Ao Li

Turn-level Multiscale Density Ratio Estimation (tlm-DRE), which assigns different weights on corresponding turns and proposes asymmetric token-level training based on the positive-negative space gaps across multiple turns of tasks, enables LLMs to perform robustly in multi-turn reasoning tasks with both in-domain and out-of-domain conditions.

Authorial expressiveness can be significantly enhanced by positioning writing styles within a unified five-dimensional space, leading to more personalized and interpretable text generation.

Multi-agent collaboration in video anomaly understanding leads to a significant boost in detection and verification accuracy, outperforming traditional methods.

Rarity-aware sampling and spatial consistency techniques enable discrete diffusion to achieve high-quality image super-resolution with fewer decoding steps than traditional methods.

Achieving state-of-the-art underwater image enhancement with a model that has only 4.23K parameters and processes images at over 600 FPS.

Real-robot trials are costly and slow, but DataLadder enables scalable evaluations and data generation through a seamless interplay between robots, simulations, and human feedback.

AIChilles reveals 49 hidden weaknesses in AI-evolved systems, challenging the assumption that AI-generated code is always superior to human-designed algorithms.

Achieve open-world 3D segmentation without manual annotation by decoupling object discovery from semantic grounding.

LLM agents can get a ~5% performance boost and 4x memory reduction by representing workflows as graphs and adaptively instantiating them based on task semantics.