Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2509.25868 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915951295856640 |
|---|---|
| author | Wang, Yindong Preiß, Martin Bugueño, Margarita Hoffbauer, Jan Vincent Ghajar, Abdullatif Buz, Tolga de Melo, Gerard |
| author_facet | Wang, Yindong Preiß, Martin Bugueño, Margarita Hoffbauer, Jan Vincent Ghajar, Abdullatif Buz, Tolga de Melo, Gerard |
| contents | The mechanisms underlying scientific confabulation in Large Language Models (LLMs) remain poorly understood. We introduce ReFACT (Reddit False And Correct Texts), a benchmark of 1,001 expert-annotated question-answer pairs with span-level error annotations derived from Reddit's r/AskScience. Evaluating 9 state-of-the-art LLMs reveals two critical limitations. First, models exhibit a dominant "salient distractor" failure mode: 61% of incorrect span predictions are semantically unrelated to actual errors. Crucially, this pattern persists across all model scales (1B to 70B), indicating a fundamental semantic grounding deficit that scaling alone fails to resolve. Second, we find that comparative judgment is paradoxically harder than independent detection, even GPT-4o's F1 score drops from 0.67 to 0.53 when comparing answers side-by-side. These findings directly challenge the reliability of LLM-as-Judge paradigms for scientific factuality. Code and data are released at https://github.com/ddz5431/ReFACT. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_25868 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | ReFACT: A Benchmark for Scientific Confabulation Detection with Positional Error Annotations Wang, Yindong Preiß, Martin Bugueño, Margarita Hoffbauer, Jan Vincent Ghajar, Abdullatif Buz, Tolga de Melo, Gerard Computation and Language The mechanisms underlying scientific confabulation in Large Language Models (LLMs) remain poorly understood. We introduce ReFACT (Reddit False And Correct Texts), a benchmark of 1,001 expert-annotated question-answer pairs with span-level error annotations derived from Reddit's r/AskScience. Evaluating 9 state-of-the-art LLMs reveals two critical limitations. First, models exhibit a dominant "salient distractor" failure mode: 61% of incorrect span predictions are semantically unrelated to actual errors. Crucially, this pattern persists across all model scales (1B to 70B), indicating a fundamental semantic grounding deficit that scaling alone fails to resolve. Second, we find that comparative judgment is paradoxically harder than independent detection, even GPT-4o's F1 score drops from 0.67 to 0.53 when comparing answers side-by-side. These findings directly challenge the reliability of LLM-as-Judge paradigms for scientific factuality. Code and data are released at https://github.com/ddz5431/ReFACT. |
| title | ReFACT: A Benchmark for Scientific Confabulation Detection with Positional Error Annotations |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2509.25868 |