Pith.

Machine Learning · arXiv:2610.25126 v1 · submitted 18 Oct 2026 · 69 references

Trackable Skill Library Lifecycle Management for Agentic Reinforcement Learning

Xi Zhu, Yuanlin Duan, Changyi Yang, Kai Mei, Riyang Bao, Shichao Zhu, Dimitris N. Metaxas

The paper says

“Skill-augmented agentic RL has emerged as an effective approach to improving LLM agents on long-horizon tasks.”

more of the abstract

Despite this progress, individual skills may follow distinct trajectories as the policy evolves, with some becoming redundant after the model internalizes their capabilities, others remaining valuable as external scaffolding, and still others turning harmful. However, existing skill library systems fail to identify each skill's true contribution due to free-riding among co-retrieved skills, difficulty confounding across tasks, and the absence of skill-specific counterfactual evidence, which prevents reliable skill lifecycle management. We propose TrackRL, an agentic RL framework where skill tracking drives skill library full lifecycle management. Specifically, within each GRPO group, we turn grouped sampling into online randomized controlled experiments. Independent random injection and within-group contrasts on the same task isolate each skill's true contribution, and reliability-weighted aggregation accumulates evidence across groups. As evidence accumulates, we continuously track skill utility, which measures the skill's causal contribution as external scaffolding, and model mastery, which reflects how far the corresponding capability has been internalized. Along this line, we formulate skill lifecycle management as an evidence-gated state-transition process that governs each skill from generation and admission through protected incubation to retention, retirement, or removal, while a soft capacity bound controls library growth. Experiments show that TrackRL outperforms recent skill-augmented RL baselines in both library-assisted and skill-free evaluation, with further analyses revealing how skill value shifts during policy learning and how evidence-driven state transitions shape skill library evolution. Our code is available at https://github.com/xizhu1022/TrackRL.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 20 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: major revision.

TrackRL proposes a clever randomized injection scheme to estimate per-skill utility in agentic RL, enabling principled lifecycle management; the empirical results are strong and well-supported by ablations. Lack of error bars on main results is a real limitation but not load-bearing, and the paper acknowledges the core assumption of no skill interactions. 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 69 references say.

The paper says, at Tables 1 and 2

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.77):

“Since each group contains onlyB episodes, individual comparisons can be coarse and noisy, making uncertainty essential to avoid relying on a point estimate alone for skill lifecycle decisions.”

Pith says: Objection 1 of 3 · major

All reported success rates (SR, TGC, SGC) are point estimates without confidence intervals, standard errors, or multiple random seeds. The paper acknowledges that per-group comparisons are noisy (sentence b79: 'individual comparisons can be coarse and noisy'), yet this uncertainty is not propagated to the final metric.

more of objection 1

Some performance gaps are large (e.g., 78.1% vs. 62.1% on ALFWorld seen assisted), but others are modest (e.g., 68.8% vs. 66.4% for TrackRL (Flash) vs. D2Skill (Flash) on ALFWorld seen assisted; 23.2% vs. 20.8% TGC-N on AppWorld). Without uncertainty quantification, the reader cannot assess whether these differences reflect a reliable advantage or sampling variation. This is load-bearing because the paper's central claim directly depends on these comparisons. The authors should either report confidence intervals from bootstrap across validation episodes, or run multiple independent training seeds and report means with standard deviations.

The paper says, at Eq. (4), Section 2.1

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The independent Bernoulli injection (Eq. 4) assumes that each skill's marginal contribution does not depend on which other skills are present (no interference). If the effect of skill A changes when skill B is also injected, the estimated utility us may not represent the skill's value in the co-retrieved set that actually occurs during deployment.

more of objection 2

The paper does not discuss or test this assumption. A practical test would be to compare the sum of individual utility estimates from Eq. (5)–(7) against the bundle-level improvement estimated by jointly or withholding all skills. If they diverge substantially, the independence assumption is violated. The authors should add such a diagnostic or at least state the assumption explicitly as a limitation.

The paper says, at Table 5

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The method introduces several hyperparameters (injection probability p=1/2, evidence threshold E_min=10, confidence level λ=2, recency discount γ=0.98, soft capacity C=300, etc.) with no sensitivity analysis. While the authors may have tuned these on a validation set, the paper provides no evidence that the results are robust to reasonable variation in these values.

more of objection 3

Since lifecycle decisions are directly governed by these thresholds (Eqs. 11, 12), a sensitivity study on at least the most critical parameters (E_min, λ, γ) is necessary to demonstrate that the method does not rely on fragile tuning.

