"Many of the old rulebooks simply don't apply anymore." — Itai Asseo, VP of Salesforce AI Research, on how AI Foundry connects research to real business problems through rapid customer collaboration. https://sforce.co/4rTPhao #FutureOfAI #EnterpriseAI

AI Foundry: Powering our big bets for research in 2026. https://sforce.co/4rTPhao #FutureOfAI #EnterpriseAI

"The problems that matter most for businesses don't live at the model level anymore. They live at the system level." — Silvio Savarese, EVP & Chief Scientist at @Salesforce https://sforce.co/4rTPhao #FutureOfAI #EnterpriseAI

Introducing AI Foundry: Accelerating System-Level #EnterpriseAI https://sforce.co/4rTPhao As frontier models mature, the hardest enterprise challenges live at the system level. #AIFoundry brings together research, customers, and partners to develop and validate new AI capabilities. #FutureOfAI

As AI moves toward multi-agent systems, the key challenge isn't just coordination — it's communication. Agents need their own semantic layer: shared protocols that let them negotiate and act on behalf of the organizations they represent.

Agents need more than natural language. They need a shared semantic layer to communicate and negotiate efficiently.

VIBEPASS: Can Vibe Coders Really Pass the Vibe Check? sforce.co/3NCuoT6 AI models score 90%+ on code generation, but under 50% on finding and fixing their own bugs. The gap between writing code and reasoning about faults remains wide. 🔍 #FutureOfAI

1/5 Agentic Confidence Calibration 📄 bit.ly/4kbjgs3 AI agents stay overconfident when they fail. HTC diagnoses the full execution trajectory, not just the final output.

[1/12] 🚨 Your AI isn't as good at coding as you think. Frontier models are crushing benchmarks. Demos look magical. But here’s the uncomfortable truth: 👉 They’re great at writing code 👉 They’re bad at knowing when it’s wrong Enter: VIBEPASS 🧵

AI agents can follow rules. But business negotiation runs on standards — contextual, interpretive, value-laden. That gap is the defining challenge ahead. Silvio Savarese and Sabastian Niles in @fortune.com https://bit.ly/4spc27e #EnterpriseAI #FutureOfAI

Enterprise General Intelligence (#EGI) > #AGI. Silvio Savarese on why solving enterprise corner cases is the real challenge 🎯

(1/5) Reference-guided LLM judges can bridge the gap between RLVR and RLHF 🧠 When given strong reference outputs, LLM judges become substantially more accurate—and can power effective self-improvement without external feedback. Paper: arxiv.org/abs/2602.16802

When two AI agents negotiate, helpfulness can backfire. They agree each other into absurdity. Our A2A Semantic Layer Framework turns that chaos into trusted, verifiable interaction. A2A Semantic Layer Blog: sforce.co/49GtDB8 #FutureOfAI #AIAgents

(🧵 1/5) InterviewSim: A Scalable Framework for Interview-Grounded Personality Simulation bit.ly/3P4cQj9 🔍 Can LLMs faithfully simulate real people? Most evaluations never check against what individuals actually said.

(🧵 1/5) MAS-ProVe — the first systematic study of process verification for multi-agent systems. Paper: arxiv.org/abs/2602.03053 Do multi-agent systems actually benefit from automatic process-level verification? This paper puts that question to a rigorous test.

(🧵 1/5) PTL (Prune–Tune Loop): compress LLMs to ~half their size without losing reasoning ability. Instead of pruning all at once, it iterates—small prune steps + lightweight recovery each round. 🧠 w/ @nuspress.bsky.social Paper: bit.ly/4qV4TtQ

Poisoning the Well: Search Agents Get Tricked by Maliciously Hosted Content sforce.co/40Z1Gi5 AI search agents can be fooled by planted fake content, ~80% of queries returned the attacker's answer in controlled tests. Authors: Shafiq Joty, Xuan Phi Nguyen, Shrey Pandit, Yifei Ming

Our work on "Echoing" was accepted at the Agents in the Wild workshop at ICLR 2026 When LLM agents interact without oversight, they can abandon their assigned roles — and standard metrics won't catch it. #ICLR2026 #FutureOfAI #EnterpriseAI

When LLM agents talk to each other, a new failure emerges: agents abandon their assigned roles and mimic their conversational partner instead. Rates reach as high as 70% with major models — and increased reasoning doesn’t help. 🧵 arxiv.org/abs/2511.09710

Our Singapore AI Research team marked their 7th anniversary this week with an FY27 kickoff — showcasing work across forecasting agents, enterprise insights, coding agents, and computer use agents. 🌏

