Hongbo Liu

LLMs show promise in interpreting cryptographic schemes but fall short in generating and transforming formal proofs, with the best model scoring just 48.7 out of 100.

LLM watermarks can now survive fine-tuning, quantization, and distillation thanks to a new method that embeds them in a stable functional subspace.

By enforcing graph isomorphism across counterfactual inputs, UGID reveals that debiasing LLMs can be achieved by directly manipulating internal representations and attention mechanisms.

By unifying layout-to-image generation and image grounding with a novel cycle-consistent learning approach, EchoGen achieves state-of-the-art results in both tasks, proving that solving two problems at once can be better than solving them separately.

Autoregressive video generation gets a 1.8x speed boost and avoids temporal drift by denoising all blocks hierarchically at the same noise level.

Iteratively prompting a graph neural network at test time to amplify out-of-distribution signals dramatically improves OOD detection accuracy.