Questions for Theory in the New Age of Machine Learning

Questions for Theory in the New Age of Machine Learning

by Tom Mitchell

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning
🎙 Tom Mitchell 👥 75K 📅 May 27, 2026 ⏱ 42 min 👁 7K 📄 expert opinion 🧭 2026-08-03
Available in: English (current) Français

Keywords

LLMlearning agentsPAC learninginductive biastheory

Summary

Tom Mitchell, a pioneer in machine learning, delivers a keynote at the Simons Institute workshop on the role of theoretical computer science (TCS) in modern ML. He challenges two long-held assumptions: that learning primarily involves tuning parameters, and that it requires big data and sophisticated statistics. He argues that LLMs enable a new paradigm of reasoning-driven learning, where the model generates justifications for training examples and distills them into a concise rubric of principles. He illustrates this with a classification task from an online education platform, where the LLM learns to identify which hint is more helpful for students. The learned rubric, expressed in natural language, improves classification accuracy. This shifts the theoretical framework: the target function is described in natural language, the hypothesis class is the set of possible rubrics, and the LLM acts as an oracle. He raises questions about sample complexity, hypothesis class complexity, and ambiguity in function descriptions. He also hints at other examples, but the talk focuses on this one. The talk is forward-looking and calls for new theoretical foundations for ML.

177 words

Critical Evaluation

The talk is intellectually stimulating and timely, given the rapid advances in LLMs. Mitchell, a respected figure, challenges conventional wisdom and proposes a new direction for ML theory. The main example is compelling: using an LLM to generate justifications and distill them into a rubric is a concrete demonstration of reasoning-driven learning. The approach is novel and could inspire new theoretical frameworks. However, the talk is more of a position piece than a rigorous analysis. The example is illustrative but not a full-scale experiment; there is no detailed evaluation of the rubric’s quality or generalizability. The theoretical questions raised are important but not formally developed. Mitchell does not provide a formal model or proofs, which is expected for a keynote but limits the depth. The talk also touches on other examples (e.g., generating new hints) but does not elaborate. The sources are not explicitly cited in the talk, but the description links to the Simons Institute page. Overall, the talk is valuable for its insights and potential to guide future research, but it lacks the rigor of a formal paper. The adéquation between title and content is good, as it indeed raises questions for theory. The talk does not include any advertising or sponsorship segments.

205 words

Title / Content Match

The title accurately reflects the content: the talk raises theoretical questions for machine learning in the era of LLMs.

Quality & Reliability

8/10

The talk is given by a renowned expert in machine learning, Tom Mitchell, at a prestigious venue (Simons Institute). It presents novel ideas and examples, but it is an opinion piece rather than a peer-reviewed study. The claims are plausible and supported by illustrative experiments, but not fully rigorous.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

Contribution & Novelties

The talk introduces a novel paradigm for machine learning where LLMs serve as reasoning engines, enabling learning from few examples via natural language justifications. This challenges traditional PAC learning assumptions and opens new theoretical questions.

Pour aller plus loin :

  • PAC learning — Foundational framework for learning theory, relevant to the discussion of sample complexity.
  • Large language models — Overview of LLMs, the core technology discussed.
  • Inductive bias — Concept of prior knowledge in learning, central to the talk’s argument.

80 words

Radar Profile

The radar profile shows high scores in quality and reliability, reflecting the expert status and reputable venue, but moderate scores in quantity and technical depth, indicating a concise talk with limited formal detail.

Reliability 8/10