New Paper Argues LM Harnesses Act as Compositional Generalizers for LLMs
a1zhang · x · 2026-11-21
The new paper 'LM Harnesses are Compositional Generalizers' argues that harnesses are no longer a crutch for weak LLMs — they compose existing model capabilities to generalize beyond training environments.
- Harness abstractions enable generalization over task length, task strategies, and recursive task depth.
- Keeping individual LLM calls locally in-distribution (LID) makes training far more data-efficient, especially on long tasks.
- Commenters including omarsar note custom harnesses show a large, underexplored edge; RLMs are a recent example, and post-training that co-evolves harnesses and models may explain why this works.
More from coding & agent
- Reverse-engineering Apple Watch: a single Claude Code prompt builds a physics watch face in 30 minutes — mtizard · 2026-11-21
- Agent RAG tweak: skip lossy summaries, extract answers, fetch 10 pages — 75% on validation — colinmcnamara · 2026-11-21
- us-legal-mcp: an MCP server exposing comprehensive US legislation to AI agents — modelcontextprotocol · 2026-11-21
- Maindex launches Expert Memory Service, a full-control memory API for AI agents — modelcontextprotocol · 2026-11-21
- GPT-6.2 Ultrafast in Codex: Up to 4.5x Faster Tokens, Dev Says It Restores Flow — dkundel · 2026-11-21
- AI-built oil painting engine: 2,300 commits in 8 days as Claude Code runs a 'studio' of sessions — FotoRe_store · 2026-11-21