Xin Wang

Standard RL rollouts can effectively provide world modeling supervision, leading to significant performance gains in language agents.

SentGuard detects 90.5% of unsafe content within two sentences, revolutionizing real-time moderation for large language models.

Predicting drug synergy for novel compounds just got a whole lot better with a new GraphLLM that bridges the gap between molecular structure and semantic understanding.

Spotting unfaithful reasoning in LLMs just got easier: a new method efficiently compares a model's internal computations against its stated rationale.

Synthesized breast MRIs can now better mimic real-world lesion complexity thanks to a diffusion model that explicitly handles multi-scale features and heterogeneous enhancement.

Distilling black-box video generators into autoregressive models doesn't require teacher scores or complex alignment—just cleverly paired rollouts and a discriminator.

LLM agents struggle to generalize from experience to reusable skills, often performing worse than simply replaying past trajectories, revealing a critical gap in current abstraction methods.

Reconstructing 3D scenes from a single view gets a boost by explicitly recognizing and leveraging repeated object instances, like chairs and tables, to inform and refine the reconstruction.

LLMs are surprisingly bad at fixing real-world logging security vulnerabilities, despite being moderately effective at detecting them.

RL unlocks genuinely new tool-use capabilities in LLMs by enabling compositional strategies that surpass what's achievable through mere re-sampling, challenging the notion that RL only improves reliability.

Stop wasting compute: a learned policy can intelligently allocate LLM inference budgets, boosting accuracy by up to 12.8% compared to uniform allocation.

VLMs can be easily fooled in the real world by strategically manipulating lighting, causing them to misinterpret scenes and hallucinate nonsensical captions.

Breast cancer risk prediction from MRI just got a whole lot faster and more interpretable, thanks to a novel 2.5D approach that beats both 2D and 3D models.

The best deepfake audio detectors are surprisingly biased by audio quality, speaker gender, and language, undermining their real-world reliability.

Injecting demonstrations with a carefully annealed probability can drastically improve exploration in RLVR, even for tasks requiring novel reasoning or domain-specific knowledge.

Forget struggling with cryptic SQL: a new LLM fine-tuned with human preferences generates comments so good, they beat Qwen3-14B by up to 13% on standard metrics.