Abstention Direction in Large Language Models: Mechanistic Interpretability of Abstention Across Query Ambiguity, Knowledge Limits, and Safety Alignment
Summary
The paper extends the refusal direction framework to study abstention in LLMs across query ambiguity, model knowledge limits, and safety alignment, showing that steering middle-layer activations can increase or decrease abstention and revealing a shared directional component across the three perspectives.
View Cached Full Text
Cached at: 11/19/26, 03:42 AM
Paper page - Abstention Direction in Large Language Models: Mechanistic Interpretability of Abstention Across Query Ambiguity, Knowledge Limits, and Safety Alignment
Source: https://huggingface.co/papers/2611.20441
Abstract
Abstention in large language models arises from query ambiguity, model knowledge limits, and value misalignment, and steering internal activations in middle layers can increase or decrease abstention across these causes.
Abstention, the withholding of a response by alarge language model (LLM)underuncertaintyorsafety risks, is essential for trustworthy models. The causes ofabstentionfall into three perspectives: ambiguity in the Query (Q), limits of Model Knowledge (MK), and misalignment with Human Values (HV). However, prior analyses ofinternal representationshave focused on HV, using arefusal directioncomputed fromharmful and harmless queries. No study has analyzed the three perspectives jointly. We therefore extend therefusal directionframework to all three perspectives. Specifically, for each perspective, we constructcounterfactual pairsthat share a template and differ only in theabstention-inducing factor, and compute anabstention directionby normalizing the difference inmean activationsbetween the two polarities.Activation steeringexperiments show that intervening inmiddle layersincreases or decreasesabstentionacross all three perspectives. Moreover, steering yieldscross-perspective effects, indicating that theabstention directions share a commondirectional component.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2611\.20441
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2611.20441 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2611.20441 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2611.20441 in a Space README.md to link it from this page.
Collections including this paper1
Similar Articles
Two Axes of LLM Abstention: Answer Correctness and Question Answerability
This paper investigates the two axes of LLM abstention: answer correctness and question answerability. It shows that a single confidence threshold conflates these two failure modes, and proposes a three-class selective acceptance framework with separate budgets. Experiments across five instruction-tuned models reveal that answerability is internally legible but poorly captured by output confidence or self-assessments.
Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions
This paper investigates why large language models answer structurally unanswerable questions, concluding that models recognize impossibility but fail to route that recognition to refusal, highlighting a routing failure rather than an encoding failure in their decision-making.
Agentic Abstention: Do Agents Know When to Stop Instead of Act?
This paper defines agentic abstention, the problem of deciding when an LLM agent should stop acting under uncertainty, and evaluates it across web shopping, terminal environments, and question answering. It introduces convolve, a context engineering method that improves timely abstention without updating model parameters.
Abstention as an Action Can Kill Both the Reward Gradient and the KL Anchor: Collapse Law and Repair for Error-Penalized Reinforcement Learning
This paper proves that using error-penalized scoring rules with abstention as a discrete action can kill both the reward gradient and the KL anchor, causing models to collapse toward refusing everything. It proposes a structural repair — training a mandatory confidence report — and validates the mechanism with simulations and language model experiments.
Mitigating LLM Over-Refusal via Dynamic Semantic Routing Calibratione
The paper presents a mechanistic analysis of over-refusal in large language models and proposes Semantic Routing Calibration (SRC), a lightweight, training-free inference framework to dynamically suppress hypersensitive safety heads and mitigate over-refusal while preserving intrinsic safety.