🏢 Hong Kong University of Science and Technology

Perovskite-LLM: a new knowledge-enhanced system boosts perovskite solar cell research by integrating a domain-specific knowledge graph, high-quality datasets, and specialized LLMs for superior knowled…

mmMamba: a novel framework creates linear-complexity multimodal models via distillation, drastically improving efficiency without sacrificing performance.

Atom of Thoughts (AOT) revolutionizes LLM test-time scaling by decomposing complex reasoning into independent sub-questions, drastically reducing computation while maintaining high accuracy.

FinMTEB: A new benchmark reveals that general-purpose embedding models struggle in the finance domain; domain-specific models excel, and surprisingly, simple BoW outperforms sophisticated models on ce…

ThinkDiff empowers text-to-image diffusion models with multimodal reasoning by aligning vision-language models to an LLM decoder, achieving state-of-the-art results on in-context reasoning benchmarks.

CODEI/O: Condensing reasoning patterns from code into LLM training data for enhanced reasoning.

CustomVideoX: Zero-shot personalized video generation, exceeding existing methods in quality & consistency via 3D reference attention and dynamic adaptation.

LLMs excel at complex reasoning but struggle with planning; this paper introduces a test-time scaling approach that enhances LLMs’ PDDL reasoning, enabling high-quality PDDL domain generation, outperf…

FlashVideo: Generate stunning high-resolution videos efficiently using a two-stage framework prioritizing fidelity and detail, achieving state-of-the-art results.

Llasa, a novel single-Transformer TTS model, achieves state-of-the-art performance by scaling both training and inference compute, improving naturalness, prosody and emotional expressiveness.

W2SD: A novel framework boosts diffusion model quality by using the difference between weak and strong models to refine sampling trajectories, achieving state-of-the-art performance.

GaussianAvatar-Editor enables photorealistic, text-driven editing of animatable 3D heads, solving motion occlusion and ensuring temporal consistency.

Diffusion as Shader (DaS) achieves versatile video control by using 3D tracking videos as control signals in a unified video diffusion model, enabling precise manipulation across diverse tasks.

TransPixar generates high-quality videos with transparency by jointly training RGB and alpha channels, outperforming sequential generation methods.

VideoAnydoor: High-fidelity video object insertion with precise motion control, achieved via an end-to-end framework leveraging an ID extractor and a pixel warper for robust detail preservation and fi…

Android Agent Arena (A3): A novel evaluation platform for mobile GUI agents offering diverse tasks, flexible action space, and automated LLM-based evaluation, advancing real-world AI agent research.

Edicho: a novel training-free method for consistent image editing across diverse images, achieving precise consistency by leveraging explicit correspondence.

M-STAR: a novel self-evolving training framework significantly boosts multimodal reasoning in large models without human annotation, achieving state-of-the-art results.

B-STAR dynamically balances exploration and exploitation in self-taught reasoners, achieving superior performance in mathematical, coding, and commonsense reasoning tasks.

MegaPairs synthesizes 26M+ high-quality multimodal retrieval training examples, enabling state-of-the-art zero-shot performance and surpassing existing methods trained on 70x more data.