Research map / Large language models / Language, safety and interpretability
Hallucination Reduction with CASAL: Contrastive Activation Steering for Amortized Learning
Wannan Yang, Xinchi Qiu, Lei Yu, Yuchen Zhang, Aobo Yang, Narine Kokhlikyan, Nicola Cancedda, Diego Garcia-Olano
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 CASAL (amortized steering) instead of SFT/DPO/GRPO | Llama-3.1-8B-Instruct · TriviaQA to EntityQA (OOD) unknown queries | hallucination rate: 11.72% vs SFT 23.13%, DPO 18.30%, GRPO 12.83% | Improved |
| train CASAL on Wikipedia-style TriviaQA, test on web | Llama-3.1-8B-Instruct · TriviaQA (Wiki to Web) | hallucination rate: 50.74% -> 32.42% | Improved |
| train CASAL on PopQA Group 1, test Group 2 | Llama-3.1-8B-Instruct · PopQA | hallucination rate: 74.35% -> 23.42% | Improved |
| use CASAL instead of DPO/GRPO for known-query refusal | Llama-3.1-8B-Instruct · TriviaQA, EntityQA | refusal rate: 7.29% vs DPO 14.37%, GRPO 17.77%; 6.84% vs DPO 17.66% | Improved |
| CASAL vs baseline on PopQA known-query accuracy | Llama-3.1-8B-Instruct · PopQA | accuracy: 91.08% -> 85.11% | Hurt |
| apply CASAL to general capability benchmarks | Llama-3.1-8B-Instruct · MMLU, GSM8K, GPQA | accuracy: MMLU 68.01 -> 68.04; GSM8K 77.48 -> 77.02; GPQA 33.31 -> 33.18 | No clear effect |
| apply CASAL to MT-Bench coherence | Llama-3.1-8B-Instruct · MT-Bench | score: 7.38 -> 7.57 | Improved |
| train CASAL on 640 examples versus baselines on 12,800 | Llama-3.1-8B-Instruct · hallucination (TriviaQA) | hallucination rate: matches or surpasses SFT, DPO, GRPO trained on 12,800 examples (20x data efficiency) | Improved |
| use CASAL instead of LoRA training | Llama-3.1-8B-Instruct · compute cost | FLOPs per token: over 30x fewer FLOPs per token than LoRA | Improved |
| apply CASAL to vision-language model | Qwen2.5-VL-7B-Instruct · WorldCuisines-VQA | hallucination rate: 72.35% -> 33.34% | Improved |
| apply CASAL to Landmark-VQA | Qwen2.5-VL-7B-Instruct · Landmark-VQA | hallucination rate: 75.78% -> 31.25% | Improved |
| apply CASAL to MoE model | OLMoE · unknown queries | hallucination rate: drops by 42.9% | Improved |
Closest papers by meaning
- Decoupling Contrastive Decoding: Robust Hallucination Mitigation in Multimodal Large Language Models (NeurIPS 2025 · similarity 0.91)
- Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models (ICLR 2026 · similarity 0.91)
- ASCD: Attention-Steerable Contrastive Decoding for Reducing Hallucination in MLLM (AAAI 2026 · similarity 0.90)
- Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models (ICLR 2025 · similarity 0.90)
- Look Carefully: Adaptive Visual Reinforcements in Multimodal Large Language Models for Hallucination Mitigation (ICLR 2026 · similarity 0.90)
- Hallucination-aware Intermediate Representation Editing in Large Vision-Lanugage Models (ICLR 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.