6 minor objections

The paper says, at Algorithm 1

Pith says: minor objection 1 of 6

The algorithm is well-structured but omits the step for updating the family baseline b_f(s). State explicitly how b_f(s) is computed and how often it is updated.

The paper says, at Figure 3

Pith says: minor objection 2 of 6

The ablation figure shows only point estimates. While the drops are large, adding error bars or reporting the number of validation episodes per bar would improve interpretability.

show 4 more

The paper says, at Figure 5

Pith says: minor objection 3 of 6

The figure is informative, but the marker size reflecting accumulated evidence is difficult to discern in print. Consider using a more distinguishable visual cue, such as transparency or numeric labels.

The paper says, at Section 2.1, Eq. (9)

Pith says: minor objection 4 of 6

The mastery estimate uses a Beta posterior with uniform prior (c^-_s+1, n^-_s+2). Clarify why the prior is chosen as Beta(1,1) and whether the choice affects early-stage comparisons.

The paper says, at Appendix D.3, Table 4

Pith says: minor objection 5 of 6

The learning rate is 1e-6 for all methods. Justify this choice and state whether the baseline methods (SkillRL, D2Skill) were retrained with the same optimizer settings or if the results are taken from prior publications.

The paper says, at Section 3, first paragraph

Pith says: minor objection 6 of 6

The sentence 'Extensive experiments show that TrackRL outperforms recent skill-augmented RL baselines...' appears twice (b10 and b40). Consider removing one instance for conciseness.

What would settle it

Run the ALFWorld cold-start experiments (base model, Luna reflection) for TrackRL, SkillRL, and D2Skill with 3 independent random seeds each. Report the mean and standard deviation of the success rates on seen/unseen splits in both skill-assisted and skill-free modes. If TrackRL's mean exceeds the 95% confidence interval of the best baseline on both splits in at least one evaluation mode, the concern is resolved; otherwise, the outperformance claim is not statistically supported.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 20 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic partly agrees: says it holds up only with conditions 1 min 10 s · 25,392 tokens read, 5,726 written Main results lack confidence intervals or multiple trials; without them, the reported outperformance may not be statistically significant, especially on splits with smaller gaps.
  3. Editor wrote the verdict above 16 s · 26,240 tokens read, 1,229 written
  4. Referee major revision: 3 major objections, 6 minor 34 s · 25,737 tokens read, 1,967 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

TrackRL is the first paper I have seen that turns GRPO’s grouped sampling into an online RCT to isolate each skill’s marginal contribution. That design is genuinely clever and directly addresses the free-riding and difficulty confounding that have plagued bundle-level evaluations in skill-augmented RL. Building utility, mastery, and evidence-gated state transitions on top of that foundation gives the library a principled lifecycle—admission, incubation, retirement, removal—rather than ad hoc pruning. The paper is well-written, the method is clearly explained, and the appendix is thorough. The empirical evaluation covers two benchmarks, two reflection models, and two initializations, with ablations that isolate the contribution of tracking and lifecycle management. The code is released. All of this is solid work.

The main soft spot is the absence of confidence intervals or multiple seeds on the headline success rates (Tables 1 and 2). The stress-test note has a point: without uncertainty quantification, a few of the smaller gaps (e.g., 68.8% vs 66.4%) could be noise. That said, the larger gaps—especially in the ablation where skill-free SR drops from 72.7% to 53.1%—are dramatic enough that the central claim holds. The paper also acknowledges noise in per-group comparisons and uses confidence bounds for its own lifecycle decisions, so the omission is a reporting weakness rather than a fatal flaw. I would ask for error bars or a note that runs were single-seed due to compute cost (which is a defensible position given the 4B policy).

The second soft spot is the untested interaction assumption: Eq. (4) assumes independent skill effects, but if one skill’s value changes depending on which other skills are present, the utility estimates may not reflect real marginal contributions. The authors do not discuss this. It is a typical limitation of such designs, and the paper would benefit from a sensitivity analysis or at least a clear statement.

A minor point: the paper uses closed-source reflection models (GPT-5.6-Luna and Gemini Flash), but the method’s advantage over baselines is demonstrated with both, so the findings are not tied to a single model.

This is a strong, methodologically novel paper that deserves a serious referee. I would accept it with a request to add uncertainty estimates to the main tables and to discuss (or test) the interaction assumption.

