MIT Paper: LM Harnesses Are Compositional Generalizers, Not Crutches
Researcher Alex Zhang (a1zhang) posted a thread of tweets on 11-20 introducing the new paper "LM Harnesses are Compositional Generalizers." Core claim: since 2026 models are strong enough that a harness's value is no longer a "crutch" patching weak models, but rather composing LLMs' existing capabilities and generalizing beyond training environments—hence the proposal of Recursive Language Models (RLM).
Confirmed
- Motivation: most tasks decompose (e.g., MapReduce), but mainstream harnesses (CodeAct, MemGPT, etc.) use "trajectory-as-a-prompt," stuffing all history into context; LLMs generalize unreliably on such OOD inputs, causing context rot—and it's not just subtasks: even the orchestrator coordinating subtasks goes OOD.
- Results: across six long-context benchmarks—MRCrv2 (64k→2M), GraphWalks (256k→1M), LongBench-Pro, OOLONG, OOLONG-Pairs, and Ada-LEval—RLM training curves lead LM+REPL across the board.
- Naming: another set of experiments shows a harness recursively calling itself stays "locally in-distribution," at which point the harness is abstracted as the sub-LLM from the original setup—hence the name RLM; they train the harness as a model via tail-recursion gating chains.
- Surrogate harness training: the paper's most practically valuable technique. For extremely long tasks (rollouts take days) with sparse rewards where direct RL is infeasible, it exploits compositionality to train orchestration LLMs in synthetic environments. Using Claude Opus 5.5 to generate fully synthetic tasks (some with fake sub-calls) to train Qwen3.8-27B, average long-task scores rose from 30.5 to 58.5; naive RL struggles to generalize to longer tasks, while targeted harness training does.
Why It Matters
This work reframes agent/harness research from "patching weak models" to "compositional generalization," and provides a reusable synthetic-data training path for long-horizon agent training, with direct practical value for extremely long tasks requiring multi-day rollouts.
2026-11-20 ~ 2026-11-21 · 10 related posts
Primary sources
- [source] New paper: LM harnesses are compositional generalizers, +26% on long tasks — a1zhang · 2026-11-20
- Why orchestrator LLMs also go OOD: how CodeAct/MemGPT/RLM curb context rot — a1zhang · 2026-11-20
- [source] RLM harness beats baselines across six long-context generalization benchmarks — a1zhang · 2026-11-20
- [source] Surrogate harness training: RL-training long-horizon agents entirely on synthetic environments — a1zhang · 2026-11-20
- Synthetic surrogate training lifts Qwen3.8-27B RLM long-task average from 30.5 to 58.5 — a1zhang · 2026-11-20
- RLMs recurse into themselves: tail-recursive gated chains turn harnesses into models — a1zhang · 2026-11-20
- Why 'Recursive Language Models': a harness invoking harnesses is itself a model — a1zhang · 2026-11-20
- New paper: compositional harnesses like RLM yield better returns on RL compute — CShorten30 · 2026-11-20
- MIT CSAIL paper: LM harnesses act as compositional generalizers during RL training — lateinteraction · 2026-11-21
1 near-duplicate retellings: RLanceMartin