JevSpawn: Adaptive Agentic Inference through Compositional Action Spaces

简体中文

L1 mechanism diagnostic of finite-action posterior updates and top-branch retention; no Jev checkpoint.

Paper information

Field	Details
Paper	arXiv v1
Institution	The paper's first page does not list an affiliation for the first author.
First public release	2026-09-30 (arXiv v1)
Author code	No public author implementation found as of 2026-10-03.
Local adapter or method key	jev-spawn
Local reproduction code	src/auto_research/agent_research/latest_20261003.py

Paper method and evidence

Build a finite compositional action space from the natural-language task. Use a Jev distribution to keep multiple branches in parallel, then update the posterior from feedback and retain alternatives to recover from failure.

flowchart LR
  T[Natural-language task] --> A[Finite compositional actions]
  A --> B[Retain parallel branches]
  B --> P[Feedback updates posterior]
  P --> R[Keep recovery alternatives]

Original paper figure

Figure 2 shows the architecture and how its components connect. From the original paper; copyright remains with the authors.

Core equation

\(p'(a)\propto p(a)e^{feedback(a)}\).

Reported results

The authors evaluate speed, cost, and success rate across eight tasks and seven agent baselines.

Local reproduction

Compare with the mechanism disabled/default state. The three-seed diagnostic is diagnostic_only=true; no formal gain or capability ranking.