aiqbench analysis
This study challenges the prevailing assumption that linear probes on LLM hidden states reveal distinct reasoning modes, demonstrating instead that apparent separations are artifacts of task format confounds like source identity and response length. By showing that residualizing these variables reduces probe accuracy to chance, the authors argue that current mechanistic interpretability claims require rigorous deconfounding to avoid misattributing structural differences to computational reasoning types. The findings underscore the necessity of controlling for superficial input features before drawing conclusions about a model's internal cognitive architecture.
Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States
arXiv:2606.02907v1 Announce Type: cross Abstract: Linear probing of large language model (LLM) hidden states is widely used to claim that models learn distinct representations for different reasoning types. We test this by probing Qwen3-14B on three benchmarks spanning the classical trichotomy: LogiQA 2.0 (deductive), ARC-Challenge (inductive), and $\alpha$NLI (abductive). At layer 32 of 40, linear probes achieve 100\% cross-validated accuracy with well-separated geometry (intrinsic dimensionalities: 20.6, 28.5, 33.6; convex hull contamination $\leq$1.5\%). However, this separation is entirely driven by format confounds. Residualizing source identity, option count, and response length reduces accuracy to chance. Trace-anchor similarity indicates largely shared reasoning across tasks (42.5\% agreement vs.\ 33.3\% chance), and causal steering with random controls ($n=20$) shows no functional link between geometry and reasoning mode ($p=0.286$). Thus, high probe accuracy reflects task format rather than computational structure, motivating routine format deconfounding in mechanistic interpretability.
Abstract:Linear probing of large language model (LLM) hidden states is widely used to claim that models learn distinct representations for different reasoning types. We test this by probing Qwen3-14B on three benchmarks spanning the classical trichotomy: LogiQA 2.0 (deductive), ARC-Challenge (inductive), and $\alpha$NLI (abductive). At layer 32 of 40, linear probes achieve 100\% cross-validated accuracy with well-separated geometry (intrinsic dimensionalities: 20.6, 28.5, 33.6; convex hull contamination $\leq$1.5\%). However, this separation is entirely driven by format confounds. Residualizing source identity, option count, and response length reduces accuracy to chance. Trace-anchor similarity indicates largely shared reasoning across tasks (42.5\% agreement vs.\ 33.3\% chance), and causal steering with random controls ($n=20$) shows no functional link between geometry and reasoning mode ($p=0.286$). Thus, high probe accuracy reflects task format rather than computational structure, motivating routine format deconfounding in mechanistic interpretability.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2606.02907 [cs.CL]
(or arXiv:2606.02907v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.02907 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 232 words
Related coverage
Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States
arXiv · Jun 6, 2026
PIRAMID: Progress and Plans
LessWrong · Jul 27, 2026
Improving CoT Monitorability of Evaluation Awareness via Verbalization Training
LessWrong · Oct 2, 2026
Comment on Measuring Reward-Seeking by Instilling Contrastive Beliefs paper from mechanistic interpretability perspective
LessWrong · Jul 22, 2026