HarnessBandit: Joint Learnability-Transferability Scheduling for Multi-Harness Agentic Reinforcement Learning

简体中文

L1 online harness scheduling only; no large-model RL.

Paper information

Field	Details
Paper	HarnessBandit: Joint Learnability-Transferability Scheduling for Multi-Harness Agentic Reinforcement Learning
Institution	Harbin Institute of Technology / Alibaba Cloud (first author).
First public release	2026-09-12 (arXiv v1)
Author code	No public author repository found as of 2026-09-16.
Local adapter or method key	harness-bandit
Local reproduction code	src/auto_research/agent_research/latest_20260916.py

Paper method and evidence

Allocate multi-harness training budget by combining learnability, transferability and an exploration bonus. The reference kernel retains gating and weight/state updates and exposes intermediate values for matched-budget comparisons.

flowchart LR
  I[Current harness statistics] --> S[Combine learnability, transfer and exploration]
  S --> B[Allocate training budget]
  B --> U[Update scheduler state]

Original paper figure

Figure 1 shows the multi-harness scheduler. From the original paper; copyright remains with the authors.

Core equation

The local kernel reports decisions and state changes at each allocation; fixed seeds and budget establish invariants only.

Reported results

Paper capability, latency, and cost results are not directly comparable to these local fixtures.

Local reproduction

Three-seed mechanism results are diagnostic_only=true. Fixed-budget fixtures verify the defining operator and auditable intermediate states; they are not a benchmark, production result, or capability-ranking entry.