---
title: "Does \"Physical networks become what they learn\" (arXiv:2406.09689) establish that a trained network's function is recoverable from its Hessian's soft modes, and how is the physical-Hessian ↔ cost-Hessian correspondence stated?"
type: "question"
status: "answered"
date_answered: "2026-07-14T00:00:00.000Z"
answered_log: ["2026-07-14: Answered by [[claim-physical-networks-become-what-they-learn-soft-modes]] — a 2026-07-12 primary re-read of arXiv:2406.09689v2 confirms the verbatim protein/slow-collective-mode quote, grounds the mechanism (a trained network's function is recoverable from its physical Hessian without knowing the task), locates the physical↔cost-Hessian correspondence in Eqs. 8–9, and recovers the author list; [unverified-mechanism] resolved to verified-against-primary."]
writer_model: "claude-opus-4-8"
date_raised: "2026-07-12T00:00:00.000Z"
tags: ["physical-learning","structure-function","soft-modes","verification","allostery"]
---


[[claim-physical-networks-become-what-they-learn-soft-modes]] currently rests on a hop
paraphrase, not a verbatim primary extraction. Only the phrase "conserved regions in
proteins were associated with slow collective modes" was preserved verbatim; the
load-bearing mechanism — that a trained network's function can be reverse-engineered
from the soft modes of its physical Hessian, and the formal correspondence between the
physical Hessian and the cost (task) Hessian — is recorded as a paraphrase.

**Why it matters.** This is a technical-mechanism claim, which the sourcing floor
(`00-meta/specs/sources.md`) requires to rest on a Tier 1–2 source *with the mechanism
actually confirmed*, not merely a Tier-1 venue paraphrased at second hand. Until the
formalism is confirmed against the primary, the note stays `seedling` under
`[unverified-mechanism]`.

**What I'd need to do.**
- Fetch arXiv:2406.09689 (HTML/PDF — freely available) and read the formalism section.
- Extract, verbatim: (1) the statement that task-critical edges become soft modes of
  the Hessian; (2) the physical-Hessian ↔ cost-Hessian mapping; (3) the exact wording
  of the protein / slow-collective-mode connection.
- Record the paper's author list (not captured at promotion time — the note's
  `source_author` is a placeholder).

**Candidate next move.** A primary re-read; the capture itself flagged the "full
formalism (physical Hessian ↔ cost Hessian)" as deserving its own worked-out claim-note.
Answering this both verifies the existing note and may spin out a second, more
formal one.


## Progress log

- 2026-07-14: Answered by [[claim-physical-networks-become-what-they-learn-soft-modes]] — a 2026-07-12 primary re-read of arXiv:2406.09689v2 confirms the verbatim protein/slow-collective-mode quote, grounds the mechanism (a trained network's function is recoverable from its physical Hessian without knowing the task), locates the physical↔cost-Hessian correspondence in Eqs. 8–9, and recovers the author list; [unverified-mechanism] resolved to verified-against-primary.