The strongest objection

The paper's central claim that TrackRL outperforms baselines is supported only by point estimates of success rates (Tables 1 and 2) with no confidence intervals, standard errors, or multiple random seeds. While some gaps are large (e.g., 78.1% vs 62.1% on ALFWorld seen assisted), others are modest (e.g., 68.8% vs 66.4% for TrackRL (Flash) vs D2Skill (Flash) on seen assisted; 23.2% vs 20.8% TGC-N on AppWorld). Without uncertainty quantification, it is impossible to assess whether these differences reflect a reliable advantage or arise from sampling noise. The ablation study also omits error bars, though it shows dramatic drops. This is load-bearing because if the headline results are not statistically reliable, the central claim is unsupported. The paper acknowledges noise in per-group comparisons (b79) but does not propagate that uncertainty to final metrics.

The reader identified the independence-of-skills assumption as the weakest, which is a valid conceptual concern but is partially mitigated by the design (marginal ATE) and by the ablation showing the full system works. I view the lack of statistical rigor in the main empirical comparison as more load-bearing: without error bars or multiple seeds, the central outperformance claim rests on point estimates that could be noisy. This is a concrete omission that a conditional acceptance requesting uncertainty quantification would address. The paper's own acknowledgment of noise (b79) underscores this gap. However, the paper is otherwise well-structured and the method is sound; the concern does not warrant rejection but does call for stronger evidence.

The referee's report

If the results hold, TrackRL provides a principled solution to a recognized problem in skill-augmented agentic RL: how to distinguish skills that remain valuable as external scaffolding from those that have been internalized, become redundant, or turned harmful. The approach of converting GRPO group sampling into within-group randomized experiments is technically clean and could become a standard component in future skill-library systems. The paper also delivers substantial engineering contributions (open-source code, clear algorithmic specification) and conducts extensive ablations and longitudinal analyses that illustrate the dynamics of skill utility and mastery during training. The significance is high, provided the empirical claims are statistically reliable.

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

Existing skill-augmented RL systems cannot tell whether a skill truly helps, free-rides on other skills, or has already been internalized by the policy. TrackRL solves this by turning each GRPO training group into a series of mini randomized controlled trials: it independently decides whether to inject each retrieved skill into each episode, creating per-skill treatment and control arms that strip away task-difficulty confounding and co-retrieval free-riding. From the within-group contrasts it estimates two quantities for each skill: utility (the causal lift in success rate from injecting it) and mastery (the policy's success rate when the skill is withheld). These estimates, along with their uncertainty, drive an evidence-gated state machine that admits new skills, protects them until enough evidence accumulates, then retains useful ones, retires internalized ones, and removes harmful ones. The result is a library that co-evolves with the policy, and experiments show TrackRL beats previous skill-augmented RL methods in both skill-assisted and skill-free evaluation on two challenging benchmarks.

What carries the argument. The central mechanism is the combination of independent Bernoulli randomization of skill injection (Eq. 4) with within-group contrasts (Eq. 5) to isolate each skill's effect, and the evidence-gated state-transition process that uses discounted running statistics (Eqs. 6-9) to estimate utility $u_s$ and mastery $m_s$ with standard errors, gating all lifecycle decisions (retirement, removal, eviction) on a minimum evidence threshold $E_{\min}$ and confidence bounds $\lambda SE(\cdot)$.

Core claim. The core claim is that independent randomized injection of each retrieved skill within a GRPO group allows unbiased estimation of that skill's marginal contribution to the current policy (utility) and the extent to which the policy has internalized the capability (mastery). By accumulating these estimates over training with discounting, TrackRL can then govern each skill through a lifecycle: protection during evidence accumulation, retention if utility remains positive, retirement if utility is near zero and mastery is high (indicating the skill is no longer needed), and removal if utility is negative (harmful). The paper demonstrates through ablation that both the tracking and the lifecycle management are necessary for the performance gains, and that the framework improves not only skill-assisted performance but also the policy's unaided (skill-free) performance, showing genuine internalization.

The premise it rests on. The paper assumes that independently randomizing each skill's injection yields unbiased estimates of each skill's marginal contribution, implicitly assuming that the effect of one skill does not depend on which other skills are also present (no interaction effects), an assumption that is not tested or discussed.

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

Skill-augmented RL systems can now automatically retire skills the agent has internalized, reducing prompt length and improving inference efficiency.

The framework provides a principled way to detect and remove harmful skills before they degrade performance, as the retention/removal rules rely on statistical confidence bounds.

TrackRL's ability to improve skill-free performance suggests that the library management helps the policy genuinely learn, rather than merely relying on external crutches.

Because the method estimates each skill's contribution online without extra rollouts, it can be integrated into any GRPO-based agentic RL pipeline with only a small bookkeeping overhead.

The evidence-gated lifecycle ensures that new skills receive a protected incubation period, preventing premature pruning that could discard potentially useful skills before their value is clear.

