---
id: "2026-09-28-harness-learning-test-time-adaptation"
url: "https://postcutoff.com/e/2026-09-28-harness-learning-test-time-adaptation/"
as_of: "2026-10-12T15:02:00+02:00"
date: "2026-09-28"
date_precision: day
category: research
importance: 2
confidence: high
status: [Confirmed]
sources: 1
editor: Adam Bicz
human_review: null
version: "2026-10-12"
---

As of: 2026-10-12 15:02 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/e/2026-09-28-harness-learning-test-time-adaptation/

# Harness learning: a small model trained with RL learns to rewrite an agent's harness from execution feedback

A preprint by Alvin Zhang, Ruslan Salakhutdinov, Daniel Khashabi, Andrea Zanette and co-authors trains a "proposer" model to edit the harness (the program that organizes an agent's model calls and tool use) of a frozen "solver" model, using how the revised harness performs as the reward. At test time nothing is retrained: the proposer reads the execution report of each run on a new task and rewrites the code again. The skill transfers to task families and benchmarks it never saw in training.

## Key facts
- Framing: meta-learning over executable programs, 'with harness revisions playing the role of weight updates in gradient-based adaptation'
- Reasoning Gym: Qwen3.5-4B as both proposer and solver; supervised fine-tuning on revisions by a 35B teacher (Qwen3.6-35B-A3B), then RL; 21 task families unseen in training
- On the 21 unseen families, RL raises the mean score of a single revision from 0.32 (base proposer) to 0.62, and the trained 4B proposer beats the 35B teacher on average
- Multi-hop QA: Qwen3-4B proposer and Qwen3-8B solver, trained on HotpotQA only, transfers to MuSiQue and 2WikiMultihopQA
- Sequential revision matters on MuSiQue: about 0.15 exact match for independent single revisions (only marginally above the seed harness) versus 0.27 after ten rounds
- A proposer trained only on single revisions keeps improving harnesses over several rounds; training on whole revision sequences helps in some settings and not in others
- Typical learned edits: answer validation with a retry, an extra retrieval hop, searching again when evidence is missing
- Small open models and academic compute only (Johns Hopkins ARCH, Oak Ridge OLCF); no frontier model is trained or modified

## What happened

An agent is a model plus a harness: the code that decides when the model is called, which tools it uses and how
information flows between steps. Harnesses are normally written by engineers. This preprint (posted 28 Sept 2026)
trains a second model, the proposer, to write the edits instead. The proposer sees the task description, the current
harness and a report of what happened when it ran, and outputs a code change. The revised harness is run, and its
task score is the reward for reinforcement learning. Only the proposer is trained; the solver model stays frozen.

At test time both models are frozen. Adaptation to a new task happens entirely by rewriting the harness over
successive rounds, without any weight update. The authors compare this to learning-to-learn methods where the
"update" is a program edit instead of a gradient step.

## Why it matters

It is a concrete, small-scale form of an agent improving how it works from its own experience, which is the same
idea as the larger self-improvement results of 2026, but with open 4B–8B models and standard benchmarks, so others
can check it. The reported gains are on Reasoning Gym puzzles and multi-hop question answering; nothing here shows
the method working on long real-world tasks or with frontier models. The authors' affiliations (Carnegie Mellon and
Johns Hopkins) are from their known positions and the acknowledgements; the PDF's first page does not list them.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 151 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 90 days after its cutoff
- Gemini 3.8 Flash (training cutoff March 2026): 181 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 120 days after its cutoff

## Sources

1. [arXiv 2609.35738: Harness Learning Enables Generalizable Test-Time Adaptation](https://arxiv.org/abs/2609.35738) (arxiv.org, paper)

## Changes

- 2026-10-12 (filed): Created (flagged by the site owner; no conjecture wording)

## Related

- 2026-09-14: [Google's Dream-RSI](https://postcutoff.com/e/2026-09-14-google-dream-rsi/index.md)