MIT CSAIL RLM: Harness-Driven Length Generalization — 64K to 2M Tokens¶

Ch05.125 MIT CSAIL RLM: Harness-Driven Length Generalization — 64K to 2M Tokens¶

📊 Level ⭐⭐ | 3.6KB | entities/mit-csail-rlm-harness-length-generalization.md

MIT CSAIL RLM: Harness-Driven Length Generalization¶

MIT CSAIL researchers propose using the harness as an explicit variable for compositional generalization in recursive language models (RLMs). By training only on 64K-token tasks, the system generalizes to ~2M token evaluations — a 32× length extrapolation — outperforming transformer baselines by approximately 10× in long-task evaluation gains.

Core Insight: Harness as a Generalization Variable¶

The key innovation is repositioning the harness from a peripheral engineering framework to a carrier of high-level inductive biases. When tasks share similar decomposition structures (even with different domains and text surfaces), the harness can produce approximately isomorphic model trajectories — the root model learns reusable organizational strategies across tasks. This is formalized through the concept of task equivalence classes.

Locally In-Distribution (LID) Design¶

The core architectural principle is Locally In-Distribution (LID): while the overall long task may be out-of-distribution, each individual model call is kept close to the training input distribution through two mechanisms:

1. Context Unloading — Task data is stored as variables in the external program environment. The root model first encounters the task structure, while domain-specific content remains external.
2. Programmatic Sub-Agent Calls — Sub-agents and tools are treated as function calls in a code environment. Results, intermediate calculations, and tool outputs are stored in variables rather than accumulating in the main context.

Results¶

Using a Qwen3-30B-A3B base with RL training (only updating the root language model), the RLM system approaches or exceeds GPT-5.5 reference results on four long-task benchmarks. Transformer baselines (including YaRN-extended) show improved training rewards on short tasks but limited long-task improvement — confirming that without harness-level generalization design, length extrapolation remains fundamentally constrained.

Implications for Agent Systems¶

This work elevates the harness from a deployment concern to a first-class generalization variable in agent system design. It suggests that agent architectures capable of decomposing long tasks into locally in-distribution calls — through well-designed harnesses — can achieve length generalization impossible through model architecture alone. This aligns with the broader harness engineering movement in agent development.

→ 原文存档