Skip to content
Computational Pathology Reading Roadmap

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

← ICML 2026 oral writing corpus

Stella Biderman, Mohammad Aflah Khan, Niloofar Mireshghallah, Catherine Arnett, Fazl Barez, Naomi Saphra · Official ICML 2026 oral record

Figure, table and section ledger

Editorial contribution type: Position & research agenda (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF
17 physical pages
Before references / appendix
≈ 8.79 page-equivalents
Main figures / tables
0 / 0
Attached appendix figures / tables
0 / 0
Attached appendix span
No boundary detected; not assumed zero
Before references / appendix: ≈ 51.71%References / intervening material: ≈ 48.29%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level headingStarts on PDF pageEstimated spanShare of entire PDF
1. Introduction1≈ 1.25 pp.7.35%
2. What would a “Science of AI” Look Like?2≈ 2.73 pp.16.06%
3. Case Studies5≈ 3.48 pp.20.47%
4. Alternative Views8≈ 0.62 pp.3.65%
5. Conclusion9≈ 0.29 pp.1.71%
Acknowledgments9≈ 0.09 pp.0.53%
References9≈ 8.21 pp.48.29%
All detected figure and table locations

Main text

    No recognized caption labels.

    Attached appendix

      No recognized caption labels.

      Argument and organization

      1. Problem and contribution

        Argues that post-hoc fixes cannot substitute for predictive theories of training dynamics, then specifies what a scientific explanation should accomplish.

      2. Argument sequence

        Normative criteria lead to case studies and named open problems; an explicit Alternative Views section addresses engineering sufficiency, theory skepticism, automation and safety pragmatism before the conclusion.

      3. Evidence and verification

        The data-attribution example exposes how fixing a realized run versus retraining on reshuffled data changes the question. Case studies use cited results to motivate unresolved predictive problems, not claim new measurements.

      Problem, claim and contribution

      Argues that post-hoc fixes cannot substitute for predictive theories of training dynamics, then specifies what a scientific explanation should accomplish.

      Introduction and Section 2, pp. 1-4 · Main paper

      Argument and section order

      Normative criteria lead to case studies and named open problems; an explicit Alternative Views section addresses engineering sufficiency, theory skepticism, automation and safety pragmatism before the conclusion.

      Sections 2-5, pp. 2-9 · Main paper

      Figures, captions and visual explanation

      Instead of a result leaderboard, colored example and open-problem boxes separate concrete illustrations from the continuous thesis; box labels make unresolved questions visibly distinct from supporting cases.

      Data-attribution box, p. 4; memorization/OpenFold boxes, p. 8; both PNGs inspected · Main paper

      Tables, comparisons and evidence

      The data-attribution example exposes how fixing a realized run versus retraining on reshuffled data changes the question. Case studies use cited results to motivate unresolved predictive problems, not claim new measurements.

      Section 2.3, p. 4; Sections 3.3-3.4, pp. 7-8 · Main paper

      Appendix and reproducibility

      There is no supplemental evidence appendix in the inspected artifact; the extensive bibliography supplies traceable destinations for the literature-based argument.

      References, pp. 9-17 · Main paper

      Clarity and writing conventions

      Counterarguments receive partial concessions: short-term safeguards and engineering progress are acknowledged, then the thesis is restated as a complementary long-term scientific agenda.

      Alternative Views, pp. 8-9 · Main paper

      What to reuse—and what not to infer

      Transferable lesson: Turn a position into a testable agenda with explicit criteria, carefully scoped examples and fair concessions to alternative priorities.

      Boundary: The paper's causal interpretations and research priorities are argued positions, not experimentally established findings of this manuscript; cited studies need independent checking before reuse.

      Apply this lesson using the argument and evidence worksheet.

      Inspection record and sources

      Checked
      2026-10-07
      Reading status
      Writing-focused manuscript scan + selected visuals inspected
      Main text
      Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-5, pp. 1-9: thesis, criteria for a science of AI, prediction/intervention/design ladder, case studies, alternative views and conclusion.
      Appendix / supplement
      Writing-focused appendix scan; code, data and external assets not executed. No appendix is attached to this 17-page proceedings PDF; pp. 9-17 are acknowledgments/references. Later pages were checked for the document boundary, not treated as experimental supplements.
      Visually inspected PDF pages
      4, 8

      This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.