This week, we're highlighting how Turing delivered over 50,000 structured multimodal preference evaluations that go beyond surface-level comparison by enforcing factuality-first verification and dimension-level metadata. We discuss how the human role is shifting from execution to verification as agents scale.
We also cover three major developments: OpenAI's GPT-Rosalind for life sciences research, Meta's Muse Spark multimodal reasoning model with its rebuilt pretraining stack and multi-agent orchestration, and GrandCode, the first AI system to beat all humans in live competitive programming.
What we're doing
This week, we’re highlighting how Turing delivered over 50,000 structured preference evaluations across image-grounded prompts, spanning scientific reasoning, mathematical problem-solving, descriptive analysis, and information-seeking tasks.
Here’s what we delivered:
- 50,000+ head-to-head preference evaluations across diverse multimodal reasoning tasks
- A multi-dimensional framework assessing factual accuracy, instruction-following, coherence, signal-to-noise ratio, and honesty
- A structured QA pipeline with star-rated review, error trend tracking, and calibrated annotator training to maintain signal quality at scale
Image-grounded evaluation requires more than surface-level comparison. By enforcing factuality-first verification, evidence-based preference reasoning, and dimension-level metadata, this dataset captures which response is better and why, enabling stronger multimodal RLHF and better model iteration.
What we're saying
🗣️ Every Organization Was Built for a Pre-Superintelligence World
In a recent post, Jonathan Siddharth shared a view on what changes next: as model capability compounds, the human role shifts from doing the work to verifying it. Agents will make tool calls to software, to other agents, and to humans, and designing that human verification layer may be the defining challenge of the next phase.
What we're reading
- Introducing GPT‑Rosalind for Life Sciences Research
OpenAI has released GPT-Rosalind, a frontier reasoning model purpose-built for biology, drug discovery, and translational medicine. The model is optimized for multi-step scientific workflows including literature synthesis, experimental planning, and tool use across 50+ public databases. On BixBench, it achieves leading performance among published models, and in a collaboration with Dyno Therapeutics on RNA sequence-to-function tasks, best-of-ten submissions ranked above the 95th percentile of human experts on prediction. The model launches through a trusted-access program for qualified U.S. enterprise customers, alongside a freely available Life Sciences Research Plugin for Codex. OpenAI positions this as the first in a planned series of domain-specific life sciences models. - Introducing Muse Spark: Scaling Towards Personal Superintelligence
Meta Superintelligence Labs introduces Muse Spark, a natively multimodal reasoning model with support for tool use, visual chain of thought, and multi-agent orchestration. The model is the product of a rebuilt pretraining stack that achieves equivalent capability with over an order of magnitude less compute than Llama 4 Maverick. A "Contemplating mode" orchestrates multiple parallel agents, reaching 58% on Humanity's Last Exam and 38% on FrontierScience Research. RL training shows smooth, log-linear scaling with predictable generalization, and a thinking-time penalty induces a phase transition where the model compresses its reasoning into fewer tokens before extending again for harder problems. - GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learning
This paper introduces GrandCode, a multi-agent RL system that became the first AI to consistently beat all human participants in live Codeforces competitions, placing first in three consecutive rounds in March 2026, including victories over legendary grandmasters. The system orchestrates specialized modules for hypothesis generation, solving, summarization, and adversarial test-case generation, all jointly optimized through a new method called Agentic GRPO, which addresses delayed rewards and off-policy drift in long, multi-stage agent rollouts by combining immediate reward updates with delayed correction. Built on Qwen 3.5-397B, the full pipeline progresses from continued pretraining through SFT and multi-component RL to test-time RL with LoRA adaptation during live contests. On internal benchmarks, the system reaches 85% acceptance rate and solves 15 out of 20 hardest-tier problems, up from 64% and 4/20 for the base model.
Where we’ll be
ICLR- The International Conference on Learning Representations
🔹 LLM Researchers Happy Hour During ICLR- April 23
📍 Rio de Janeiro, Brazil | 🗓️ April 23 - 27
📌 Booth #301
ICLR focuses on cutting-edge research in deep learning, highlighting advancements in representation learning, optimization, and AI theory.
AI Dev 26- The AI Developers Conference
🔹 LLM Researchers Happy Hour During AI Dev- April 28
📍 San Francisco, California | 🗓️ April 28 - 29
AI Dev brings together developers for hands-on AI workshops, expert talks, startup showcases, and live demos focused on real-world AI systems.
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.