acceptodds
Under review as a conference paper at ICLR 2027

SkillAge: Lifecycle-Aware Skill Aging and Safe Pruning for Self-Evolving Agents

Abstract

Self-evolving Large Language Model (LLM) agents excel at accumulating procedural skills for complex tasks. Yet, unchecked skill accumulation inevitably triggers severe repository bloat, where redundant or obsolete routines inflate retrieval overhead and disrupt execution fidelity. To tackle this challenge, we propose SkillAge, an online lifecycle-aware framework for verifiable skill governance and safe pruning. SkillAge models procedural knowledge as dynamic assets transitioning across controllable operational states. From trajectory feedback, it extracts multidimensional evidence, including exposure-weighted performance evidence, semantic redundancy, and context cost, to parameterize an evidence-grounded retention score. This score guides a constrained Markov decision process (CMDP) reinforcement learning policy to optimize repository compactness without future-task annotations. Crucially, irreversible deletion is excluded from online actions; instead, candidates are staged in quarantine where a dual-gate replay protocol certifies permanent removal by bounding both cumulative and worst-case capability degradation. Evaluations across continuous lifelong benchmarks demonstrate that SkillAge achieves up to 89.5% repository compression and 92.0% token exposure reduction while reliably preserving critical future capabilities.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
LLM Agents, Lifelong Learning, Skill Pruning, Continual Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gzFkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses skill repository bloat in self-evolving LLM agents. It proposes SkillAge, a lifecycle governance framework that computes a six-dimensional handcrafted retention score, models skill state transitions (Retain, Decay, Refresh, Protect, Quarantine) as a Constrained Markov Decision Process (CMDP) trained via primal-dual reinforcement learning, and verifies candidate deletions through historical task replay (Algorithm 1). Experiments on a 238-task stream (158 initial skills) across three LLM backbones report 87%-90% repository compression while maintaining task utility.

Critical strengths

Practical regression-testing intuition for skill pruning. Treating skill deletion as a reversible process gated by historical task replay (Algorithm 1) is a sensible engineering practice, analogous to regression testing for software plugins. The ablation in Table 6 (- Quarantine Replay, where degradation rises from 0.031 to 0.142) demonstrates that outcome-based verification before permanent deletion effectively prevents the accidental removal of useful skills.

Critical weaknesses

W1. Unjustified complexity of the CMDP/RL controller and incomplete MDP specification. Skills in modern agent frameworks are modular tools whose maintenance primarily requires deduplication, usage tracking, and regression testing. In SkillAge, the authors already define a comprehensive handcrafted retention score ρi,t∈[0,1]\rho_{i,t} \in [0,1] (Eq. 7), enforce a hard quarantine threshold ρi,t<0.5\rho_{i,t} < 0.5 (Eq. 8), and deterministically sort candidates by ρi,t\rho_{i,t} during replay (Algorithm 1). With these explicit heuristics and hard verification gates already in place, the necessity of training a 3-layer MLP policy via primal-dual RL is unconvincing. Notably, the paper omits the most fundamental ablation: removing the RL policy and directly triggering Algorithm 1 whenever ρi,t<0.5\rho_{i,t} < 0.5 (or equipping baseline heuristics with Algorithm 1). Moreover, the MDP formulation itself is incomplete: (i) the Decay action has no mathematical definition anywhere in the paper or appendices and never enters the retrieval scoring function; and (ii) Table 2 and Figure 9 report 1,500 RL policy updates on an online stream of only 238 tasks, which is impossible in a true single-pass online setting without multi-epoch offline training on the test stream.

W2. Theoretical over-packaging and a direct contradiction between Theorem 1 and Table 6. Several theoretical claims formalize elementary algebra as heavy theorems. Theorem 1 (Appendix B.1) simply states that if each step enforces Dt≤BtD_t \le B_t and updates Bt+1=Bt−DtB_{t+1} = B_t - D_t, then ∑t=1HDt=B−BH+1≤B\sum_{t=1}^H D_t = B - B_{H+1} \le B, which is a trivial accounting identity. Similarly, Proposition 1 invokes Clarke's generalized subdifferential merely to differentiate a 1D piecewise-linear hinge function with slopes 1/6 and 1/12, and Corollary 1 is a simple rearrangement of 1/1+n>Ω1/\sqrt{1+n} > \Omega. More importantly, Theorem 1 directly contradicts the empirical results in Table 6: as long as Algorithm 1 (Quarantine Replay) is active, the hard gate Di≤BremD_i \le B_{\text{rem}} must bound cumulative replay degradation by B=0.05B = 0.05 regardless of how candidates are proposed. Yet in Table 6, three ablations that retain Quarantine Replay (- Retention Barrier, - Epistemic Buffer, and Fixed Dual Multiplier) report replay degradations of 0.068, 0.079, and 0.071, all violating the 0.05 bound.

