Harness Learning paper: a 4B model trained to rewrite an agent's harness code beats its 35B teacher on average across 21 unseen reasoning tasks

Why it matters

【Key points】

On Sept 28, Alvin Zhang, Ruslan Salakhutdinov, Daniel Khashabi, Yuda Song, Andrea Zanette and co-authors posted "Harness Learning Enables Generalizable Test-Time Adaptation" (arXiv 2609.35738). It treats an agent as model plus harness, the code that organizes model calls, tools and information flow, and trains a proposer model with RL to rewrite a solver's harness from execution feedback. At test time no weights change; only the code does.

After DAIR.AI's weekly list on Oct 11, Pedro Domingos posted "the model swallows the harness" (~353K views) and Elon Musk replied that Tesla FSD's network likewise ate the C++ code (~246K views).

【Verifiable】

Paper: https://arxiv.org/abs/2609.35738

On Reasoning Gym (Qwen3.5-4B proposer and solver, SFT on a 35B teacher's revisions, then RL), mean single-step revision score on 21 unseen families rises from 0.32 to 0.62, above the teacher's 0.56; under oracle best-of-eight the teacher still wins. A proposer trained only on HotpotQA transfers to MuSiQue and 2WikiMultihopQA; on MuSiQue, exact match is 0.15 for independent revisions versus 0.27 after ten rounds.

X: Domingos https://x.com/pmddomingos/status/2109420932955804040 , Musk https://x.com/elonmusk/status/2109468545054880123 ; ~150 related posts in 24h, 24-33 per hour over the last five hours.

【Our read】

1. "The model swallows the harness" is half backwards. The harness is not folded into weights; another model rewrites it, and it stays readable, reviewable code. That is the opposite of FSD replacing code with a network, and better for auditing.

2. The novelty is making harness revision a trainable skill that transfers to unseen tasks, on the same line as Sakana's MASS (e/820).

3. Evidence is small: 4B/8B models on puzzles and QA, not frontier coding agents. Whether it holds on long real tool tasks decides if this is a method shift or just a good paper.