Gespeichert in:
| Hauptverfasser: | , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | https://arxiv.org/abs/2604.14324 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866917411753558016 |
|---|---|
| author | An, Hao Lou, Yibin Guo, Jiayi Xu, Yang |
| author_facet | An, Hao Lou, Yibin Guo, Jiayi Xu, Yang |
| contents | 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 the **GeoDe** (**Geo**metric **De**noising) framework for abstention fine-tuning. This method constructs a truth hyperplane using linear probes and performs "geometric denoising" by employing geometric distance as a confidence signal for abstention decisions. 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 GeoDe significantly enhances model truthfulness and demonstrates strong generalization in out-of-distribution (OOD) scenarios. Code is available at https://github.com/Notbesidemoon/GeoDe. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_14324 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Purging the Gray Zone: Latent-Geometric Denoising for Precise Knowledge Boundary Awareness An, Hao Lou, Yibin Guo, Jiayi Xu, Yang Computation and Language 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 the **GeoDe** (**Geo**metric **De**noising) framework for abstention fine-tuning. This method constructs a truth hyperplane using linear probes and performs "geometric denoising" by employing geometric distance as a confidence signal for abstention decisions. 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 GeoDe significantly enhances model truthfulness and demonstrates strong generalization in out-of-distribution (OOD) scenarios. Code is available at https://github.com/Notbesidemoon/GeoDe. |
| title | Purging the Gray Zone: Latent-Geometric Denoising for Precise Knowledge Boundary Awareness |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2604.14324 |