4arXiv cs.CL (Computation and Language)·
Linear readouts of LLM hidden states decode causal reasoning about diagnostic evidence
Researchers introduce a paired-prompt benchmark testing whether language models can correctly match diagnostic evidence to causal claims that vary by population, estimand, or identifying assumptions — a task where surface-level cues can mislead. Using linear probes on final-token hidden states from Qwen2.5-7B, Qwen3-8B, and Llama-3.1-8B, they find balanced accuracy of 0.654–0.659 on a 49-pair benchmark spanning nine diagnostic families, exceeding permutation nulls and text-only baselines. The key finding is that hidden states contain linearly decodable information about causal relevance that is not fully captured by output logits or surface features.
Related guides (2)
Related events (8)
Linear probes reveal LLMs encode remaining output length in hidden states
Researchers trained minimal linear probes on frozen hidden states of three open-weight 7-8B models and found that total response length is linearly decodable from the prompt's final hidden state before any output is generated. The probe directions transfer across natural-language and synthetic datasets, and per-position estimates shift upward when models retract and restart partial solutions. The authors interpret this as evidence that LLMs maintain a plan-like internal representation of remaining generation length, distinct from exact-counting, though causality is not established.
Linear probes on chain-of-thought hidden states can predict reasoning non-convergence early
A new arXiv preprint studies the bimodal convergence pattern in chain-of-thought models like DeepSeek-R1-Distill-Qwen-7B, where generations either complete within a token budget (90.3% accuracy on AIME) or exhaust it without concluding (6.6% accuracy), with a 62% overall convergence rate. The authors train linear probes on hidden-state activations at early token positions (50-300) and find that layer-20 activations at token 150 achieve AUC 0.608, reliably above chance and outperforming behavioral baselines from token entropy and repetition statistics. The results suggest convergence fate is partially encoded in intermediate representations early in generation, pointing toward early-exit inference and adaptive compute allocation strategies. Statistical evidence is modest (permutation test p=0.063), limiting strong conclusions.
Causal tracing reveals mid-to-late decoder layers drive LLM stance sensitivity to linguistic construction
A new arXiv preprint investigates whether linguistic construction choices (beyond lexical variation) systematically shift LLM stance judgments, using political stance classification as a case study. The authors create six controlled rewrite types that preserve or invert meaning, then apply activation patching across four open-weight models to localize where these shifts originate. Results show that mid-to-late decoder layers, particularly block outputs at the final prompt position, provide the strongest restoration signal for the original stance distribution. The work extends mechanistic interpretability methods to a socially sensitive domain and highlights a largely overlooked axis of LLM prompt sensitivity.
Mechanistic interpretability study reveals geometric structure of LLM-as-judge bias in hidden states
A new arXiv preprint investigates LLM-as-judge scoring bias at the representation level rather than the input-output level, studying seven judge models across seven bias types and nine benchmarks. The authors find that biased inputs are displaced along low-dimensional, type-specific subspaces in activation space, and that steering hidden states along these subspaces causally controls scoring direction. A linear projection onto bias-direction features predicts judge failures on unseen benchmarks, substantially outperforming text-based alternatives. The work provides a mechanistic account that unifies geometric structure, causal control, and operational prediction within a single framework.
Language models linearly encode a 'value axis' tracking expected goal success, study finds
Researchers construct a 'value axis' in Qwen3-8B's activation space using synthetic in-context RL data, finding that this axis distinguishes high vs. low confidence, backtracking vs. non-backtracking rollouts, and correct vs. corrupted code. Steering along this axis causally modulates self-correction behavior and verbosity, while DPO training shifts the internal value of rewarded behaviors. Applied to real-world settings, the axis reveals that Qwen assigns low internal value to politically sensitive queries post-training and that SFT increases domain-specific confidence. The findings suggest LLMs linearly encode an estimate of expected goal success that shapes their generative behavior.
Question-Answering as Hidden State Probing for Test-Time Reasoning Intervention
This paper proposes using question-asking as an inference-time intervention to surface information about an LLM's hidden state during chain-of-thought reasoning. The authors train a probe on a student model's hidden states before and after question generation, finding it predictive of final answer correctness even before the teacher responds—suggesting self-diagnosis during question generation carries meaningful signal. They frame question-asking as a sequential decision problem with a gating policy, but find a gap between detection and recovery: interventions are as likely to harm correct trajectories as to fix incorrect ones. The results have implications for the limits of LLM self-refinement under uncertainty.
Analysis-driven transformer linearization outperforms prior baselines on LLaMA and Qwen up to 32B
A new arXiv paper analyzes why post-hoc linearization of causal self-attention degrades model quality, identifying key-dependent rank-1 orthogonal projections as the mechanism softmax relies on and explaining why delta-style networks outperform gated accumulation. The authors introduce structural interventions—sink tokens, short convolutions, and fixed-budget cache routing—applied in a frozen-backbone regime. Scaling across LLaMA and Qwen models up to 32B parameters, the approach outperforms prior post-hoc linearization baselines on MMLU and matches long-context retrieval of adaptive-caching frameworks.
ThReadMed-QA benchmark reveals frontier LLMs degrade sharply on medical misconception correction across multi-turn conversations
Researchers introduce ThReadMed-QA, a multi-turn medical dialogue dataset of 2,437 conversation threads (8,204 QA pairs) derived from real AskDocs patient interactions, designed to evaluate LLM ability to detect and correct embedded misconceptions over multiple turns. Evaluation of five LLMs using an LLM-as-a-Judge rubric finds that even frontier models like GPT-5 and Claude Haiku drop from ~85% misconception correction on initial questions to ~50% within two follow-up turns. Oracle analysis attributes much of this degradation to error propagation from prior model outputs rather than context length alone. The findings highlight a significant safety gap in current evaluation frameworks, which do not capture multi-turn conversational dynamics in medical settings.