Salesforce AI Research
@sfresearch.bsky.social
We advance state-of-the-art #AI techniques paving the path for innovative products at @Salesforce. Focus areas: #AIAgents, #EnterpriseAI, #EGI, and #TrustedAI.
Authors: Xiaoqiang Lin, Jun Hao Liew, Silvio Savarese, and Junnan Li at Salesforce AI Research.
February 11, 2026 at 3:12 PM
Authors: Xiaoqiang Lin, Jun Hao Liew, Silvio Savarese, and Junnan Li at Salesforce AI Research.
(7/7) Authors: Revanth Gangi Reddy, Ye Liu, Wenting Zhao, JaeHyeok Doo, Tarun Suresh, Daniel Lee, Caiming Xiong, Yingbo Zhou, Semih Yavuz, Shafiq Joty, University of Illinois at Urbana-Champaign, Salesforce AI Research and KAIST AI
#FutureOfAI #EnterpriseAI
#FutureOfAI #EnterpriseAI
February 10, 2026 at 2:27 AM
(7/7) Authors: Revanth Gangi Reddy, Ye Liu, Wenting Zhao, JaeHyeok Doo, Tarun Suresh, Daniel Lee, Caiming Xiong, Yingbo Zhou, Semih Yavuz, Shafiq Joty, University of Illinois at Urbana-Champaign, Salesforce AI Research and KAIST AI
#FutureOfAI #EnterpriseAI
#FutureOfAI #EnterpriseAI
(6/7) The agentic approach mirrors developer workflows: broad exploration → context building → converging to root cause. Combines efficiency of specialized retrievers with depth of agentic reasoning.
📄 arxiv.org/abs/2512.20482
📄 arxiv.org/abs/2512.20482
February 10, 2026 at 2:27 AM
(6/7) The agentic approach mirrors developer workflows: broad exploration → context building → converging to root cause. Combines efficiency of specialized retrievers with depth of agentic reasoning.
📄 arxiv.org/abs/2512.20482
📄 arxiv.org/abs/2512.20482
(5/7) Results show significant improvements: → SWE-PolyBench: 58.28% Acc@10 → SWE-Bench-Multilingual: 64.10% Acc@10 → Multi-turn reasoning: 3-6 point boost over single-pass → Effective for low-overlap cases
📄 arxiv.org/abs/2512.20482
📄 arxiv.org/abs/2512.20482
February 10, 2026 at 2:27 AM
(5/7) Results show significant improvements: → SWE-PolyBench: 58.28% Acc@10 → SWE-Bench-Multilingual: 64.10% Acc@10 → Multi-turn reasoning: 3-6 point boost over single-pass → Effective for low-overlap cases
📄 arxiv.org/abs/2512.20482
📄 arxiv.org/abs/2512.20482
(4/7) SWERANKAGENT uses iterative, multi-turn reasoning for complex bugs: → Progressive searches across turns → Memory buffer of relevant functions → Query refinement from intermediate results → Evidence aggregation before final ranking
February 10, 2026 at 2:27 AM
(4/7) SWERANKAGENT uses iterative, multi-turn reasoning for complex bugs: → Progressive searches across turns → Memory buffer of relevant functions → Query refinement from intermediate results → Evidence aggregation before final ranking
(3/7) SWERANKMULTI extends code ranking to 10 languages using SWELOCMULTI dataset (155K instances, 4K+ repos): → Bi-encoder retriever → LLM reranker → JavaScript, Java, TypeScript, Ruby, Rust, Go, PHP, C, C++, Python
📄 arxiv.org/abs/2512.20482
📄 arxiv.org/abs/2512.20482
February 10, 2026 at 2:27 AM
(3/7) SWERANKMULTI extends code ranking to 10 languages using SWELOCMULTI dataset (155K instances, 4K+ repos): → Bi-encoder retriever → LLM reranker → JavaScript, Java, TypeScript, Ruby, Rust, Go, PHP, C, C++, Python
📄 arxiv.org/abs/2512.20482
📄 arxiv.org/abs/2512.20482
(2/7) Software issue localization maps bug reports to code functions needing fixes. As codebases grow across languages, manual localization becomes infeasible. SWERANK+ addresses this with multilingual ranking and agentic search.
February 10, 2026 at 2:27 AM
(2/7) Software issue localization maps bug reports to code functions needing fixes. As codebases grow across languages, manual localization becomes infeasible. SWERANK+ addresses this with multilingual ranking and agentic search.