What if AI could turn a research goal into a better model—and prove the improvement is real?
That’s what we’re building with TrainForge: an autonomous R&D system that generates ideas, implements them, runs training and evaluation, and learns from the outcome. Built by Salesforce AI Research.
Autonomous R&D
No improvement means no promotion.
Why this is new
TrainForge discovers training ideas, tests them, and turns every result into the next, better proposal—while verified code and checkpoint-bound evidence keep the loop under control.
Existing methods are references, not a fixed menu. Agents search broadly, debate alternatives, and redesign the full improvement loop around each target. Each proposal remains a novelty candidate until prior-work checks and experiments support it.
search broadly → design end to end → test noveltyNo repository or recipe is provided. TrainForge chooses what the codebase needs, Codex writes it from an empty workspace, and checks verify that the running code still matches the idea.
freeze intent → write from scratch → verifyIdeas start with a fast LoRA test. Only promising ones move to a short full-parameter run and then full training.
LoRA test → short full-parameter run → full trainingEvery result—success or failure—changes the next proposal, so the method improves across rounds.
test → learn → regenerateTasks and evaluations
You provide the target. TrainForge chooses the research direction and trains on separate training data; the datasets below are frozen, held-out evaluations only. The effects shown are averages across the listed benchmarks.
Task
One research direction chosen by TrainForgeCan models preserve reasoning structure—not just final answers?
Select an approach to view its details.
Task
One research direction chosen by TrainForgeCan a repair preserve program semantics—not just pass visible tests?
Select an approach to view its details.
Task
One research direction chosen by TrainForgeCan a research agent learn what to search, what to trust, which steps mattered, and when to stop?
Select an approach to view its details.
Guided demonstration
See the complete story in one interface: diagnose the gap, generate and challenge ideas, steer the system, verify the code, train, evaluate, and promote.
The idea-evolution lineage is illustrative. The two rejected screens and the successful Qwen3-4B reference result are real measurements and are labeled in the demo.