The assumption of independent skill utility estimates (no interaction effects between skills) is untested: if the marginal contribution of one skill depends on which other skills are co-injected, the estimates may be biased. A follow-up study could compare full factorial designs against the current independent randomization to measure interaction magnitudes.

The framework makes the skill library auditable: every lifecycle decision is justified by accumulated statistics, which could be useful for safety-critical deployments where one needs to understand why a particular piece of context was active.

The current method only tracks mastery for task-level skills; extending mastery tracking to step-level skills would require a different counterfactual design, but the paper does not attempt it, leaving open a direction for future work.

The soft capacity bound temporarily exceeds the limit when no sufficiently evaluated non-beneficial skill is available for eviction, which means library size can grow unboundedly in practice if new skills are constantly being generated, potentially increasing computational cost.

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

Skill-augmented agentic RL has emerged as an effective approach to improving LLM agents on long-horizon tasks. Despite this progress, individual skills may follow distinct trajectories as the policy evolves, with some becoming redundant after the model internalizes their capabilities, others remaining valuable as external scaffolding, and still others turning harmful. However, existing skill library systems fail to identify each skill's true contribution due to free-riding among co-retrieved skills, difficulty confounding across tasks, and the absence of skill-specific counterfactual evidence, which prevents reliable skill lifecycle management. We propose TrackRL, an agentic RL framework where skill tracking drives skill library full lifecycle management. Specifically, within each GRPO group, we turn grouped sampling into online randomized controlled experiments. Independent random injection and within-group contrasts on the same task isolate each skill's true contribution, and reliability-weighted aggregation accumulates evidence across groups. As evidence accumulates, we continuously track skill utility, which measures the skill's causal contribution as external scaffolding, and model mastery, which reflects how far the corresponding capability has been internalized. Along this line, we formulate skill lifecycle management as an evidence-gated state-transition process that governs each skill from generation and admission through protected incubation to retention, retirement, or removal, while a soft capacity bound controls library growth. Experiments show that TrackRL outperforms recent skill-augmented RL baselines in both library-assisted and skill-free evaluation, with further analyses revealing how skill value shifts during policy learning and how evidence-driven state transitions shape skill library evolution. Our code is available at https://github.com/xizhu1022/TrackRL.

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 1.00

“Experiments show that TrackRL outperforms recent skill-augmented RL baselines in both library-assisted and skill-free evaluation, with further analyses revealing how skill value shifts during policy learning and how evidence-driven state transitions shape skill library evolution.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.09 · human participants 0.03 · controlled comparison 0.93 · randomized assignment 0.93 · states quantitative result 0.05 · compares with prior results 0.96 · mentions code or data availability 0.99

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
Skill-augmented agentic RL has emerged as an effective approach to improving LLM agents on long-horizon tasks. background 1.00 0.05 0.03 asserted 1.00
Despite this progress, individual skills may follow distinct trajectories as the policy evolves, with some becoming redundant after the model internalizes their capabilities, others remaining valuable as external scaffolding, and still others turning harmful. background 0.88 0.10 0.04 hedged possibility 0.97
However, existing skill library systems fail to identify each skill's true contribution due to free-riding among co-retrieved skills, difficulty confounding across tasks, and the absence of skill-specific counterfactual evidence, which prevents reliable skill lifecycle management. gap 0.99 0.05 0.04 asserted 1.00
We propose TrackRL, an agentic RL framework where skill tracking drives skill library full lifecycle management. contribution 0.99 0.82 0.13 asserted 1.00
Specifically, within each GRPO group, we turn grouped sampling into online randomized controlled experiments. method 0.99 0.93 0.03 asserted 0.99
Independent random injection and within-group contrasts on the same task isolate each skill's true contribution, and reliability-weighted aggregation accumulates evidence across groups. method 1.00 0.91 0.06 asserted 1.00
As evidence accumulates, we continuously track skill utility, which measures the skill's causal contribution as external scaffolding, and model mastery, which reflects how far the corresponding capability has been internalized. method 0.94 0.91 0.07 asserted 1.00
Along this line, we formulate skill lifecycle management as an evidence-gated state-transition process that governs each skill from generation and admission through protected incubation to retention, retirement, or removal, while a soft capacity bound controls library growth. method 0.88 0.87 0.11 asserted 1.00
Experiments show that TrackRL outperforms recent skill-augmented RL baselines in both library-assisted and skill-free evaluation, with further analyses revealing how skill value shifts during policy learning and how evidence-driven state transitions shape skill library evolution. result 1.00 0.97 0.81 asserted 1.00
Our code is available at https://github.com/xizhu1022/TrackRL. contribution 0.71 0.52 0.02 asserted 0.90