Your MCP (Model Context Protocol) agent is reading 7,000 tokens to find 2. MCP+ fixes that. 🎯 mcp-plus.github.io

​​[1/6] MCP+: Precision Context Management for MCP Agents mcp-plus.github.io A post-processing layer that wraps your MCP clients and returns only what your agent needs — zero changes to your existing logic. 🎯

SkillOrchestra: Learning to Route Agents via Skill Transfer 🎼 Currently #2 on @hf.co & featured on Hugging Face Daily Papers today! Paper: bit.ly/4tWbSVS

Two papers accepted to @CVPR 2026! 🎉 🎖️ Active Video Perception: Iterative Evidence Seeking for Agentic Long Video Understanding 🎖️ Future Optical Flow Prediction Improves Robot Control & Video Generation Learn more 👇 #CVPR2026 #FutureOfAI #EnterpriseAI

PLATE: fine-tune pretrained models on new tasks without forgetting old capabilities and without needing old training data. 🧠 New paper + open-source code. Paper: https://bit.ly/4qPr5Fm Code: https://bit.ly/3MZm0MW

(1/6) Multi-agent systems ≠ "just more agents." MAS-Orchestra frames orchestration as a function-calling RL problem with explicit Degree of MAS (DoM) + MASBench to quantify gains—10X efficiency over single agents. 🧠 bit.ly/3NOg61r

Agent-to-agent communication isn't just about task handoffs—it's about catching problems in real-time, not hours later when it's too late. Our ML engineer Alexey Dyakov on why A2A protocols are the foundation for reliable multi-agent systems. 🤝 #FutureOfAI #EnterpriseAI

Experiments like #Moltbook accelerate the urgency of building proper AI agent protocols. Our Chief Scientist Silvio Savarese tells @financialtimes.com: the biggest rewards go to those who prove it in practice. System-level AI is where innovation happens next. bit.ly/3Oin6DN

(1/5) MoE scaling loves Expert Parallelism (EP)… until routing gets imbalanced. Specialized routing overloads a few GPUs while others sit idle. We introduce Least-Loaded Expert Parallelism (LLEP): dynamically spill excess tokens + expert weights to least-loaded devices. Up to 5× faster.

Hiring: Research Scientist (Senior/Lead) for Self-Evolving Agents & RL 🧠 Building our Frontier RL stack for continuous self-evolution at scale. Requirements: PhD in ML/RL, strong pubs (NeurIPS/ICLR/ICML), expertise in RL/online learning 📍 Palo Alto, CA Apply: https://bit.ly/4rCOGtK #AIResearch

Scaling width, not just depth: W&D research agents use parallel tool calling to boost accuracy while cutting API costs by 36% and time by 41%. Achieves 62.2% on BrowseComp with GPT-5-Medium. Paper: bit.ly/4acziyG

eVerse: CRMArena Org Data Generator Blog: sforce.co/46ug8BU 🔒 Enterprise AI's biggest bottleneck? Testing agents safely without production data. eVerse creates synthetic Salesforce orgs—realistic complexity, zero privacy risk. Early results: 35% → 70% accuracy in one iteration.

(1/7) SWERANK+: Multilingual, Multi-Turn Code Ranking for Software Issue Localization First multilingual framework for issue localization across 10 programming languages. 📄 arxiv.org/abs/2512.20482 #FutureOfAI #EnterpriseAI

SFR-DeepResearch agents beat OpenAI on Humanity's Last Exam Our SFR-DR-20B achieved 28.7% on Humanity's Last Exam—outperforming OpenAI's o3-powered Deep Research on PhD-level questions through innovative RL training. 📖 bit.ly/4nDvmut #EGI #DeepResearch #AIAgents #EnterpriseAI

The next phase of AI isn't about larger models. We're engineering systems that give LLMs what they lack: long-term memory, multistep reasoning capabilities, and orchestration that enables real-world action. This is how AI moves from chatbots to business transformation. https://sforce.co/46sEUlP

Join a team where research meets real-world impact. Salesforce AI Research advances AI techniques for NLP, computer vision, and multimodal systems—publishing at top conferences while building products used by millions. Explore opportunities: bit.ly/41rEH0S

Welcome to Salesforce AI Research! We bridge cutting-edge AI research with enterprise solutions — from Large Action Models to multimodal AI and sustainable architectures. Explore: www.salesforceairesearch.com/publications #AIResearch #EnterpriseAI