This week, we explore the growing complexity behind how models learn and reason—from training data quality and attribution to decentralized agent collaboration and emotion interpretation in vision-language models.

What we're thinking

This week, we’ve been thinking deeply about the role of training data in shaping model performance—not just in aggregate, but at the level of attribution, quality, and impact.

Which data actually drives performance? Methods like data ablation and attribution studies can help pinpoint which subsets of data meaningfully improve results on key benchmarks—and which don’t.
Balancing data coverage and domain depth: Adding more data isn't always better. Measuring coverage across tasks, formats, and domains is critical to avoid overfitting to common distributions while underrepresenting edge cases.
Defining “good” data: From truthfulness and bias to knowledge density and style, data quality remains hard to measure—but directly influences generalization, reasoning, and hallucination behavior.
Evaluating the training mix: Optimizing mixtures across task types (e.g., text, math, code) and choosing the right ratios may matter more than increasing scale alone.

As models scale, so do the consequences of what they’re trained on. Thoughtful, structured evaluation of training data is becoming just as important as how we evaluate the models themselves.

What we're saying

Insights from Turing’s leadership on the latest AGI developments and industry shifts:

The News: Anthropic launched Claude for Education, introducing a “Learning Mode” that uses Socratic prompts to foster critical thinking. OpenAI made ChatGPT Plus free for college students in the U.S. and Canada through May 31 to broaden access to AI tools.

Sam Ho, Product Leader:
"Making ChatGPT free for students is a powerful move—but the real shift is in how tools like Claude now guide students through questions instead of giving answers. As a new parent, I'm especially focused on how we design AI to teach thinking, not just facts."

What we're reading

Welcome to the Era of Experience
This paper introduces the “era of experience”, where AI agents learn by interacting with the world rather than relying solely on human data. By grounding rewards in real-world outcomes and optimizing over long-term interactions, agents can adapt, reason, and discover novel strategies. The authors argue this shift will unlock more general and superhuman capabilities through large-scale, autonomous learning.
SkillFlow: Efficient Skill and Code Transfer Through Communication in Adapting AI Agents
SkillFlow is a decentralized framework that lets AI agents acquire new skills from one another—enabling dynamic adaptation beyond static tool use. In benchmarks, it cut task time by 46.4% through local execution of shared skills. Inspired by biological systems, it points toward a future where agents evolve by collaborating, not just scaling.
Why We Feel: Breaking Boundaries in Emotional Reasoning with Multimodal Large Language Models
This paper introduces Emotion Interpretation (EI)—a new benchmark task focused not on labeling emotions, but explaining why they arise. The authors present EIBench, a multimodal dataset paired with a novel Coarse-to-Fine Self-Ask (CFSA) method that guides Vision-Language Models to reason about emotional triggers. Experiments show that while current LLMs perform well on basic EI, they still struggle in complex, multi-perspective scenarios—highlighting the need for more nuanced emotional understanding in AI.

Where we’ll be

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

ICLR 2025 [Singapore | Apr 24 – 28]
A top-tier deep learning conference covering representation learning, AI optimization, and theoretical advancements.
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.

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]