Can AI agents rebuild websites through browser exploration? The 70 WebsiteBench tasks for the Harbor framework.
AI & ML interests
Natural Language Processing, Image Generation
Recent Activity
Papers
ReGE: Stronger Agents through Experience Synthesis
VIEScore2: Unified Image Evaluation with Spatially Grounded Explanations
Benchmark dataset (V1+V2), live leaderboard Space, and full V1 execution traces — everything you need to run, regrade, or compare on ClawBench.
SWE-Next turns real repository changes into self-verifying software engineering tasks and reusable execution environments for scalable agent training
MMMU Dataset
OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis
Building Multi-Language Visualization Coding Agents
EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing
Crique-Coder
MMLU-Pro
Reasoning in the pixel space
Advancing LLMs' general reasoning capabilities
Video Mamba
A collection of models and datasets from ABC: Achieving Better Control of Multimodal Embeddings using VLMs.
PixelWorld
The dataset and models for CritiqueFineTuning
-
TIGER-Lab/WebInstruct-CFT
Viewer • Updated • 654k • 1.33k • 60 -
TIGER-Lab/Qwen2.5-Math-7B-CFT
Text Generation • 8B • Updated • 65 • • 8 -
TIGER-Lab/Qwen2.5-32B-Instruct-CFT
Text Generation • 33B • Updated • 40 • 6 -
Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitate
Paper • 2501.17703 • Published • 57
The generalist image editing model
The VLM2Vec embedding models.
The datasets and models for the MAmmoTH project
Imagenhub
The structure knowledge grounded language model
Mantis model family optimized for multi-image reasoning with interleaved text/image format
Building Automatic Metrics to Simulate Fine-grained Human Feedback for Video Generation
Function-aware FIM mid-training for coding agents: 6 checkpoints (7B/8B/14B) + the 400K/2.6B-token corpus. github.com/TIGER-AI-Lab/FIM-Midtraining
A Reasoning Reward Model that Scale Image Generation Both Training and Test Time
-
TIGER-Lab/RationalRewards-8B-T2I
Image-to-Text • 9B • Updated • 40 • 6 -
TIGER-Lab/RationalRewards-8B-Edit
Image-to-Text • 9B • Updated • 45 • 5 -
TIGER-Lab/RationalRewards-EvalData-GenAIBench-MMRB2-ERBench
Viewer • Updated • 1k • 160 • 1 -
TIGER-Lab/RationalRewards-SFTData
Viewer • Updated • 348k • 238 • 1
An execution-based framework that evaluates physical reasoning by generating executable simulator code from visual inputs.
-
TIGER-Lab/VisPhyBench-Data
Updated • 698 -
VisPhyWorld: Probing Physical Reasoning via Code-Driven Video Reconstruction
Paper • 2602.13294 • Published • 13 -
TIGER-Lab/VisPhyWorld-Sub-Generated-Videos
Viewer • Updated • 2.52k • 435 -
VisPhyWorld Comparator
🎞Compare ground-truth and generated videos sample by sample
Stress Testing Image Generation Models
Browseragent: An agent that can interact with browser to complete tasks
VideoScore2
web explorer model
Converting VLM to a general embedding model
Unleashing the Reasoning Potential of Pre-trained LLMs by Critique Fine-Tuning on One Problem
The pioneering work in Dialogue-driven Movie Shot Generation
SoTA VLM for Reasoning
-
TIGER-Lab/VL-Rethinker-72B
Visual Question Answering • 73B • Updated • 24 • 5 -
VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning
Paper • 2504.08837 • Published • 46 -
TIGER-Lab/VL-Rethinker-7B
Image-Text-to-Text • 8B • Updated • 653 • 15 -
TIGER-Lab/VL-Reasoner-72B
Visual Question Answering • 73B • Updated • 29 • 3
Scaling up MM data
-
TIGER-Lab/VisualWebInstruct-Recall
Viewer • Updated • 361k • 1.58k • 4 -
TIGER-Lab/VisualWebInstruct-Seed
Viewer • Updated • 60.3k • 731 • 19 -
TIGER-Lab/VisualWebInstruct
Viewer • Updated • 1.91M • 3.53k • 45 -
VisualWebInstruct: Scaling up Multimodal Instruction Data through Web Search
Paper • 2503.10582 • Published • 25
The dataset and model for MAmmoTH-VL
Video Augmentation for Synthetic Video Instruction-following Data Generation
-
TIGER-Lab/VISTA-LongVA
Video-Text-to-Text • 8B • Updated • 35 • 2 -
TIGER-Lab/VISTA-Mantis
Video-Text-to-Text • 8B • Updated • 20 • 1 -
TIGER-Lab/VISTA-VideoLLaVA
Video-Text-to-Text • 7B • Updated • 30 -
VISTA: Enhancing Long-Duration and High-Resolution Video Understanding by Video Spatiotemporal Augmentation
Paper • 2412.00927 • Published • 28
List of model variates of TIGEREScore checkpoints and the associated dataset
-
TIGER-Lab/TIGERScore-13B
Text Generation • 13B • Updated • 87 • 18 -
TIGER-Lab/MetricInstruct
Viewer • Updated • 42.5k • 222 • 13 -
TIGER-Lab/TIGERScore-7B
Text Generation • 7B • Updated • 17 • 2 -
TIGERScore: Towards Building Explainable Metric for All Text Generation Tasks
Paper • 2310.00752 • Published • 3
The dataset and model for UniIR project
ConsistI2V Image-to-Video generation models
Scaling up instruction data from the web for to build better LLMs
Long-context research projects
Can AI agents rebuild websites through browser exploration? The 70 WebsiteBench tasks for the Harbor framework.
