Research infrastructure
Furong Lab builds the evidence layer for intelligent systems: datasets that expose meaningful structure, benchmarks that test consequential capabilities, and evaluation environments that reveal where models fail.
Featured releases
Selected resources span physical interaction, multimodal reasoning, scientific judgment, latent safety, and content provenance.
2026Task-driven spatial-functional scene graphs and a six-capability evaluation suite for embodied planning and scene understanding.
Scale 1,050 task-oriented subgraphs; 6,278 multiview images; 350+ household scenes
2026A cross-embodiment corpus and held-out evaluation suite for learning transferable world models from human and robot videos.
Scale 123K videos; 1.8M observation-trace-language triplets; five evaluation environments
2026Released evaluation episodes, normalization statistics, and model artifacts for transferable 3D interaction-trace prediction.
Scale Evaluation episodes spanning human and robot embodiments
2026
2025
2024
2025Agentic red-teaming environments surface latent behavioral risks that may remain hidden in single-turn safety tests.
2024A standardized benchmark and stress-testing toolkit for image-watermark detection and identification under diverse attacks.
Browse by scientific question
The collection is organized by the capability or failure mode being measured—not merely by paper venue or publication date.
Data and evaluation for scene understanding, interaction traces, physical reasoning, memory, and transferable robot behavior.
5 resources →02Benchmarks that probe scientific judgment, multimodal reasoning, task difficulty, temporal structure, and generalization.
7 resources →03Stress tests for latent risk, hallucination, false refusal, unlearning, watermark robustness, and generative-model trustworthiness.
8 resources →Complete collection
Search by capability, modality, or resource name, then follow the direct links to datasets, leaderboards, code, and papers.
2026Task-driven spatial-functional scene graphs and a six-capability evaluation suite for embodied planning and scene understanding.
Scale 1,050 task-oriented subgraphs; 6,278 multiview images; 350+ household scenes
2026An evaluation setting for persistent visual-semantic memory across multiple questions and continuous embodied operation.
20262026A cross-embodiment corpus and held-out evaluation suite for learning transferable world models from human and robot videos.
Scale 123K videos; 1.8M observation-trace-language triplets; five evaluation environments
2026A multi-task, multimodal time-series reasoning benchmark for evaluating generalist models across heterogeneous temporal tasks.
2026Released evaluation episodes, normalization statistics, and model artifacts for transferable 3D interaction-trace prediction.
Scale Evaluation episodes spanning human and robot embodiments
2025A controlled dataset for evaluating whether vision-language models can reliably unlearn fictitious facial identities.
20252025Agentic red-teaming environments surface latent behavioral risks that may remain hidden in single-turn safety tests.
2025
2025A dynamic benchmarking platform for evaluating trustworthiness across generative language, image, and vision-language models.
2025
2024Automatically generates diverse visual-reasoning stress tests for diagnosing hallucination failures in vision-language models.
2024
2024A NeurIPS competition, dataset, and evaluation toolkit for stress-testing image watermarks under black-box and beige-box attacks.
2024A diagnostic benchmark that disentangles language hallucination from visual illusion in large vision-language models.
2024
20242024A standardized benchmark and stress-testing toolkit for image-watermark detection and identification under diverse attacks.
2021Controlled image-classification datasets for studying how models generalize from easy training examples to harder test examples.
Try a broader keyword or clear one of the filters.
Evaluation philosophy
Datasets make states, difficulty, interactions, and failure conditions explicit enough to study.
Evaluation asks whether systems generalize across environments, embodiments, modalities, and levels of difficulty.
Agentic and adversarial tests probe behavior beyond comfortable average-case distributions.