New paper: compositional harnesses like RLM yield better returns on RL compute
CShorten30 · x · 2026-11-20
- New paper "LM Harnesses are Compositional Generalizers" (Alex Zhang et al.), amplified by lateinteraction (Charles Packer, Letta).
- Core claim: harnesses are shifting from crutches for weak LLMs to mechanisms that let strong models compose existing capabilities and generalize beyond training environments — over task length, strategies, and recursive depth.
- During RL post-training, RLMs easily learn to reshape messy observations into compositions of familiar problems; the paper isolates key harness abstractions: context offloading, programmatic tool calling, and sub-agent calling.
- These keep trajectories locally in-distribution (LID), making long-task RL training far more data-efficient; benchmark charts show RLM beating LM+REPL, Pi, and Base+YaRN baselines on OOLONG and cross-domain OBLIQ-Bench. Compositional harnesses may deliver better returns on RL compute than standard coding agents.
Related event: MIT Paper: LM Harnesses Are Compositional Generalizers, Not Crutches(10 posts)→
More from coding & agent
- Build a 'company brain': one shared knowledge base for your whole team's AI — VibeMarketer_ · 2026-11-21
- LangChain ships four updates: LangSmith Agent Platform, trace data controls, Tool Gateway — LangChain · 2026-11-21
- Stanford CS329Z: Full AI Agents Course Free With Slides, Videos, Homework — stanfordnlp · 2026-11-21
- Grok Bot Gets Its Own Email Addresses to Sign Up and Handle Promos for You — elonmusk · 2026-11-21
- CapCut × ChatGPT plugin turns chat into a video generator and editor — SimplyAnnisa · 2026-11-21
- Princeton & Meta's Strategist evolves inference strategies, beating baselines on 24/30 benchmarks at 39% lower cost — 141_1337 · 2026-11-21