Entity · paper

Knowledge Knows, Verbalization Tells: Disentangling Latent Directions for Mathematical Solvability in LLMs

paperactiveprovisionalknowledge-knows-verbalization-tells-disentangling-latent-directions-for-mathematical-solvability-in-llms-5d4059e1·1 events·first seen 40h ago

Aliases: Knowledge Knows, Verbalization Tells: Disentangling Latent Directions for Mathematical Solvability in LLMs

More like this (12)

Operads for compositional reasoning in LLMs Operadic consistency: a label-free signal for compositional reasoning failures in LLMs Conversable Complexity: Agentic LLM Collectives as Interpretable Substrates Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning Janus: A Benchmark for Goal-Conditioned Information Distortion in LLMs Attractor States Emerge in Multi-Turn LLM Conversations code synthesis LLMs Which Models Are Our Models Built On? Auditing Invisible Dependencies in Modern LLMs LMs as Task-Specific Knowledge Bases: An Interpretability Analysis Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs Follow the Latent Roadmap: Navigating Revocable Decoding for Diffusion LLMs with Anchor Tokens Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs

Recent events (1)

5arXiv · cs.CL·40h ago·source ↗

Disentangling solvability knowledge from verbalization in LLM hidden states

A new arXiv preprint probes LLM internal representations to separately decode 'solvability knowledge' and 'verbalization' as distinct linear directions in hidden states. The authors find that fabrication (hallucination of solutions to unsolvable problems) is primarily driven by shifts in verbalization representations rather than underlying knowledge. Prompting with unsolvability cues and activation steering can mechanistically shift verbalization to improve model abstention. The work advances mechanistic understanding of why LLMs confabulate on unsolvable math problems.

Evaluation and Benchmarking AI Safety Research Knowledge Knows, Verbalization Tells: Disentangling Latent Directions for Mathematical Solvability in LLMs