VICO
A co-evolutionary post-training framework for vision-language models. It introduces an EnvRewriter component that dynamically generates training samples with calibrated difficulty, enabling the visual environment and model capabilities to evolve together without requiring additional human annotation.
Timeline (last 90 days)
The VICO paper points out that when model capability improves, fixed training tasks deviate from its learning frontier, causing the learning signal to collapse.
VICO (Visual Environments Co-Evolving) is a post-training framework that enables visual environments and vision-language models to co-evolve.
VICO introduces a component called the EnvRewriter, which is jointly trained with the actor.
VICO uses a pass-rate-based reward to calibrate the difficulty of generated tasks, keeping tasks at the model's learning frontier.
VICO first edits structured intermediate representations and then renders them into images, thereby generating valid training samples without relying on manual annotation.
VICO was validated on nine multimodal benchmarks, covering two categories of tasks: mathematical reasoning and visual grounding understanding.
VICO achieves performance comparable to RLVR methods specifically optimized for charts, while using 16 to 160 times fewer annotated samples.
VICO shifts VLM post-training from supervised manual annotation toward image-editing co-evolution, providing a scalable path beyond static-corpus RLVR.
VICO's co-evolution can be seen as a fusion of curriculum learning and self-play ideas in the vision-language modality, with the EnvRewriter playing the role of a teacher that dynamically generates questions based on the student's capability.
All Facts (9)
The VICO paper points out that when model capability improves, fixed training tasks deviate from its learning frontier, causing the learning signal to collapse.
50%UnverifiedVICO (Visual Environments Co-Evolving) is a post-training framework that enables visual environments and vision-language models to co-evolve.
50%UnverifiedVICO introduces a component called the EnvRewriter, which is jointly trained with the actor.
50%UnverifiedVICO uses a pass-rate-based reward to calibrate the difficulty of generated tasks, keeping tasks at the model's learning frontier.
50%UnverifiedVICO first edits structured intermediate representations and then renders them into images, thereby generating valid training samples without relying on manual annotation.
50%UnverifiedVICO was validated on nine multimodal benchmarks, covering two categories of tasks: mathematical reasoning and visual grounding understanding.
50%UnverifiedVICO achieves performance comparable to RLVR methods specifically optimized for charts, while using 16 to 160 times fewer annotated samples.
50%UnverifiedVICO shifts VLM post-training from supervised manual annotation toward image-editing co-evolution, providing a scalable path beyond static-corpus RLVR.
50%UnverifiedVICO's co-evolution can be seen as a fusion of curriculum learning and self-play ideas in the vision-language modality, with the EnvRewriter playing the role of a teacher that dynamically generates questions based on the student's capability.
50%