Multimodal models can caption charts, but analyst-grade reasoning is a different problem. This week, we highlight how Turing delivered 12,000+ chart-grounded Q&A pairs sourced from real enterprise documents, enforcing a zero-inference, zero-approximation standard that maps answers directly to visual encodings. We also cover new research on AI-driven formal proof search, omnimodal world models for physical AI, and executive-level skill optimization for self-evolving agents.
What we're doing
This week, we're highlighting how Turing delivered a chart understanding dataset for multimodal AI training, sourced from real-world documents. Unlike standard visual QA benchmarks focused on captioning and object detection, this dataset trains models to reason over charts the way analysts do: drawing only on what is visible, staying grounded in the source document, and producing answers that hold up without additional context.
Here's what we delivered:
- 12,000+ structured chart Q&A pairs across business reports, financial documents, government files, academic papers, and more
- A zero-inference, zero-approximation standard enforced at every stage with chart-anchored question design that maps training signal to visual encodings rather than document metadata, preventing shortcut learning
- 100% client acceptance rate across all delivered tasks, backed by a multi-layer QA system combining automated structural checks with expert human review
💡 By enforcing literal, chart-grounded answers at scale across real enterprise documents, this dataset gives models the supervision signal they need to reason like analysts.
What we're saying
🗣️ Jonathan Siddharth, Founder & CEO
Jonathan sat down with POLITICO's Digital Future Daily to share his thinking on where AI progress actually comes from and where the hype is misplaced.
Three things worth reading closely: real-world deployment data is underrated as a source of model improvement; pure synthetic data is overhyped and the best pipelines are hybrid; and education needs to shift from narrow specialists to generalist problem solvers, because superintelligence makes that breadth possible for the first time.
What we're reading
- Advancing Mathematics Research with AI-Driven Formal Proof Search
Google DeepMind introduces AlphaProof Nexus, an AI-driven formal proof system that combines LLMs, Lean verification, evolutionary search, and AlphaProof to tackle open mathematical research problems. Unlike natural-language theorem proving, every proof is formally verified in Lean, eliminating hallucinations and ensuring correctness. - In the largest evaluation of its kind, the system autonomously solved 9 of 353 open Erdős problems, proved 44 of 492 open OEIS conjectures, resolved a 15-year-old open algebraic geometry problem, improved an optimization convergence bound, and contributed to graph theory, additive combinatorics, and quantum optics research. The full-featured agent solved these problems at a cost of only a few hundred dollars per problem.
- Cosmos 3: Omnimodal World Models for Physical AI
NVIDIA introduced Cosmos 3, the first fully open omnimodal world model designed for Physical AI. Built on a novel mixture-of-transformers architecture, it unifies vision reasoning, world simulation, and action generation in a single system capable of processing and generating text, images, video, audio, and actions. - Cosmos 3 can function as a vision-language model, world simulator, video generator, forward and inverse dynamics model, and robot policy model. This allows developers to train and evaluate robots, autonomous vehicles, and vision AI systems using a shared foundation model rather than fragmented AI stacks.
- Among open models, Cosmos 3 ranks #1 on Artificial Analysis, Physics-IQ, PAI-Bench, and R-Bench for world generation, RoboLab and RoboArena for robot policy learning, and VANTAGE-Bench and TAR for vision understanding.
- SkillOpt: Executive Strategy for Self-Evolving Agent Skills
This paper introduces SkillOpt, a framework that treats an agent’s skill document as a trainable external state, optimizing it with the same discipline used for model training. Instead of one-shot prompts or ad hoc revisions, SkillOpt uses a frontier optimizer model to propose bounded add/delete/replace edits, accepting changes only when they improve a held-out validation score. - Across 6 benchmarks, 7 models, and 3 execution harnesses (direct chat, Codex, Claude Code), SkillOpt was best or tied-best on all 52 evaluated settings. On GPT-5.5, it improved average performance by +23.5 points in direct chat, +24.8 points in Codex, and +19.1 points in Claude Code, outperforming human-written skills, GEPA, TextGrad, Trace2Skill, and EvoSkill.
- A key finding is that the learned skills remain compact (300–2,000 tokens) and transfer across models, benchmarks, and agent frameworks without retraining.
Where we’ll be
🔹 CVPR 2026 — IEEE/CVF Conference on Computer Vision and Pattern RecognitionCVPR 2026 — IEEE/CVF Conference on Computer Vision and Pattern Recognition
📍 Denver, Colorado | 🗓️ June 3-7
CVPR is the world's premier conference that brings together researchers and practitioners to share significant advancements in computer vision, pattern recognition, and AI.
Stay ahead with AGI Advance
Turing is leading the charge in bridging AI research with real-world applications. Subscribe to AGI Advance for weekly insights into breakthroughs, research, and industry shifts that matter.
Ready to Optimize Your Model for Real-World Needs?
Partner with Turing to fine-tune, validate, and deploy models that learn continuously.