RESEARCHInvestigateNEXT 12 MONTHS
Not All LLM Reasoning is Visible in the Chain-of-Thought
arXiv cs.CL — Computation and Language
Factual evidence
What the source reports
Research shows frontier LLMs use 'invisible reasoning' with semantically irrelevant filler tokens to improve performance on synthetic tasks.
OneBench interpretation
Institutional assessment
So what
LLM explainability and model risk frameworks must account for non-transparent reasoning pathways, impacting validation and responsible AI.
Do what
Add this finding to the Q3 model risk working group agenda for impact assessment on current explainability techniques.