AI Paper Atlas

Research map / Large language models / Alignment and preferences

RLSF-V: Mitigating Hallucinations in MLLMs via Fuzzy Semantic Self-Feedback

Changhao He, Shuhao Yan, Shuxian Li, Xi Peng, Peng Hu

ICML 2026 · Large language models › Alignment and preferences · approach: Robust and calibrated DPO

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 changedWhereMeasuredEffect
Train with RLSF-V self-feedback preference data instead of base modelLLaVA-v1.5-7B · AMBER GenerativeHalRate: 35.9 -> 5.1Improved
Apply RLSF-V to stronger backbonesInternVL3.5-8B-HF · AMBER DiscriminativeAcc/F1: 86.7 -> 86.7 acc; 89.4 -> 89.4 F1No clear effect
Apply RLSF-V to Qwen2.5-VL-3BQwen2.5-VL-3B-Instruct · AMBER GenerativeHalRate: 49.1 -> 36.5Improved
Add POS-aware token selection (ablation)RLSF-V · AMBER DiscriminativeF1: 62.4 -> 87.6 (LLaVA-v1.5-13B)Improved
Add top-R token pooling (ablation)RLSF-V · AMBER GenerativeHalRate: 50.3 -> 5.1 (LLaVA-v1.5-7B)Improved
Use fuzzy semantics instead of energy score (ablation)RLSF-V · AMBER GenerativeHalRate: 46.7 -> 5.1Improved
Use fuzzy semantics instead of Shannon entropy (ablation)RLSF-V · AMBER GenerativeHalRate: 5.8 -> 5.1 (7B); 5.8 -> 4.4 (13B)Improved
Vary top-R pooling size from 10 to 100 (ablation)RLSF-V · AMBER and HallusionBenchhallucination metricsMixed
Apply RLSF-V to Qwen3-VL-8B for general benchmarksQwen3-VL-8B-Instruct · 10 general benchmarksaccuracy: MMStar 62.88 -> 63.78; OCRBench 83.00 -> 82.40Mixed
RLSF-V on Relation-type hallucinationsLLaVA-v1.5-13B · AMBER Discriminative per-type F1F1Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy