Microsoft Research (India)

In Paper 17.4, the architecture and implementation of Microsoft's MAIA AI silicon, a reticle-scale 750W AI SoC, is presented. Innovation across power delivery, thermal management, physical design, validation, and packaging is detailed, resulting in a scalable approach to high-performance AI acceleration.

Spreadsheet programming is challenging. Programmers use spreadsheet programming knowledge (e.g., formulas) and problem-solving skills to combine actions into complex tasks. Advancements in large language models have introduced language agents that observe, plan, and perform tasks, showing promise for spreadsheet creation. We present TableTalk, a spreadsheet programming agent embodying three design principles—scaffolding, flexibility, and incrementality—derived from studies with seven spreadsheet programmers and 85 Excel templates. TableTalk guides programmers through structured plans based on professional workflows, generating three potential next steps to adapt plans to programmer needs. It uses pre-defined tools to generate spreadsheet components and incrementally build spreadsheets. In a study with 20 programmers, TableTalk produced higher-quality spreadsheets 2.3 times more likely to be preferred than the baseline. It reduced cognitive load and thinking time by 12.6%. From this, we derive design guidelines for agentic spreadsheet programming tools and discuss implications on spreadsheet programming, end-user programming, AI-assisted programming, and human-agent collaboration.

PurposeTo evaluate the utility of CataractBot, an LLM (Large Language Model)-powered chatbot that provides doctor-verified answers to patient questions about cataract surgery. We examine its use by both end-users (patients and attendants) and medical experts.MethodsA 24-week study was conducted to evaluate CataractBot among patients, their attendants, doctors, and patient coordinators. The bot responded instantly to questions by querying a knowledge base curated by medical professionals. Each response was asynchronously verified by an ophthalmologist (for medical questions) or a patient coordinator (for logistical questions), and their edits contributed to updating the knowledge base, thereby minimizing future expert intervention. A mixed-methods analysis was conducted on interaction logs, including patient and attendant questions, chatbot answers, and expert verifications.ResultsA total of 318 patients and attendants sent 1,992 messages, and LLM-generated answers were verified by five doctors and two coordinators. Questions asked pre-surgery were significantly more than post-surgery . Participants asked significantly more medical than logistical questions . Doctors rated 84.5% of CataractBot's answers to medical questions as accurate and complete. Their edits, which mainly involved adding information, increased the acceptance of the bot's answers by 19.0% over time.ConclusionCataractBot was predominantly used to address medical questions. It incorporated expert corrections to improve its answers and reduce the experts' bot-related workload over time. This study highlights the potential of LLM-powered chatbots to support patient-provider communication in ophthalmology.

Large multimodal models (LMMs) demonstrate impressive capabilities in understanding images, videos, and audio beyond text. However, efficiently serving LMMs in production environments poses significant challenges due to their complex model architectures and heterogeneous characteristics across their multi-stage inference pipelines and modalities.

