CEDAR-GRPO: Process-Aware RL Enhances Abductive Reasoning in LLMs

A new research paper introduces CEDAR-GRPO, a process-aware reinforcement learning framework designed to improve abductive reasoning in large language models (LLMs). Abductive reasoning, or inference to the best explanation, is crucial for making sense of uncertain situations, from everyday problem-solving to scientific discovery. However, prior LLM research has largely focused on narrow, task-specific benchmarks, leaving it unclear whether improvements generalize beyond those benchmarks. The authors propose that reinforcement learning (RL) post-training can enhance abduction as a transferable capability. CEDAR-GRPO combines final-answer correctness with abductive rewards that assess evidence coverage and the directionality from evidence to explanation. Four open-weight LLMs were post-trained on a controlled, domain-neutral mix of abductive hypothesis-generation and hypothesis-selection tasks. The models were then evaluated on 11 unseen tasks covering hypothesis selection, missing-fact generation, defeasible inference, and long-context investigation. The paper is available on arXiv under the identifier 2608.14791.