Decompose, Look, and Reason: Reinforced Latent Reasoning for VLMs
This addresses the challenge of complex visual reasoning for VLMs, offering a novel method with superior interpretability, though it is incremental in improving existing reasoning approaches.
The paper tackles the problem of visual information loss in textual chain-of-thought reasoning for Vision-Language Models by proposing DLR, a reinforced latent reasoning framework that dynamically decomposes queries and extracts visual latents, resulting in consistent outperformance of strong baselines on vision-centric benchmarks.
Vision-Language Models often struggle with complex visual reasoning due to the visual information loss in textual CoT. Existing methods either add the cost of tool calls or rely on localized patch-based embeddings that are insufficient to extract semantics in multi-step reasoning. We propose \emph{"Decompose, Look, and Reason" (DLR)}, a reinforced latent reasoning framework that dynamically decomposes queries into textual premises, extracts premise-conditioned continuous visual latents, and deduces answers through grounded rationales. We introduce a three-stage training pipeline and propose a novel Spherical Gaussian Latent Policy to enable effective exploration in the latent space. Extensive experiments on vision-centric benchmarks show that DLR consistently outperforms strong baselines, including text-only, interleaved multimodal CoT, and latent reasoning methods, while providing superior stepwise interpretability.