AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

HyLaR: Hybrid Latent Reasoning with Decoupled Policy Optimization

Tao Cheng, Shi-Zhe Chen, Hao Zhang, Yixin Qin, Jinwen Luo, Zheng Wei

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 5 citations (OpenAlex)

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 latent visual reasoning (vs text-only SFT) (ablation)HyLaR-SFT · V*accuracy (%): 69.11 -> 80.63Improved
add latent visual reasoning (vs text-only SFT) (ablation)HyLaR-SFT · HRBench-8Kaccuracy (%): 61.00 -> 67.19Improved
add DePO RL stage over SFTHyLaR-SFT · V*accuracy (%): 80.63 -> 83.77Improved
use DePO instead of GRPO (ablation)HyLaR RL · V*accuracy (%): 78.53 -> 83.77Improved
use GRPO RL on top of SFT (ablation)HyLaR-SFT · V*accuracy (%): 80.63 -> 78.53Hurt
use DAPO RL on top of SFT (ablation)HyLaR-SFT · V*accuracy (%): 80.63 -> 79.06Hurt
use VLPO RL on top of SFT (ablation)HyLaR-SFT · average over 4 benchmarksaccuracy (%): 72.58 -> 72.65No clear effect
add latent loss weight alpha=0.5 (vs 0) (ablation)HyLaR-7B · V*accuracy (%): 80.63 -> 83.77Improved
raise latent loss weight to 2.0 (vs 0.5) (ablation)HyLaR-7B · V*accuracy (%): 83.77 -> 80.63Hurt
use latent clipping 0.05 instead of 0.2 shared with text (ablation)HyLaR-7B · V*accuracy (%): 80.10 -> 83.77Improved
use no latent clipping (0.0) vs 0.05 (ablation)HyLaR-7B · V*accuracy (%): 79.58 -> 83.77Improved
use vMF instead of Gaussian latent distribution (ablation)HyLaR RL · SeedBench2Plusaccuracy (%): 67.28 -> 70.32Improved

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