SkeMex
Paper

Research projectNeurIPS 2026arXiv v3

Experience becomes
skill.

SkeMex gives medical agents a way to improve after deployment. It turns useful interaction trajectories into reusable skills, retrieves them by context and utility, and governs the memory as it grows.

Self-evolving skill memoryfor generalizable medical agent reasoning
Explore the evidence Read the paper
00

Reported averages
from the paper

+7.88 ppoffline ID
+13.78 ppoffline OOD
78.56%online epoch@3
+10.25 pppooled transfer

01The method

Memory that learns
what to keep.

Raw trajectories contain both useful procedures and noise. SkeMex distills experience into structured skills, estimates their context-dependent utility, and keeps the repository healthy over time.

01

Value-aware retrieval

Retrieve skills using semantic relevance, learned utility, and memory strength — so the agent sees experience that is useful for the current clinical context.

similarityutilitymemory strength

02What the paper shows

Structure turns experience
into transfer.

A

Skills over traces

Trajectory → reusable procedure

Trajectory-to-skill distillation compresses recurring procedures into branch-aware memory that can be recalled across related clinical contexts.

B

Utility guides reuse

Outcome → context-dependent value

Feedback estimates which memories help in context, guiding retrieval and repository decisions instead of treating every stored trace equally.

C

Closed-loop evolution

Review → promotion, merge, removal

Selective buffering, draft review, promotion, merging, and removal keep the skill repository compact and safer to reuse.

03Results explorer

Move through
the evidence.

Values are copied from arXiv v3. Select a view to compare the reported averages and transfer settings.

DEEPSEEK-V3.2 · OFFLINE ID

Seven-benchmark average

ReAct SkeMex
SkeMex improves the average from 48.20% to 56.08%.

04Design check

Each part of the loop
earns its place.

Removing selection or review lets noisier experience enter the repository. The full design reaches the highest reported average.

See the ablation
Full SkeMex
53.22
Single-prompt
50.97
w/o draft review
48.82
w/o buffer gating
47.56
02550

05Continue exploring

Read the method.
Run the memory.

The public repository contains the implementation, configuration guides, data preparation notes, and selected skill examples.

PaperarXiv v3 · full method and resultsCodebaseConfigurations, tools, and entry pointsData guideOfficial sources and preparation instructions