作者位次 #3/5ACL 2026Poster· Natural Language Processing
The Paradox of Outcome Optimization: A Causal Information-Theoretic Bound on Reasoning Shortcuts in LLM s
Zihan Chen · Yiming Zhang · Wenxiang Geng · Zenghui Ding · Yining Sun
摘要
Large Language Models (LLMs) aligned via outcome-based Reinforcement Learning (RL) frequently exhibit a critical failure mode: they achieve high performance on in-distribution benchmarks while demonstrating brittle reasoning capabilities on out-of-distribution (OOD) tasks. We term this phenomenon...