Research map / Large language models / Language, safety and interpretability
CoFact: Dynamic Coordination of Attention Heads for Improving Factual Consistency in LLMs
Shike Li, Xiaokai Wang, Xiaofeng Liu, Xin Tong, Hu Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CoFact head coordination | LLaMA2-7B-Chat · TruthfulQA open-ended | True*Info (%): 44.56 -> 59.86 | Improved |
| add CoFact head coordination | Gemma2-9B-it · TruthfulQA open-ended | True*Info (%): 58.33 -> 76.87 | Improved |
| add CoFact head coordination | Mistral-7B-Instruct-v0.3 · TruthfulQA | Info (%): 98.04 -> 97.31 | Hurt |
| add CoFact head coordination | Mistral-7B-Instruct-v0.3 · TruthfulQA | MC accuracy: 47.43 -> 46.95 | Hurt |
| add CoFact head coordination | LLaMA3-8B-Instruct · TriviaQA | True (%): 64.02 -> 65.31 | Improved |
| add CoFact head coordination | LLaMA3-8B-Instruct · NQ | True (%): 57.78 -> 59.91 | Improved |
| add ITI activation steering | LLaMA3-8B-Instruct · TriviaQA | True (%): 64.02 -> 58.56 | Hurt |
| add redundancy penalization (conflict arbiter) (ablation) | CoFact, LLaMA2-7B-Chat · TruthfulQA | True*Info (%): 45.11 -> 59.86 | Improved |
| add semantic utility scoring (semantic valuator) (ablation) | CoFact, LLaMA2-7B-Chat · TruthfulQA | True*Info (%): 46.23 -> 59.86 | Improved |
| add adaptive utility-aware weights instead of uniform head weighting (ablation) | CoFact, LLaMA2-7B-Chat · TruthfulQA | True*Info (%): 42.01 -> 59.86 | Improved |
| add adaptive weights instead of top-16 head selection (ablation) | CoFact, LLaMA2-7B-Chat · TruthfulQA | True*Info (%): 50.82 -> 59.86 | Improved |
| add noisy prompt prefix at inference (ablation) | CoFact, LLaMA2-7B-Chat · TruthfulQA (50 samples) | True*Info (%): 59.86 -> 47.41 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TruthFlow — TruthFlow: Truthful LLM Generation via Representation Flow Correction (ICML 2025) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators (AAAI 2025 · similarity 0.88)
- FLAME : Factuality-Aware Alignment for Large Language Models (NeurIPS 2024 · similarity 0.87)
- Cognitive Mirrors: Exploring the Diverse Functional Roles of Attention Heads in LLM Reasoning (NeurIPS 2025 · similarity 0.87)
- Reallocating Attention Across Layers to Reduce Multimodal Hallucination (CVPR 2026 · similarity 0.87)
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.87)
- One SPACE to Rule Them All: Jointly Mitigating Factuality and Faithfulness Hallucinations in LLMs (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
