SPLIT-RL
A staged training method for visual-language reasoning that separates visual reasoning and language reasoning into distinct training phases and introduces a Claim-Level Advantage mechanism (CLA-GRPO). This approach addresses the problem of ambiguous attribution of multimodal capabilities during reinforcement learning training of vision-language models.
Timeline (last 90 days)
The arXiv paper "SPLIT-RL: Staged Perception-Language Reasoning Training with Claim-Level Advantages" proposes a solution to the attribution difficulty in visual-language reasoning capabilities.
The core idea of SPLIT-RL is to separate the training of visual reasoning and language reasoning into two non-overlapping stages.
Each stage of SPLIT-RL is optimized using a phase-specific reward.
The policy trained by SPLIT-RL is used during inference in the same way as an ordinary GRPO model, requiring only a single CoT call and adding no inference cost.
SPLIT-RL proposes Claim-Level Advantage (CLA-GRPO), which decomposes rollouts from the visual reasoning stage into atomic visual claims and provides fine-grained advantage signals for each claim individually.
On the Qwen3-VL series of models (from 2B to 30B-A3B) and on InternVL3.5-8B, SPLIT-RL improves average accuracy by 1.4 to 6.1 percentage points compared to standard GRPO.
SPLIT-RL brings substantial improvements in both the visual reasoning and language reasoning dimensions.
All Facts (7)
The policy trained by SPLIT-RL is used during inference in the same way as an ordinary GRPO model, requiring only a single CoT call and adding no inference cost.
50%UnverifiedThe core idea of SPLIT-RL is to separate the training of visual reasoning and language reasoning into two non-overlapping stages.
50%UnverifiedEach stage of SPLIT-RL is optimized using a phase-specific reward.
50%UnverifiedSPLIT-RL brings substantial improvements in both the visual reasoning and language reasoning dimensions.
50%UnverifiedSPLIT-RL proposes Claim-Level Advantage (CLA-GRPO), which decomposes rollouts from the visual reasoning stage into atomic visual claims and provides fine-grained advantage signals for each claim individually.
50%UnverifiedOn the Qwen3-VL series of models (from 2B to 30B-A3B) and on InternVL3.5-8B, SPLIT-RL improves average accuracy by 1.4 to 6.1 percentage points compared to standard GRPO.
50%UnverifiedThe arXiv paper "SPLIT-RL: Staged Perception-Language Reasoning Training with Claim-Level Advantages" proposes a solution to the attribution difficulty in visual-language reasoning capabilities.
50%