiTQ and HiTab, while the baseline SkillOpt collapses out-of-distribution.

Smaller backbones receive a larger share of benefit from the external validation gates: the 4B model shows fewer writes per diagnosis and stronger realized compression under closed-loop training plus Prox, yielding stable gains across seeds.

The forward gate's reliance on same-batch re-execution selects for falsifiable edits, changes whose behavioral effect can be measured immediately, so SkillProx likely biases evolved skills toward testable, executable rules and away from vague meta-advice; the paper's case study of an actionable dynamic-reference rule replacing a 'trace an example' template supports this reading.

Because the utility audit is frozen before Prox and estimated by full deletion, a natural extension would be to update utility estimates after each consolidation, allowing interaction effects between knowledge units to be discovered rather than assumed away.

The forward-backward split is not specific to skills: any text-space optimizer that can re-run a candidate and decompose its artifact into editable units could adopt the same outcome-gated forward step and validation-gated shrinkage, so the framework may transfer to prompt bundles or tool libraries.

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

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

Reviews by people

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

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

LLM agents increasingly adapt to recurring tasks by accumulating procedural knowledge in skills. These skills are lightweight, reusable textual artifacts that are loaded into the agent's context without weight updates. Recent methods refine skills through iterative task execution, failure diagnosis, and trajectory-guided text-space updates. However, existing frameworks lack explicit diagnosis–outcome feedback and treat deletion as a generic edit operation rather than a dedicated mechanism for consolidating accumulated knowledge. We introduce SkillProx, a proximal-gradient-inspired forward–backward framework that couples closed-loop diagnostic evolution with utility-aware proximal refinement. Motivated by a composite objective balancing task loss and skill complexity, the forward stage re-executes diagnosis-driven edits on the same task batch, rolls back regressions, and feeds measured outcomes into subsequent diagnoses. The backward stage decomposes the resulting skill into auditable knowledge units, estimates their contributions using a frozen leave-one-out utility audit, and applies validation-gated consolidation, demotion, or removal. Experiments on in-distribution and out-of-distribution benchmarks across multiple backbone LLMs show that SkillProx improves average accuracy by 3.0 percentage points over the strongest gradient-based baseline. Component ablations demonstrate the complementary effects of closed-loop diagnosis and proximal refinement.

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 · 2026-09-23

The authors' main claim, as the classifier picked it out 1.00

“Experiments on in-distribution and out-of-distribution benchmarks across multiple backbone LLMs show that SkillProx improves average accuracy by 3.0 percentage points over the strongest gradient-based baseline.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.11 · states limitation 0.09 · 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.

SentenceRoleThe authors' own workStates the main conclusionStrength
LLM agents increasingly adapt to recurring tasks by accumulating procedural knowledge in skills. background 1.00 0.05 0.02 asserted 1.00
These skills are lightweight, reusable textual artifacts that are loaded into the agent's context without weight updates. background 1.00 0.07 0.02 asserted 1.00
Recent methods refine skills through iterative task execution, failure diagnosis, and trajectory-guided text-space updates. background 1.00 0.04 0.02 asserted 1.00
However, existing frameworks lack explicit diagnosis--outcome feedback and treat deletion as a generic edit operation rather than a dedicated mechanism for consolidating accumulated knowledge. gap 1.00 0.08 0.03 asserted 1.00
We introduce SkillProx, a proximal-gradient-inspired forward--backward framework that couples closed-loop diagnostic evolution with utility-aware proximal refinement. contribution 0.97 0.89 0.13 asserted 1.00
Motivated by a composite objective balancing task loss and skill complexity, the forward stage re-executes diagnosis-driven edits on the same task batch, rolls back regressions, and feeds measured outcomes into subsequent diagnoses. method 1.00 0.87 0.05 asserted 0.99
The backward stage decomposes the resulting skill into auditable knowledge units, estimates their contributions using a frozen leave-one-out utility audit, and applies validation-gated consolidation, demotion, or removal. method 0.99 0.91 0.05 asserted 0.97
Experiments on in-distribution and out-of-distribution benchmarks across multiple backbone LLMs show that SkillProx improves average accuracy by 3.0 percentage points over the strongest gradient-based baseline. result 1.00 0.97 0.69 asserted 1.00
Component ablations demonstrate the complementary effects of closed-loop diagnosis and proximal refinement. result 0.82 0.94 0.07 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99

