The Paradox of Outcome Optimization: A Causal Information-Theoretic Bound on Reasoning Shortcuts in LLMs 文章

ArXiv CS.AI2026-06-02NEWSen作者: Zihan Chen, Yiming Zhang, Wenxiang Geng, Zenghui Ding, Yining Sun

查看原文 →

The Paradox of Outcome Optimization: A Causal Information-Theoretic Bound on Reasoning Shortcuts in LLMs · 相关技术

相关技术

shortcut learning Markovian Screening stochastic gradient descent Information Bottleneck structural causal model Reward-Induced Manifold Collapse reinforcement learning large language models