
Looking for the cutting-edge of AI research? Follow Salesforce AI Research to see how we're transforming enterprise technology through advanced innovations. From world models to agentic systems, discover the future of AI before it hits the market.

We advance state-of-the-art #AI techniques paving the path for innovative products at @Salesforce. Focus areas: #AIAgents, #EnterpriseAI, #EGI, and #TrustedAI.

Looking for the cutting-edge of AI research? Follow Salesforce AI Research to see how we're transforming enterprise technology through advanced innovations. From world models to agentic systems, discover the future of AI before it hits the market.

Most leaders know how to lead change. Innovation asks for something else entirely. @HarvardHBS’s Linda A. Hill on why the future gets co-created, and what that demands of the people in charge. Watch the full session on @salesforce+ → sforce.co/4zsYqe6 #FutureOfAI

📢 Service Cloud Fall Release: Moirai, our time-series foundation model, now powers agentic Workforce Engagement Management, helping service leaders predict demand and staffing for hybrid teams of human reps and AI agents. sforce.co/4ew92Ru #EnterpriseAI #AgenticAI

SFR-AutoR&D is a set of autonomous agents that run the research loop end to end. Applied to SFR-Embedding V3, the model behind Data Cloud retrieval, it compressed weeks of research into ~80 hours. An early look at self-evolving systems in #EnterpriseAI. sforce.co/4zFRevN

Ambient intelligence is one of our big bets at @SFResearch, and SOV-Transcribe-1 is where it meets the real world: open-weight models that hear accents, noisy rooms and domain terms, and turn speech into something agents can act on.
Welcome Missionforce Sovereign AI SOV-Transcribe-1: Salesforce’s new open-weight speech-augmented language model built for transcription. - accurate on emails, phone numbers & URLs - robust against accents & noisy audio - context biasing for names & domain terms #1 among open-weight models on Artificial Analysis (#7 overall). Built for real-world speech to text. Congrats @SFResearch! ❤️ cryptobriefing.com/salesforc…

The recent backlash over AI and data centers has been loud lately. Here’s how we think about developing AI responsibly and our thoughts on cost-efficient, open-source strategies.
"It's a wild time in AI," says Mick Costigan, VP, Salesforce Futures. There's been backlash in the news recently about AI and data centers for a variety of reasons. But there is a way to go about AI development that focuses on real business problems and maintains trust and safety. Hear from Costigan and Itai Asseo, VP, Strategy & Incubation - AI Research at Salesforce about what's behind this post-midterm AI backlash and how Salesforce is approaching AI in this time of change.

17/17 Attributing Agent Failures to Their History Needs an Executed Counterfactual: Toward a Reporting Standard (Workshop) A proposed standard for tracing agent failures back to their history. #NeurIPS2026 #AgenticAI #FutureOfAI

16/17 Bookkeeping, Composition, or Unreachable Gold? Reading MemoryAgentBench’s Conflict-Resolution Scores Against a Frozen Last-Write Resolver (Workshop) Rereads memory benchmark scores against a simple baseline. arxiv.org/abs/2610.09193

15/17 Does an Agent’s History Tell You When Compaction Will Hurt? A Modest, Bounded Effect on the TRACE Paired-Replay Corpus (Workshop) Tests whether an agent’s history predicts when context compaction will hurt performance. arxiv.org/abs/2610.08722

14/17 MemoryVCD: Benchmarking Multimodal Self-Evolving Memory for Personalized Vision-Language Agents (Workshop) Measures how vision-language agents build and evolve memory to personalize over time.

13/17 Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails Evolving the agent harness alongside the model, with on-policy correction closing gaps imitation leaves open. arxiv.org/abs/2609.09134

12/17 User-in-the-Loop World Simulation for Evaluating Streaming Interactive Agents Simulated worlds with users in the loop for evaluating agents that interact in real time.

11/17 Cerebro: An Enterprise Language Model for Agentic Tool Use A language model built for enterprise agents that need to call tools reliably inside real business workflows. arxiv.org/abs/2609.15066

10/17 Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction–Reality Gaps Agents predict outcomes, compare them with what actually happens, and use that gap to calibrate RL training. arxiv.org/abs/2610.02740

9/17 PACE: A Proxy for Agentic Capability Evaluation A lightweight proxy for estimating agentic capability without the cost of running full-scale agent evaluations. arxiv.org/abs/2607.02032

8/17 CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling Web agents verify their own work with conformal guarantees, strengthening both training and test-time scaling. arxiv.org/abs/2610.06829