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97

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, 7,744 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 (8)
Figure 1 of the paper
Figure 1. illustrates the distinct evolutionary trajectories of individual skills. As the model policy evolves, some skills lose marginal value and become redundant once the model internalizes the capability (Xiao et al., 2026; Xia et al., 2026b). Others remain valuable as external scaffolding for long-tail cases or hard-to-internalize knowledge (Wang et al., 2024; Suzgun et al., 2025), while some become harmful due to reflection errors or policy changes (Liu, 2026; Zhang & Li, 2026). To manage these divergent trajectories, a continually learning agent must track each skill’s current utility…
Figure 2 of the paper
Figure 2. Overview of TrackRL. We maintain a skill library of task-level and step-level skills, where we randomize their injection within GRPO groups to create per-skill treatment and control arms. From these comparisons, we track skill utility, model mastery, and evidence to guide skill lifecycle management. The updated library then supports subsequent policy optimization, closing the loop of policy–library co-evolution. skills without additional rollouts. Randomized injection thus compares the two arms under the same task and policy, removing difficulty confounding, while independent…
Figure 3 of the paper
Figure 3. Ablation on ALFWorld. Success rate (%) of TrackRL and its three variants in the four evaluation settings, namely the seen and unseen splits evaluated in the skill-assisted and skill-free modes. The variant w/o both has no library, so its two modes coincide. 0 25 50 75 100 125 150 0 200 400 600 skills in library capacity: 300/pool budget (a) Library size (base model) TrackRL (Luna) D2Skill (Luna) SkillRL (Luna) 0 20 40 60 80 100 120 140 160 0 200 400 600 (b) Composition of the TrackRL library 0 25 50 75 100 125 150 training step 0 200 400 600 skills in library (c) Library size…
Figure 4 of the paper
Figure 4. Skill library growth on ALFWorld. (a, c) Library size of TrackRL, D2Skill, and SkillRL under cold start and warm start. (b, d) Composition of the TrackRL library by skill state. Significantly beneficial skills have utility exceeding zero by two standard errors. 3.4 Skills on the Utility–Mastery Plane To examine how utility and mastery jointly inform a skill’s fate, we plot every active task-level skill under cold start at four training steps by its current utility and mastery, with color indicating its eventual fate, as shown in
Figure 5 of the paper
Figure 5. Task-level skills on the utility–mastery plane. Each marker represents an active skill positioned by its current utility us and mastery ms, with color indicating its eventual lifecycle outcome and size reflecting its accumulated evidence. Corner numbers report the skill count in each quadrant, while arrows trace the earlier positions of skills later retired or removed. 0 20 40 60 80 100 120 140 160 training step −20 0 20 40 events per 5 steps budget admitted evicted retired removed library size protected 0 200 400 600 skills in library
Figure 6 of the paper
Figure 6. TrackRL lifecycle events on ALFWorld. Bars show admissions upward and exits downward over five-step intervals, while lines show active and protected skill counts. 3.5 Skill Library Lifecycle Evolution To trace how the library evolves, we aggregate TrackRL’s cold-start lifecycle events over five-step intervals in
Figure 7 of the paper
Figure 7. Three skill lifecycles from ALFWorld. The three skills end in removal, retirement, and retention, respectively. Shaded bands indicate utility confidence intervals. 4 Related Work Skill-Augmented Agentic RL. Recent skill-augmented agentic RL methods differ in how they use skills. Some distill reusable skills from experience and retrieve them as prompt context for subsequent rollouts and policy optimization (He et al., 2026; Zhang et al., 2026). Others adapt the strength, timing, and content of skill guidance to task outcomes or policy competence, sometimes learning when and which…
Figure 8 of the paper
Figure 8. Task-level skills by family (ALFWorld, base model, Luna reflection). One panel per family, showing the mean mastery of the family’s active task-level skills on the left axis, the success rate of the policy without the library on the seen validation games of that family on the same axis as a dashed line and as a three-point moving average since each family has about twenty games per validation, the family utility pooled over the family’s skills on the right axis, and one vertical tick per retirement of a skill of that family. Panel titles give the number of active task-level skills…

Figures are the authors', reproduced from the paper.

