RLM harness beats baselines across six long-context generalization benchmarks

a1zhang · x · 2026-11-20

Additional experiment charts from a1zhang's paper: across MRCrv2 (64k→2M), GraphWalks (256k→1M), LongBench-Pro, OOLONG, OOLONG-Pairs, and Ada-LEval, the trained RLM generally outperforms LM+REPL, Pi, and Base+YaRN baselines on longer evaluation tasks, demonstrating short-to-long compositional generalization.

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

Original post →

More from coding & agent

coding & agent channel →