Turing

When people talk about AI in HR, the conversation usually centers on efficiency. But that's only part of the story. At Turing, AI is helping our People team move faster while creating more space for high-value human work. Since adopting AI tools, we've reduced help desk response times by 33%, and AI assistants now handle 80% of HR support tickets. We're also building internal AI agents, cutting SaaS spend, and working alongside frontier AI labs and our Production Engineering teams to explore what's possible in HR. Have we gotten everything right? Of course not. We've tested ideas, learned from failures, improved what worked, and kept building. That's what meaningful AI adoption looks like. Thank you to Google Workspace for featuring Taylor Bradley and sharing Turing's journey. Video link in the comments.

Who is teaching the world’s most advanced AI models to think? At Turing, we’ve spent years building the human intelligence infrastructure behind frontier AI. Today, Turing works with many of the world’s leading AI companies, helping power the development of coding, reasoning, tool use, and multimodal capabilities that are shaping the next generation of AI systems. As Turing CEO Jonathan Siddharth explains: “A lot of problems in life are math problems framed the right way, and code is a way to express and solve math problems. If you can solve coding, tool use, reasoning, and multimodality, you have the keys to superintelligence.” These four capabilities are more than technical milestones. They are the building blocks that enable AI to move beyond simple prediction and toward meaningful problem solving, decision making, and real-world impact. The path to AGI isn’t just about larger models. It’s about developing the capabilities that allow AI to understand, reason, act, and create. Watch the full conversation with Siddhartha Ahluwalia, managing partner at Neon Fund, to hear his perspective on what comes next and why these four pillars matter. To watch the interview in its entirety: https://lnkd.in/eKGesTEK

Rubrics-Graded Reasoning is here. Today, Turing is releasing the Advanced PhD Reasoning Rubrics Data Pack, a new rubric-based reasoning dataset built for frontier AI teams developing and evaluating advanced reasoning models. The dataset spans Computer Science, Data Science, and Chemistry and includes 1,106 expert-authored PhD-level tasks paired with weighted atomic rubrics and golden answers. 𝐖𝐡𝐲 𝐝𝐨𝐞𝐬 𝐭𝐡𝐢𝐬 𝐦𝐚𝐭𝐭𝐞𝐫? Most benchmarks measure whether a model reached the correct answer. They tell you what happened, but not how the model reasoned. This dataset goes deeper by evaluating: • Intermediate reasoning steps • Derivations and calculations • Scientific mechanisms and structures • Code and data workflows • Methodological decisions • Edge-case handling • Structured outputs and multi-step pipelines Each task is designed to transform expert evaluation into machine-verifiable training signals, making it particularly valuable for: -Reinforcement Learning (RL) -Reward Modeling -Post-Training -Process-Level Evaluation -Benchmarking & Regression Testing -Reasoning Failure Analysis -Scientific & Engineering QA To ensure the dataset remains challenging for frontier systems, it was calibrated across 16 evaluation rounds, with pass rates ranging from 0% to 50% on state-of-the-art models. This release also showcases the quality bar behind the reasoning datasets Turing delivers to frontier-model partners. If you're working on model reasoning, evaluation, or RL infrastructure, we'd love to hear what you think. Explore the dataset: On Hugging Face:  https://lnkd.in/e-HW5SrV Learn more: https://lnkd.in/eC2px-sS

Most AI benchmarks focus on text quality. Enterprise adoption depends on something much more practical: whether models can reliably generate the presentations, spreadsheets, reports, PDFs, and structured documents teams use every day. To evaluate that capability, Turing built a benchmark spanning more than 1,500 validated artifacts across nine output formats, including PPTX, DOCX, PDF, HTML, JSON, Excel, CSV, TXT, and infographics. The benchmark evaluated leading AI providers across four levels of prompt complexity, from basic artifact generation to highly specific enterprise workflows with detailed formatting, content, and citation requirements. Every artifact underwent format-specific QA validation. Every failure was classified using a structured taxonomy covering issues such as wrong-format outputs, download failures, citation loss, clarification loops, and file-handling errors. The project delivered: • 1,500+ validated artifacts • Full complexity coverage across document formats • Provider, model, timing, and QA metadata for every run • 99.9% artifact acceptance rate One of the clearest lessons was that artifact generation success cannot be measured with simple pass/fail metrics. Different models fail in different ways, and understanding those patterns is critical for enterprise deployment decisions. The result is a model-aware, format-aware benchmark that provides a far more realistic view of enterprise AI performance than traditional evaluations. Full case study: https://lnkd.in/eY6tvcM2

You have spent the day at [ICML] Int'l Conference on Machine Learning, absorbing cutting-edge research. Come decompress with the people building what comes next. Turing is hosting a Happy Hour on Wednesday, July 8 from 6:00 to 9:00 PM in Seoul, and we want you there. The evening is designed for genuine conversation between AI researchers advancing today's state-of-the-art foundation models and enterprise leaders driving real-world AI deployment. We will share Turing's perspective on the evolving role of LLMs and the future of AI, but mostly we want to hear from you. Drinks and hors d'oeuvres will be served. Turing is a superintelligence accelerator that closes the loop between frontier AI labs and enterprise deployment, turning real enterprise work into the training inputs that make frontier models better in the world. If that mission resonates with you, this is the room to be in. Register here to see the address and secure your spot. Space is limited: https://lnkd.in/eEy4uH9u #ICML2026

