Research map / Multi-agent systems and games / LLM agents

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SkillOpt, a text-space optimizer that trains a single agent skill document via bounded add/delete/replace edits proposed by a frontier optimizer model, accepted only on held-out validation gains.

Problem

Existing agent-skill adaptation hand-writes, generates once, or grows skills via loosely controlled self-revision; there is no controllable, validated optimization of a skill document for frozen agents.

Approach

Treat the skill document as trainable external state with deep-learning-style controls: a textual learning-rate edit budget, minibatch reflection, held-out acceptance gate, rejected-edit buffer, and epoch-wise slow/meta update, with zero added inference cost.

Key results

• On GPT-5.5 direct chat, SkillOpt raises six-benchmark average from 58.8 (no skill) to 82.3, +23.5 points, best or tied on 52 of 52 evaluated (model, benchmark, harness) cells.
• On GPT-5.5 Codex harness SkillOpt gains +24.8 over no skill and +14.0 over EvoSkill; under Claude Code it gains +19.1 over no skill and +3.2 over EvoSkill.
• Cross-harness transfer of a Codex-trained SpreadsheetBench skill to Claude Code yields 22.1 to 81.8 (+59.7); final skills are 379 to 1,995 tokens after 1-4 accepted edits.

Surprising findings

• Gains are driven by very few accepted edits (1-4); LiveMath and OfficeQA gains of +29.3 and +39.0 come from a single accepted edit each.
• A target-matched optimizer (same model as target) recovers 56-74% of the gain from a strong GPT-5.5 optimizer, so the loop contributes beyond distillation.

Limitations

• Requires scored rollouts and reliable verifiers or metrics; open-ended tasks need stronger evaluation.
• Training costs rollouts and optimizer API calls, which may be unattractive for one-off tasks.
• Single skill document per domain may be insufficient for highly heterogeneous domains.

Experimental setup

Datasets: SearchQA SpreadsheetBench OfficeQA DocVQA LiveMathematicianBench ALFWorld OlympiadBench Omni-MATH

Benchmarks: SearchQA SpreadsheetBench OfficeQA DocVQA LiveMathematicianBench ALFWorld Omni-MATH

Metrics: accuracy exact match hard score

Baselines: No skill Human skill One-shot LLM skill Trace2Skill TextGrad GEPA EvoSkill

Builds on: GEPA TextGrad Trace2Skill EvoSkill

Closest papers by meaning

• From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills (NeurIPS 2026 · similarity 0.90)
• AgentRefine: Enhancing Agent Generalization through Refinement Tuning (ICLR 2025 · similarity 0.89)
• SkillMaster: Toward Autonomous Skill Mastery in LLM Agents (NeurIPS 2026 · similarity 0.89)
• PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.89)
• Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.89)
• Self-Challenging Language Model Agents (NeurIPS 2025 · similarity 0.88)