W3. Unrealistic online oracle assumptions and inconsistent overhead arithmetic. Algorithm 1 requires an automated ground-truth task verifier during online deployment to re-evaluate historical tasks. Assuming test-time access to ground-truth evaluators is unrealistic for open-ended online agents, and evaluating baselines (e.g., LRU, Cosine Dedup, ASAP, MemGPT) without this oracle rollback mechanism makes the comparison unfair. Furthermore, the claim that replay verification adds only 4.1% amortized token compute is arithmetically inconsistent with the reported statistics. On a stream of 238 tasks, triggering replay in 14.7% of rounds (≈35\approx 35 rounds) with an average of 8.4 tasks per replay requires at least 35×8.4≈29435 \times 8.4 \approx 294 full task re-executions (and up to 140×8.4≈1,176140 \times 8.4 \approx 1,176 executions if pruned candidates are evaluated sequentially as written in Line 8 of Algorithm 1). Executing 294 to 1,176 extra tasks on a 238-task stream implies 123% to 490% additional task runs, which cannot plausibly correspond to a 4.1% token overhead.

Questions for authors

Q1. How does a simple rule-based variant that bypasses the RL policy---directly passing skills with ρi,t<0.5\rho_{i,t} < 0.5 to Algorithm 1---perform in Table 3 and Table 6? Also, where is the mathematical definition of how the Decay mode alters skill retrieval or task execution?

Q2. Given that Algorithm 1 enforces a hard budget gate (Di≤BremD_i \le B_{\text{rem}}) guaranteeing ∑t=1HDt≤B=0.05\sum_{t=1}^H D_t \le B = 0.05 (Theorem 1), how can the three ablations in Table 6 that retain Quarantine Replay (- Retention Barrier, - Epistemic Buffer, and Fixed Dual Multiplier) yield replay degradations of 0.068, 0.079, and 0.071?

Q3. How were 1,500 RL policy updates executed on a 238-task online stream without multi-epoch passes over the evaluation tasks? Please also provide the exact token accounting that reconciles at least 294 replay task executions (238×14.7%×8.4238 \times 14.7\% \times 8.4) with the claimed 4.1% token overhead.

Reviewer_bDjwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses skill bloat in self-evolving agents: append-only skill libraries fill up with redundant or stale skills that waste retrieval slots and context. SkillAge tracks several signals per skill (usage, success, estimated contribution, redundancy, cost) and combines them into a retention score. An RL policy, trained under constraints, then chooses lifecycle actions such as decay, protect or quarantine. A quarantined skill is deleted only if replaying the past tasks it supported stays within a degradation budget. The authors present this as a safety certificate and back it with theory. Experiments on SkillFlow with three backbones compare SkillAge against Random, LRU, LFU, ASAP, combined deletion, AgeMem and an oracle. The paper reports up to 89.5% library compression and 92% token reduction, and claims that critical capabilities are preserved.

Critical strengths
  • Separating reversible lifecycle actions from deletion is a good design. Quarantined skills are deleted only after replay on the past tasks that used them, and each candidate is tested jointly with the ones already accepted (Section 4.5, Eqs. 13-14). This targets mutually redundant fallback skills that a one-at-a-time test would miss.
  • The quarantine mask (Section 4.3) requires prior exposure and aging pressure above retention evidence, and Lemma 1 shrinks the contribution estimate toward zero at low exposure. Together these give a simple, interpretable way to keep rarely used skills out of quarantine.
Critical weaknesses
  • The main-text results contradict each other and the headline claim. In Table 1, Loss equals 1 - Util. in every row, so SkillAge loses 0.319, 0.328 and 0.305 of utility, while the Oracle loses at most 0.018 at higher compression (93.7 vs 88.4 on Qwen). This is hard to square with "reliably preserving critical future capabilities" (abstract) and "within the safety budget".
  • The safety certificate is never checked against future utility. Theorem 1 holds by construction: a deletion passes only if the replay drop is at most B_t, and B_t+1 = B_t - D_t telescopes. The main text reports no B, epsilon_max or realized D_t. The "zero safety breaches" claim rests on Fig. 9 in the appendix, and the out-of-distribution claim is never tested.
Questions for authors
  • Which numbers are correct for SkillAge (Full), Table 1 or Fig. 8, and which backbone does Fig. 8 use?
  • What values of B and epsilon_max were used, and what was the realized sum of D_t? How does it compare with the utility loss in Table 1?
  • Are the 1,500 iterations (line 431) passes over the same stream that is evaluated, and how many seeds or stream orders are behind Table 1? How is deterministic replay obtained, and how many extra agent runs does it need?
  • Could you add a baseline that keeps the replay gate but quarantines by a threshold on the retention score, without RL?
Reviewer_W17TOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

