LLaDA-PRM uses bidirectional attention to improve reasoning evaluation
September 22, 2026
The LLaDA-PRM 8B model utilizes bidirectional attention to evaluate reasoning steps, outperforming the 34B ReasonEval-Llemma by up to 11.3 F1 points on MR-MATH-invalid. By looking at downstream consequences of earlier steps, the evaluator provides more consistent step-level verification.
HOW THIS AFFECTS YOU
●
builderYou can use more accurate evaluators to prune reasoning traces in your agentic workflows.
●
researcherBidirectional attention masks provide a superior alternative to causal attention for step-level process supervision.