Research map / Large language models / Retrieval and knowledge
Train for Truth, Keep the Skills: Binary Retrieval-Augmented Reward Mitigates Hallucinations
Tong Chen, Akari Asai, Luke Zettlemoyer, Hannaneh Hajishirzi, Faeze Brahman
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 |
|---|---|---|---|
| train with binary retrieval-augmented reward instead of continuous VeriScore reward (ablation) | Qwen3-8B GRPO · ALPACAEVAL (length-controlled) | win rate: 42.2 -> 53.9 | Improved |
| use binary RAR reward instead of VeriScore reward (ablation) | Qwen3-8B GRPO · Short-form average (POPQA, GPQA) | hallucination rate: 42.3 -> 27.6 | Improved |
| online RL with Binary RAR on top of base model | Qwen3-8B · Long-form average | hallucination rate: 61.9 -> 37.5 | Improved |
| add SFT on most factual response | Qwen3-8B · Long-form average | hallucination rate: 61.9 -> 60.9 | No clear effect |
| add DPO on factuality preference pairs | Qwen3-8B · Long-form average | hallucination rate: 61.9 -> 53.4 | Improved |
| use general LM-judge reward for RL | Qwen3-8B · Long-form average | hallucination rate: 61.9 -> 65.4 | Hurt |
| use Binary RAR with Qwen3-4B base | Qwen3-4B · Short-form average (POPQA, GPQA) | hallucination rate: 68.7 -> 41.9 | Improved |
| apply Binary RAR to a different model family | Tulu3-8B · BIOGRAPHY | hallucination rate: 68.1 -> 60.1 | Improved |
| apply Binary RAR to Tulu3-8B | Tulu3-8B · ALPACAEVAL | win rate: 24.3 -> 23.4 | No clear effect |
| evaluate POPQA and GPQA with no abstention after Binary RAR | Qwen3-8B · POPQA, GPQA (no abstention) | accuracy: POPQA 20.2 -> 20.6; GPQA 48.2 -> 48.8 | No clear effect |
| threshold VeriScore at 0.5 as a binary reward (claim-level) instead of response-level binary RAR (ablation) | Qwen3-8B RL · ALPACAEVAL vs BIOGRAPHY FactScore | hallucination-utility trade-off | Hurt |
| use conflict-only claim reward (fraction of non-contradictory claims) instead of binary RAR (ablation) | Qwen3-8B RL · ALPACAEVAL | win rate | Hurt |
Closest papers by meaning
- BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (NeurIPS 2026 · similarity 0.93)
- Token-Guard: Towards Token-Level Hallucination Control via Self-Checking Decoding (ICLR 2026 · similarity 0.92)
- Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.91)
- Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling (NeurIPS 2025 · similarity 0.91)
- The Mirage of Performance Gains: Why Contrastive Decoding Fails to Mitigate Object Hallucinations in MLLMs? (NeurIPS 2025 · similarity 0.90)
- ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability (ICLR 2025 · similarity 0.90)
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.
