Research map / Large language models / Reasoning and chain of thought
Inference-Time Intervention: Eliciting Truthful Answers from a Language Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 ITI to LLaMA-7B | LLaMA-7B · TruthfulQA | True*Info (%): 30.5 -> 43.5 | Helped |
| add ITI on top of few-shot prompting | Few-shot prompting LLaMA-7B · TruthfulQA | True*Info (%): 49.5 -> 51.4 | Helped |
| add ITI to Alpaca | Alpaca · TruthfulQA | True*Info (%): 32.5 -> 65.1 | Helped |
| add ITI to Vicuna | Vicuna · TruthfulQA | True*Info (%): 51.5 -> 74.0 | Helped |
| add ITI to LLaMA-7B (zero-shot transfer) | LLaMA-7B · MMLU | accuracy: 35.71 -> 40.16 | Helped |
| add ITI to LLaMA-7B (zero-shot transfer) | LLaMA-7B · Natural Questions | MC accuracy: 46.6 -> 51.3 | Helped |
| add ITI to LLaMA-7B (zero-shot transfer) | LLaMA-7B · TriviaQA | MC accuracy: 89.6 -> 91.1 | Helped |
| use mass mean shift direction instead of probe weight direction (ablation) | ITI · TruthfulQA (2-fold CV) | True*Info (%): 34.8 -> 42.3 | Helped |
| use probe-derived truthful direction instead of CCS direction (ablation) | ITI · TruthfulQA (2-fold CV) | True*Info (%): 33.4 -> 34.8 | Helped |
| use mass mean shift instead of random direction (ablation) | ITI · TruthfulQA (2-fold CV) | True*Info (%): 31.2 -> 42.3 | Helped |
| restrict intervention to top-K heads instead of all heads (ablation) | ITI · TruthfulQA (2-fold CV) | True*Info (%): 35.4 -> 42.3 | Helped |
| head-wise selection instead of point-wise selection (ablation) | ITI · TruthfulQA (2-fold CV) | True*Info (%): 39.2 -> 42.3 True*Info; True 55.3 vs 45.1; KL 1.95 -> 0.27 | Mixed |
Closest papers by meaning
- Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning (AAAI 2024 · similarity 0.89)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- GRATH: Gradual Self-Truthifying for Large Language Models (ICML 2024 · similarity 0.88)
- An Architecture Search Framework for Inference-Time Techniques (ICML 2025 · similarity 0.87)
- DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models (ICLR 2024 · similarity 0.87)
- Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.87)
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.