SLVR
A research paper proposing a Structured Latent Visual Reasoning framework. It introduces typed latent reasoning stages—planning, localization, evidence selection, and reasoning integration—combined with a two-stage training strategy to enhance the visual reasoning capabilities of multimodal large language models.
Timeline (last 90 days)
SLVR (Structured Latent Visual Reasoning) organizes multimodal reasoning into several typed latent stages corresponding to four steps: planning, localization, evidence selection, and reasoning integration.
SLVR adopts a two-stage training strategy: the first step masks text that leaks the answer and uses visual distractors to force the model to rely on the image, and the second step applies structured stage-wise supervision.
SLVR's localization stage uses bounding-box supervision, evidence selection corresponds to visual evidence, integration corresponds to the final reasoning rationale, and planning corresponds to the plan.
SLVR is built on top of Qwen2.5-VL-7B.
SLVR achieves an absolute improvement of +9.4 on the MMVP benchmark.
SLVR achieves an absolute improvement of +14.2 on the BLINK Relation benchmark.
SLVR also shows varying degrees of improvement on benchmarks such as V*, MathVista, and ChartQA.
SLVR differs from Coconut (continuous chain-of-thought), which freely iterates in the hidden space without imposing stage divisions, whereas SLVR explicitly requires each latent state to align with a specific semantic function.
SLVR is a new research preprint from arXiv.
The researchers believe that structured latent supervision can improve fine-grained visual reasoning capabilities while avoiding the decoding overhead introduced by textual chain-of-thought.
All Facts (10)
SLVR (Structured Latent Visual Reasoning) organizes multimodal reasoning into several typed latent stages corresponding to four steps: planning, localization, evidence selection, and reasoning integration.
50%UnverifiedSLVR adopts a two-stage training strategy: the first step masks text that leaks the answer and uses visual distractors to force the model to rely on the image, and the second step applies structured stage-wise supervision.
50%UnverifiedSLVR's localization stage uses bounding-box supervision, evidence selection corresponds to visual evidence, integration corresponds to the final reasoning rationale, and planning corresponds to the plan.
50%UnverifiedSLVR is built on top of Qwen2.5-VL-7B.
50%UnverifiedSLVR achieves an absolute improvement of +9.4 on the MMVP benchmark.
50%UnverifiedSLVR achieves an absolute improvement of +14.2 on the BLINK Relation benchmark.
50%UnverifiedSLVR also shows varying degrees of improvement on benchmarks such as V*, MathVista, and ChartQA.
50%UnverifiedSLVR differs from Coconut (continuous chain-of-thought), which freely iterates in the hidden space without imposing stage divisions, whereas SLVR explicitly requires each latent state to align with a specific semantic function.
50%UnverifiedSLVR is a new research preprint from arXiv.
50%UnverifiedThe researchers believe that structured latent supervision can improve fine-grained visual reasoning capabilities while avoiding the decoding overhead introduced by textual chain-of-thought.
50%