AISFR-AutoR&D aims to automate the AI research loop
SFResearch says its system combines code optimization, training-method discovery and scalable infrastructure. It reports TrainForge gains of 11–14 percentage points on held-out math, code and research benchmarks.
TLDR
SFResearch introduced SFR-AutoR&D as a step toward agents that choose experiments and test whether they work. It says AutoR&D-Engineer optimizes code, reporting 3.14× faster generation in its reinforcement-learning stack.
The team says TrainForge writes new training methods from an empty workspace, tests ideas and promotes a checkpoint only if it beats the baseline on frozen, held-out evaluations. Reported gains average 11 percentage points for math, 12 for code and 14 for research agents across each task’s held-out benchmarks.
SFResearch says AutoInfra supplies the infrastructure for those experiments, including full-parameter training of 1-trillion-parameter models with 1-million-token context windows.
