Vernier: Probing Representational Misalignment Behind Lexical Gaps in Causal Reasoning

Instruction-tuned language models can answer the same causal-reasoning question differently when English variable names are swapped for type-preserving placeholders, according to a study from the Vernier lab posted to arXiv on 14 June 2026 [1][2]. The work argues the gap reflects representational misalignment rather than information loss. The paper, titled “Vernier: Probing Representational Misalignment Behind Lexical Gaps in Causal Reasoning,” examines why models change their answers even though the underlying structural causal model and the gold answer remain unchanged [1][2]. The authors introduce a paired-view weight update as an instrument and then inspect the mechanism that remains after the lexical gap is closed [2]. In the regimes where the method works, the evidence points to representational misalignment [2]. A variable-name probe becomes more accurate on the placeholder view, and activation patching on Qwen-7B, Qwen-14B, and Llama-3.1-8B shows that the decision-token representation can transfer answer identity between the original and placeholder views [1][2]. The update that realigns the two views is counterfactual augmentation over the original and placeholder prompts, while the answer-subspace KL divergence mainly sharpens intermediate answer-belief agreement [2]. Success is constrained by model family, scale, and task [2]. CRASS transfer is reliable across Qwen scales and Llama, but performance on the e-CARE benchmark remains weak [1][2]. Preliminary experiments on non-causal rename tasks exhibit a similar qualitative pattern, suggesting the phenomenon may extend beyond causal reasoning [2]. The study was posted on arXiv, an open-access repository that hosts electronic preprints across fields including computer science and mathematics [6]. As of November 2024, the repository was receiving roughly 24,000 submissions per month [6]. The Vernier paper appears within the Computation and Language category, a subfield of computer science that deals with large language models—machine learning systems trained on vast amounts of text for natural language processing tasks [8].