Function-aware FIM mid-training for coding agents: 6 checkpoints (7B/8B/14B) + the 400K/2.6B-token corpus. github.com/TIGER-AI-Lab/FIM-Midtraining
Benchmark dataset (V1+V2), live leaderboard Space, and full V1 execution traces — everything you need to run, regrade, or compare on ClawBench.
A Reasoning Reward Model that Scale Image Generation Both Training and Test Time
-
TIGER-Lab/RationalRewards-8B-T2I
Image-to-Text • 9B • Updated • 40 • 6 -
TIGER-Lab/RationalRewards-8B-Edit
Image-to-Text • 9B • Updated • 45 • 5 -
TIGER-Lab/RationalRewards-EvalData-GenAIBench-MMRB2-ERBench
Viewer • Updated • 1k • 160 • 1 -
TIGER-Lab/RationalRewards-SFTData
Viewer • Updated • 348k • 238 • 1
SWE-Next turns real repository changes into self-verifying software engineering tasks and reusable execution environments for scalable agent training
MMMU Dataset
An execution-based framework that evaluates physical reasoning by generating executable simulator code from visual inputs.
-
TIGER-Lab/VisPhyBench-Data
Updated • 698 -
VisPhyWorld: Probing Physical Reasoning via Code-Driven Video Reconstruction
Paper • 2602.13294 • Published • 13 -
TIGER-Lab/VisPhyWorld-Sub-Generated-Videos
Viewer • Updated • 2.52k • 435 -
VisPhyWorld Comparator
🎞Compare ground-truth and generated videos sample by sample
OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis
Stress Testing Image Generation Models
Building Multi-Language Visualization Coding Agents
Browseragent: An agent that can interact with browser to complete tasks
EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing
VideoScore2
Crique-Coder
web explorer model
MMLU-Pro
Converting VLM to a general embedding model
Unleashing the Reasoning Potential of Pre-trained LLMs by Critique Fine-Tuning on One Problem
Reasoning in the pixel space
The pioneering work in Dialogue-driven Movie Shot Generation
Advancing LLMs' general reasoning capabilities
SoTA VLM for Reasoning
-
TIGER-Lab/VL-Rethinker-72B
Visual Question Answering • 73B • Updated • 24 • 5 -
VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning
Paper • 2504.08837 • Published • 46 -
TIGER-Lab/VL-Rethinker-7B
Image-Text-to-Text • 8B • Updated • 653 • 15 -
TIGER-Lab/VL-Reasoner-72B
Visual Question Answering • 73B • Updated • 29 • 3
Video Mamba
A collection of models and datasets from ABC: Achieving Better Control of Multimodal Embeddings using VLMs.
Scaling up MM data
-
TIGER-Lab/VisualWebInstruct-Recall
Viewer • Updated • 361k • 1.58k • 4 -
TIGER-Lab/VisualWebInstruct-Seed
Viewer • Updated • 60.3k • 731 • 19 -
TIGER-Lab/VisualWebInstruct
Viewer • Updated • 1.91M • 3.53k • 45 -
VisualWebInstruct: Scaling up Multimodal Instruction Data through Web Search
Paper • 2503.10582 • Published • 25
PixelWorld
The dataset and models for CritiqueFineTuning
-
TIGER-Lab/WebInstruct-CFT
Viewer • Updated • 654k • 1.33k • 60 -
TIGER-Lab/Qwen2.5-Math-7B-CFT
Text Generation • 8B • Updated • 65 • • 8 -
TIGER-Lab/Qwen2.5-32B-Instruct-CFT
Text Generation • 33B • Updated • 40 • 6 -
Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitate
Paper • 2501.17703 • Published • 57
The dataset and model for MAmmoTH-VL
Video Augmentation for Synthetic Video Instruction-following Data Generation
-
TIGER-Lab/VISTA-LongVA
Video-Text-to-Text • 8B • Updated • 35 • 2 -
TIGER-Lab/VISTA-Mantis
Video-Text-to-Text • 8B • Updated • 20 • 1 -
TIGER-Lab/VISTA-VideoLLaVA
Video-Text-to-Text • 7B • Updated • 30 -
VISTA: Enhancing Long-Duration and High-Resolution Video Understanding by Video Spatiotemporal Augmentation
Paper • 2412.00927 • Published • 28
The generalist image editing model
The VLM2Vec embedding models.
List of model variates of TIGEREScore checkpoints and the associated dataset
-
TIGER-Lab/TIGERScore-13B
Text Generation • 13B • Updated • 87 • 18 -
TIGER-Lab/MetricInstruct
Viewer • Updated • 42.5k • 222 • 13 -
TIGER-Lab/TIGERScore-7B
Text Generation • 7B • Updated • 17 • 2 -
TIGERScore: Towards Building Explainable Metric for All Text Generation Tasks
Paper • 2310.00752 • Published • 3
The datasets and models for the MAmmoTH project
The dataset and model for UniIR project
Imagenhub
The structure knowledge grounded language model
ConsistI2V Image-to-Video generation models
Mantis model family optimized for multi-image reasoning with interleaved text/image format
Scaling up instruction data from the web for to build better LLMs
Building Automatic Metrics to Simulate Fine-grained Human Feedback for Video Generation
Long-context research projects