Salesforce AI Research

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.

Correction: Congratulations to @PhilippeLaban!

Our Infrastructure & DevOps team was featured by @GoogleCloud. Lavanya Karanam and Avinash Gudagi on how @Salesforce is solving one of the hardest problems in large-scale AI training: keeping powerful GPUs fully fed. bit.ly/4n6HDc3 #FutureOfAI #EnterpriseAI

3 AI trends shaping #EnterpriseAI: simulation environments, ambient intelligence, and agent-to-agent communication. @Digitai79 on our focus on Enterprise General Intelligence (#EGI) in @gregorojstersek's #TDX26 recap. bit.ly/4mT4irW #FutureOfAI

Don't Stop Early: Scalable Enterprise Deep Research with Controlled Information Flow and Evidence-Aware Termination Paper: bit.ly/49fk2zQ Enterprise deep research often fails to produce decision-ready reports due to uneven coverage, context explosion, and premature stopping. This work introduces an EDR architecture built on three mechanisms: → 🗂️ Coverage-driven task decomposition via outline generation with reflection → 🔗 Dependency-guided execution with localized contexts, so information flows only when explicitly required → ✅ Evidence-based termination criteria that keep agents gathering until sufficiency conditions are met 🏆 Results: 82.09 HAA and 4.31 coverage on internal sales enablement, outperforming Gemini DR, OpenAI DR, ODR, and DeerFlow. Highly competitive in comprehensiveness (54.85) and insightfulness (55.63) on DeepResearch Bench. 👥 Authors: Prafulla Kumar Choubey, Kung-Hsiang Huang @steeve__huang, Pranav Narayanan Venkit, Jiaxin Zhang @jxzhangjhu, Vaibhav Vats, Yu Li, Xiangyu Peng, Chien-Sheng Wu @jasonwu0731 #FutureOfAI #EnterpriseAI #AgenticAI

Heading to the DATA-FM workshop at #ICLR2026? 🦾 Catch my invited talk at 1:30 PM today in Room 203A+B: "Agentic Ambient Intelligence: Perception, Reasoning & Action." I’ll be diving into 4 new research pillars for moving AI into the physical world. Come say hi! 📍

Congratulations to our researchers on winning an Outstanding Paper Award at #ICLR2026 for LLMs Get Lost in Multi-Turn Conversation! 🏆 Authors: Philippe Laban @plaban, Hiroaki Hayashi @hiroakiLhayashi, Yingbo Zhou @yingbozhou_ai, Jennifer Neville bit.ly/3ZoC6T1https:/…

Attending #ICLR2026? Meet us at @Salesforce Booth #203 to hear the latest on how we're evolving enterprise general intelligence. For more, follow our journey at ICLR: bit.ly/4vLy9qohttps:/… #FutureOfAI #EnterpriseAI

📍 Live from Rio: 21 papers at #ICLR2026, starting today. Our work spans agent architectures, reasoning, evaluation, deep research reliability, and scalable RL, tackling what matters most for #EnterpriseAI. sforce.co/4vRgqOs #FutureOfAI

Can AI coding assistants maintain effectiveness as codebases scale 100×? LoCoBench-Agent evaluates agents across 10K to 1M token contexts, spanning 8,000 scenarios in 10 programming languages and four difficulty tiers. sforce.co/4txCDj4 👥: @_Jason_Q @huan__wang #AgenticAI

The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillation: bit.ly/48iccVY On-policy distillation (OPD) improves task accuracy but systematically traps models in severe overconfidence. We trace this to an information mismatch between training and deployment, and introduce CaOPD to fix it. → Identifies a pervasive Scaling Law of Miscalibration: even frontier LLMs exhibit massive calibration gaps that scale does not resolve → Formalizes how privileged teacher context induces entropy collapse and optimism bias in the student → Replaces self-reported confidence with a student-grounded empirical target, decoupling what the model answers from how certain it should be → Achieves Pareto-optimal calibration without the capability tax of RL-based methods, enabling a compact 8B model to rival frontier LLMs on reliability Code: bit.ly/4cUtCKO Authors: Jiaxin Zhang @jxzhangjhu, Xiangyu Peng @beckypeng6, Qinglin Chen, Qinyuan Ye @qinyuan_ye, Caiming Xiong @CaimingXiong, Chien-Sheng Wu @jasonwu0731 #FutureOfAI #EnterpriseAI

At #TDX26, Itai Asseo @Digitai79 on Enterprise General Intelligence: refining generic LLMs into reliable enterprise agents through AI Foundry and Agentforce Labs, including eVerse, the Learning Engine, and Agent Startup. sforce.co/3OrRYCb #FutureOfAI #EnterpriseAI #EGI

(7/7) Lost in Translation: Do LVLM Judges Generalize Across Languages? This work examines whether large vision-language model judges maintain evaluation reliability across languages, investigating cross-lingual generalization gaps in automated assessment systems. Authors: Md Tahmid Rahman Laskar, Mohammed Saidul Islam, Mir Tafseer Nayeem, Amran Bhuiyan, Mizanur Rahman, Shafiq Joty @JotyShafiq, Enamul Hoque, Jimmy Huang Accepted to #ACL2026

(6/7) Discovering the Gems in Early Layers: Accelerating Long-Context LLMs with 1000x Input Token Reduction: bit.ly/4mQq0gj GemFilter uses early LLM layers as filters to identify and compress relevant input tokens before full processing. The method achieves a 2.4x speedup and 30% reduction in GPU memory usage compared to state-of-the-art approaches, with negligible accuracy loss. Authors: Zhenmei Shi, Yifei Ming @ming5_alvin, Xuan-Phi Nguyen @xuanphinguyen, Yingyu Liang, Shafiq Joty @JotyShafiq Accepted to #ACL2026

(1/7) @Salesforce AI Research has 6 papers accepted to ACL 2026, advancing work across web agent evaluation, LLM reasoning verification, uncertainty quantification, long-context efficiency, and multilingual judge systems. ACL 2026 takes place July 2-7 in San Diego, California. #ACL2026 #FutureOfAI #EnterpriseAI

That's a wrap on #TDX26! The @Salesforce AI Research team was on the ground in San Francisco, connecting with the community and sharing the latest from our #AIFoundry. From research demos to conversations about what's next in #EnterpriseAI, it was great to be part of the energy!