Purging the Gray Zone: Latent-Geometric Denoising for Precise Knowledge Boundary Awareness

ACL ARR 2026 January Submission935 Authors

Keywords: Hallucination, Abstention, Probing, QA, Knowledge boundary

Abstract: Large language models (LLMs) often exhibit hallucinations due to their inability to accurately perceive their own knowledge boundaries. Existing abstention fine-tuning methods typically partition datasets directly based on response accuracy, causing models to suffer from severe label noise near the decision boundaries and consequently exhibit high rates of abstentions or hallucinations. This paper adopts a latent space representation perspective, revealing a "gray zone" near the decision hyperplane where internal belief ambiguity constitutes the core performance bottleneck. Based on this insight, we propose GeoDe (Geometric Denoising) framework for abstention fine-tuning. This method constructs a truth hyperplane using linear probes and performs ``geometric denoising'' by employing geometric distance as a continuous abstention decision confidence metric. This approach filters out ambiguous boundary samples while retaining high-fidelity signals for fine-tuning. Experiments across multiple models (Llama3, Qwen3) and benchmark datasets (TriviaQA, NQ, SciQ, SimpleQA) demonstrate that our method significantly enhances model truthfulness and exhibits outstanding generalization capabilities in out-of-distribution (OOD) scenarios.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: fine-tuning

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 935