Systems and benchmarks

Projects

Research systems, benchmarks, and efficient inference methods with their public artifacts.

Training-Free Multimodal LLM Orchestration paper first page
Multimodal Systems · ICML 2026

Training-Free Multimodal LLM Orchestration

Problem. How can an interactive omni-modal assistant coordinate heterogeneous expert models without retraining a monolithic model?

Method. LLM-based routing composes off-the-shelf modality experts with persistent cross-modal memory and interruption-aware streaming interaction.

Role
First author
Status
Paper, code, and research note available
OrchestrationOmni-modalStreaming
SocialOmni paper first page
Benchmark · arXiv

SocialOmni

Problem. Existing omni-model benchmarks underrepresent time-sensitive audio-visual social interaction.

Method. A standardized benchmark and dataset evaluate social perception, temporal grounding, and interactive response behavior across audio and video.

Role
First author
Status
Paper, code, and dataset available
Social InteractionAudio-VisualEvaluation
WFS-SB paper first page
Long Video · CVPR 2026

Long-video Frame Selection

Problem. Long-video models spend substantial computation on redundant frames while missing sparse semantic events.

Method. Event anchors and wavelet-detected semantic boundaries guide adaptive frame allocation without training an additional selector.

Role
Fourth author on WFS-SB
Status
WFS-SB accepted; EFS preprint available
Long VideoFrame SelectionTraining-Free
Correct While Verifying paper preview
Efficient Inference · AAAI 2027

Correct While Verifying

Problem. Speculative decoding discovers incorrect draft tokens only after final verification, then discards the entire suffix following the first rejection.

Method. Intermediate target-model states predict correct continuations early, seed correction branches, and verify original and corrected drafts in parallel.

Role
Fifth author
Status
AAAI 2027 submission; project note available
Speculative DecodingDraft CorrectionEfficient Inference