Di Liang

Traditional relevance proxies fail on hard negatives, but a new causal measurement approach reveals how to allocate context for up to 20.5% better recall in generative search.

User feedback prediction can be learned, revealing critical performance gaps in AI assistants that traditional evaluation methods overlook.

Forget sifting through walls of text – now you can pinpoint exactly where the AI found its answer, down to the pixel, even in complex visuals like charts and diagrams.

Parameter importance isn't forever: dynamically adapting which parameters are frozen during fine-tuning significantly improves generalization and reduces forgetting in LLMs.

Safety-aligned LLMs are so consistently risk-averse that a single, transferable "poison" document can now block up to 96% of queries across different RAG systems, even without access to the target model.