Yuhang Zang

Heterogeneous diffusion models can be seamlessly stitched together, unlocking new capabilities by sharing a common latent language.

LVLMs can be trained to actively mentalize spatial scenes from novel viewpoints, boosting their reasoning accuracy by up to 20% on complex tasks.

Monolithic video evaluation dilutes critical action cues, but chunking interactive rollouts into action-aligned visual evidence allows targeted reward models to outperform GPT-5.5 at scoring world model dynamics.

HPSD enables TI2V models to internalize high-quality visual cues, resulting in a remarkable boost in text-to-video performance while simultaneously enhancing image-to-video generation.

Intern-S2-Preview-397B not only excels in multimodal scientific reasoning but also enhances biological instruction performance without altering its foundational architecture.

Forgetting earlier observations, not decision-making flaws, is the primary source of errors in multimodal LLMs navigating complex tasks.

CapRL++ redefines caption quality through utility, enabling models to produce high-fidelity descriptions without the constraints of traditional supervised fine-tuning.

Streaming spatial intelligence remains a significant hurdle for multimodal LLMs, with top models trailing human experts by 27 points in allocentric mapping tasks.

LLM agents can achieve a remarkable 82 wins out of 402 games by leveraging structured pseudocode, transforming how they interact with skill libraries.

Forget textual rules and coarse embeddings: a multimodal reward model that directly compares rendered visuals unlocks significant gains in vision-to-code RL.

Text-to-image flow models can achieve superior preference alignment by augmenting the condition space, creating a "dense" reward mapping that better captures inter-sample relationships.

Diffusion models can now reason their way through complex spatial tasks with near-perfect accuracy, thanks to a new framework that unlocks chain-of-thought reasoning within the latent space.

Hallucinations in RL-based image editing and generation are tamed with FIRM, a new framework that trains robust reward models on curated datasets to provide more accurate guidance.

By mimicking how humans use visual anchors, ChartVSR lets models iteratively correct their own visual perception errors, leading to more accurate chart parsing.

Ditch the clunky pipelines: SongGen generates complete songs from text in a single pass, offering unprecedented control over musical elements and voice cloning.