MIT's PDDL-INSTRUCT Boosts Llama-3-8B Planning Accuracy from 28% to 94%
mdancho84 · x · 2026-09-25
MIT researchers propose PDDL-INSTRUCT, claiming it teaches LLMs genuine logical reasoning and planning rather than pattern matching.
- Training: fine-tune on correct and incorrect plans with explanations.
- External verification: an outside verification process checks each step of the LLM's reasoning for soundness.
On their planning benchmarks, Llama-3-8B jumped from 28% to 94% accuracy — a step-change rather than incremental gain. This is a third-party thread summarizing the paper; details warrant verification.
More from Research
- Inside LFM2.5-2.6B's post-training recipe: SFT, RL, multi-domain distillation — helloiamleonie · 2026-09-25
- Multi-domain on-policy distillation lets one model learn many skills at once — helloiamleonie · 2026-09-25
- On-policy distillation explained: teachers grade every token of student trajectories — helloiamleonie · 2026-09-25
- RLVR refresher: verifiable rewards work but the signal is sparse — helloiamleonie · 2026-09-25
- Liquid AI's "Antidoom Training" cuts model doom loops from 10% to 1% — helloiamleonie · 2026-09-25
- Liquid AI explains on-device small model design: antidoom training cuts doom loops from ~10% to 1% in LFM2.5 — helloiamleonie · 2026-09-25