This week, we explore why precision training data, self-correcting agents, and human-in-the-loop collaboration are becoming critical pillars for the next generation of applied AGI systems. From medical image caption rewrites to agents that learn from human interventions, the future of AGI looks more grounded—and more human.

What we're thinking

As LLMs grow more capable across modalities, we’ve been reflecting on the need for structured, domain-grounded data—especially in high-stakes areas like healthcare and science.

Key insights from our 10,000+ medical caption rewrite campaign:

Most errors weren’t technical—they were factual: Over 64% of model-generated captions required correction, often due to subtle misinterpretations in anatomy, diagnosis, or measurement.
Fine-grained SFT works best with structure: Using layered QA rubrics, domain-specific rewrite reasons, and visual-grounded workflows helped us reduce hallucinations and standardize quality at scale.
Difficulty calibration matters: We introduced a taxonomy to classify tasks by visual + caption complexity and mapped them to appropriate education levels—ensuring the right level of rewrite for the right content.

In high-stakes domains, alignment isn’t just a prompt engineering problem—it’s a precision problem. Reliable outputs require training data grounded in expert judgment, not just web-scale patterns. As we push toward AGI, success will depend as much on human-labeled, domain-specific datasets as on model architecture or scale.

What we're saying

📑 The Research: The new Era of Experience paper argues that the future of AI won’t be driven by more static data—but by models learning through long-term, grounded interaction with the world.

🗣️Sam Ho, Product Leader:
"AI won’t transform healthcare by diagnosing faster—it’ll do it by learning continuously. We’re entering an era where models can track, adapt, and optimize treatments in real time, far beyond the limits of a five-minute doctor visit. The real breakthrough is using AI to personalize medicine based on experience—not just data."

What we're reading

LLManager: Self-Reflective, Human-in-the-Loop Approval for Agents
LLManager is a LangGraph-based workflow designed to manage approval requests by combining dynamic prompt generation, few-shot learning, and human review. It uses past reflections and examples to improve decision-making over time, and integrates human edits to fine-tune future responses. By leveraging reasoning reports, editable final answers, and targeted reflection after mistakes, LLManager creates a self-correcting loop for scalable, trustworthy approval flows.
CODESIM: Multi-Agent Code Generation and Problem Solving through Simulation-Driven Planning and Debugging
CODESIM introduces a multi-agent framework for code generation, blending planning, coding, and step-by-step debugging through input/output simulation. Inspired by how humans reason about algorithms, CODESIM verifies plans and code without relying heavily on external tools. It outperforms prior methods like MapCoder across major benchmarks (HumanEval, MBPP, APPS) and demonstrates strong generalization on open-source models—delivering both higher accuracy and lower token usage.
COWPILOT: A Framework for Autonomous and Human-Agent Collaborative Web Navigation
COWPILOT introduces a framework where LLM agents and humans work together to navigate websites, blending agent suggestions with human intervention when needed. In experiments across major sites, COWPILOT’s collaborative mode achieved 95% task success—outperforming both full autonomy and human-only baselines. It shows that mixing human intuition with agent automation can boost accuracy while reducing human effort, pointing toward a future of co-adaptive agents.

Where we’ll be

Turing will be at two major AI conferences in the coming months—join us to discuss the future of AGI:

MLSys 2025 [Santa Clara, CA | May 12 – 15]
A major event focused on the intersection of machine learning and systems, discussing efficient AI model training, distributed learning, and AI hardware innovations.
ICML 2025 [Vancouver Convention Center, Canada | July 13 – 19]
The International Conference on Machine Learning (ICML) is a leading international conference focused on the advancements in machine learning and its applications.

If you’re attending, reach out—we’d love to connect and exchange insights!

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.

[Subscribe & Read More]