Research map / Large language models / Language, safety and interpretability
Hallucinate Less by Thinking More: Aspect-Based Causal Abstention for Large Language Models
Vy Nguyen, Ziqi Xu, Jeffrey Chan, Estrid He, Feng Xia, Xiuzhen 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 aspect conditioning (vs pooling all CoTs, Collapsed-X) (ablation) | ABCA, GPT-4.1 · TruthfulQA | Acc: .835 -> .914 | Improved |
| add aspect conditioning (vs No-X) (ablation) | ABCA, GPT-4.1 · TruthfulQA | Acc: .869 -> .914 | Improved |
| use dual-agent debate for aspect discovery instead of a single agent (ablation) | ABCA, GPT-4.1 · TruthfulQA | Acc: .871 -> .914 | Improved |
| use aspect-aware weights instead of uniform weights (ablation) | ABCA, GPT-4.1 · TruthfulQA | Acc: .851 -> .914 | Improved |
| use causal-effect significance instead of uniform effects (ablation) | ABCA, GPT-4.1 · KUQ | Acc: .746 -> .768 | Improved |
| use full iterative debate and AIPW sampling instead of Lite (T=K=N=1) (ablation) | ABCA, GPT-4.1 · TruthfulQA | Acc: .895 -> .914 | Improved |
| use discovered aspects instead of fixed language aspects (Fixed-X) (ablation) | ABCA, GPT-4.1 · KUQ | Acc: .757 -> .768 | Improved |
| add ABCA over zero-shot prompting | GPT-4.1 · TruthfulQA | Acc: .838 -> .914 | Improved |
| add ABCA over zero-shot prompting | GPT-4.1 · TruthfulQA | U-Ac: .476 -> .964 | Improved |
| add ABCA on AbstainQA (MMLU) | GPT-4.1 · AbstainQA (MMLU) | Acc: .693 (CFMAD) -> .696 | No clear effect |
| add ABCA to Mistral-NeMo 12B on AbstainQA | Mistral-NeMo 12B · AbstainQA (MMLU) | Acc: .341 -> .403 | Improved |
| add aspect conditioning to CoT sampling | ABCA · NLI diversity of CoTs | NLI Diversity: GPT-4.1 TruthfulQA 0.65 (+0.26 over Self-Consistency) | Improved |
Closest papers by meaning
- Mitigating Hallucinations in Large Language Models via Causal Reasoning (AAAI 2026 · similarity 0.91)
- In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation (ICML 2024 · similarity 0.89)
- The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining (NeurIPS 2026 · similarity 0.89)
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.89)
- Hallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the Answer (NeurIPS 2026 · similarity 0.89)
- Mitigating Modality Prior-Induced Hallucinations in Multimodal Large Language Models via Deciphering Attention Causality (ICLR 2025 · similarity 0.89)
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.
