Research map / Large language models / Language, safety and interpretability
TruthFlow: Truthful LLM Generation via Representation Flow Correction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TruthFlow, which trains a flow-matching model to map each query's hidden state to a query-specific truthful correction vector, then projects it onto top singular directions and adds it to LLM activations at inference to improve TruthfulQA truthfulness.
Problem
Existing representation intervention methods for truthfulness apply one universal correction vector to all queries, which cannot fit diverse queries' hallucinated-to-truthful directions.
Approach
Use flow matching to learn a query-conditioned transport from query last-token hidden states to their truthful correction vectors (correct minus incorrect answer mean activations), then purify the generated vector by projecting onto the top-k SVD singular subspace of training corrections before intervening at one layer.
Key results
- On TruthfulQA with Llama3, TruthFlow reaches 64.79 True and 61.15 True*Info versus 52.32 and 47.97 for the base model and 54.52 and 49.32 for ITI.
- TruthFlow improves True*Info over the base model across all six tested LLMs on open-ended TruthfulQA generation, e.g. Llama2-7B 44.56 to 54.91 and Gemma2 58.33 to 73.35.
- Transfer to unseen HaluEval, NQ and TriviaQA with Llama3 keeps or improves True and True*Info, whereas ITI degrades them substantially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply query-specific TruthFlow correction instead of the base model | Llama3 · TruthfulQA open-ended | True (%) | Helped |
| apply TruthFlow correction instead of the base model | LLMs tested · TruthfulQA open-ended | True*Info (%) | Helped |
| apply TruthFlow on unseen datasets | Llama3 · HaluEval, NQ, and TriviaQA | True and True*Info | Helped |
| add truthful subspace projection to query-specific correction (ablation) | TruthFlow on Gemma2 · TruthfulQA | True (%): 70.17 -> 76.52 | Helped |
| add truthful subspace projection to query-specific correction (ablation) | TruthFlow on Gemma2 · TruthfulQA | Info (%): 92.18 -> 95.84 | Helped |
| add truthful subspace projection to query-specific correction (ablation) | TruthFlow on Gemma2 · TruthfulQA | True*Info (%): 64.68 -> 73.35 | Helped |
| add truthful subspace projection to query-specific correction (ablation) | TruthFlow on Llama3 · TruthfulQA | True (%): 53.55 -> 64.79 | Helped |
| add truthful subspace projection to query-specific correction (ablation) | TruthFlow on Llama3 · TruthfulQA | Info (%): 89.98 -> 94.38 | Helped |
| add truthful subspace projection to query-specific correction (ablation) | TruthFlow on Llama3 · TruthfulQA | True*Info (%): 48.18 -> 61.15 | Helped |
| apply TruthFlow at layer 12 rather than layer 11 (ablation) | Llama3 · TruthfulQA | True: 62.10 -> 64.79 | Helped |
| apply TruthFlow at layer 12 rather than layer 13, 14, or 15 (ablation) | Llama3 · TruthfulQA | True: 60.15 -> 64.79; 56.23 -> 64.79; 53.30 -> 64.79 | Helped |
Surprising findings
- Applying the flow-generated correction without SVD projection gives a smaller gain on Llama3 (True*Info 48.18 vs base 47.97 in the ablation), so the projection step carries most of the improvement there.
- ITI hurts open-ended generation on transfer benchmarks (e.g. HaluEval True 25.76 vs base 36.74).
Limitations
- Gains on multiple-choice MC1/MC2 are mixed; some baselines and the base model remain competitive on MC2 for Mistral models.
- Info score sometimes drops because truthful answers such as 'I have no comment' are judged uninformative.
- Intervention strength alpha, layer, and k are tuned per model, with different alpha used for ablations, which may limit fair comparison.
Experimental setup
Datasets: TruthfulQA HaluEval Natural Questions TriviaQA
Benchmarks: TruthfulQA HaluEval Natural Questions TriviaQA
Metrics: BLEURT True score Info score True*Info MC1 MC2
Baselines: DoLa Activation Decoding ITI NL-ITI TruthX
Builds on: ITI rectified flow flow matching NL-ITI DoLa
Closest papers by meaning
- Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning (AAAI 2024 · similarity 0.88)
- Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension (ICML 2024 · similarity 0.87)
- Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators (AAAI 2025 · similarity 0.86)
- HonestLLM: Toward an Honest and Helpful Large Language Model (NeurIPS 2024 · similarity 0.86)
- DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models (ICLR 2024 · similarity 0.86)
- LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations (ICLR 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.