SkillAge constructs skill states based on usage frequency, success rate, exposure-weighted contribution, semantic redundancy, context cost, and evidence scarcity. It then employs constrained reinforcement learning to select operations: retain, decay, refresh, protect, or isolate. Before permanent deletion, historical task replay is used to examine average degradation and worst-case single-task degradation. Evaluated on three model backbones of SkillFlow, the paper reports up to 89.5% skill library compression and a 92.0% reduction in prompt token exposure.

Critical strengths

1.The problem bears practical value. Accumulated skills may introduce retrieval contention, redundant prompts, and outdated execution logic, while low-frequency skills may be critical for key tasks. Research on skill retention and eviction mechanisms is meaningful. 2.The deletion pipeline is reasonably designed. It separates reversible maintenance from permanent deletion, and validates the cumulative deletion set upon each tentative deletion. This can identify fallback dependencies where "two skills are individually deletable but cannot be deleted simultaneously". 3.The reported results are promising and methodological details are provided. Table 1 shows that SkillAge improves task utility by approximately 7.9–8.0 percentage points compared with AgeMem, while achieving a higher compression ratio. The appendix includes state definitions, algorithm pseudocode, ablation studies, and overhead analysis.

Critical weaknesses

1.Table 1 reports SkillAge utility ranging from 0.672 to 0.695, yet it does not present results for the unpruned skill library under identical retrieval configurations. This makes it impossible to quantify the capability change introduced by pruning. Outperforming other pruning strategies is insufficient to demonstrate "reliable preservation of future critical capabilities". In particular, the paper does not report separate results for rare critical tasks. Comparisons between pruned and unpruned average performance should be conducted on a fixed held-out task set, and degradation for critical tasks should be reported separately.

2.In Figure 8, the full method achieves 92.5% Utility and 4.1% Loss, whereas Table 1 and Table 6 report ranges of 67.2%–69.5% and 30.5%–32.8% for these two metrics, respectively. The paper does not explain these differing settings. Moreover, the three ablations in Table 6 (removal of retention barriers, removal of scarcity buffers, and fixed multipliers) yield certified degradation between 0.061 and 0.112, exceeding the 0.05 budget. If these variants still run the hard replay gate in Algorithm 1, such exceedance should not occur. These two unexplained inconsistencies directly undermine the assessment of capability preservation and the effectiveness of the safety mechanism.

3.Experiments are conducted on only one benchmark, with 166 tasks, 238 rounds, and an initial set of 158 skills. The scale of newly added skills is not clearly reported, and targeted results for shifting task distributions or the reappearance of rare tasks after long absence are missing. The three model backbones cannot substitute for validation on these scenarios. Although the appendix states that held-out dynamic exposure is adopted, the exact split among training, replay, and future evaluation is not specified, nor are the number of independent replicates and result uncertainty reported. Therefore, the current experiments provide limited support for continuously growing skill libraries and long-term capability preservation.

Questions for authors

1.Relative to the unpruned baseline under the same retrieval configuration, what is the capability retention rate of the final skill library? Do rare critical tasks exhibit substantial degradation masked by averaged results? 2.Are Figure 8 and Tables 1/6 obtained under different experimental setups? How is certified degradation calculated for the three ablations in Table 6, and do they follow the same hard replay gate? If yes, why is the budget exceeded? 3.How many new skills are actually added during experiments? What is the performance under shifting task distributions and when rare tasks reappear after a long gap? Please specify the exact split of training, replay, and future evaluation, together with results from independent replicates.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.