Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents

简体中文

L1 public-observation failure signatures and repair gate; no harness code generation or execution.

Paper information

Field	Details
Paper	Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents
Institution	Chengdu Institute of Computer Applications, Chinese Academy of Sciences (first author).
First public release	2026-09-10
Author code	cuiyu-ai/Ecdysis.
Local adapter or method key	ecdysis
Local reproduction code	src/auto_research/agent_research/latest_20260914.py

Paper method and evidence

Aggregate recurring failures across tasks to distinguish one-off model mistakes from systematic harness defects. Multi-role FDCR diagnosis turns those patterns into repair specifications.

flowchart LR
  O[Public observations] --> S[Aggregate repeated failure signatures]
  S --> D[Separate incidental and systematic failures]
  D --> R[FDCR repair specification]

Original paper figure

The original figure describes the runtime-harness training workflow. From the original paper; copyright remains with the authors.

Core equation

The deterministic reference emits grouping and trigger statistics, with tests for decision invariants and boundaries.

Reported results

The paper reports up to 1.84× faster harness training and up to 18.56% higher inference accuracy. These are not local or online results.

Local reproduction

Three-seed metrics use the same public fixtures, budget and seeds and verify only the state transition.