Program Learning with Verifiable Rewards: Symbolic Backpropagation for Post-Training LLMs
The paper argues reasoning gains should be learned as inspectable programs, not hidden weight updates.
PLVR builds explicit programs from deterministic and neural primitives using input-output examples. Its “symbolic backpropagation” sends typed requirements backward through program layers, turning credit assignment into a derivation checked by per-step contracts. The author reports 30B base models beating matched-budget RL by 27.8 points on average across LiveCodeBench v6 and Tau2Bench, and outperforming much larger frontier models by 13.6 points. The paper says the library and conformance checker are being released for use with other primitive libraries. ArXiv · AI/CL/LG's note
PLVR builds explicit programs from deterministic and neural primitives using input-output examples. Its “symbolic backpropagation” sends typed requirements backward through program layers, turning credit assignment into a derivation checked by per-step contracts. The author reports 30B base models beating matched-budget RL by 27.8 points on average across LiveCodeBench v6 and Tau2Bench, and outperforming much larger frontier models by 13.6 points. The paper says the library and conformance checker are being released for use with other primitive libraries. ArXiv · AI/CL/LG's note
score 5