Vision-Language Models (VLMs) achieve strong multimodal understanding but suffer from high inference costs due to redundant vision tokens. Existing hard-pruning methods discard low-scoring tokens directly, inevitably sacrificing fine-grained visual context. We propose SPack, a training-free token compression method that reframes pruning as information-energy redistribution rather than deletion. Instead of discarding pruned tokens, SPack transfers their semantic mass to retained tokens, enforcing a global principle we term Information Energy Conservation (IEC): the total information contained in redundant tokens is strictly conserved and fully transferred as information gain to the retained components. To realize this conservation, we introduce Optimal Transport (OT) theory, which treats removed energy as mass density and retained tokens as receiver slots. Solving the optimal transport plan under the Wasserstein distance yields a minimum-cost mapping that precisely transfers geometric structure and higher-order moments according to semantic affinity. Concretely, we estimate token importance via relation attention, semantic affinity, and value-space energy, then adaptively route pruned tokens to multiple retained tokens using OT, producing hyper-dense tokens that preserve both local semantics and global context. Extensive experiments demonstrate that SPack consistently achieves a superior accuracy-efficiency trade-off compared to state-of-the-art methods. The code will be released upon acceptance.

Recent advances in Large Language Models (LLMs) have shifted toward latent reasoning, where intermediate processes are represented as continuous vectors rather than text tokens. While this paradigm reduces inference cost, existing approaches often suffer from instability on complex, out-of-distribution tasks where robust reasoning is critical. To address this, we propose Bayesian Latent Thought Optimization (BLTO), a training-free framework that enhances frozen LLMs entirely at test time. BLTO treats latent thought tokens as stochastic variables and formulates reasoning as a Bayesian inference problem. We introduce a variational posterior over latent vectors and optimize it by maximizing a reward-weighted objective that balances reasoning quality with KL regularization toward a standard Gaussian prior. The reward is derived from the model’s predictive consistency via output entropy, enabling fully self-supervised optimization without external supervision. BLTO employs reparameterized sampling and reward-weighted updates within a variational inference framework, together with an adaptive controller that dynamically determines inference steps based on reward improvement and uncertainty convergence. Extensive experiments on challenging reasoning benchmarks demonstrate that BLTO consistently improves frozen LLM performance.

Mental health disorders represent a burgeoning global public health challenge. While Large Language Models (LLMs) have demonstrated potential in psychiatric assessment, their clinical utility is severely constrained by benchmarks that lack ecological validity and fine-grained diagnostic supervision. To bridge this gap, we introduce MentalDx Bench, the first benchmark dedicated to disorder-level psychiatric diagnosis within real-world clinical settings. Comprising 712 de-identified electronic health records annotated by board-certified psychiatrists under ICD-11 guidelines, the benchmark covers 76 disorders across 16 diagnostic categories. Evaluation of 18 LLMs reveals a critical paradigm misalignment: strong performance at coarse diagnostic categorization contrasts with systematic failure at disorder-level diagnosis, underscoring a gap between pattern-based modeling and clinical hypothetico-deductive reasoning.In response, we propose MentalSeek-Dx, a medical-specialized LLM trained to internalize this clinical reasoning process through supervised trajectory construction and curriculum-based reinforcement learning. Experiments on MentalDx Bench demonstrate that MentalSeek-Dx achieves state-of-the-art (SOTA) performance with only 14B parameters, establishing a clinically grounded framework for reliable psychiatric diagnosis. The dataset and code are available.