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

Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 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 SkeMex, a non-parametric self-evolving skill-memory framework for medical agents that writes, utility-values, retrieves and governs skills across general, task, and action branches without updating model weights.

Problem

Medical agents need to reuse experience across evolving clinical cases, but raw-trace memories are redundant and noisy, lack usefulness estimates, and prior self-improving methods require parameter updates.

Approach

Distills interaction trajectories into structured skills organized in three branches, estimates context-dependent utility from environment feedback via window-level advantage-based credit, uses value-aware retrieval, and runs a Read-Write-Assess-Govern lifecycle that promotes, merges, and removes skills without touching backbone weights.

Key results

• Offline with DeepSeek-V3.2, SkeMex raises average from 48.20% (ReAct) to 56.08% (+7.88), beating the strongest memory baseline by 3.84 points.
• Offline with Qwen3.6-Plus, SkeMex reaches 59.22% average vs 48.63% ReAct (+10.59) and 3.14 points above the best competing memory method.
• Online ablations show removing skill writing costs 7.88-9.06 average points and removing utility valuation costs 6.06-6.87 points.

Surprising findings

• Several memory baselines fall below the memory-free ReAct agent on out-of-domain benchmarks such as MediQ, suggesting negative transfer from less structured memory.

Limitations

• Relies on closed API backbones and an LLM judge (Gemini-3-Flash) for rubric-based scoring.
• Utility estimation depends on environment feedback and window-level aggregation, which may be noisy for small or heterogeneous task sets.

Experimental setup

Datasets: AgentClinic LiveClin MedJourney LiveMedBench HealthBench MediQ MedXpertQA MMMU MMMU-Pro

Benchmarks: LiveClin MedXpertQA HealthBench LiveMedBench AgentClinic MediQ MMMU MMMU-Pro MedJourney

Metrics: accuracy rubric-based score

Baselines: ReAct CoT Reflexion CRITIC Voyager DILU ExPeL Generative Memory (GM) Memp SkillWeaver AWM Agent KB

Builds on: ReAct Reflexion Agent Workflow Memory Memory-based MDP Ebbinghaus forgetting curve

Closest papers by meaning

• MedMemoryBench: Benchmarking Agent Memory in Personalized Healthcare (NeurIPS 2026 · similarity 0.89)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.88)
• MedEvoEval: Evaluating Continual Evolution of Doctor Agents through Simulated Clinical Episodes (NeurIPS 2026 · similarity 0.88)
• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.88)
• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.88)
• MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs (NeurIPS 2026 · similarity 0.87)