30x faster analytics, GPU kernels generated automatically from SQL, AI matched to lab-grown tumor models for cancer treatment, and LLMs that learn across tasks without retraining. Dive into the latest issue of Research Focus: https://t.co/X0A6tEHLuw https://t.co/QLHpdWMYnP

Hardware-Accelerated Analytics, AI-Powered Cancer Research, and the Future of Knowledge Work

Source: https://www.linkedin.com/pulse/hardware-accelerated-analytics-ai-powered-cancer-research-nznve In this issue:

• CoddSpeed: A hardware acceleration framework delivering up to 30x faster analytics across GPUs, FPGAs, and other accelerators in Microsoft Fabric
• GIRAFFE: An AI-powered wildlife re-identification system running 120x faster than baseline methods with over 0.9 accuracy across nine metrics
• Adopt vs. Adapt: A longitudinal study of 12,000 Copilot users finds individual LLM usage habits change very little over time
• ConDABench: A new benchmark framework capturing the real-world complexity and interactivity that existing LLM data analysis evaluations miss
• EvoLib: A test-time learning framework that lets LLMs accumulate and evolve knowledge across tasks without updating model parameters
• DEPART: A diagnostic framework that explains multilingual LLM performance gaps using observable language properties and model-conditional probes
• Browsing Large Graphs: A new browser-native approach to visualizing large graphs uses tile pyramids and sleeve routing to keep key labels readable at every zoom level, no dedicated server required
• HARNESS-LM: A three-phase training recipe that recovers 98% of a large retriever’s precision at up to 27x lower latency for sponsored search
• GPU Scalar Acceleration: An automated compiler pipeline converts CPU-based SQL scalar functions into optimized GPU kernels, closing a major performance gap in analytical databases without hand-written code
• OptiMind: A specialized language model that translates natural-language business problems into solver-ready mathematical formulations
• Project Ex Vivo: Microsoft researchers are combining AI with lab-grown tumor models to understand why patients with similar cancer diagnoses have vastly different outcomes and improve treatment matching
• AI and the Democratization of Knowledge Work: A new article outlining the technical and policy steps needed for AI to deliver broad, equitable benefits

NEW RESEARCH

This research introduces CoddSpeed, a hardware acceleration framework for executing analytics workloads on heterogeneous accelerators. Built for Microsoft Fabric, Microsoft’s unified cloud-native data platform, CoddSpeed allows analytics engines to leverage evolving accelerator technologies, including GPUs, FPGAs, ASICs, NVLink, and InfiniBand.

The authors describe a multi-year effort to redesign Fabric for accelerated analytics workloads while ensuring flexibility across hardware generations. They present a GPU-based execution engine derived from Microsoft’s Tensor Query Processor (TQP) and a data movement system optimized for high-speed interconnects.

Test results show substantial improvement over CPU-based execution in multiple production and benchmark scenarios, including up to 30× acceleration on TPC-H 1TB workloads. The work offers practical lessons for building large-scale, accelerator-native analytics systems, highlighting strategies for performance and portability in the emerging era of heterogeneous computing.

NEW RESEARCH

The research presents a system for automated re-identification of wildlife, which is critical for a wide range of ecological studies. GIRAFFE (Generalized Image-based Re-identification using AI for Fauna Feature Extraction), focuses on giraffes and is extensible to other species. It uses local feature matching to identify known individuals and partition unknown individuals for label annotation at scale. The authors also developed a user interface that enables both technical and non-technical users to curate large datasets and analyze repeat survey data.

Unlike existing methods that require manual labeling to facilitate individual re-identification, GIRAFFE automates key steps in the re-identification pipeline, reducing manual effort while maintaining accuracy and interpretability. Validated on real-world datasets, the system achieves over 0.9 accuracy across nine standard metrics, with some metrics achieving near perfect score. It runs 120 times faster than baseline methods and delivers a 132-fold improvement in cost-effectiveness.

Accurate wildlife re-identification is critical for a wide range of ecological studies, including density estimation via capture–recapture, demographic analyses, and behavioral research. This technology supports endangered species tracking, improves analysis of population dynamics and movement patterns, and, ultimately, allows for the implementation of data-driven conservation strategies.

NEW RESEARCH

This research analyzes how individual users of large language models (LLMs) change their behavior over time, a dimension largely missing from current research, which paints a static picture of LLM usage.

The authors address this gap by examining the conversational trajectories of 12,000 randomly sampled Microsoft Bing Copilot users and compare these with data from WildChat-4.8M. While the Copilot data contains significant population-level trends, the evidence around individual user trajectories is much weaker. User habits tend to change very little over time.

The research also finds stark differences between users of different activity levels: more active users have more successful conversations and use LLMs for more complex and professionally oriented tasks. The results suggest that existing user behavior is difficult to change and that LLM usage varies considerably across individuals.

NEW RESEARCH

ConDABench is a framework for generating conversational data analysis (ConDA) benchmarks and evaluating external tools on those generated benchmarks. This addresses a problem with real-world data analysis tasks, which often come with under-specified goals and unclean data. User interaction is necessary to understand and disambiguate a user’s intent, which makes it essential to solving these complex tasks. Existing benchmarks for evaluating LLMs on data analysis tasks do not capture these complexities or provide first-class support for interactivity.

Evaluation of state-of-the-art LLMs on these benchmarks reveals that while the new generation of models are better at solving more instances, they are not necessarily better at solving tasks that require sustained, long-form engagement. ConDABench is an avenue for model builders to measure progress towards truly collaborative models that can complete complex interactive tasks.

