Research map / Large language models / Retrieval and knowledge
TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning
Zhepei Wei, Xiao Yang, Kai Sun, Jiaqi Wang, Rulin Shao, Jingxiang Chen, Mohammad Kachuee, Teja Gollapudi, Yiwei Liao, Nicolas SCHEFFER, Rakesh Wanga, Anuj Kumar and 3 more
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 |
|---|---|---|---|
| use ternary reward instead of binary reward (add abstention-neutral reward) (ablation) | TruthRL · average of 4 benchmarks, Llama3.1-8B | truthfulness: 4.5 -> 25.6 | Improved |
| use ternary reward instead of binary reward (ablation) | TruthRL · average of 4 benchmarks, Llama3.1-8B | hallucination rate: 47.7 -> 18.8 | Improved |
| use ternary reward instead of binary reward (ablation) | TruthRL · average of 4 benchmarks, Llama3.1-8B | accuracy: 52.2 -> 44.4 | Hurt |
| add knowledge-enhanced reward to binary reward (ablation) | TruthRL (binary) · average of 4 benchmarks | truthfulness: 4.5 -> 11.7 | Improved |
| add knowledge-enhanced reward to ternary reward (ablation) | TruthRL · average of 4 benchmarks | truthfulness: 25.6 -> 23.2 | No clear effect |
| use neutral abstention reward (0) instead of negative (-0.5) (ablation) | TruthRL · CRAG, Llama3.1-8B | truthfulness: 30.8 -> 37.2 | Improved |
| use neutral abstention reward (0) instead of positive (0.5) (ablation) | TruthRL · CRAG, Llama3.1-8B | truthfulness: 34.7 -> 37.2 | Improved |
| use model-based verifier instead of rule-based verifier (ablation) | TruthRL · CRAG, Llama3.1-8B | truthfulness: -3.6 -> 37.2 | Improved |
| train with online GRPO instead of offline DPO (ablation) | TruthRL · average of 4 benchmarks, Llama3.1-8B | truthfulness: -10.1 -> 25.6 | Improved |
| run iterative DPO for 4 iterations instead of 3 (ablation) | Iterative DPO · average of 4 benchmarks | truthfulness: 12.6 -> -2.0 | Hurt |
| add additive reasoning reward to outcome reward (ablation) | TruthRL · CRAG, Llama3.1-8B | truthfulness: 37.2 -> 36.1 | Mixed |
| apply TruthRL to Llama3.2-3B-Instruct | TruthRL · CRAG (with retrieval) | truthfulness: 1.9 -> 27.4 | Improved |
Closest papers by meaning
- Learning to Reason for Factuality (ICML 2026 · similarity 0.90)
- Reasoning Models Hallucinate More: Factuality-Aware Reinforcement Learning for Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
- FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees (ICML 2025 · similarity 0.89)
- LongRLVR: Long-Context Reinforcement Learning Requires Verifiable Context Rewards (ICLR 2026 · similarity 0.89)
- TTRL: Test-Time Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Mitigating Factual Hallucination in Large Reasoning Models via Mixed-Mode Advantage Regularization (NeurIPS 2026 · similarity 0.88)
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.
