# Salesforce's SFR-AutoR&D Accelerates AI Model Development and Efficiency

> Salesforce's new SFR-AutoR&D agents autonomously enhance AI models, marking a pivotal shift in AI development. This technology promises to drastically reduce the time and effort required for model refinement.

**Source**: salesforce.com | **Published**: 2026-11-12 | **Type**: article

## Key Facts

- Salesforce's SFR-AutoR&D reduced model improvement time from weeks to 80 hours, enhancing agility.
- The nDCG@10 score increased by 6.13 points, indicating significant gains in retrieval relevance.
- 36% throughput increase and 17% lower latency demonstrate cost efficiency in AI operations.
- Autonomous improvements could shift competitive dynamics, favoring firms with advanced AI capabilities.
- Trust and ethical constraints in AI training may become key differentiators in enterprise software.

## Summary

Salesforce AI Research has unveiled SFR-AutoR&D, a groundbreaking set of autonomous agents designed to enhance AI model development and improve codebases. This innovation allows the agents to autonomously formulate hypotheses, conduct experiments, and implement improvements based on empirical evidence. As enterprises increasingly rely on AI for operational efficiency, this development signals a significant shift in how AI models are refined and deployed, potentially reshaping competitive dynamics in the AI market.

The SFR-AutoR&D agents have already demonstrated their capabilities by autonomously improving Salesforce’s own SFR-Embedding V3 model, which underpins data retrieval in its Data Cloud. The system also enhanced the open-source model EmbeddingGemma 2, resulting in the release of SFR-Embedding V4, which offers measurable improvements in both quality and latency. This advancement illustrates a broader trend where AI systems can now autonomously enhance the software and models that businesses depend on, reducing the time and human effort traditionally required for such tasks.

Historically, refining embedding models was a labor-intensive process, often taking weeks of human experimentation. With SFR-AutoR&D, this cycle has been compressed to approximately 80 hours. Silvio Savarese, Salesforce's EVP and chief scientist, noted that this shift represents a move from human-driven progress to a more rapid, autonomous research system that adheres to the scientific method. The implications for businesses are profound; faster iteration cycles can lead to more agile responses to market demands and improved operational efficiencies.

The agents operate within a framework that includes two specialized components: AutoR&D-Engineer, which focuses on algorithmic improvements, and TrainForge, which explores new training methodologies. Both components leverage shared GPU resources through AutoInfra, enabling extensive experimentation. This structured approach allows for rigorous validation of improvements, ensuring that only those enhancements that surpass established benchmarks are implemented.

In practical terms, the improvements achieved by SFR-AutoR&D are significant. Evaluated against ten public retrieval tasks, the system improved the model's relevance score, nDCG@10, by 6.13 points, indicating a marked enhancement in the quality of results returned to users. Additionally, a subsequent session yielded a 36% increase in throughput and a 17% reduction in latency, demonstrating a tangible impact on performance and cost-effectiveness for businesses utilizing these models.

Trust remains a cornerstone of Salesforce's model training. The company ensures that all datasets used in training have undergone ethical and legal reviews, integrating these considerations into the problem definition rather than treating them as afterthoughts. This commitment to trust and safety is crucial, particularly as businesses navigate the complexities of AI deployment in sensitive environments.

Looking ahead, the emergence of self-improving AI systems like SFR-AutoR&D suggests a future where the role of researchers and engineers will evolve. Instead of being deeply involved in the minutiae of model refinement, their focus will shift to setting parameters and constraints for autonomous frameworks. As this trend accelerates, companies that leverage these advancements will gain a competitive edge, positioning themselves as leaders in the rapidly evolving landscape of enterprise AI. The shift from weeks of human effort to hours of computational efficiency not only enhances productivity but also sets the stage for a new era of AI-driven innovation in business operations.

## Entities

- **Companies**: Salesforce
- **Products**: SFR-AutoR&D, SFR-Embedding V3, SFR-Embedding V4, EmbeddingGemma 2, pandas, hnswlib, USearch
- **Technologies**: AI, embedding models, AutoInfra
- **People**: Silvio Savarese, Shafiq Joty
- **Organizations**: Salesforce AI Research

## Key Concepts

autonomous AI research, self-improving systems, empirical validation, trust in AI models, research loop, algorithmic improvements, controlled experiments, enterprise AI

## Definitions

- **SFR-AutoR&D**: A set of agents developed by Salesforce AI Research to autonomously improve codebases and train AI models.
- **embedding models**: Models that convert text into numerical representations, enabling systems to retrieve relevant information from large datasets.
- **nDCG@10**: A metric that measures the relevance of the top 10 results returned by a model, giving more weight to higher-ranked results.
- **AutoInfra**: The training backbone used by SFR-AutoR&D agents to carry out their experiments across shared GPU capacity.
- **controlled experiments**: Experiments where variables are controlled to validate hypotheses and measure improvements against established baselines.

## Use Cases

- Training and improving production AI models
- Enhancing public codebases
- Optimizing embedding models for better retrieval quality
- Reducing latency and increasing throughput of AI models
- Ensuring ethical and legal compliance in model training
- Developing autonomous frameworks for self-improvement in AI

## Frequently Asked Questions

**What is SFR-AutoR&D?**

SFR-AutoR&D is a set of autonomous agents created by Salesforce AI Research to improve codebases and train AI models efficiently. It formulates hypotheses and conducts experiments without human intervention.

**How does SFR-AutoR&D improve AI models?**

The system autonomously tests various methods and retains only those improvements that are supported by evidence and proven to be accurate and safe. This significantly speeds up the research process.

**What role do embedding models play in enterprise AI?**

Embedding models are crucial as they convert text into numerical formats, allowing systems to retrieve relevant information from vast datasets. They are essential for maintaining operational efficiency in businesses.

**How does Salesforce ensure trust in its AI models?**

Salesforce ensures trust by using datasets that have passed ethical and legal reviews for training models. This is integrated into the problem definition, ensuring that models meet production requirements before deployment.

**What are the future implications of self-improving AI systems?**

Self-improving AI systems are expected to become standard, shifting the role of researchers to setting goals and constraints. This evolution will enhance the efficiency and reliability of AI development in enterprises.

## Links

- [Read on Welcome.AI](https://welcome.ai/content/salesforces-sfr-autord-accelerates-ai-model-development-and-efficiency)
- [Original source](https://www.salesforce.com/news/stories/agents-that-improve-ai-models/)

---

Source: Welcome.AI | https://welcome.ai/content/salesforces-sfr-autord-accelerates-ai-model-development-and-efficiency