This paper introduces TRACE, an FP4 quantization framework for reinforcement learning (RL) training of Mixture-of-Experts (MoE) language models. TRACE addresses the limitation of existing FP4 RL methods by using rollout-side quantization outcomes to guide training-side FP4 rounding decisions, reducing train-rollout discrepancy. The framework also includes an efficient quantization-information caching scheme to reduce storage and communication overhead. Evaluation on four large-scale MoE language models shows that TRACE achieves RL performance comparable to BF16 rollout with up to 5.4x rollout speedup and strong final FP4 performance compared to post-hoc FP4 quantization of BF16-trained policies.
The chart will appear after repeat observations. The current metric comes from the source.
11 source votesReal observations only. History before source connection is not reconstructed.
This method could be useful for improving the efficiency of reinforcement learning training for large language models by reducing computational and memory overhead while maintaining performance.
The paper introduces WING, a framework that transfers interaction knowledge from human egocentric videos to robot policies by separating observer-induced motion from hand-object interactions and using spectral analysis to identify shared temporal structures between human and robot behaviors. It achieves high success rates on multiple benchmarks and demonstrates strong performance in real-world tasks.
5 days agoDaily PapersRealCompanion introduces a benchmark for evaluating AI's ability to understand humans through long-term conversations, featuring 27,218 messages from 10 real relationships. The study reveals that most queries don't require long-term memory and that current methods struggle to identify when memory is needed.
6 days agoDaily PapersEgo2Act is a benchmark for evaluating goal-directed manipulation in egocentric video generation, featuring 2,640 videos from 110 real-world tasks. It assesses whether video generation models can create realistic egocentric videos of a hand performing multi-step object manipulations to achieve high-level goals.
6 days agoDaily PapersThe paper investigates on-policy distillation (OPD) in cross-tokenizer settings, focusing on alignment coverage and supervision reliability. It finds that strict 1:1 token alignment covers most student-generated tokens despite vocabulary mismatches, and that restricting reverse KL to a top-16 subset of shared vocabulary achieves comparable accuracy to full shared-vocabulary OPD, while adding span supervision reduces accuracy.
yesterdayDaily PapersThis paper introduces ProWAM, a progressive world action model that predicts actions and sparse visual sub-goals to guide robotic control. It improves long-horizon planning by using a single video-backbone pass for efficient action generation and demonstrates strong performance on simulation and real-world benchmarks.
6 days agoLessWrongThe text discusses a scenario where an AI must respond to the question 'What's the date?' without access to real-time data. It explores the challenge of providing an accurate date without hallucinating, considering OpenAI's training focus on avoiding false information. The text speculates on potential model versions and knowledge cutoff dates but acknowledges uncertainty due to lack of explicit information.
6 days ago