SFR-AutoR&D: autonomous research and development from Salesforce AI Research

How SFR-AutoR&D works

Start with an R&D goal, an artifact to work on, and a budget. The agent develops a method, tests it, and uses the evidence to decide what comes next.

Input

A research goal

What could work better?

ArtifactRepository or model

TargetA measurable gain

BoundsBudget & constraints

SFR-AutoR&D Autonomous R&D

Output

Verified progress

A better solution, with proof.

A new method
Patch or checkpoint
Reproducible evidence

Today, we’re introducing two specialized R&D agents.

SFR-AutoR&D

AutoR&D-Engineer

Explore AutoR&D-Engineer

Discover better methods. Implement them. Prove the improvement.

Input Repository + research goalSystem, target & constraints

AutoR&D-Engineer

Output Verified implementationPatch + benchmark evidence

Research highlights

pandasNullable-dtype reductionsup to1.84×faster reductions Salesforce SFR-RLTwo improvements · separate workloads Serving · up to3.14×generated tokens / secondTraining−39%peak training memory

Vector search

USearchSingle-core · GCC 11 · 1M vectors+17.5%more queries / second hnswlibMedian across 3 index builds+4.1%more queries / second

EmbeddingGemma 2Embedding model trainingMTEB (eng, v2) · 10 retrieval tasks+6.13retrieval nDCG@10 points

Gains are measured against each workload’s baseline. Open a result for its setup, correctness checks, and benchmark evidence.

SFR-AutoR&D

TrainForge

Explore TrainForge

Invent training methods. Run the experiments. Keep the best model.

Input Baseline model + targetTask, evaluation & compute budget

TrainForge

Output Best verified checkpointModel + evaluation record

Research highlights

Qwen3-4BHard mathematical reasoningHeld-out · Omni-MATH · AIME 2025 · AIME 2026up to+11percentage points Qwen3.5-35B-A3BCode generation & repository repairHeld-out · LiveCodeBench Hard · SWE-bench Liveup to+12percentage points

GPT-OSS-20BLong-horizon research qualityHeld-out · FRAMES · Official GAIA · HLE · BrowseCompup to+14percentage points Guided end-to-end demonstrationQwen3-4B · guided walkthroughWatch demo

Each result comes from a novel training strategy discovered by TrainForge. Open a program for its held-out benchmarks and method-level results.

Shared training infrastructure

AutoInfra supports the research loop.

AutoInfra provides shared training and inference infrastructure for SFR-AutoR&D. TrainForge uses it to launch and compare model experiments in parallel. It also supports training the models that power TrainForge and AutoR&D-Engineer, while validated improvements from either agent can be incorporated into AutoInfra for the next research cycle.

Explore AutoInfra

What it means in practice and for Salesforce.

Improve codebases and solutions

Use AutoR&D-Engineer to discover and implement better algorithms in existing systems, from data pipelines to retrieval engines. Validate gains in speed and memory efficiency while preserving correctness.

Train better models

Use AutoR&D-TrainForge to explore training methods and optimizers through controlled experiments. Identify what improves model quality and training efficiency within a defined compute budget.

Advance scientific discovery in AI Research

Use AutoR&D to explore open research questions, form hypotheses, and develop new algorithms and methods. Help Salesforce AI Research turn promising ideas into reproducible findings that inform the next experiment.