SPLIT-RL Teaches VLMs to See and Reason Separately: +1.4–6.1 pp Over GRPO

The authors of arXiv preprint 2610.10889 propose SPLIT-RL, a post-training scheme that first teaches a vision-language model to extract information from an image correctly, and only then to derive an answer from it. Average accuracy rose by 1.4–6.1 points over standard GRPO. The authors tested Qwen3-VL models from 2B to 30B-A3B and InternVL3.5-8B. Nothing changes at inference: it is a single chain-of-thought call, just like with a regular GRPO model.

One reward for two skills

VL reasoning has two parts. The visual part (VR) is where the model pulls the necessary, precise details out of the image. The language part (LR) is where it builds an answer from those details. Reinforcement learning with verifiable rewards (RLVR) usually pushes both parts through a single reasoning chain, with a reward only for the final answer. Every CoT token receives the same sequence-level advantage. If the model misreads a value on a chart but the logic is flawless, the correct steps get penalized too. If perception was right but the conclusion was wrong, perception takes the penalty as well. The signal does not show which capability actually failed.

Two phases and claim-level advantage

SPLIT-RL splits training into non-overlapping phases: VR on its own, then LR on its own, each with its own reward. The logic builds on how GRPO works, where rollouts within a group are compared against each other. If, within a phase, rollouts differ in only one capability, group-relative advantage isolates exactly that capability. The abstract does not give the specific reward formulas.

For the perception phase, the authors added Claim-Level Advantage (CLA-GRPO). Rollouts are decomposed into atomic visual claims such as "the image shows three red objects," and advantage is computed for each claim separately. Groups are formed by the visual type of the claim. Instead of one scalar for the whole text, the model gets a detailed signal about what exactly it saw correctly or incorrectly.

Key parameters of the work:

• Method: two-phase post-training (VR, then LR) with a separate reward for each phase.
• New algorithm for the VR phase: CLA-GRPO with advantage at the level of visual claims.
• Gain: +1.4–6.1 points of average accuracy over GRPO.
• Models: Qwen3-VL (2B–30B-A3B) and InternVL3.5-8B.
• Inference: a single CoT call, with the same architecture and evaluation protocol as a GRPO model.

Diagnosis: perception does not move under GRPO

The most interesting part of the paper is not the headline gain but the diagnosis. The authors evaluated each capability separately using an oracle approach. Their conclusion: GRPO with an answer-only reward leaves perception unchanged, while SPLIT-RL improves both VR and LR. It follows that a noticeable share of the "reasoning progress" under standard RLVR actually comes down to the language part, while the visual part stays at the base model's level.

The limitations are also visible. The 4.7-point spread between the smallest and largest gain means the effect depends heavily on the model, and the abstract does not say where the effect is largest. Training also gets more complicated: it needs two phases and a mechanism for splitting answers into verifiable visual claims. The abstract does not estimate the cost of this.

What to take away

Teams that fine-tune VLMs with RLVR should add separate evaluation of perception and reasoning to their pipeline, before and after training. If benchmark accuracy rises while perception stays flat, the model is probably just getting better at deriving answers from the same inaccurate reading of the image. This matters most for tasks involving documents, charts, and tables, where a misread number invalidates any amount of logic. The extra cost of SPLIT-RL falls only on training; serving stays the same. Before adopting it, wait for the full benchmark tables and the code, then test the method on your own data.