SLVR Adds 14.2 Points on BLINK Relation to Qwen2.5-VL-7B Without Text CoT

The authors of SLVR show how to make a multimodal model rely on the image instead of guessing the answer from the text of the question, without paying for it with a long reasoning chain at output. On top of Qwen2.5-VL-7B, the method gained +9.4 points on MMVP and +14.2 on BLINK Relation. Improvements were also recorded on V*, MathVista and ChartQA. The model generates no text reasoning when it answers.

Four latent stages instead of a long CoT

The problem is a familiar one. Multimodal LLMs often answer visual questions from linguistic priors rather than from what is actually in the picture. Text chain-of-thought partly fixes this: the model breaks the question into evidence-seeking steps. But those steps are generated autoregressively, and inference cost grows with the length of the chain. Latent reasoning is cheaper because there is no explicit text. The trade-off is that it is unclear what the intermediate states actually encode, and there is no way to attach a separate training signal for planning, region grounding or evidence selection.

SLVR tries to take the best of both approaches. Training runs in two stages. First, the model is taught to look at the image: fragments of the text that give away the answer are masked, and the correct answer is contrasted with visually plausible distractors. Then the reasoning is decomposed into typed latent stages: planning, grounding, visual evidence selection and integration. Each stage gets its own supervision: plans, boxes, visual evidence and final rationales, respectively. The latent steps thus receive a functional structure, and inference does not require generating a text chain.

What the benchmarks show

The authors tested the method on a single base model. The reported results:

• base: Qwen2.5-VL-7B;
• MMVP: +9.4 points absolute;
• BLINK Relation: +14.2 points absolute;
• V*, MathVista, ChartQA: improvements, with no magnitude given in the abstract;
• inference: no text CoT decoding.

A few caveats. The abstract gives no absolute scores and does not say what the gains are measured against: the original model, ordinary fine-tuning, or a variant with text CoT. This matters a lot for interpretation. MMVP is small, so the gain on it may be noisy. There is also no data on real inference savings: the abstract states that there is no decoding overhead, but it reports no latency or token measurements. The work is still an arXiv preprint with no independent verification. A project page has been published.

Who this is useful for

The idea matters most to teams that fine-tune open VLMs for tasks requiring precise visual analysis: reading charts, spatial relations, finding small objects in an image. Where text CoT is too expensive in latency, a structured latent approach looks like a reasonable alternative.

The catch is data. Training requires plans, boxes, visual evidence and rationales, which means annotation that is much richer than ordinary "image–answer" pairs. Before moving the method into production, it is worth estimating how much of this annotation can be obtained automatically, and checking the gain on your own data rather than only on public benchmarks. Debugging is a separate question. Typed stages should in theory be easier to analyze than unstructured hidden states, but the abstract offers no confirmation of this. It is worth checking in the full paper and the code.