Yuning Han Yuning Han

Verifying the many candidate trajectories a software engineering agent generates at test time can cost as many tokens as generating them. LatentSift replaces the LLM-based first verification stage with a token-free, execution-free filter over hidden states the policy already produced — reasoning, observation and function-call states compared against banks from successful and unsuccessful training trajectories. On SWE-bench Verified, hybrid Best@16 matches or improves on each agent's reference workflow.

A self-speculative decoding framework for memory-limited LLM inference on edge devices: a lightweight draft adapter proposes a candidate token tree, a shallow verifier cheaply prunes it, and the target model confirms the survivors in a single pass — cutting target-model forward calls without raising peak memory.

A universal, imperceptible backdoor trigger for diffusion models built from universal adversarial perturbations: image- and model-agnostic, it drives a high attack success rate on triggered inputs while keeping generation quality high on clean data, and evades state-of-the-art backdoor defenses (Elijah, TERD).