Research map / Large language models / Language, safety and interpretability
LLMInertia: Adaptive Counter-Inertial Reasoning to Improve Evidence Faithfulness in Large Language Models
Xinxin You, Xien Liu, Chenwei Yan, Siqi Song, Chen Ning, Kaiyin Zhou, Shaohui Liu, Ji Wu
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 adaptive counter-inertial reminders (LLMInertia) to the base model | LLaMA-3-Instruct-8B · CNN/DailyMail | UniEval Consistency: 86.40 -> 90.33 | Improved |
| add LLMInertia to the base model | Qwen-2.5-Instruct-7B · SQuAD v2 | Anah-v2: 17.22 -> 11.22 | Improved |
| add LLMInertia to the base model | LLaMA-3-Instruct-70B · SQuAD v2 | Anah-v2: 11.52 -> 9.56 | Improved |
| add LLMInertia instead of SFT | LLaMA-3-Instruct-8B · HaluEval | Anah-v2: 8.27 vs 8.21 | No clear effect |
| add entity extraction (vs generic reminder for all samples) (ablation) | LLMInertia · CNN/DailyMail | UniEval Consistency: 87.54 -> 90.33 | Improved |
| add the LLM inertia probe (vs sliding-window co-occurrence pairs) (ablation) | LLMInertia · CNN/DailyMail | UniEval Consistency: 85.76 -> 90.33 | Improved |
| add the LLM inertia probe (vs sliding-window co-occurrence pairs) (ablation) | LLMInertia · SQuAD v2 | Anah-v2: 17.48 -> 12.53 | Improved |
| add adaptive reminders (vs fixed generic reminder) (ablation) | LLMInertia · SQuAD v2 | Anah-v2: 14.99 -> 12.53 | Improved |
| increase probe output from 1 to 5 co-occurring entities per entity (ablation) | LLMInertia · CNN/DailyMail | AlignScore: 86.73% -> 88.24% | Improved |
| increase probe output from 3 to 5 entities (ablation) | LLMInertia · Medical induction (25%) | accuracy: peaks at 3, declines slightly at 5 | Hurt |
| increase proportion of co-occurrence data in continued pretraining | LLaMA-3-8B (medical) · Contradiction test | hallucination induction rate: 15.71% -> 44.56% | Hurt |
| increase proportion of co-occurrence data in continued pretraining | LLaMA-3-8B (medical) · Contradiction test | accuracy: decreases by 9.38% to 37.70% | Hurt |
Closest papers by meaning
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.90)
- More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models (NeurIPS 2025 · similarity 0.89)
- Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.89)
- Mitigating Hallucinations in Large Language Models via Causal Reasoning (AAAI 2026 · similarity 0.89)
- LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations (ICLR 2025 · similarity 0.89)
- InEx: Hallucination Mitigation via Introspection and Cross-Modal Multi-Agent Collaboration (AAAI 2026 · 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.
