Jie Jiang

The results show that the key to memory-augmented repair is not simply accumulating more experiences, but retrieving the right experiences for the right repair context, and that relevance, quality, and redundancy matter more than raw memory volume.

Scalar rewards in on-policy distillation tell tokens to gain or lose probability without specifying where that mass should actually go—framing distillation as explicit pairwise probability transport eliminates this background leakage and beats reverse-KL across reasoning tasks.

R^3 achieves a groundbreaking balance between compliance and semantic intent preservation, outperforming existing methods in video ad rectification.

TimeThink revolutionizes video reasoning by enabling models to pinpoint relevant temporal evidence with unprecedented accuracy, outperforming existing approaches.

NOVA reduces silent failures in recommender systems by over 13x while boosting performance metrics by up to 2.02% in online A/B testing.

Explainable detection of hateful videos is now possible, revealing the nuanced reasoning behind classifications that traditional methods overlook.

ForeMoE achieves a remarkable 1.45× speedup in RL post-training by anticipating load imbalances, transforming how we manage expert resources in large language models.

Action verification can now be reliably performed in VLA models, reducing the risk of grasp failures and task errors in real-world robotic applications.

LLM agents can learn to use tools more efficiently and accurately by explicitly learning when *not* to use them, leading to a 25% increase in tool productivity.

Revitalizing target-specific control within latent-query architectures for sequential CTR prediction yields consistent performance gains across diverse datasets and backbones, especially when combined with a simple position-aware reference.

Agentic RAG gets a 7.7 point accuracy boost thanks to Search-P1's path-centric reward shaping, which extracts learning signals even from failed reasoning attempts.

Dramatically reduce hallucination in industrial RAG systems by jointly optimizing retrieval and generation with graph-aware retrieval and reinforcement learning, leading to a 92.7% reduction in URL hallucination in a real-world advertising QA system.

Spatial relationship hallucinations in image inpainting can be significantly reduced by directly optimizing for preferences on background plausibility, even when foregrounds are identical.