Sharon Li

TRACE transforms how long-horizon agents are trained, leading to a remarkable performance increase on complex tasks without the need for supervised fine-tuning.

OAT achieves up to 5000 times faster failure attribution for LLM agents without the need for costly step-level supervision.

LLM agents struggle with exploration in multi-agent settings, leading to poor coordination and increased regret, but a new framework can turn this around.

Progress advantage reveals a powerful, annotation-free scoring mechanism that outperforms traditional reward models in LLM agentic settings.

Discovering when a robot's about to fail just got easier: Hide-and-Seek pinpoints failure signals in VLA trajectories using only coarse, trajectory-level labels, ditching the need for expensive step-by-step annotations.

Multimodal LLMs struggle with in-context learning not because they can't see, but because they can't reason across modalities, leading to a breakdown in transferring learned task mappings.

Stop reimplementing multimodal models: TorchUMM offers a unified codebase for evaluation, analysis, and post-training, streamlining research across diverse architectures and tasks.

RLVR models exhibit "Early Correctness Coherence" under noisy supervision, suggesting a surprising opportunity for self-correction via dynamic label refinement.

Data laundering can't hide forever: this new technique lets rights owners detect misuse of their content in LLMs, even when the data has been heavily transformed.

LVLMs can now better judge their own vision-based answers, thanks to a new method that focuses on how much they actually "see" in the image.

Forget tweaking prompts – understanding how retrieved context warps an LLM's hidden states is the key to unlocking better RAG performance.