Jia Li

Memory overload in autoregressive video generation can be tackled by absorbing historical context into model weights, achieving up to 50% cache reduction with minimal quality loss.

Ling-2.6 and Ring-2.6 achieve unprecedented efficiency in agentic intelligence, enabling instant responses and deep reasoning at trillion-parameter scale.

Vision-language models can be transformed to recognize multiple objects in images without any labels, challenging the conventional focus on iconic representations.

GCD, designed to enhance code reliability, can paradoxically serve as a vulnerability that enables LLMs to produce malicious code through a novel attack called CodeSpear.

A learnable missing token representation allows MRAF to achieve 100% accuracy in polyglot speaker identification, even when face information is missing.

FlashMemory-DeepSeek-V4 slashes GPU memory usage by over 90% for ultra-long contexts while enhancing model accuracy.

Cross-tokenizer On-Policy Distillation achieves superior efficiency and flexibility, enabling knowledge transfer between diverse model families without the constraints of shared tokenizers.

Bayesian-Agent transforms how LLM agents evolve skills, achieving up to 100% success on complex benchmarks through a novel posterior-guided optimization approach.

Accurately predicting battery health from early data just got a whole lot better: BatteryMFormer leverages multi-level Transformer learning to substantially outperform existing methods.

Standard multimodal fusion can hurt performance in emotion recognition, but this new approach knows when to drop modalities, leading to state-of-the-art results.