MIT: LM harnesses are compositional generalizers, not just scaffolding
akbirthko · x · 2026-10-27
An MIT CSAIL article by Alex Zhang and Omar Khattab argues modern post-training is brute-force because Transformers generalize poorly; the harness — the program encoding environment state into LLM calls — should carry higher-level inductive bias, shaping each call to be locally in-distribution so harnesses achieve compositional generalization. Commenter notes harness-vs-no-harness is a tradeoff against environment feedback rate.
More from coding & agent
- Planning a Master-Agent Architecture: LangGraph vs Microsoft Agent Framework — DonXendon · 2026-10-27
- Score centering lands in TRL: default-on for async GRPO, stabilizes RL under mismatch — _lewtun · 2026-10-27
- Cloakroom: Open-Source Layer That Masks Keys in .env and Logs Before Coding Agents See Them — Shanjeev · 2026-10-27
- Claude Code adds env var to skip even its last safety block, like rm -rf ~ — lydiahallie · 2026-10-27
- Codex ban update: users jailbreak AI agents to bypass Daybreak filter for reverse engineering — FlolightC · 2026-10-27
- GPT-6.1 Terra hits GA in GitHub Copilot, stronger terminal coding at lower cost than GPT-6.1 Sol — 0xkarasy · 2026-10-27