Benchmark scores are climbing. But are AI models actually improving scientific work? At [ICML] Int'l Conference on Machine Learning 2026 in Seoul, Turing's Charlotte Tao and Tristan Tager will take on one of the most important questions in frontier AI evaluation: the growing disconnect between what benchmarks measure and what scientists actually need. SciCode rose from 4.6% to 59% in a year. HLE went from 8% to 47%. But scientific work is not paper progress. Through real-world examples from Turing’s frontier data work, Charlotte and Tristan will examine why models can appear strong on isolated tasks while still struggling with full scientific workflows. Attendees will leave with a practical lens for distinguishing benchmark progress from workflow progress and what scientific AI evaluations need to measure next. The talk will also explore what comes next for frontier data: how the field may evolve from datasets toward modular, composable capability infrastructure. Proud to see Turing contributing to this conversation at ICML 2026. 𝐀𝐝𝐯𝐚𝐧𝐜𝐢𝐧𝐠 𝐅𝐫𝐨𝐧𝐭𝐢𝐞𝐫 𝐒𝐜𝐢𝐞𝐧𝐭𝐢𝐟𝐢𝐜 𝐂𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬, 𝐓𝐨𝐝𝐚𝐲 𝐚𝐧𝐝 𝐓𝐨𝐦𝐨𝐫𝐫𝐨𝐰 𝐂𝐡𝐚𝐫𝐥𝐨𝐭𝐭𝐞 𝐓𝐚𝐨 𝐚𝐧𝐝 𝐓𝐫𝐢𝐬𝐭𝐚𝐧 𝐓𝐚𝐠𝐞𝐫, 𝐓𝐮𝐫𝐢𝐧𝐠 𝐈𝐂𝐌𝐋 2026, 𝐒𝐞𝐨𝐮𝐥, 𝐉𝐮𝐥𝐲 6-11.

Most AI benchmarks focus on text. The real world runs on tables. Financial statements, regulatory filings, research papers, healthcare records, and enterprise reports all contain critical information embedded in structured data. Yet table reasoning remains one of the most challenging capabilities for modern AI systems. To help advance this frontier, Turing partnered on a project to build more than 70,000 table reasoning Q&A pairs from real-world documents for AI training. The challenge was not simply extracting information. Models must learn to: • Compare values across rows and columns • Perform multi-step calculations • Connect information across tables and surrounding text • Interpret complex document structures • Generate accurate answers grounded in evidence As AI adoption accelerates across industries, the quality of training data is becoming a key differentiator. Better models require datasets that reflect how information exists in production environments, not just in simplified benchmarks. This project highlights an important reality: advancing AI is not only about larger models. It is also about building the high-quality datasets that teach those models how to reason. Read our latest case study: https://lnkd.in/e-NbRMHG

[ICML] Int'l Conference on Machine Learning 2026 is four weeks away, and Turing will be there as a Diamond Sponsor in Seoul, Korea. We are getting ready for what is shaping up to be one of the most consequential gatherings in machine learning research this year. Frontier model capabilities, agentic systems, reasoning at scale -- the work being presented at COEX this July reflects where AI is actually heading, and we are proud to be part of that conversation at the highest level. If you are planning to attend, stay tuned. More details on where to find us (Booth #8406), who to talk to, and what we will be sharing are coming soon. 📍Seoul. July 6-11. See you there. #ICML2026

Teaching AI to think like an ad director is harder than it sounds. When a multimodal model needs to generate product video concepts from image inputs, it can't just describe what it sees. It has to understand advertising logic: how a wide establishing shot gives way to a feature close-up, how a voice-over line earns credibility by tying directly to what's on screen, and how 15 sequential shots build a coherent brand story in under two minutes. That's the dataset Turing just delivered. 500+ structured ad storyboard tasks. 7,500+ original shot descriptions, each grounded in real product images and descriptions with zero invented features. 6,000+ voice-over lines matched to specific product benefits visible in the corresponding shot. 90%+ quality score maintained across the full dataset. Every shot was built around a defined advertising arc, from product introduction through feature demonstration, lifestyle relevance, and brand reinforcement. Camera motion was specified from a standardized taxonomy covering static, pan, tilt, tracking, arc, and handheld directions. Quality was enforced through a four-dimension rubric covering conceptual accuracy, visual creativity, ad structure, and technical adherence, with rework required before delivery. Storyboards were also tested against SOTA video generation models including VEO and Seedance to verify that descriptions translated into coherent video sequences without hallucinations or generation artifacts. If you're building or evaluating multimodal models for product video generation, this is the kind of training data that makes the difference between a model that describes products and one that understands how to sell them. Read the full case study: https://lnkd.in/e6vBBXhx