publications
publications in reverse chronological order.
Rethinking LLM Ensembling from the Perspective of Mixture Models
Published in ICML (Spotlight), 2026
Mixture-model-like Ensemble (ME) is a training-free, plug-and-play ensembling method that reinterprets LLM ensembling as a mixture model and samples from the same ensemble distribution while invoking only one model per step. ME is mathematically equivalent to sampling from the ensemble distribution and only requires evaluating one model per step, making it 1.78x-2.68x faster than conventional ensembling.
d2Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching
Published in ICLR, 2026
Dual aDaptive Cache (d²Cache) is a training-free approximate KV cache framework for accelerating dLLM inference. d²Cache features a two-stage fine-grained selection strategy to identify tokens and adaptively update their KV states at each decoding step, while caching the KV states of the remaining tokens for reuse. Furthermore, d²Cache naturally offers a more reliable decoding alternative, which can enable quasi left-to-right generation and mitigate premature overconfidence in tokens at the end of the sequence. Extensive experimental results on two representative dLLMs (i.e., LLaDA and Dream) demonstrate that d²Cache not only achieves substantial inference speedups, but also yields consistent improvements in generation quality.
Mimic In-Context Learning in Multimodal Tasks
Published in CVPR, 2025
MimIC is a novel framework that mimics in-context learning for multimodal tasks by injecting lightweight, query-conditioned shift vectors after each attention head. Applied to Idefics1-9B, MimIC achieves up to +3.46% accuracy improvement on VQAv2, +3.57% on OK-VQA, and +9.00 CIDEr on image captioning, compared to standard 32-shot in-context learning. Moreover, MimIC effectively mitigates hallucination commonly introduced by conventional ICL approaches, while incurring inference overhead comparable to zero-shot inference.
Fast Large Language Model Collaborative Decoding via Speculation
Published in ICML, 2025
Collaborative decoding via Speculation (CoS) is a novel framework that accelerates the ensemble of any number of LLMs without sacrificing performance. It could reach 1.11x-2.23x over standard ensemble techniques on two-model or three-model pairs.