MIT CSAIL paper: LM harnesses act as compositional generalizers during RL training

lateinteraction · x · 2026-11-21

A new MIT CSAIL paper by Alex Zhang, Souradip Chakraborty and Omar Khattab, "LM Harnesses are Compositional Generalizers," pushes back on the assumption that RL naturally generalizes — generalization depends on learned representations that may not transfer. By programmatically managing context so every individual LM call stays locally in-distribution (LID), harnesses let RL learn reusable decomposition strategies that compose existing capabilities to solve unseen tasks, generalizing over task length, strategy, and recursive depth while making long-task training more data-efficient. The takeaway: harnesses are not inference-time scaffolding but deeply intertwined with model learning, and agent training should be designed accordingly.

Related event: MIT Paper: LM Harnesses Are Compositional Generalizers, Not Crutches(10 posts)→

Original post →

More from coding & agent

coding & agent channel →