MIT CSAIL paper: stronger LLMs won't make harnesses obsolete — they're compositional generalizers
MIT_CSAIL · x · 2026-11-21
An MIT CSAIL paper, Language Model Harnesses Can Be Compositional Generalizers, argues against the claim that stronger models will make harnesses obsolete — the opposite appears true.
Core idea: neurosymbolic harnesses like the Recursive Language Model (RLM) externalize and manage context programmatically, decomposing complex problems so each observation the neural LM sees stays locally in-distribution even on novel tasks.
Key findings:
- With RL post-training, RLM-style harnesses readily learn decompositions: training only on short tasks generalizes to tasks 2–64× longer, while the base LM improves far less
- Training on one domain transfers to novel domains sharing latent structure
- Surrogate harness training: generate easier synthetic environments with fake tool/sub-calls to target individual LLM calls within a harness
- The resulting RLM-27B improves Qwen3.8-27B by 26.9% on average across hard long-context suites (CTC-Bench, ExpertLongBench) while using 2× fewer tokens
- Harnesses can compose learned decompositions via recursive sub-harness calls
Related event: MIT CSAIL Paper: LM Harnesses Are Compositional Generalizers(14 posts)→
More from coding & agent
- uivoid turns any API into a hosted, authenticated MCP server with zero code — Necessary_Client_114 · 2026-11-21
- Your personal agent can badger airlines over every past delay until they pay up — nickbaumann_ · 2026-11-21
- MIT's Sutherland Verifies Riemann Hypothesis Lean Proof: All 175,187 Modules Compile — altryne · 2026-11-21
- Dev Laments: AI Coding Offloaded the Fun Parts, Leaving Only Reviewing and Correcting — pritisinghhhh · 2026-11-21
- Route agents by task type, not fixed rules: not every request needs a frontier LLM call — rohanpaul_ai · 2026-11-21
- OpenClaw's Skill Workshop 2.0 lets your agent review and improve its own skills — heyneighbor · 2026-11-21