Megadose The AI curve, compressed.

ModHarness-RL: Recasting Harnesses as Learnable and Evolving Programs for Agentic RL

· ArXiv · AI/CL/LG ·
Learned context and guidance programs beat a fixed harness on held-out performance in most tested settings.

ModHarness-RL stores agent interactions as typed blocks, then uses one program to assemble context and another to supply guidance when behavior deviates. During reinforcement learning, candidate programs are proposed from recent trajectories, checked by execution and replayed over past trajectories. The study examines CodeSearch in depth, with further analysis on ALFWorld and BIRD SQL, using models from 1.5B to 8B parameters. The reported gains support learning and updating the harness alongside the model, though improvements were not universal.

ArXiv · AI/CL/LG's note

score 5

Categories: Research