Vision–language reasoning · Reinforcement learning

SPLIT-RL

Staged Perception–Language Reasoning Training with Claim-Level Advantages

See accurately. Reason from evidence. Train each capability in turn.

Raja Kumar^1,*Rajat Koner²Ritwick Chaudhry²Zhuowei Li²Nishant Sankaran²Yifan Xing²

¹ University of Southern California ² Amazon AGI

* Work done during an internship at Amazon AGI.

arXiv

+1.4–6.1accuracy points over GRPO

2 stagesone shared policy

1 callat inference

01 / The training scaffold

Two stages. Different credit.

The description is the bottleneck. The reasoner cannot see the image.

View the full paper diagram ↗

The full SPLIT-RL pipeline: Stage 1 updates description tokens with recall, answer, and claim-level signals. Stage 2 updates reasoning and answer tokens without the image.

02 / Claim-level advantage

Put credit where the claim is.

Same image. Three rollouts. Compare claims within each visual type.

Object

metal circle✓

pulley device✓

“a pulley device”0

Text label

“460”✓

“480”✕

“460”✓

“the number 460”+

Spatial

circle in middle✓

not stated—

circle at the left✕

“circle at the left”−

0 no disagreement+ correct claim− incorrect claim

CLA-GRPO → type-relative advantages on rollout 3’s claim tokens.

Schematic credit assignment; claim-level credit is added to the rollout reward advantage.

03 / Six benchmarks · Five backbones

Stronger across scales.

Standard single-call inference. Same prompt and decoding as GRPO.

Backbone

GRPOSPLIT-RL

Accuracy (%) · Right: change over GRPO in percentage points.

Qwen3-VL-2B · six-benchmark mean48.0 → 54.1 (+6.1 pts)

All benchmark results ↓

Accuracy (%) · AVG: six-benchmark mean.
Backbone	Method	MathVista	MathVision	MathVerse	WeMath	LogicVista	DynaMath	AVG
Qwen3-VL-2B	Base	60.3	18.8	38.8	56.5	34.2	55.9	44.1
	GRPO	60.6	35.9	44.4	54.1	36.7	56.5	48.0
	SPLIT-RL	65.4	42.8	46.7	62.1	44.1	63.8	54.1
Qwen3-VL-4B	Base	77.1	59.2	61.7	75.6	63.1	73.4	68.4
	GRPO	78.0	61.8	66.4	79.8	56.4	74.7	69.5
	SPLIT-RL	78.7	68.4	69.3	84.0	62.6	76.1	73.2
Qwen3-VL-8B	Base	77.2	63.2	63.8	79.9	60.4	74.4	69.8
	GRPO	77.9	64.5	67.8	84.6	65.6	75.3	72.6
	VLM-CapCurr	80.0	44.7	67.6	81.2	60.9	74.6	68.2
	PEARL	77.9	58.2	68.9	84.8	61.5	72.8	70.7
	SPLIT-RL	79.8	69.4	69.8	87.4	66.0	79.1	75.3
Qwen3-VL-30B-A3B	Base	79.6	59.2	65.7	75.5	60.4	78.1	69.8
	GRPO	80.8	64.9	74.7	76.8	64.1	75.1	72.7
	SPLIT-RL	80.0	66.4	73.0	80.3	64.2	80.9	74.1
InternVL3.5-8B	Base	79.2	55.9	60.9	83.9	57.3	70.5	68.0
	GRPO	79.4	62.2	63.2	81.6	57.3	73.6	69.5
	SPLIT-RL	80.6	63.5	64.0	84.6	60.1	77.1	71.6

Improve both capabilities.

Qwen3-VL-8B · Oracle-isolated gains over the base model

Other role fixed to Kimi-K3 · Each axis: mean accuracy gain in points.

04 / Qualitative examples

One detail changes the answer.

GRPO misreads A as (5,8) and reaches the token limit. SPLIT-RL reads A as (6,8), rotates to (6,3), and chooses the correct segment 5. — GRPO: A = (5, 8), no answer. SPLIT-RL: A = (6, 8), correct answer.

Is the food half eaten? GRPO says mostly intact and answers No. SPLIT-RL identifies that half remains and correctly answers Yes. — GRPO: “mostly intact” → No. SPLIT-RL: “only half remains” → Yes.

Citation

BibTeX

@misc{kumar2026splitrl,
  title = {{SPLIT-RL}: Staged Perception-Language Reasoning Training with Claim-Level Advantages},
  author = {Raja Kumar and Rajat Koner and Ritwick Chaudhry and Zhuowei Li and Nishant Sankaran and Yifan Xing},
  year = {2026},
  eprint = {2610.10889},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  doi = {10.48550/arXiv.2610.10889},
  url = {https://arxiv.org/abs/2610.10889}
}