← Back to all news
arXiv Machine Learning By Fredrik R{\o}mming, Mantas Bak\v{s}ys, Martin S. Fixman, Sean B. Holden

Imitation Learning for Connection-Tableau Construction

Read the original on arXiv Machine Learning →

The paper presents an approach to automated theorem proving by framing the construction of clausal connection tableaux as a policy in a transition system. It introduces a graph neural network that scores proof edits based on structure, trained via imitation learning from existing proofs. Experiments on M2k, MPTP2078-bushy, and TPTP v9.2.1 show that the learned policies solve up to 46% more problems than leanCoP and find proofs in an order of magnitude fewer steps.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

Related stories

arXiv AI
Aug 28

ProofEvolve: Neuro-Symbolic Evolution for Formal Automated Theorem Proving

ProofEvolve is a neuro‑symbolic framework that evolves formally verified symbolic proof structures alongside neural models to expand the knowledge boundary in automated theorem proving. The neural component proposes variation operators such as decompositions, repairs, and schema recombinations, while the Lean kernel verifies every proof transition, ensuring formal soundness. Across three competition‑level Lean benchmarks, ProofEvolve achieves the highest average solve rate among evaluated proof systems.

By Wenqian Ye, Ziwei Guan, Eric Xie, Bohan Liu, Shivani Modi, Buyun Zhang, Ellie Dingqiao Wen, Henry Kautz, Aidong Zhang
benchmarks
More like this →
arXiv AI
2d ago

Trust the Critic More

The paper introduces Actor‑Critic with Action Chunking (AC2), a method that assigns credit to short action chunks instead of entire trajectories, enabling policy updates without waiting for terminal rewards. AC2 employs local readiness, reference solutions, and 10k‑token chunks to make critic‑based credit assignment reliable. Experiments on Qwen3‑4B with FineProofs‑RL show AC2 surpasses GRPO’s peak validation score while using 2.5× fewer decoding FLOPs and fewer training steps.

By Kaiyue Wen, Luke Bailey, Arvind Mahankali, Tengyu Ma
llmsragreinforcement-learning
More like this →
arXiv AI
Aug 14

Exploiting Symbolic Heuristics for the Synthesis of Domain-Specific Temporal Planning Guidance using Reinforcement Learning

arXiv:2505. 13372v2 Announce Type: replace Abstract: Recent work investigated the use of Reinforcement Learning (RL) for the synthesis of heuristic guidance to improve the performance of temporal planners when a domain is fixed and a set of training problems (not plans) is given.

By Irene Brugnara, Alessandro Valentini, Andrea Micheli
reinforcement-learning
More like this →