Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

Research presented at TrustNLP 2026 reveals that linear probes, commonly used to infer distinct reasoning representations in large language model (LLM) hidden states, primarily detect task format rather than actual reasoning modes. The study, using Qwen3-14B, tested three benchmarks: LogiQA 2.0 (deductive), ARC-Challenge (inductive), and αNLI (abductive). While probes at layer 32 of 40 achieved 100% cross-validated accuracy with well-separated geometry (intrinsic dimensionalities: 20.6, 28.5, 33.6; convex hull contamination ≤1.5%), this separation vanished when format confounds like source identity, option count, and response length were residualized, reducing accuracy to chance. Further analysis showed largely shared reasoning across tasks (42.5% agreement vs. 33.3% chance) and no functional link between geometry and reasoning mode (p=0.286) via causal steering.

Key takeaway

For NLP Engineers developing mechanistic interpretability tools, you should routinely deconfound for task format variables like source identity, option count, and response length when using linear probes. Failing to do so risks misinterpreting probe accuracy as evidence of distinct reasoning modes, when it merely reflects superficial task characteristics. Validate findings with causal steering to confirm functional links.

Key insights

LLM linear probes detect task format, not reasoning mode, due to format confounds.

Principles

• High probe accuracy can mask format confounds.
• Reasoning modes may share representations.

Method

Probed Qwen3-14B on LogiQA 2.0, ARC-Challenge, and αNLI. Residualized format confounds, used trace-anchor similarity, and causal steering with random controls (n=20).

In practice

• Deconfound format in interpretability studies.
• Validate probe findings with causal steering.

Topics

• Linear Probing
• LLM Interpretability
• Reasoning Modes
• Task Format
• Qwen3-14B
• Trustworthy NLP

Best for: AI Scientist, Research Scientist, NLP Engineer

Related on AIssential

• Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States — Linear probes in LLMs often detect task format confounds, not distinct reasoning mode representations, challenging common interpretability claims.
• Rhetorical Questions in LLM Representations: A Linear Probing Study — LLMs encode rhetorical questions via multiple linear directions, emphasizing distinct cues rather than a single shared representation.
• LMs as Task-Specific Knowledge Bases: An Interpretability Analysis — Language models encode factual knowledge in a task-specific manner, undermining the consistent "knowledge base" analogy.
• Probing the Misaligned Thinking Process of Language Models — Linear probes can reliably detect LLM misalignment indicators in internal activations.
• Attention Probes — Attention probes offer a multi-headed, attention-based alternative to traditional pooling for classifying language model states.
• H-Probes: Extracting Hierarchical Structures From Latent Representations of Language Models — LLMs geometrically represent hierarchical structures like tree depth and pairwise distance in low-dimensional latent subspaces.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.