NEW RESEARCH

This research introduces EvoLib, a test-time learning framework that enables LLMs to self-improve by accumulating and evolving knowledge across tasks without parameter updates. This addresses one of the biggest shortcomings of LLMs today—most of them don’t learn from experience from one conversation to the next.

Instead of adapting model parameters, this approach maintains a shared library of knowledge abstractions, including modular skills and reflective insights, automatically extracted from the model’s own inference trajectories. To support continual improvement, the authors introduce a principled weighting and consolidation mechanism that jointly optimizes for immediate utility and long-term value. This allows simple, instance-specific abstractions to evolve into more general and reusable ones over time. Across challenging benchmarks in mathematical reasoning, code generation, and multi-turn agentic environments, EvoLib improves substantially over the top test-time scaling and learning methods without ground-truth feedback.

NEW RESEARCH

This research reframes multilingual LLM evaluation from a passive mapping of performance gaps across the languages they serve into an explainable, diagnostic framework. This addresses a deeper question: how much of the gap can be predicted before running a benchmark, versus how much remains as an irreducible per-checkpoint residual? The answer is derived using observable language properties (script, family, typological distance, resource class) and cheap model-conditional probes (tokenizer fertility, internal representation similarity to English).

This paper recasts multilingual evaluation from passive performance mapping into an explainable, diagnostic framework with concrete levers for targeting the root drivers of language disparity.

NEW RESEARCH

This research addresses challenges in visualizing graphs in web browsers without a dedicated server. The authors create a tile pyramid enabling a semantic zoom view in which the most important node labels remain readable at every zoom level, similar to how major geographic labels persist on digital maps. The edges are routed by a new method called “sleeve routing”, which searches the dual graph of a Constrained Delaunay Triangulation to select a sequence of triangles through the free space, then applies the funnel algorithm to compute a shortest path inside the selected sleeve. Several heuristics are deployed to speed up the routing.

NEW RESEARCH

This paper presents HARNESS-LM (HLM), a new method aimed at balancing retrieval quality with production latency in the competitive landscape of sponsored search. HLM is a three-phase training framework for transferring the capabilities of large-scale retrievers into compact, cost-efficient models. This approach comprises: (1) training a high-performance reference (“teacher”) retriever by fine-tuning a billion-parameter-scale small language model (SLM); (2) aligning query representations via an L2 objective to distill knowledge into a sub-600M parameter student encoder; and (3) applying a final contrastive refinement stage to optimize the student for retrieval performance.

The authors also present a comprehensive empirical study of key design choices, including alignment objectives, embedding dimensionality, model scale, architecture, and optimization strategies, to identify configurations that are most effective in production settings. On a real-world Bing Ads evaluation benchmark, HLM recovers over 98% of the reference retriever’s precision across multiple settings, while delivering up to 27x lower online query-encoder latency and 20x higher throughput on NVIDIA A100 GPUs. Online A/B testing on Bing Ads shows gains in revenue, impression, and click uplift over current retrievers, clearly highlighting the practical efficacy of the HLM recipe in a real-world sponsored search setting.

NEW RESEARCH

This work introduces an automated compiler pipeline that converts CPU-based SQL scalar functions into optimized GPU kernels, addressing a major gap in GPU-accelerated databases. While prior systems focused mainly on accelerating relational operators, scalar functions often remain a key performance bottleneck in analytical queries.

The system uses an LLVM/MLIR-based compiler to translate hundreds of production database scalar functions into GPU code while preserving their original semantics and support for data types and database context. The authors evaluate the approach on a production workload and an enhanced TPC-H benchmark designed to stress scalar functions in standard benchmarks.

Integrated into a state-of-the-art GPU database system, the approach achieved significant performance improvements over a leading CPU database. The generated GPU kernels performed comparably to hand-optimized implementations, demonstrating a practical and scalable path for accelerating complex SQL analytics on GPUs.

NEW RESEARCH

OptiMind is a specialized language model designed to offer domain expertise in complex business settings. Released as anexperimental model in Foundry, it translates natural-language business problems into solver-ready mathematical formulations.

Unlike other generative AI models, OptiMind solves a fundamentally different class of problem around optimization and decision-making under real-world operational constraints, where models that look impressive in clean environments may come up short. The people using these systems are rarely researchers, but may be schedulers, plant managers, and engineers who need reliable, practical solutions inside systems where constraints fluctuate and tradeoffs carry real consequences.

NEW RESEARCH

Even with more precise cancer treatments, patients with similar diagnoses can have very different outcomes. Microsoft’s Project Ex Vivo focuses on how cell behavior can and should impact how cancer is categorized and treated. By combining AI with lab-grown tumor models, researchers are learning why these differences occur and why lab results don’t always match patient outcomes, helping move toward better-matched, more effective treatments.

ARTICLE

This research explores the extent to which AI democratizes knowledge work and describes technical and social interventions needed to ensure that AI delivers the broadest benefits possible.

Related recent research shows that when people use generative AI for work, they primarily use it for high-complexity, knowledge-work tasks. However, this leaves open the question of who are the users and beneficiaries of such tools.

The authors recommend the following:

Technical priorities

1. Improve privacy and security so AI can be safely used in sensitive domains.
2. Invest in post-training for creative and open-ended work to expand the set of tasks where AI’s use benefits novice versus expert users, respectively.
3. Expand performance and safety across languages and cultures to ensure global accessibility and fairness.

Policy priorities

1. Protect worker agency by integrating AI in ways that augment, rather than replace, human decision-making.
2. Share AI-driven productivity gains through workforce training, reskilling, and support for displaced workers.
3. Broaden AI access and literacy while continuing to expand internet and broadband infrastructure.