Back to homepage

Research

Projects

Peer-reviewed and ongoing work in multimodal reasoning, knowledge augmentation, and instruction-driven adaptation. Each page follows a consistent academic layout with paper metadata and figures.

Multimodal · Efficiency

ResAdapt: Adaptive Resolution for Efficient Multimodal Reasoning

Input-side budget allocation before visual encoding · CAPO · Video QA & grounding

Learning how much visual budget each frame receives via a lightweight allocator paired with a frozen MLLM, improving efficiency–accuracy trade-offs under aggressive compression.

Open project page
QA · Knowledge

Awakening Augmented Generation (IAG)

COLING 2025 · NQ, TQA & WQ · T5 & Llama2

Awakening latent knowledge in LLMs through explicit and implicit imagination modules without retrieval-heavy pipelines, evaluated on NQ, TQA, and WQ.

Open project page
Meta-learning · Adapters

TAGI: Task Adapters Generation from Instructions

NeurIPS 2024 · Super-Natural Instructions & P3

Generating task-specific adapters from instructions via hypernetworks and distillation, improving cross-task generalization with lower compute than instance-only training.

Open project page