Are Language Models Efficient Reasoners? A Perspective from Logic Programming

Bernhard Schölkopf (MPI for Intelligent Systems, Tübingen) · Andreas Opedal (Department of Computer Science, ETHZ - ETH Zurich) · Yanick Zengaffinen (ETHZ - ETH Zurich) · Haruki Shirakami (EPFL - EPF Lausanne) · Clemente Pasti (ETHZ - ETH Zurich) · Mrinmaya Sachan (ETH Zurich) · Abulhair Saparov (Purdue University) · Ryan Cotterell (ETH Zürich)
accuracy declineaxiomsdeductive reasoningdomain-consistent distractionsempirical evaluationinference alignmentirrelevant informationlanguage modelslogic programmingmath word problemsmodel assessmentnatural language proofsproof detoursreasoning frameworksemantic overlap

Modern language models (LMs) exhibit strong deductive reasoning capabilities, yet standard evaluations emphasize correctness while overlooking a key aspect of human-like reasoning: *efficiency*. In real-world reasoning scenarios, much of the available information is irrelevant, and effective deductive inference requires identifying and ignoring such distractions. We propose a framework for assessing LM reasoning efficiency through the lens of logic programming, introducing a simple method to align proofs written in natural language