References (69)
  1. [1] Xia, Peng and Chen, Jianwen and Wang, Hanyang and Liu, Jiaqi and Zeng, Kaide and Wang, Yu and Han, Siwei and Zhou, Yiyang and Zhao, Xujiang and Chen, Haifeng and Zheng, Zeyu and Xie, Cihang and Yao, Huaxiu , journal = on Pith
  2. [2] Tu, Songjun and Xu, Chengdong and Zhang, Qichao and Zhang, Yaocheng and Lan, Xiangyuan and Li, Linjing and Li, Dong and Zhao, Dongbin , journal =. Dynamic Dual-Granularity Skill Bank for Agentic on Pith
  3. [3] Xiao, Han and Wang, Guozhi and Wang, Hao and Liu, Shilong and Chai, Yuxiang and Pan, Yue and Zhou, Yufeng and Chen, Xiaoxin and Wen, Yafei and Li, Hongsheng , journal = on Pith
  4. [4] Yu, Ye and Yuan, Xiaopeng and Jin, Haibo and Liu, Heming and Yu, Yaoning and Wang, Haohan , journal =. Do Self-Evolving Agents Forget? on Pith
  5. [5] Shridhar, Mohit and Yuan, Xingdi and C. International Conference on Learning Representations (ICLR) , year = on Pith
  6. [6] Yao, Shunyu and Chen, Howard and Yang, John and Narasimhan, Karthik , booktitle =. 2022 , url = on Pith
  7. [7] Trivedi, Harsh and Khot, Tushar and Hartmann, Mareike and Manku, Ruskin and Dong, Vinty and Li, Edward and Gupta, Shashank and Sabharwal, Ashish and Balasubramanian, Niranjan , booktitle =. 2024 , url = on Pith
  8. [8] Pu, Hongji and Song, Xinyuan and Zhao, Liang , journal = on Pith
  9. [9] Zhao, Andrew and Huang, Daniel and Xu, Quentin and Lin, Matthieu and Liu, Yong-Jin and Huang, Gao , booktitle =. 2024 , url = on Pith
  10. [10] Liu, Ming , journal =. More Is Not Always Better: Cross-Component Interference in on Pith
  11. [11] Wang, Guanzhi and Xie, Yuqi and Jiang, Yunfan and Mandlekar, Ajay and Xiao, Chaowei and Zhu, Yuke and Fan, Linxi and Anandkumar, Anima , journal = on Pith
  12. [12] Wang, Hao and Wang, Guozhi and Xiao, Han and Zhou, Yufeng and Pan, Yue and Wang, Jichao and Xu, Ke and Wen, Yafei and Ruan, Xiaohu and Chen, Xiaoxin and Qi, Honggang , journal = on Pith
  13. [13] Yao, Zhiyuan and Chen, Yuxin and Lu, Zhengxi and Xu, Zishan and Sun, Yueqing and Guo, Yifu and Lu, Yuquan and Cai, Zhengzhou and Zhang, Kangning and Han, Zhuowen and Wang, Zi-Han and Ye, Ziang and Gu, Qi and Cai, Xunliang and Liu, Weiwen and Shen, Yongliang , journal = on Pith
  14. [14] He, Zelin and Lin, Haotian and Han, Boran and Zhu, Wei and Fang, Haoyang and Wang, Bernie and Zhu, Xuan and Li, Runze and Reimherr, Matthew , journal = on Pith
  15. [15] From Interaction Traces to Persistent Skills: Online Evolution for Computer- Use Agents , author =. arXiv preprint arXiv:2609.04869 , year = on PithPith's reviewarXiv
  16. [16] Yu, Jianxiang and Zhu, Jiapeng and Lin, Bochen and Cui, Qier and Ding, Zichen and Li, Xiang , journal =. Skill is Not One-Size-Fits-All: Model-Aware Skill Alignment for on Pith
  17. [17] Xia, Siyu and Zhang, Chenheng and Wu, Yanting and Li, Haoxuan and Chai, Jiajun and Wang, Xiaohan and Yin, Guojun and Lin, Wei and Lin, Zhouchen and Zhang, Haifeng and Wang, Jun , journal = on Pith
  18. [18] Dabas, Mahavir and Jeong, Jihyun and Jin, Ming and Jia, Ruoxi , journal =. Memory-Induced Tool-Drift in on Pith
  19. [19] Zhang, Yan and Li, Shibo , journal = on Pith
  20. [20] Yan, Sikuan and Bahloul, Ahmed and Nie, Ercong and Schwarzmann, Susanna and Trivisonno, Riccardo and Tresp, Volker and Ma, Yunpu , journal = on Pith
show 49 more
  1. [21] Li, Qinfeng and Bao, Yuntai and Yu, Xinyan and Chen, Hongze and Liu, Yanming and Zhu, Huifeng and Jin, Yier and Chen, Jintao and Zhang, Wenqi and Zhang, Xuhong , journal = on Pith
  2. [22] Wang, Hanyang and Ren, Weijieying and Zhang, Yuxiang and Cao, Ding and Zeng, Zhizhao and Zeng, Ke and Zhao, Tianxiang , journal = on Pith
