arXiv:2604.03307v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable success, yet they remain prone to perception-related hallucinations in fine-grained tasks. This vulnerability arises from a fundamental limitation: their reasoning is largely restricted to the language domain, treating visual input as a static, reasoning-agnostic preamble rather than a dynamic participant. Consequently, current models act as passive observers, unable to re-examine visual details to ground their evolving reasoning states. To overcome this, we propose V-Reflection, a framework that transforms the MLLM into an active interrogator through a "think-then-look" visual reflection mechanism. During reasoning, latent states function as dynamic probes that actively interrogate the visual feature space, grounding each reasoning step for task-critical evidence. Our approach employs a two-stage distillation strategy. First, the Box-Guided Compression (BCM) module establishes stable pixel-to-latent targets through explicit spatial grounding. Next, a Dynamic Autoregressive Compression (DAC) module maps the model's hidden states into dynamic probes that interrogate the global visual feature map. By distilling the spatial expertise of the BCM teacher into the DAC student, V-Reflection internalizes the ability to localize task-critical evidence. During inference, both modules remain entirely inactive, maintaining a purely end-to-end autoregressive decoding in the latent space with optimal efficiency. Extensive experiments demonstrate the effectiveness of our V-Reflection across six perception-intensive benchmarks, significantly narrowing the fine-grained perception gap. Visualizations confirm that latent reasoning autonomously localizes task-critical visual evidence.

arXiv:2509.22258v5 Announce Type: replace-cross Abstract: Recent advances in vision-language models (VLMs) have achieved remarkable performance on standard medical benchmarks, yet their true clinical reasoning ability remains unclear. Existing datasets predominantly emphasize classification accuracy, creating an evaluation illusion in which models appear proficient while still failing at high-stakes diagnostic reasoning. We introduce Neural-MedBench, a compact yet reasoning-intensive benchmark specifically designed to probe the limits of multimodal clinical reasoning in neurology. Neural-MedBench integrates multi-sequence MRI scans, structured electronic health records, and clinical notes, and encompasses three core task families: differential diagnosis, lesion recognition, and rationale generation. To ensure reliable evaluation, we develop a hybrid scoring pipeline that combines LLM-based graders, clinician validation, and semantic similarity metrics. Through systematic evaluation of state-of-the-art VLMs, including GPT-4o, Claude-4, and MedGemma, we observe a sharp performance drop compared to conventional datasets. Error analysis shows that reasoning failures, rather than perceptual errors, dominate model shortcomings. Our findings highlight the necessity of a Two-Axis Evaluation Framework: breadth-oriented large datasets for statistical generalization, and depth-oriented, compact benchmarks such as Neural-MedBench for reasoning fidelity. We release Neural-MedBench at https://neuromedbench.github.io/ as an open and extensible diagnostic testbed, which guides the expansion of future benchmarks and enables rigorous yet cost-effective assessment of clinically trustworthy AI.

arXiv:2510.15746v2 Announce Type: replace-cross Abstract: Ideal or real - that is the question.In this work, we explore whether principles from game theory can be effectively applied to the evaluation of large language models (LLMs). This inquiry is motivated by the growing inadequacy of conventional evaluation practices, which often rely on fixed-format tasks with reference answers and struggle to capture the nuanced, subjective, and open-ended nature of modern LLM behavior. To address these challenges, we propose a novel alternative: automatic mutual evaluation, where LLMs assess each other's output through self-play and peer review. These peer assessments are then systematically compared with human voting behavior to evaluate their alignment with human judgment. Our framework incorporates game-theoretic voting algorithms to aggregate peer reviews, enabling a principled investigation into whether model-generated rankings reflect human preferences. Empirical results reveal both convergences and divergences between theoretical predictions and human evaluations, offering valuable insights into the promises and limitations of mutual evaluation. To the best of our knowledge, this is the first work to jointly integrate mutual evaluation, game-theoretic aggregation, and human-grounded validation for evaluating the capabilities of LLMs.