Get Started
Home
Topics
Search
Library
Research questionHow can we distinguish decodable logical validity from reasoning that actually drives a language model’s answers?A logical-validity signal can be readily decoded from hidden states even when the model answers incorrectly. The central difficulty is determining whether that signal is expressed in behavior and causally influences the answer.
AI
Evaluation & Benchmarks
Machine Learning
Mechanistic Interpretability
Natural Language Processing
Reasoning
Research Paper
Latest papersRecent research connected to this question, newest first.When Chain-of-Thought Fails, the Solution Hides in the Hidden StatesThe evidence comes from language models solving GSM8K, using activation patching to transfer token-level hidden states from chain-of-thought runs into direct-answer runs. It compares correct and incorrect traces and examines token types and model layers; it does not establish behavior beyond this task and analysis setting.research paper · Sep 2, 2026When Decodability Is Not Enough: Logical Validity Representations, Behavioral Dissociation, and Causal Tests in Language ModelsThe evidence covers five open-weight transformer models evaluated on matched valid–invalid premise–claim pairs spanning multiple inference families, semantic domains, templates, and difficulty levels. Validity remains decodable across held-out conditions and some behaviorally incorrect examples, while leave-one-out tests reveal limits and probe-direction interventions produce weak, nonspecific effects relative to random controls.research paper · Sep 2, 2026
Related questions
How can language models perform complex logical reasoning without accumulating token-level errors?How can language models produce reliable educational answers with reasoning that is verifiable?How can LLMs interleave reasoning with reliable step-level self-critique without a separate verifier?How can we quantify and reduce divergent, nonsensical reasoning in large language models?