Position: Don't Just "Fix it in Post": A Science of AI Must Study Training Dynamics

arXiv cs.AI Papers

Summary

This position paper argues that a scientific understanding of AI must go beyond post-hoc analysis and instead study the training dynamics that shape model behavior, with implications for predicting, intervening, and designing training procedures for desired properties like capabilities and safety.

arXiv:2606.06533v1 Announce Type: new Abstract: What would it mean to have a scientific understanding of AI? Models are not static objects: they are snapshots of time-evolving processes shaped by data, objectives, architectures, and optimization dynamics. Yet much of AI research treats models as fixed artifacts, analyzing behaviors after training rather than asking why they emerge. This position paper argues that a science of AI must move beyond post-hoc fixes and study the training dynamics that produce model behavior. Such a science should support progressively stronger forms of understanding: predicting outcomes from early training signals, intervening when trajectories go wrong, and ultimately designing training procedures that more reliably produce desired properties. Scaling laws have made prediction routine for loss; the challenge is extending this success to capabilities, biases, robustness, and safety-relevant behaviors. We articulate requirements for such theories grounded in the history and philosophy of science, examine progress in mechanistic interpretability, fairness, memorization, and simplicity bias, and identify concrete open problems.
Original Article
View Cached Full Text

Cached at: 06/08/26, 09:13 AM

# Position: Don't Just "Fix it in Post": A Science of AI Must Study Training Dynamics
Source: [https://arxiv.org/abs/2606.06533](https://arxiv.org/abs/2606.06533)
[View PDF](https://arxiv.org/pdf/2606.06533)

> Abstract:What would it mean to have a scientific understanding of AI? Models are not static objects: they are snapshots of time\-evolving processes shaped by data, objectives, architectures, and optimization dynamics\. Yet much of AI research treats models as fixed artifacts, analyzing behaviors after training rather than asking why they emerge\. This position paper argues that a science of AI must move beyond post\-hoc fixes and study the training dynamics that produce model behavior\. Such a science should support progressively stronger forms of understanding: predicting outcomes from early training signals, intervening when trajectories go wrong, and ultimately designing training procedures that more reliably produce desired properties\. Scaling laws have made prediction routine for loss; the challenge is extending this success to capabilities, biases, robustness, and safety\-relevant behaviors\. We articulate requirements for such theories grounded in the history and philosophy of science, examine progress in mechanistic interpretability, fairness, memorization, and simplicity bias, and identify concrete open problems\.

## Submission history

From: Stella Biderman \[[view email](https://arxiv.org/show-email/12356afe/2606.06533)\] **\[v1\]**Wed, 3 Jun 2026 17:58:14 UTC \(97 KB\)

Similar Articles

Interactive Evaluation Requires a Design Science

Hugging Face Daily Papers

This position paper argues that interactive AI evaluation should be treated as a design science paradigm, proposing a two-axis taxonomy and reporting standards for assessing dynamic system behavior through trajectories.