AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

DEVA: Fine-tuning Multimodal Large Language Models for Visual Perception Tasks

Debasmit Das, Munawar Hayat, Fatih Porikli

CVPR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

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
add GFlowNet diversity loss to GRPO (ablation)GRPO · LISAavg reward std: 0.234 -> 0.262Improved
add GFlowNet diversity loss to DAPO (ablation)DAPO · LISAavg reward std: 0.201 -> 0.240Improved
add GFlowNet diversity loss to BNPO (ablation)BNPO · LISAavg reward std: 0.193 -> 0.221Improved
add diversity loss to ViRFT (ablation)ViRFT (GRPO) · LISA testmIoU (Qwen2-VL-2B): 37.6 -> 43.6Improved
add global entropic exploration regularization (ablation)ViRFT + diversity loss · LISA testmIoU (Qwen2-VL-2B): 43.6 -> 44.7Improved
add alignment volume reward (ablation)ViRFT + diversity + exploration · LISA testmIoU (Qwen2-VL-2B): 44.7 -> 46.7Improved
add harmonic reward aggregation (ablation)ViRFT + diversity + exploration + volume · LISA testmIoU (Qwen2-VL-2B): 46.7 -> 48.9Improved
add all DEVA components to ViRFTViRFT · COCO 4-shot detectionmAP (Qwen2-VL-2B): 40.6 -> 47.3Improved
add all DEVA components to ViRFTViRFT · COCO 4-shot detectionmAP (Qwen2-VL-7B): 54.3 -> 60.0Improved
add diversity loss to ViRFT (ablation)ViRFT · COCO 4-shot detectionmAP (Qwen2-VL-2B): 40.6 -> 43.9Improved
use CoT-style data in SFT (ablation)SFT · COCO 4-shot detectionmAP (Qwen2-VL-2B): 25.2 -> 29.7Improved
replace pairwise alignment reward with alignment volume reward (ablation)DEVA · LISA testmIoUImproved

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