• Pusa N Doctor app enhances real-time N prescription in maize, ensuring yield parity. • N50PK + App shows optimal yields, high N uptake, and reduced energy consumption. • N50PK + App boosts N and energy use efficiency while minimizing GHG emissions. • N50PK + App outperforms RDF with higher N recovery and lower nitrate leaching. • Balanced N application with Pusa N Doctor reduces CO2 and N2O emissions effectively. During previous implementation of dark green colour index (DGCI), conventional tools were found inadequate for providing accurate nitrogen (N) recommendations in maize. In contrast, camera-based DGCI methods demonstrated greater effectiveness in predicting the in-season N requirements. To address this incongruity, maize leaf images were captured; resized; white balance corrected; followed by selection of region of interest; normalization; red, green, blue channel extraction; conversion into hue, saturation, brightness spectrum and calculation of DGCI. Simultaneously, NDVI, SPAD, LCC and leaf N% data were collected; and correlated with DGCI; followed by performance analysis; DGCI-based N prescription algorithm development; and its incorporation in “Pusa N Doctor” app developed using Android studio with JNI and Android NDK. The app validation was carried out using basal N dose of 0 kg ha −1 (N 0 PK), 50 kg ha −1 (N 50 PK), & 75 kg ha −1 (N 75 PK) including 2 split applications of N at 35 & 45 days after sowing (DAS) as per Pusa N Doctor & GreenSeeker™ recommendation. The treatments directed by the mobile application (app-guided) (N 50 PK + App and N 75 PK + App) were evaluated with reference to standard RDF as well as those managed using GreenSeeker™ (N 50 PK + GS TM and N 75 PK + GS TM ). The N 50 PK + App showed at par yield attributes, stover yield (9.34 t ha −1 ), total biomass yield (17.17 t ha −1 ), grain protein yield (646.17 kg ha −1 ), total N uptake (145.96 kg ha −1 ) and remobilized vegetative N (89.8 kg ha −1 ) into grain with the RDF. Partial factor productivity (PFP N ) and apparent recovery efficiency (ARE N ) of N in N 50 PK + App was 23% and 22.1% higher than RDF respectively. The B:C in N 50 PK + App (2.43) was at par with N 75 PK + App (2.44). N 50 PK + App had the lowest energy consumption (9.72% lower than RDF), highest N energy use efficiency (24.3% higher than RDF) along with the maximum energy profitability, productivity and use efficiency. N 50 PK + App treatment resulted in marked reductions in GHG emission compared to RDF, with 11.1% lower CO2-eq. emission ha −1 , 29.4% lower N 2 O emission ha −1 , and 11.14% lower CO2-eq. emission t −1 . Thus, basal application of 50 kg ha −1 N with two splits of N (35 & 45 DAS) as per Pusa N Doctor can provide at par yield with RDF and GS TM , while simultaneously promoting N and energy use efficiency thereby, minimising GHG emission.

Recent processors rely on huge pages to reduce the cost of virtual-to-physical address translation. However, huge pages are notorious for creating memory bloat – a phenomenon wherein the OS ends up allocating more physical memory to an application than its actual requirement. This extra memory can be reclaimed by the OS via de-bloating at runtime. However, we find that current OS-level solutions either lack support for dynamic memory de-bloating, or suffer from performance and fairness pathologies while de-bloating. We address these issues with EMD (Efficient Memory De-bloating). The key insight in EMD is that different regions in an application’s address space exhibit different amounts of memory bloat. Consequently, the tradeoff between memory efficiency and performance varies significantly within a given application e.g., we find that memory bloat is typically concentrated in specific regions, and de-bloating them leads to minimal performance impact. Hinged on this insight, EMD employs a prioritization scheme for fine-grained, efficient, and fair reclamation of memory bloat. EMD improves performance by up to 69% compared to HawkEye — a state-of-the-art OS-based huge page management system. EMD also eliminates fairness concerns associated with dynamic memory de-bloating.

We introduce MindlessGen, a Python-based generator for creating chemically diverse, “mindless” molecules through random atomic placement and subsequent geometry optimization. Using this framework, we constructed the MB2061 benchmark set, containing 2061 molecules with high-level PNO-LCCSD(T)-F12 reference data for dissociation reactions. This set provides a challenging benchmark for testing, validation, and training of density functional approximations (DFAs), semiempirical methods, force fields, and machine learning potentials using molecular structures beyond the conventional chemical space. For DFAs, we initially hypothesized that highly parameterized functionals might perform poorly on this set. However, no consistent relationship between fitting strategy and accuracy was observed. A clear Jacob’s ladder trend emerges, with ωB97X-2 achieving the lowest mean absolute error (MAE) of 8.4 kcal·mol−1 and r²SCAN-3c offering a robust cost-efficient alternative (19.6 kcal·mol−1). Furthermore, we discuss the performance of selected semiempirical methods and contemporary machine learned interatomic potentials.