Sixun Dong (董思勋)

Reframes long-video VLM efficiency as a joint allocation over frame count, per-frame resolution, and front-end decoding latency, not just which tokens to keep. LoHi is a training-free, single-pass framework that pairs a dense low-resolution video stream with a few high-resolution frames (picked from codec I-frames or by a query-aware CLIP DPP). It gains +10.6% over the native-resolution baseline at matched token budget and +5.2% over the strongest prior efficiency methods, while cutting front-end decoding latency by up to 7x on hour-scale videos. Joint work with UCF, Axon & Meta.

Shows that summarization-based recommenders suffer from temporal aliasing: short-lived intent, medium-term interests, and long-term preferences get entangled in one compressed user memory. MARS writes the full history into gated recurrent (linear-attention-style) state tracks with different half-life priors, and a Top-k router lets each cached seed read only the relevant time scales, keeping candidate scoring fixed-size. It beats strong baselines such as VISTA on Amazon, KuaiRand-1K, and KuaiRec, with the largest gains for long-history users. Joint work with Duke & Meta.

A multi-agent 'film crew' that orchestrates off-the-shelf video generators for long-form narrative-to-film generation, coordinated through FilmDSL—a film-oriented domain-specific language that makes cinematic constraints (shots, camera, assets, character personas) explicit. Generation and critic agents plan, synthesize, and repair clips to keep both visual identity and character behavior consistent across the film. Joint work with UMass & MIT.

Analyzes VLM end-to-end latency and reveals that output token length dominates inference cost. While large models with short outputs outperform small models with long generations, reasoning remains essential for complex tasks. We bridge this by proposing a multi-agent framework where a small model computes the reasoning tokens and transfers them to a large model, achieving large-model accuracy with minimal latency.

Proposes an LLM-agent-based post-ASR correction framework for dysarthric speech recognition that focuses on semantic accuracy rather than just minimizing Word Error Rate (WER).

Proposes MMTok to efficiently select informative vision tokens via a multimodal coverage maximization strategy, significantly accelerating VLM inference while maintaining model performance.

Introduces an automated data generation framework and evaluation benchmark for complex logical instructions to assess and enhance the multi-step reasoning capabilities of LLMs.

Proposes TimesCLIP, innovatively aligning time series data with visual and textual multi-modal perspectives to effectively enhance both short-term and long-term forecasting performance.

Proposes a reward-guided hierarchical diffusion model that "sculpts" data representations from noise to generate optimal feature transformations for specific downstream tasks.

Introduces the MLLM-Tool framework and contributes a specialized dataset to empower Multimodal Large Language Models (MLLMs) with the ability to understand, learn, and invoke external tools.

Proposes a weakly supervised video representation learning approach that utilizes unaligned text for sequential videos, significantly reducing the reliance on fine-grained video-text alignment annotations.

Introduces the RepCount dataset and pioneers the use of regression density maps alongside a Transformer-based architecture to encode multi-scale temporal correlations, significantly improving repetitive action counting.