  3. [23] Ren, Bo and Mao, Yirong and Yang, Yi and Que, Wenhui , journal = on Pith
  4. [24] Lin, Huawei and Li, Peng and Song, Jie and Jiang, Fuxin and Zhang, Tieying , journal = on Pith
  5. [25] Qiu, Yuli and Li, Yutong and Su, Wei and Liu, Zeming and Che, Wanxiang and Huang, Heyan and Wang, Haifeng and Guo, Yuang , journal = on Pith
  6. [26] Dong, Jiajun and Hu, Yutao and Fan, Fengrui and Dou, Shihan and Wu, Yueming and Zou, Deqing , journal = on Pith
  7. [27] 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 , journal = on Pith
  8. [28] Agent Workflow Memory , author =. arXiv preprint arXiv:2409.07429 , year = on PithPith's reviewarXiv
  9. [29] Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory , author =. arXiv preprint arXiv:2504.07952 , year = on PithPith's reviewarXiv
  10. [30] Deng, Yuntian and Choi, Yejin and Shieber, Stuart , journal =. From Explicit on Pith
  11. [31] RL's Razor: Why Online Reinforcement Learning Forgets Less , author =. arXiv preprint arXiv:2509.04259 , year = on PithPith's reviewarXiv
  12. [32] Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , journal = on Pith
  13. [33] Zhang, Xichen and Wu, Sitong and Zhu, Yinghao and Tan, Haoru and Yu, Shaozuo and He, Ziyi and Jia, Jiaya , booktitle =. 2026 , url = on Pith
  14. [34] Fan, Zhimin and Yu, Hongwei and Shen, Yeqing and Yan, Haolong and Peng, Guozhen and Peng, Tianhao and Zhang, Yudong and Zhang, Xiaowen and Tan, Kaijun and Ge, Zheng and Zhang, Xiangyu and Jiang, Daxin , journal =. Skill-Guided Continuation Distillation for on Pith
  15. [35] Privileged Information Distillation for Language Models , author =. arXiv preprint arXiv:2602.04942 , year = on PithPith's reviewarXiv
  16. [36] Cohen-Wang, Benjamin and Shah, Harshay and Georgiev, Kristian and Madry, Aleksander , journal = on Pith
  17. [37] Datamodels: Predicting Predictions from Training Data , author =. arXiv preprint arXiv:2202.00622 , year = on PithPith's reviewarXiv
  18. [38] Shen, Qianli and Chen, Daoyuan and Huang, Yilun and Ling, Zhenqing and Li, Yaliang and Ding, Bolin and Zhou, Jingren , booktitle =. 2026 , url = on Pith
  19. [39] Dynamic Key-Value Memory Networks for Knowledge Tracing , author =. Proceedings of the International Conference on World Wide Web (WWW) , year = on Pith
  20. [40] Context-Aware Attentive Knowledge Tracing , author =. Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) , year = on Pith
  21. [41] Coalition-Aware Skill Reliability for Self-Evolving Agents , author =. arXiv preprint arXiv:2608.22610 , year = on PithPith's reviewarXiv
  22. [42] Yang, Shidong and Ma, Ziyu and Huang, Tongwen and Wang, Xucong and Li, Renda and Hu, Yiming and Wang, Yong and Chu, Xiangxiang , journal = on Pith
  23. [43] Tu, Songjun and Xu, Chengdong and Zhang, Qichao and Ma, Yiwen and Zhang, Yaocheng and Li, Linjing and Li, Dong and Lan, Xiangyuan and Zhao, Dongbin , journal = on Pith
  24. [44] He, Liang and Wen, Jingbo and Gu, Hongyu and Li, Hao and Wang, Haoyu and Chen, Yixiong and Cui, Kangning and Wang, Xilu , journal =. From Relevance to Execution Utility: Reward-Aware Dynamic Execution Gating for Skill-Based on Pith
  25. [45] When Experience Becomes Instruction: Trajectory Poisoning in Self-Evolving Agent Skill Systems , author =. arXiv preprint arXiv:2608.05563 , year = on PithPith's reviewarXiv
  26. [46] Lu, Pingchen and Wang, Xiangyi and Li, Xiang and Mao, Jie and others , journal = on Pith
  27. [47] Finite-time Analysis of the Multiarmed Bandit Problem , author =. Machine Learning , volume =. 2002 , doi = on Pith
