Salesforce AI Research (@SFResearch) header
Salesforce AI Research (@SFResearch) profile picture

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

Palo Alto, CA
Joined September 2014
Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)Photo by Salesforce AI Research (@SFResearch)
Salesforce AI Research retweeted
📢 Excited to share our latest work from @SFResearch: Learning Generalizable Behaviors for Terminal Agents 🤔 Q1: How does RL actually help terminal agents generalize? → Not by teaching new skills. It reshapes behavior. • Skills (write the regex, call the tool) are already there from pre-training + SFT • Behaviors (inspect first, verify, stop looping) are what RL shapes, and what predicts success 🔍 Q2: Can you trust your synthetic RL environments? → Often not. The reward can be wrong in both directions. • The cleanest public collection we audited: 35.8% clean • Reward 1 for copying a leaked answer. Reward 0 for a correct solution 🌊 So we built RIVER → Filter defective envs, penalize repetitive loops, then run RL. • River-8B: best among evaluated open RL-trained 8B models on all 4 terminal benchmarks • 2B→27B with <30% of the envs: RL gains +106% on Terminal-Bench-Lite, +30% on Terminal-Bench-v2.1 • Same 3.5K-env budget: filtered 19.4 vs. randomly sampled 17.7 ✅ Takeaway: audit your verifiers before you scale your envs. 🙌 Led by @YYYao45 and @bo_pang0 , with @xuanphinguyen , @zhao__ding and @JotyShafiq 📄 Paper: arxiv.org/abs/2608.22631 🌐 Website: self-evolving-agents.salesfo… 🧵👇
3
7
2
34
6,444
Salesforce AI Research retweeted
We announced Koa at #DF26 this week. #EnterpriseAI runs on governed, repeatable work, and that takes domain-specific reasoning. Built on @Nvidia's model, @SFResearch trained Koa on 27 years of enterprise workflows. @justjayesh's and my piece for more: sforce.co/4rhI9WA
1
4
1
22
2,223
Salesforce AI Research retweeted
Excited to be at #CVPR2026 this week and to present my internship work with @SFResearch: Active Video Perception: Iterative Evidence Seeking for Agentic Long Video Understanding In this work, we study how multimodal agents can actively reason over long videos by iteratively seeking the most relevant evidence, rather than passively processing all video content at once. If you’re attending CVPR, feel free to stop by our poster! 📍 #245, Findings Posters, ExHall A 📅 Sunday, June 7 🕢 7:30 – 9:00 AM Project page: activevideoperception.github… Looking forward to connecting at CVPR! #CVPR2026 #ComputerVision #MultimodalAI #VideoUnderstanding #AIAgents
1
5
14
1,479
Salesforce AI Research retweeted
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! 📍
2
1
6
3,017
Salesforce AI Research retweeted
AI got real in 2025. 💫 One-off agents evolved into the Agentic Enterprise, AI systems began working together in real business workflows, and so much more. Our top 10 AI new stories of the year: sforce.co/3KRcASN
3
4
9
2,782
Trust is the #1 bottleneck for scaling AI. Salesforce leaders agree that the future belongs to companies that can integrate trust natively across their agentic architecture. See what execs think this will look like in 2026: sforce.co/4pJvJVS
1
1
4
591
Salesforce is leading the shift to the Agentic Enterprise by bringing together apps, data, and agents into one deeply unified platform. Discover the product updates, investments, and acquisitions that propelled the company forward in Q3: salesforce.com/news/stories/…
1
5
895
Salesforce AI Research retweeted
Salesforce AI Research (@SFResearch) needed a trusted partner to verify multi-step reasoning chains and the evaluation required: -Strict binary judgments (Correct/Incorrect) for each reasoning step -Zero tolerance for carry-forward errors; if one step was incorrect, all dependent steps were marked incorrect - Short written justifications for each step, explaining correctness judgments -High reviewer alignment due to the interpretive difficulty of parallel versus dependent reasoning branches Turing’s role was to execute high-precision annotation, reviewer alignment, and delivery at scale, following Salesforce’s internal evaluation rubric. -The Approach: Turing evaluated over 200 long-form math responses, each containing 8–15 structured reasoning steps, generated by frontier reasoning models. For every step, annotators issued a binary Correct/Incorrect judgment and authored a short written justification, following a strict zero-tolerance rubric: - Key Results: Annotated 200+ model responses, each broken down and evaluated step-by-step Applied error-carry-forward logic and justification writing with full reviewer alignment Achieved 100% delivery pass-through via Salesforce’s internal 3-tier review pipeline Maintained average annotation time of 90 minutes per model output and 63 minutes per review - The Outcome: Salesforce used this dataset to: Evaluate LLM performance on open-domain math tasks with strict reasoning fidelity Detect early-stage reasoning flaws that derail multi-step logic chains Align model outputs with human-grade reasoning standards for future fine-tuning and scoring research Need a human-led verifier for complex math reasoning? Partner with Turing to annotate symbolic CoT logic with binary step grading and reasoning trace validation.
