Search “Microsoft”

Research team fine-tunes Qwen3.8-27B with RL using ThinkingBox-Bench terminal-state checks as rewards

The research team performed full-parameter RL fine-tuning of Qwen3.8-27B, using the executable checks on the final database state in ThinkingBox-Bench directly as reinforcement learning rewards. This work was done by the Microsoft Copilot Studio team in collaboration with Toloka and others. On 507 out-of-distribution tasks not used in training, the number of tasks the model passed on all 20 attempts rose from 38 (7.5%) to 107 (21.1%), pass@1 rose from 51.70% to 60.78%, and the cost per successful attempt dropped from $0.177 to $0.139. The training pool used offline-policy relative-difficulty filtering, ultimately retaining 187 tasks; the training and evaluation code has been open-sourced at microsoft/thinkingbox-training.

English translation is being prepared.

View story:English translation is being prepared.

Microsoft researchers prove the KLS conjecture

Microsoft Research researcher Xinzhi Zhang and co-authors proved the KLS conjecture proposed in 1995, giving a dimension-independent fixed bound; the previous best result only allowed the bound to grow slowly with dimension. According to Microsoft Research, the proof came from a human-guided agent search process, in which the team explored more than 100 candidate paths, which researchers evaluated, pruned, and redirected. An independent team also gave a complete proof around the same time, followed by Lean 4 formal verification. The authors state that previous unguided end-to-end attempts with multiple frontier models failed to achieve the goal. Another report states that AI agents explored about 100 research paths, the draft was about 140 pages, AI was responsible for exploring candidate methods, executing derivations, assisting in writing the manuscript, and performing internal proof checks, while the recursive method and error bounds that completed the proof are credited to the two researchers.

Microsoft researchers prove the KLS conjecture

Microsoft Research researcher Xinzhi Zhang and co-authors proved the KLS conjecture proposed in 1995, giving a dimension-independent fixed bound; the previous best result only allowed the bound to grow slowly with dimension. According to Microsoft Research, the proof came from a human-guided agent search process, in which the team explored more than 100 candidate paths, which researchers evaluated, pruned, and redirected. An independent team also gave a complete proof around the same time, followed by Lean 4 formal verification. The authors state that previous unguided end-to-end attempts with multiple frontier models failed to achieve the goal. Another report states that AI agents explored about 100 research paths, the draft was about 140 pages, AI was responsible for exploring candidate methods, executing derivations, assisting in writing the manuscript, and performing internal proof checks, while the recursive method and error bounds that completed the proof are credited to the two researchers.

View story:Microsoft researchers prove the KLS conjecture

Microsoft research team publishes test results of AI diagnostic tool on complex cases

On October 23, 2026, the Microsoft research team announced the diagnostic results of an AI diagnostic tool in tests on complex cases. Phoenix Technology reported that day that Microsoft launched the AI diagnostic system Microsoft AI Diagnostic Orchestrator (MAI-DxO), which simulates a joint consultation of physician specialists by orchestrating multiple AI agents; in a simulated evaluation based on complex cases from the New England Journal of Medicine, its accuracy reached 83%, compared with 36% for primary care general practitioners and general internists who could independently consult reference materials, while specialists were not included in the comparison; in an evaluation of common U.S. emergency department cases based on the MIMIC-IV-ED dataset, its diagnostic accuracy was 1.7 times that of human physicians. Cailian Press reported the same day that the tool is powered by GPT-5 and, through a mechanism called the "Diagnostic Orchestrator," simulates multiple clinical experts analyzing collaboratively, deciding step by step which medical histories to ask about and which tests to arrange, and continuously updating diagnostic hypotheses; general practitioners without access to reference materials had an accuracy of about 22%. The research team's next step will be to collaborate with Boston's Beth Israel Deaconess Medical Center to test emergency triage capabilities; Microsoft has not yet stated the timing or form of commercialization.

Microsoft and City University of Hong Kong propose Taste-Bench to evaluate agent decision-making

Researchers from Microsoft and City University of Hong Kong have built Taste-Bench to evaluate the long-term decision-making capabilities of AI agents on long tasks. The benchmark contains 502 binary-choice questions. According to The Decoder, the best-performing model, GPT-5.6 Sol, achieved an average accuracy of only 59.7%, while GPT-5.5 reached 59.5%. When key clues only become apparent after a large amount of subsequent work, the average accuracy of all models drops to 21%, close to random guessing, and increasing the reasoning budget brings almost no improvement. The researchers also distilled the teacher model's judgments using Qwen3.6-27B as the base model; the student's accuracy on unseen tasks rose from 30% to 47.9%, and its success rate on 41 SWE-bench Pro tasks increased from 14.6% to 33.7%.

View story:Microsoft and City University of Hong Kong propose Taste-Bench to evaluate agent decision-making

Brookings Institution Releases Study on U.S. AI Infrastructure Investment

On September 26, 2026, the Brookings Institution released a study by Columbia Business School professor Stijn Van Nieuwerburgh estimating that cumulative U.S. AI infrastructure investment from 2025 to 2032 will total approximately $10.3 trillion, averaging 3.63% of U.S. GDP annually—higher than the historical investment intensity of railroad, highway, and fiber-optic construction. The report notes that capital expenditures by five companies—Oracle, Microsoft, Amazon, Meta, and Alphabet—are projected to reach $800.5 billion in 2026, exceeding their operating cash flow of $707.1 billion for the first time, with the gap reliant on borrowing and external financing. The report also mentions that the debt-to-asset ratio of Meta's Hyperion project is close to 90%, and that large cloud providers' off-balance-sheet lease and purchase commitments total approximately $2.4 trillion; IT assets such as GPUs are calculated on a 6-year economic life, and if demand growth cannot keep pace with capacity expansion, this could create pressure for declining utilization rates and lease prices. Currently, this research has been disclosed through a report by Huxiu, with no subsequent developments.

Aim Labs discloses M365 Copilot zero-click vulnerability EchoLeak

Aim Labs disclosed the zero-click vulnerability EchoLeak in Microsoft 365 Copilot. An attacker only needs to send the victim one email to exfiltrate sensitive data in the Copilot context without the user's awareness. The attack chain consists of three vulnerabilities, which sequentially bypass the XPIA prompt injection classifier, the sanitization of external links and images, and, with the help of SharePoint and Teams URLs, bypass CSP restrictions. Aim Labs calls the core technique LLM Scope Violation and believes that similar design flaws may also exist in other RAG-based chatbots and AI agents. Aim Labs said Microsoft has confirmed that no customers were affected.

English translation is being prepared.

View story:English translation is being prepared.