Number 3 of 12 · November 18, 2026

Abstention Direction in Large Language Models: Mechanistic Interpretability of Abstention Across Query Ambiguity, Knowledge Limits, and Safety Alignment

The authors find directions inside language models that control abstention (declining to answer) for unclear questions, unknown facts and harmful requests, and these directions partly overlap.

What they found

The authors extend the "refusal direction" method, which was earlier used only for harmful requests, to three causes of abstention: unclear queries, limits of the model's knowledge, and conflict with human values. For each cause, they build matched pairs of prompts that share a template and differ only in the factor that should trigger abstention. They take the difference in average internal activations between the two kinds of prompt and normalize it to get an abstention direction. Pushing the model along this direction in middle layers (activation steering) raised or lowered abstention for all three causes. Steering on one cause also changed abstention on the others, which the authors read as a sign of a shared directional component.

Why it matters

Models that decline to answer when they should are more trustworthy. Until now, internal studies mostly looked only at refusing harmful requests. This work looks at all three causes together. The abstract gives no numbers, so it does not say how large the effects were or which models were tested.

How it changes the race

This is an incremental step in understanding SI models. If it holds, builders could use one shared internal signal to study and tune when a model abstains, but the abstract does not show how well this works in practice.

Safety and control

The work bears on control. A direction that raises or lowers abstention could help make models more careful. It could also be used to weaken refusals of harmful requests, so the same tool has risks as well as benefits.

The abstract, by the authors

Abstention, the withholding of a response by a large language model (LLM) under uncertainty or safety risks, is essential for trustworthy models. The causes of abstention fall into three perspectives: ambiguity in the Query (Q), limits of Model Knowledge (MK), and misalignment with Human Values (HV). However, prior analyses of internal representations have focused on HV, using a refusal direction computed from harmful and harmless queries. No study has analyzed the three perspectives jointly. We therefore extend the refusal direction framework to all three perspectives. Specifically, for each perspective, we construct counterfactual pairs that share a template and differ only in the abstention-inducing factor, and compute an abstention direction by normalizing the difference in mean activations between the two polarities. Activation steering experiments show that intervening in middle layers increases or decreases abstention across all three perspectives. Moreover, steering yields cross-perspective effects, indicating that the abstention directions share a common directional component.

Quoted from arXiv:2611.20441. The research and its words belong to its authors.

The authors

Koki Itai, Jun Hirako, Hiroaki Yamagiwa

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.