Published in EACL 2024, 2024
We re-evaluate topic models for content analysis, showing that model choice should be situation-specific depending on the application domain.
Published in CVPR 2024, 2024
Diagnoses entangled language hallucination and visual illusion in large vision-language models.
Published in EMNLP Findings 2024, 2024
Generates scientific diagrams from documents using multi-aspect feedback refinement.
Published in EMNLP 2024, 2024
A cheap, interpretable approach to answer equivalence evaluation that correlates well with human judgments.
Published in CVPR Workshop 2025 (Oral), 2025
A comprehensive survey of VLMs covering model architectures, alignment methods, benchmarks, and challenges including hallucination, fairness, and safety.
Published in Preprint, 2025
PrefBERT scores open-ended long-form generation in GRPO, providing better semantic reward feedback than ROUGE-L and BERTScore.
Published in ACL 2025, 2025
We measure knowledge users acquire with topic models, finding that LLM-based methods generate human-readable but overly generic topics for domain-specific data.
Published in NeurIPS 2025, 2025
A benchmark for evaluating and mitigating hallucinations in MLLMs on synthetic videos from Sora, Veo2, and Kling.
Published in Preprint, 2026
The first RL-based framework to achieve zero-data self-evolution for VLM reasoning via a multi-role training framework with a Proposer, Coder, and Solver.
Published in Preprint, 2026
An inference-time structural retrieval layer that constructs an executable skill graph offline and retrieves dependency-aware skill bundles at inference time.
Published in ICLR 2026, 2026
Vision-SR1 trains VLMs to self-reward by splitting reasoning into visual perception and language reasoning, improving without human labels or external rewards.
Published in ICLR 2026, 2026
R-Zero trains LLMs entirely without human-curated data by pitting two copies of the base model against each other in a self-evolving curriculum.
Published in Preprint, 2026
COS-PLAY introduces a co-evolution framework where LLM decision agents and skill bank agents mutually improve each other for long-horizon task completion.
Published in Preprint, 2026
An agentic harness that autonomously develops and refines image generation skills, enabling continuous improvement without manual intervention.
Published in CVPR 2026, 2026
Video models implicitly treat the first frame as a conceptual memory buffer for visual entities, enabling robust video content customization with only 20-50 training examples.
Published in CVPR 2026, 2026
A self-evolving RL framework that enables VLMs to autonomously improve reasoning abilities from large amounts of unlabeled image data.
Published in CVPR 2026, 2026
MASS injects motion-aware spatial-temporal signals into VLMs, improving physics reasoning on a new benchmark of 4,350 real-world and AI-generated videos.