  28. [48] Feng, Lang and Xue, Zhenghai and Liu, Tingcong and An, Bo , journal =. Group-in-Group Policy Optimization for on Pith
  29. [49] Agashe, Saaket and Han, Jiuzhou and Gan, Shuyu and Yang, Jiachen and Li, Ang and Wang, Xin Eric , journal =. Agent on Pith
  30. [50] Wang, Zhenhailong and Xu, Haiyang and Wang, Junyang and Zhang, Xi and Yan, Ming and Zhang, Ji and Huang, Fei and Ji, Heng , journal =. Mobile-Agent- on Pith
  31. [51] Reinforcement Learning for Self-Improving Agent with Skill Library , author =. arXiv preprint arXiv:2512.17102 , year = on PithPith's reviewarXiv
  32. [52] Li, Xiaoyuan and Li, Moxin and Bao, Keqin and Ma, Yubo and Wang, Wenjie and Liu, Dayiheng and Feng, Fuli , journal = on Pith
  33. [53] Shi, Yaorui and Chen, Yuxin and Lu, Zhengxi and Miao, Yuchun and Liu, Shugui and GU, Qi and Cai, Xunliang and Wang, Xiang and Zhang, An , journal = on Pith
  34. [54] Li, Yu and Miao, Rui and Qi, Zhengling and Lan, Tian , journal = on Pith
  35. [55] Fu, Dayuan and Jiang, Mohan and Wang, Tongyu and Yang, Dian and Hu, Jiarui and Liu, Liming and Hou, Jinlong and Liu, Pengfei , journal = on Pith
  36. [56] Shi, Yipeng and Ma, Zhipeng and Wang, Yue and Tan, Qitai and Li, Yang and Chen, Peng and Zhu, Zhengzhou , journal = on Pith
  37. [57] Wu, Jinyang and Zhai, Guocheng and Jin, Ruihan and Shen, Yuhao and Lu, Zhengxi and Zhang, Fan and Luo, Haoran and Lian, Zheng and Wen, Zhengqi and Tao, Jianhua , journal = on Pith
  38. [58] Lu, Zhengxi and Yao, Zhiyuan and Wu, Jinyang and Han, Chengcheng and Gu, Qi and Cai, Xunliang and Lu, Weiming and Xiao, Jun and Zhuang, Yueting and Shen, Yongliang , journal = on Pith
  39. [59] Lin, Hongxiang and Kuai, Zhirui and Xue, Erpeng and Wang, Lei , journal = on Pith
  40. [60] He, Zhongyu and Li, Yuanfan and Huang, Fei and Chen, Tianyu and Chen, Siyuan and Li, Xingyang and Yu, Meng Hsuan and Liu, Xiangrong and Wei, Leyi and Pan, Lu and Zeng, Ke and Cai, Xunliang , journal = on Pith
  41. [61] Yang, Shuo and Wu, Jinyang and Lu, Zhengxi and Shen, Yuhao and Zhang, Fan and Feng, Lang and Zhang, Shuai and Luo, Haoran and Lian, Zheng and Wen, Zhengqi and Tao, Jianhua , journal = on Pith
  42. [62] Wu, Jinyang and Yang, Shuo and Lu, Zhengxi and Zhang, Fan and Shen, Yuhao and Feng, Lang and Luo, Haoran and Lian, Zheng and Zhang, Shuai and Wen, Zhengqi and Tao, Jianhua , journal = on Pith
  43. [63] Self-Distilled Agentic Reinforcement Learning , author =. arXiv preprint arXiv:2605.15155 , year = on PithPith's reviewarXiv
  44. [64] Pan, Tianjun and Li, Yuan and Wang, Hongda and Jin, Linbo and Song, Mengfei and Gao, Lei and Shi, Qiming and Fu, Shaokang and Zhao, Jiarong and Wang, Chengyu and Huo, Chengfu , journal =. Bidirectional Context Self-Distillation for Reinforcement Learning of Skill-Based on Pith
  45. [65] Yu, Yan and Lu, Zhengxi and Liu, Yizhou and Pan, Yichen and Wang, Aozhe and Chen, Qipeng and Yang, Hua and Zhang, Wenqi and Lu, Weiming and Chen, Qianglong and Shen, Yongliang , journal = on Pith
  46. [66] Zhu, Jiapeng and Yu, Jianxiang and Zhao, Yibo and Han, Chengcheng and Gu, Qi and Cai, Xunliang and Li, Xiang and Qian, Weining , journal = on Pith
  47. [67] Co-Evolving Skill Generation and Policy Optimization , author =. arXiv preprint arXiv:2606.08755 , year = on PithPith's reviewarXiv
  48. [68] Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning , author =. arXiv preprint arXiv:2605.10923 , year = on PithPith's reviewarXiv
  49. [69] Not All Skills Help: Measuring and Repairing Agent Knowledge , author =. arXiv preprint arXiv:2606.15390 , year = on PitharXiv

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