1
6
1
18
31,987
Salesforce AI Research retweeted
Thrilling day keynoting @Gartner Symposium in Orlando with 10,000 customers! 🌟 The Agentic Enterprise era has begun — where AI empowers people and redefines how organizations think, work, and trust. Still riding the Dreamforce high. 🤖💙 ciodive.com/news/Salesforce-…
20
26
11
121
21,871
Salesforce AI Research retweeted
🚀 I'm hiring AI Research Interns @Salesforce for Summer 2026! Work on frontier AI problems: AI agents, multi-agent systems, RL, LLMs, and more. You’ll collaborate with our AI Research team (and me!) to publish impactful research in a top-tier AI conference. Some of our previous intern-led projects: ‣ 𝗧𝗔𝗖𝗢 🌮 (#EMNLP 2025), a multi-modal action model featured in VentureBeat and more. ‣ 𝗨𝘀𝗲𝗿𝗕𝗲𝗻𝗰𝗵 (Submitted), a new benchmark for evaluating AI agents on real-world user conversations. ‣ 𝗧𝗼𝗼𝗹𝗟𝗶𝗯𝗚𝗲𝗻 (Submitted), a new approach for creating tools for agents to use. Apply here 👉 tinyurl.com/4f7m2zym Questions? Ask below or DM.
19
43
4
381
26,379
Salesforce AI Research retweeted
Sooo, same time next year? 🗓️ #DF26: September 15–17, 2026. Save the date: sforce.co/4qdS7HL
8
19
4
88
12,578
Salesforce AI Research retweeted
1T OpenAI tokens. Adoption matters. Humans & Agents working together. #Dreamforce | #Agentforce #1 AI CRM See it at Dreamforce. #DF25
16
16
5
97
36,618
Salesforce AI Research retweeted
VLM2Vec-V2: A Unified Computer Vision Framework for Multimodal Embedding Learning Across Images, Videos, and Visual Documents VLM2Vec-V2 is a new multimodal embedding model that unifies image, video, and visual document retrieval within a single framework. Leveraging the advanced Qwen2-VL backbone—featuring dynamic resolution handling, multimodal rotary position embedding, and unified 2D/3D convolution—VLM2Vec-V2 enables effective learning and representation across diverse visual modalities for real-world tasks like multimedia search, AI assistants, and retrieval-augmented generation. Full Analysis: marktechpost.com/2025/07/27/… To evaluate and train this model, researchers developed MMEB-V2, an extensive benchmark that expands prior datasets to include five new task types: visual document retrieval, video retrieval, temporal grounding, video classification, and video question answering. Experimental results show that VLM2Vec-V2 achieves strong, balanced performance across 78 tasks covering text, images, videos, and documents, establishing a new standard for scalable and flexible multimodal representation learning Paper: arxiv.org/abs/2507.04590 GitHub Page: github.com/TIGER-AI-Lab/VLM2… Model on Hugging Face: huggingface.co/VLM2Vec/VLM2V… @SFResearch
9
28
1,414
Salesforce AI Research retweeted
🎆Headed to #ICLR25? From efficient key caches and multimodal embeddings to self-improving reasoning and faithful context adherence, we're thrilled to present and discuss a broad range of #AIResearch. Come see us at Booth # G03, and don’t forget to #BookmarkMe 🔖 💥💥💥 #ICLR25 Poster Sessions VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks 🔖 Paper: arxiv.org/abs/2410.05160 Thursday, April 24 | 3:00pm - 5:30pm BingoGuard: LLM Moderation Tools with Explicit Risk Levels 🔖 Paper: arxiv.org/abs/2503.06550 Friday, April 25 | 10:00am - 12:30pm FaithEval: Can Your Language Model Stay Faithful to Context, Even If “The Moon is Made of Marshmallows'' 🔖 Paper: arxiv.org/abs/2410.03727 Friday, April 25 | 10:00am - 12:30pm Integrating Expertise of Software Engineering Agents 🔖 Paper: arxiv.org/abs/2408.07060 Friday, April 25 | 10:00am - 12:30pm ThinK: Thinner Key Cache by Query-Driven Pruning 🔖 Paper: arxiv.org/abs/2407.21018 Saturday, April 26 | 10:00am - 12:30pm Automatic Curriculum Expert Iteration for Reliable LLM Reasoning 🔖 Paper: arxiv.org/abs/2410.07627 Saturday, April 26 | 10:00am - 12:30pm REGENESIS: LLMs can grow into reasoning generalists via self improvement 🔖 Paper: arxiv.org/abs/2410.02108v1 Saturday, April 26 | 10:00am - 12:30pm Saturday, April 26 | 3:30pm - 5:00pm SiReRAG: Indexing Similar and Related Information for Multihop Reasoning 🔖 Paper: arxiv.org/abs/2412.06206 Saturday, April 26 | 3:00pm - 5:30pm Preference Optimization for Reasoning with Pseudo Feedback 🔖 Paper: arxiv.org/abs/2411.16345 