Long (Tony) Lian

ThreadWeaver is a framework for adaptive parallel reasoning that achieves accuracy on par with cutting-edge sequential reasoning models while significantly reducing inference latency. ThreadWeaver utilizes a two-stage parallel trajectory generator, trie-based training-inference co-design, and parallelization-aware reinforcement learning (P-GRPO).

We demonstrate a new dimension of scaling—parallel reasoning—by giving LLMs spawn() and join() functions to control when to reason sequentially or in parallel, enabling lower latency and improved scalability in complex reasoning tasks.

Describe Anything Model (DAM) generates detailed descriptions for user-specified regions in images and videos, marked by points, boxes, scribbles, or masks. We introduce DLC-Bench to evaluate such region-based descriptions.

Independent partial patch reconstruction facilitates efficient representation learning.

A text-to-image generation framework that iteratively corrects inaccuracies by integrating a self-correcting LLM with diffusion models, significantly improving image alignment with complex prompts and enabling image editing tasks.

Improving text-to-video generation by using a large language model to make plans before the actual video generation, achieving realistic video generation that align with complex input prompts.

Enhancing the prompt understanding capabilities of text-to-image diffusion models with large-language models for grounding.

AutoGaze removes redundant video patches before ViT or MLLM processing, reducing visual tokens by 4x-100x and enabling efficient long, high-resolution video understanding.

V1 unifies generation and pairwise self-verification for parallel reasoners, improving test-time scaling through tournament-based ranking and reinforcement learning for joint generation-verification.

VisGym is a suite of 17 long-horizon visual interaction environments for evaluating and training multimodal agents across symbolic puzzles, real-image tasks, navigation, and manipulation.

VPBench shows that small visual prompting details, including marker color, marker size, dataset size, and JPEG compression, can substantially shift VLM benchmark performance and rankings.

Pillar-0 is a radiology foundation model trained on large-scale volumetric CT and MRI data, paired with RATE for scalable label extraction across hundreds of radiologic findings.

ECHO constructs image generation benchmarks from real-world social media evidence of model use, surfacing emerging tasks, stronger model differentiation, and community-informed quality metrics.

TULIP improves contrastive image-text learning with generative data augmentation, richer image-image and text-text contrastive learning, and reconstruction regularization for stronger fine-grained visual understanding.

Atlas is a new neural network using Multi-Scale Attention for efficient cross-scale image modeling. Atlas achieves state-of-the-art accuracy with significantly better speed and compute efficiency on high-resolution image tasks.

A novel unified framework adept at performing various instance, semantic and panoptic segmentation tasks without any supervision.

Running Stable Diffusion in 4-bit weights with high generation quality for the first time.

Segmenting objects in videos without any human annotation in any training stage.

We focus on selecting the right data to query for label for semi-supervised learning without any label or task information. Our method demonstrates that a small compute spent on careful labeled data selection brings big annotation efficiency.

We propose a novel and effective debiased learning method based on counterfactual reasoning and adaptive margins to deal with the undesired effect from naturally imbalanced pseudo-labels.

Rather than training a universal RL policy invariant to train-test distribution shift, we proposed unsupervised visual attention and invariance method (VAI) to disperse interference task-irrelevant factors towards a RL policy robust to distractions.

Leveraging bias-variance tradeoff, we proposed a universal classification framework for long-tailed distributed data with RoutIng Diverse Experts (RIDE) that improves both accuracy and inference speed.