Research
My work focuses on multimodal evidence, reasoning budgets, and AI systems whose intermediate decisions can be inspected and evaluated.
Multimodal perception
Audio-visual segmentation often mistakes visually salient but silent objects for sounding objects.
Use spectral-kinematic alignment to make audio retrieve motion evidence, then use that prior to guide object queries.
Paper metadata and full code release are being prepared.
LLM reasoning
Long reasoning is not always helpful. Some referring expressions are clear, while others need careful object-aware reasoning.
Compare zero, short, and long reasoning modes before grounding and segmentation to study when extra reasoning helps or drifts.
Reproducible codebase and documentation are available.
AI for research
Generative tools for scientific communication often stop at a first draft, even when the real work is revision.
Build agent pipelines with explicit parsing, curation, layout, rendering, review, and repair stages.
PosterMELD has released its paper, generation code, benchmark evaluation, and project page.
