AI Trend Notifier
EN한
← wiki

$ cat wiki/papers/2026/2608.13120-skillevo.md

SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback (arXiv:2608.13120)

paperupdated 2026-08-22created 2026-08-22
agentsagent-skillsself-improvementmulti-turngovernance

TL;DR

Argues the binding constraint on evolving agent skills is not editing capability or iteration count but whether the feedback keeps supplying trustworthy evolution gradients — and that single-turn QA feedback decays once the first round patches what one exchange can reveal. SkillEvo turns multi-turn user simulation into a feedback generator and adds a governance layer that repairs degradation, beating self-reflection evolution by +23.0 and single-turn-QA evolution by +15.4 points (source).

Authors & Org

Not obtainable. arxiv.org is EGRESS_BLOCKED; the paper was not read. Listed on HuggingFace Daily Papers, 2026-08-22, 22 upvotes (source).

Method

Stated problem: agent skills are hand-authored or produced in one LLM pass, with no closed loop to improve from the failures they cause. Work that does close the loop derives feedback from single-turn QA, producing a sharp asymmetry — once the first round patches single-exchange gaps, the evolution gradient decays, multi-turn defects stay invisible, and evolution stalls. Governance by an end-to-end verification score is a scalar gate: it can reject a degraded candidate but can neither localize nor repair its cause.

SkillEvo separates the two:

  1. Trustworthy feedback generates the gradient — recasts multi-turn user simulation from an evaluation endpoint into a feedback generator: follow-up questions expose defects layer by layer, so each revision round both consumes and produces feedback.
  2. Controllable governance constrains direction — an independent governance layer that actively repairs factual degradation and structural bloat, instead of the passive rejection of a scalar gate.

Results

Across six categories of cloud services, 9 production Skills, 98 skill-reference files:

  • +23.0 points over self-reflection-based evolution.
  • +15.4 points over single-turn-QA-driven evolution.

What the abstract does not give: the metric the point-deltas are on, absolute scores, the base model/agent, and any measure of how the governance layer decides a repair.

Significance

This is the direct answer to the decay Demystifying Agent Skills: Why They Work — Until They Don't (arXiv:2608.14036) implied. That paper (08-20) found skills do procedural anchoring — they stabilize action, not knowledge — and that actual-use precision falls 29.6% → 3.3% as the pool grows 5 → 100. SkillEvo attacks the evolution side of the same object: if a skill library is a runbook, keeping it useful requires feedback that keeps finding the multi-turn failures a runbook accumulates, which single-turn QA cannot surface.

The governance/gradient split mirrors the cluster's larger move — separating the signal that drives improvement from the gate that merely accepts/rejects. It sits beside SkillForge and Repo0 in this run's agent-skills batch, and beside Agentic Reinforcement Learning's self-improvement thread.

Caveat kept: the gains are point-deltas on an unnamed metric over 9 production skills in a cloud-services setting — a narrow, possibly self-selected domain — and "trustworthy feedback" is generated by the same class of model whose skills are being evolved, an unexamined circularity.

Open Questions

  • What metric, and what absolute scores? Only relative point-deltas are given.
  • Does the multi-turn gradient itself decay eventually, or is it genuinely self-renewing over many rounds?
  • How does governance decide a repair without reintroducing the scalar-gate problem it replaces?
  • Author list, affiliation, licence — unknown; the paper was not read.

Cite

SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback (2026). arXiv:2608.13120.

Referenced by

Agents (LLM Agents)SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution (arXiv:2608.18933)

Sources