5arXiv cs.CL (Computation and Language)·
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.
Related guides (2)
Related events (8)
Counterfactual context revision framework for auditing LLM-based stance simulation in online discussions
Researchers introduce a counterfactual context revision framework to audit how LLMs simulate individual users' stances in online discussions. By applying controlled text-only and multimodal (meme-based) revisions to conversational contexts, they measure how readily simulated stances shift in response to semantically independent changes. Results show effective and robust stance transitions across both revision types and polarization-preference mechanisms, raising concerns about whether LLM simulations reflect genuine user-specific beliefs or are highly context-sensitive artifacts. The work contributes an evaluation framework and highlights risks of using LLMs to model online opinion dynamics.
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.
Tracing the Emergence of Human-Like Pragmatic Reasoning in LLMs Across Languages
Researchers conducted a population-matching experiment evaluating 25 LLMs on conditional inference tasks across four languages, comparing model behavior to matched human populations. The study finds that LLMs function as accurate semantic operators but systematically fail to capture pragmatic enrichments—context-sensitive inferences beyond literal logical meaning—that humans apply effortlessly. Model performance on pragmatic reasoning is not predicted by open vs. closed weights, training orientation, or architecture type, suggesting pragmatic reasoning remains an emergent and unreliable capability. The findings contribute to ongoing debates about whether LLMs reason like humans or merely approximate surface-level linguistic patterns.
Inference-time steering methods compared for cross-lingual factual consistency in LLMs
Researchers investigate four intervention strategies for reducing cross-lingual factual inconsistency in LLMs, where models give different factual answers depending solely on prompt language. Testing on Gemma 3 12B Instruct, they compare persona prompting, Contrastive Activation Addition (CAA), and DPO-based adapters, finding that simple persona prompting outperforms more invasive methods in efficacy, safety, and generalization. The paper also introduces a novel culturally-rooted generalization benchmark to assess whether factual interventions transfer beyond the training distribution. Results suggest cross-lingual inconsistency is partly a selection problem addressable without weight modification.
Benchmark study reveals how linguistic framing of user beliefs shifts LLM context-following behavior
A new arXiv paper introduces a typology of 17 linguistically motivated expression-of-belief (EoB) types—spanning form, evidentiality, epistemic stance, and tone—to evaluate how phrasing affects whether LLMs defer to user-stated beliefs or their own prior knowledge. The authors benchmark 16 LLMs across Llama 3, Qwen3, and Gemma3 families at scales from 1B to 30B parameters, finding that larger and instruction-tuned models are systematically less context-following than smaller or base models. Specific linguistic framings (e.g., presuppositions, certainty markers) are identified as statistically more persuasive, with implications for prompt robustness and sycophancy research.
Systematic study reveals effectiveness-fluency trade-offs in LLM conditioning methods
A new arXiv paper systematically evaluates a range of LLM conditioning methods across both concept injection and removal scenarios, finding that efficient steering methods often degrade fluency significantly. A key finding is that activation steering is substantially less effective on instruction-tuned models than on base models, a previously overlooked interaction. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics are found to correlate well with expensive LLM-as-judge evaluations.
Alignment tuning installs distinct representational directions for sycophancy and cue-induced biases in LLMs
A new arXiv paper investigates where sycophancy and related cue-induced biases live inside LLMs, studying five model families and seven bias types. The authors find that alignment tuning — not pretraining — is primarily responsible for installing these biases, as base models show minimal susceptibility and no cue-specific activation signal. Each bias manifests as a distinct, causally active direction in hidden states that can be probed, transferred across datasets, and steered to recover unbiased answers. The work also demonstrates a modest debiasing intervention that reduces bias-induced errors while preserving correct answers.
Paper argues LLMs learn causal structure via difference-making logic, not Pearl/Rubin frameworks
A new arXiv preprint proposes that LLMs learn causal structure through 'variational induction' — a difference-making logic — rather than through the dominant formalisms of Judea Pearl's interventionist approach or the Neyman-Rubin potential outcomes framework. The author analyzes how this logic is realized during training and maps specific architectural features (token embeddings, self-attention) to their roles in this inductive process. The argument draws a parallel between LLM causal learning and the experimental method of systematically varying circumstances. This is a theoretical contribution to understanding how LLMs represent causal and world-model structure.