Entity · paper

Understanding the Impact of Linguistic Realization Choices on LLM Stance with Causal Tracing

paperactiveprovisionalunderstanding-the-impact-of-linguistic-realization-choices-on-llm-stance-with-causal-tracing-568037c9·1 events·first seen 3d ago

Aliases: Understanding the Impact of Linguistic Realization Choices on LLM Stance with Causal Tracing

Co-occurring entities

activation patching

More like this (12)

From Plausible to Actionable: A Position on LLM Self-Explanations Revising Context, Shifting Simulated Stance: Auditing LLM-Based Stance Simulation in Online Discussions Knowledge Knows, Verbalization Tells: Disentangling Latent Directions for Mathematical Solvability in LLMs Expert-Aware Causal Tracing of Factual Recall in Sparse MoE Language Models Conversable Complexity: Agentic LLM Collectives as Interpretable Substrates Words as Difference Makers: How Large Language Models Determine Causal Structure in Text Groc-PO: Grounded Context Preference Optimization for Truthful Multimodal LLMs Actionable Activation Directions for Detecting and Mitigating Emergent Misalignment Across Language Model Families The Masked Advantage: Uncovering Local-Language Access to Cultural Knowledge in LLMs Inside the Unfair Judge: A Mechanistic Interpretability Account of LLM-as-Judge Bias Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection The Value Axis: Language Models Encode Whether They're on the Right Track

Recent events (1)

5arXiv · cs.CL·3d ago·source ↗

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.

Evaluation and Benchmarking AI Safety Research Understanding the Impact of Linguistic Realization Choices on LLM Stance with Causal Tracing activation patching