Jingbo Zhu

SPRI achieves a remarkable 3.39 BLEU point improvement over the best existing MoE upcycling method, demonstrating that pretrained weight structures can be effectively leveraged for better expert diversity.

By explicitly structuring and managing retrieval state, EfficientGraph-RAG achieves state-of-the-art RAG performance while slashing large-model token usage by over 3x.

Stop LLMs from drifting to English when reasoning in other languages: language-adaptive RL can guide them to stay consistent without sacrificing performance.

Reasoning with LLMs just got a whole lot faster: MemoSight cuts KV cache footprint by 66% and speeds up inference by 1.56x without sacrificing CoT performance.

Multilingual question answering is harder than you think: even state-of-the-art RAG systems stumble when dealing with questions and knowledge in multiple languages.

SER models, often assumed to generalize well to synthesized speech, actually fail miserably, revealing their reliance on spurious correlations rather than genuine emotional understanding.

Supervised fine-tuning can be dramatically improved by explicitly encouraging exploration of low-confidence data and suppressing high-confidence errors, leading to sustained gains in mathematical reasoning even after extensive RLVR training.

SLMs still lag behind omni language models in multi-turn conversational style control, as revealed by the new StyleBench benchmark.