Author: Ricardo

Can a 8B-parameter language mannequin produce provably legitimate multi-step plans as an alternative of believable guesses? MIT CSAIL researchers introduce PDDL-INSTRUCT, an instruction-tuning framework that {couples} logical chain-of-thought with exterior plan validation (VAL) to carry symbolic planning efficiency of LLMs. On PlanBench, a tuned Llama-3-8B reaches 94% legitimate plans on Blocksworld, with massive jumps on…

Meta researchers launched a technique that compresses repeated reasoning patterns into quick, named procedures—“behaviors”—after which circumstances fashions to make use of them at inference or distills them by way of fine-tuning. The end result: as much as 46% fewer reasoning tokens on MATH whereas matching or bettering accuracy, and as much as 10% accuracy positive…