SERIES
Researchers set out a system-level framework for work agents, dividing digital work into categories and identifying capability groups that they say are needed for reliable execution.
The researchers built a fleet-learning system that pulls supervision from deployed robots’ own executions, a sign policy improvement can continue without new expert demonstrations.
Researchers tested six language models with six prompting interventions and found that explicit no-think controls did not reliably stop visible inference, especially on open-ended tasks.
The researchers propose TokenRouter, a serving system that routes tokens across multiple language models with asynchronous subservers and delayed batching, and they report higher decoding throughput in their tests.
Researchers proposed disaggregated quantization for large language models, using different formats and weights for prompt processing and token generation, and reported accuracy and latency gains in tests on Qwen 3 and Gemma 3.