Xipeng Qiu

Agents can now recursively evolve their own training environments, jointly building and validating tasks, contexts, and verifiers for robust long-horizon learning.

Hybrid LLMs exhibit a "Seesaw Effect" in context extension, with linear attention models thriving on continual pretraining and sliding-window models excelling at extrapolation, a distinction rooted in their positional inductive biases.

Standard visual token pruning targets the wrong encoder depth: tapping into intermediate representations while asymmetrically skipping decoder sub-layers cuts MLLM inference FLOPs by 79% without requiring any retraining.