arXiv:2606.05753v1 Announce Type: new 
Abstract: Latent visual reasoning (LVR) inserts supervised latent tokens between perception and answer generation in vision-language models (VLMs). The field uses alignment between these latents and their visual targets, i.e., cosine similarity or mean squared error (MSE), as both the training loss and the quality metric, assuming that better alignment yields a better answer. We test this with a designed matrix of five LVR variants and find the assumption inverted: cosine alignment is negatively correlated with accuracy across all five (r=-0.94). To explain this, we introduce PRISM, a pair of inference-time diagnostics: a linear probe that asks where the answer is decodable, and a corruption test that asks whether the latent is load-bearing. The supervised latents are largely bypassed. Corrupting them shifts accuracy by at most four points. The answer is decodable downstream of the latent but not at it, and the size of this decodability gap predicts how much each variant relies on its latent under perturbation. Consistent with an Information Bottleneck reading of the loss, the auxiliary objective reshapes the language model via shared parameters rather than via the latent variable it nominally optimizes.

تكشف دراسة حديثة نُشرت على arXiv أن الافتراض القائل بأن المحاذاة الكوسينية مرتبطة إيجابيًا بالدقة في نماذج الرؤية-اللغة (VLM) هو عكس ذلك، حيث تظهر ارتباطًا سلبيًا (r=-0.94). تقدم الدراسة أداة تشخيصية تُدعى PRISM، والتي تبرز الدور المحدود للرموز الكامنة المُشرفة في عملية توليد الإجابات.

Un estudio reciente publicado en arXiv revela que la suposición de que la alineación coseno es positivamente correlacionada con la precisión en los modelos de visión-lenguaje (VLM) es invertida, mostrando una correlación negativa (r=-0.94). La investigación introduce PRISM, una herramienta de diagnóstico que destaca el papel limitado de los tokens latentes supervisados en el proceso de generación de respuestas.

Une étude récente publiée sur arXiv révèle que l'hypothèse selon laquelle l'alignement cosinus est positivement corrélé à la précision dans les modèles de vision-langage (VLM) est inversée, montrant une corrélation négative (r=-0,94). La recherche introduit PRISM, un outil de diagnostic qui met en évidence le rôle limité des tokens latents supervisés dans le processus de génération de réponses.

A recent study published on arXiv reveals that the assumption of cosine alignment being positively correlated with accuracy in vision-language models (VLMs) is inverted, showing a negative correlation (r=-0.94). The research introduces PRISM, a diagnostic tool that highlights the limited role of supervised latent tokens in the answer generation process.