Multimodal Large Language Models have demonstrated remarkable capabilities in video understanding, yet face prohibitive computational costs and performance degradation from "context rot" due to massive visual token redundancy. Existing compression strategies typically rely on heuristics or fixed transformations that are often decoupled from the downstream task objectives. We propose SCORE (Surprise-augmented token COmpression via REinforcement learning), a unified framework that learns an adaptive token compression policy. SCORE introduces a lightweight policy network conditioned on a surprise-augmented state representation that incorporates inter-frame residuals to capture temporal dynamics and motion saliency, optimized via a group-wise reinforcement learning scheme with a split-advantage estimator and a static→real two-stage curriculum. SCORE achieves a 16× prefill speedup while preserving 99.5% of original performance at a 10% retention ratio.

Multimodal Large Language Models (MLLMs) have achieved impressive performance across vision-language tasks, yet their fine-grained visual perception and precise spatial reasoning remain limited. We introduce DiG (Differential Grounding), a proxy task framework where MLLMs learn fine-grained perception by identifying and localizing all differences between similar image pairs, without prior knowledge of the number of differences. To support scalable training, we develop an automated 3D-rendering-based data generation pipeline and employ curriculum learning for stable optimization. DiG significantly improves performance on RefCOCO, RefCOCO+, RefCOCOg, and general multimodal perception benchmarks.

We introduce FPEdit, a framework that leverages knowledge editing to inject semantically coherent natural language fingerprints through sparse, targeted modifications to model weights. Our Promote-Suppress Value Vector Optimization achieves 95–100% fingerprint retention under full-parameter fine-tuning and parameter-efficient adaptation, remaining robust under quantization, pruning, and stochastic decoding. FPEdit embeds 10 fingerprint pairs into LLaMA2-7B in under 2 minutes using less than 30 GB GPU memory.