Get Started
Home
Topics
Search
Library
Research questionHow should vision-language model post-training balance perception and reasoning under SFT versus reinforcement learning?Post-training can strengthen a vision-language model’s reasoning more than its ability to extract accurate visual information. This imbalance limits end-to-end visual reasoning, and its cause differs between supervised fine-tuning and reinforcement learning.
AI
Computer Vision
LLM Pretraining & Post-training
Machine Learning
Multimodal Models
Reasoning
Reinforcement Learning
Latest papersRecent research connected to this question, newest first.On Asymmetric Optimization of Reasoning and Perception in Vision-Language Model Post-TrainingEvidence comes from two synthetic tasks designed to disentangle perception from reasoning and three real-world visual-reasoning benchmarks. The analysis attributes the imbalance in supervised fine-tuning to token imbalance and in reinforcement learning to reward coupling, examining reweighted losses, perception-aware rewards, and a surrogate reward when ground-truth perception rewards are unavailable.research paper · Sep 2, 2026
Related questions
Should reasoning LLMs combine on-policy distillation and verifiable-reward RL jointly or sequentially during post-training?How can reasoning-model fine-tuning improve pass@1 without reducing pass@k coverage?How can reinforcement learning post-training prioritize useful reasoning prompts as learning signals shift?How can we diagnose vision-language-action models’ failures on spatially ambiguous, long-horizon manipulation tasks?