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, 2026-09-23: 1 request, 6,887 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 (4)
Figure 1 of the paper
Figure 1. Pipeline of SkillProx. The forward stage iteratively executes, diagnoses, and patches the current skill, using same￾batch re-execution to accept beneficial updates or roll back rejected ones with outcome feedback. After K updates, the backward stage audits the utility of knowledge units on a fixed validation split and applies validation-gated consolidation, demotion, or removal. Accepted edits produce the final skill X⋆ . satisfies an outcome-grounded performance gate. Otherwise, the update is rolled back and retried, while its measured effect and rejected edit direction are…
Figure 2 of the paper
Figure 2. Accuracy–compression Pareto frontier of Qwen3.6-
Figure 3 of the paper
Figure 3. OJ accuracy before and after validation-gated Prox consolidation. Evaluation uses 100 fixed tasks with three test cases
Figure 4 of the paper
Figure 4. Final G3D skill length versus SpreadsheetBench IID OJ hard accuracy for nine Qwen3.5-27B training seeds. Length

Figures are the authors', reproduced from the paper.

References (12)
  1. [10] Consolidate (preferred): salvage unique generalizable content into the 12most-overlapping retained section, then remove the target on Pith
  2. [62] the previous momentum_memory.md (empty at iteration 1); and on Pith
  3. [73] the current SKILL.md and references. 8 9Write: on Pith
  4. [101] an updated cross-iteration pattern record; and on Pith
  5. [111] Group overlay entries sharing a pattern on Pith
  6. [112] a per-task overlay in the same order as the diagnoses. 12 13A pattern is a class of mistake or success, not a task instance. Default to 14merging evidence into an existing pattern. Split only when the required 15operations or corrective actions genuinely differ. Preserve remed... on Pith
  7. [122] Read their pattern history and raw diagnoses on Pith
  8. [132] Demote: move excessive L2 detail into a concise L3 procedure on Pith
  9. [133] Brainstorm 2--3 candidate remedies on Pith
  10. [143] Remove: delete content that is fully redundant or misleading. 15 16Touch only the target and, when needed, one receiving section. Preserve all 17L2--L3 pointer invariants. A downstream validation gate will reject the edit 18if accuracy drops. The resulting skill must be strict... on Pith
  11. [144] Apply the simplest edit that generalizes. 15 16Use YAGNI. Prefer extending or merging an existing section over creating a 17new one. Never place task-specific columns, rows, filenames, or constants in 18L2. Every L3 file must have exactly one L2 pointer and must contain runnab... on Pith
  12. [2025] Large Language Model Agent: A Survey on Method- ology, Applications and Challenges. arXiv:2503.21460. Ma, Z.; Zhang, B.; Zhang, J.; Yu, J.; Zhang, X.; Zhang, X.; Luo, S.; Wang, X.; and Tang, J. 2024. SpreadsheetBench: TowardsChallengingRealWorldSpreadsheetManipulation. arXiv:2... on PithPith's reviewarXiv
Cited by 2 Pith papers
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 13 sentences without a confident label

Uses data or values from 3 papers
Compares with 4 papers
Cites as background 7 papers
  • SkillOpt [ref] bounds each edit with a textual learning rate and a held-out gate, while SkillGrad [ref] treats the skill package as a structured parameter updated by trajectory-level loss evidence, textual gradients, and momentum.

    SkillOpt: Executive Strategy for Self-Evolving Agent Skills · Related Works · 0.86

  • SkillOpt [ref] bounds each edit with a textual learning rate and a held-out gate, while SkillGrad [ref] treats the skill package as a structured parameter updated by trajectory-level loss evidence, textual gradients, and momentum.

    SkillGrad: Optimizing Agent Skills Like Gradient Descent · Related Works · 0.86

  • Recent LLM agents increasingly externalize procedural knowledge into skills—structured bundles of instructions, scripts, and resources that are loaded into the agent's context at inference time and improve task execution without any weight update [ref] .

    SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning · Related Works · 0.87

Challenges and discussion

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

No comments yet.

Sign in to comment