Saturday, April 26 | 3:00pm - 5:30pm Bridging the Data Provenance Gap Across Text, Speech, and Video 🔖 Paper: arxiv.org/abs/2412.17847 Saturday, April 26 | 3:00pm - 5:30pm 🔥🔥🔥 #ICLR25 Workshop Sessions Workshop on Scaling Self-Improving Foundation Modes VLM Q-Learning: Aligning Vision-Language Models for Interactive Decision-Making Sunday, April 27 | Time TBD Location: Garnet 214—215 🔖 Paper:openreview.net/pdf?id=BJMpix… Workshop on Deep Learning for Code CLOVER: A Test Case Generation Benchmark with Coverage, Long-Context, and Verification Monday, April 28 | Time TBD Garnet 218-219 🔖 Paper: arxiv.org/abs/2502.08806 Workshop on Building Trust in Language Models and Applications ToolScan: A Benchmark For Characterizing Errors In Tool-Use LLMs Monday, April 28 | 9 a.m Hall 4 #6 🔖 Paper: arxiv.org/abs/2411.13547 Workshop on Reasoning and Planning for LLMs TACO: Learning Multi-modal Models to Reason and Act with Synthetic Chains-of-Thought-and-Action Workshop time and date TBD 🔖 Paper: arxiv.org/abs/2412.05479 Workshop on Reasoning and Planning for LLMs BOLT: Bootstrap Long Chain-of-Thought in Language Models without Distillation Workshop time and date TBD Paper: arxiv.org/abs/2502.03860 🇸🇬 We'll see you soon, Singapore!
1
4
13
1,704
Salesforce AI Research retweeted
As large language models increasingly serve as judges for evaluating other models during both development and deployment, most existing benchmarks still focus on non-contextual tasks like chat completions or logical reasoning. Our team present ⚖️ ContextualJudgeBench ⚖️ — a benchmark designed to evaluate LLM-as-judge capabilities in context-rich scenarios such as RAG-based QA and summarization. In real-world business applications, context is everything—AI systems must provide accurate and timely information. But assessing their performance is challenging: judge models need to interpret both the context and the question, then evaluate accuracy, justification for refusals, and overall quality, including completeness and conciseness. 📘 Paper: arxiv.org/abs/2503.15620 📊 Data: huggingface.co/datasets/Sale… 💻 Code: github.com/SalesforceAIResea…
1
16
1
29
4,117
Salesforce AI Research retweeted
📣 Today 3/17 at 4:10 PM PT, join us for the 7th Advanced LLM Agents MOOC lecture on Multimodal Agents – From Perception to Action @CaimingXiong @SFResearch @salesforce. 🌐 Join the thriving community of the LLM Agents MOOC series, with 22K+ registered learners & 9K+ members on Discord! 🚀 Register NOW for the AgentX Competition by @BerkeleyRDI @UCBerkeley, w. sponsors @Amazon @huggingface @LambdaAPI @MistralAI @Google @GroqInc @schmidtsciences, and VC partners @Accel @BainCapVC @BessemerVP @lightspeedvp @MayfieldFund @NEA! Exciting announcements on prizes/credits/resources and more coming soon!
2
11
80
6,550
Salesforce AI Research retweeted
A Coding Guide to Build a Multimodal Image Captioning App Using Salesforce BLIP Model, Streamlit, Ngrok, and Hugging Face [COLAB NOTEBOOK INCLUDED] In this tutorial, we’ll learn how to build an interactive multimodal image-captioning application using Google’s Colab platform, Salesforce’s powerful BLIP model, and Streamlit for an intuitive web interface. Multimodal models, which combine image and text processing capabilities, have become increasingly important in AI applications, enabling tasks like image captioning, visual question answering, and more. This step-by-step guide ensures a smooth setup, clearly addresses common pitfalls, and demonstrates how to integrate and deploy advanced AI solutions, even without extensive experience.... Full Tutorial: marktechpost.com/2025/03/13/… Colab Notebook: colab.research.google.com/dr… @SFResearch @streamlit @huggingface
8
1
15
2,204
Salesforce AI Research retweeted
AI is EXPENSIVE! AI is NOT FREE! Meanwhile, @salesforce just announced that Agentforce is now available for FREE as part of Developer Orgs! This is walking the talk on LOWERING THE ENTRY BARRIER on AI!👏🏻 Well done Salesforce!🚀 #Salesforce #Agentforce #Ai #Agent #AgenticAI
7
19
1
94
3,039
Salesforce AI Research retweeted
ProGen: LLMs generate functional protein sequences across families. Congrats @SFResearch! ❤️🧬 nature.com/articles/s41587-0…
3
14
2
81
26,092