🏢 UC Berkeley

S*: Hybrid test-time scaling for code generation, boosting both coverage and selection accuracy.

Autellix: Efficient LLM Serving for Agents

ARM4R pre-trains autoregressive robotic models using low-level 4D representations from human videos, achieving efficient transfer learning and improved task performance across various environments.

LLMs can be effectively taught complex reasoning via efficient fine-tuning on demonstration data focusing on structure, not content, of the reasoning process.

ENCORE enables lifelong sequential knowledge editing in LLMs without performance loss, achieving 10,000 edits while maintaining downstream accuracy.

Reinforcement learning (RL) surpasses supervised fine-tuning (SFT) in fostering generalization in foundation models, while SFT aids RL’s stability; a comparative study across text and visual domains r…

FAST: A novel action tokenization method using discrete cosine transform drastically improves autoregressive vision-language-action models’ training and performance, enabling dexterous and high-freque…

Toto, a new autoregressive video model, achieves competitive performance across various benchmarks by pre-training on over 1 trillion visual tokens, demonstrating the effectiveness of scaling video mo…

SWE-Gym, a novel environment for training real-world software engineering agents using 2,438 real-world Python task instances, achieves new state-of-the-art performance and is publicly available.

RAPL efficiently aligns robots with human preferences using minimal feedback by aligning visual representations before reward learning.

Predicting emergent LLM capabilities is now possible by finetuning smaller models; this approach shifts the emergence point, enabling accurate predictions of future model performance, even with up to …