# Consistency Is Not Understanding: Orbit Audits for Neural Field Weight Readers [Canonical reading](https://escapemeridian.com/readings/consistency-is-not-understanding-orbit-audits-for-neural-field-48b07091) The authors of “Consistency Is Not Understanding: Orbit Audits for Neural Field Weight Readers” argue that a reader can give consistent outputs for equivalent neural-network weights without thereby showing that it understands the represented content. Hidden-neuron permutations and positive ReLU rescalings change parameter coordinates while preserving the represented function, giving the authors a way to test whether readers rely on arbitrary coordinates or functional information. Their evaluation compares a statistics reader and a normalized set reader on shape recognition, and tests frozen language models and a reasoning setup on weight-reading prompts. The abstract reports that the normalized set reader retains accuracy with unchanged predictions under joint transformations; it also reports disagreement between fresh, identical raw prompts. Canonicalization makes equivalent prompts identical, but the authors explicitly caution that this neither establishes recognition nor guarantees identical generated responses. The authors present joint correctness/agreement metrics, semantic-cue controls, fresh repeats and exact-rational rounding bounds as an inspectable evaluation protocol. They restrict their claims to the tested fields, prompts and execution configurations. The abstract’s accuracy values appear as unreplaced placeholders, so their numerical results cannot be read from Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Nov 4, 2026 (2026-11-04). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-09T12:19:55.897Z. First observed: 2026-11-05T05:00:28.404Z. Retrieved: 2026-11-09T12:19:01.687Z. - [Consistency Is Not Understanding: Orbit Audits for Neural Field Weight Readers](https://arxiv.org/abs/2611.05798) ## Limits The supplied abstract contains unreplaced placeholders for the reported accuracy and disagreement rates, so those numerical results cannot be assessed from this evidence. The evidence is the arXiv abstract and page, not the full paper; detailed methods and results beyond the abstract are unavailable here. ## Open question How well do the reported results and evaluation protocol generalize beyond the tested fields, prompts and execution configurations? ## Source-stated The authors test whether visual weight readers separate functional information from arbitrary parameter coordinates using transformations that preserve the represented function. > We audit this separation with hidden-neuron permutations and positive ReLU rescalings that preserve the represented function. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source-stated For shape recognition, the abstract reports that a normalized set reader retains accuracy under joint transformations, with unchanged predictions; the reported accuracy value is unavailable in the supplied text. > a normalized set reader retains \\ShapeNormSet\\% with unchanged predictions. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source-stated Fresh identical raw prompts disagree on some origins in the reasoning experiment; the abstract’s reported rate is an unreplaced placeholder. > fresh identical raw prompts disagree on \\FrontRawRepeatRange\\% of origins. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source-stated Canonicalization makes equivalent prompts identical, but the authors say this does not establish recognition or guarantee identical generated responses. > Canonicalization makes equivalent prompts identical, but this neither establishes recognition nor guarantees identical generated responses. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source-stated The authors describe the work as an inspectable evaluation protocol and limit its claims to the tested fields, prompts and execution configurations. > The result is an inspectable evaluation protocol for visual weight readers, with claims restricted to the tested fields, prompts, and execution configurations. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source and contributor context Cha, Sungguk, Kim, Mintae, Han, Youngsub, Jeon, Byoung-Ki, Lee, Sangyeob · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Neural network A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data. - [Definition reference](https://developers.google.com/machine-learning/glossary#neural-network) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 8ce9055a-86a5-4996-81af-6b875c5c57a2. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T12:19:55.897Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 1749edcd6352f6d939512d2ddc7f222afd064d3efee427babb8c4c4227a769db. Source text hash: 28318f65936b72c71a7eda14ad4d07002b333eb083b3802b10fca61d0d82e4f3. ## Revision and source-check history ### Version 1 — 2026-11-09T12:19:55.897Z The authors of “Consistency Is Not Understanding: Orbit Audits for Neural Field Weight Readers” argue that a reader can give consistent outputs for equivalent neural-network weights without thereby showing that it understands the represented content. Hidden-neuron permutations and positive ReLU rescalings change parameter coordinates while preserving the represented function, giving the authors a way to test whether readers rely on arbitrary coordinates or functional information. Their evaluation compares a statistics reader and a normalized set reader on shape recognition, and tests frozen language models and a reasoning setup on weight-reading prompts. The abstract reports that the normalized set reader retains accuracy with unchanged predictions under joint transformations; it also reports disagreement between fresh, identical raw prompts. Canonicalization makes equivalent prompts identical, but the authors explicitly caution that this neither establishes recognition nor guarantees identical generated responses. The authors present joint correctness/agreement metrics, semantic-cue controls, fresh repeats and exact-rational rounding bounds as an inspectable evaluation protocol. They restrict their claims to the tested fields, prompts and execution configurations. The abstract’s accuracy values appear as unreplaced placeholders, so their numerical results cannot be read from ## Original source arXiv cs.AI · Article Original source date: Nov 4, 2026 (2026-11-04). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-09T12:19:55.897Z. First observed: 2026-11-05T05:00:28.404Z. Retrieved: 2026-11-09T12:19:01.687Z. - [Consistency Is Not Understanding: Orbit Audits for Neural Field Weight Readers](https://arxiv.org/abs/2611.05798) ## Limits The supplied abstract contains unreplaced placeholders for the reported accuracy and disagreement rates, so those numerical results cannot be assessed from this evidence. The evidence is the arXiv abstract and page, not the full paper; detailed methods and results beyond the abstract are unavailable here. ## Open question How well do the reported results and evaluation protocol generalize beyond the tested fields, prompts and execution configurations? ## Source-stated The authors test whether visual weight readers separate functional information from arbitrary parameter coordinates using transformations that preserve the represented function. > We audit this separation with hidden-neuron permutations and positive ReLU rescalings that preserve the represented function. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source-stated For shape recognition, the abstract reports that a normalized set reader retains accuracy under joint transformations, with unchanged predictions; the reported accuracy value is unavailable in the supplied text. > a normalized set reader retains \\ShapeNormSet\\% with unchanged predictions. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source-stated Fresh identical raw prompts disagree on some origins in the reasoning experiment; the abstract’s reported rate is an unreplaced placeholder. > fresh identical raw prompts disagree on \\FrontRawRepeatRange\\% of origins. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source-stated Canonicalization makes equivalent prompts identical, but the authors say this does not establish recognition or guarantee identical generated responses. > Canonicalization makes equivalent prompts identical, but this neither establishes recognition nor guarantees identical generated responses. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source-stated The authors describe the work as an inspectable evaluation protocol and limit its claims to the tested fields, prompts and execution configurations. > The result is an inspectable evaluation protocol for visual weight readers, with claims restricted to the tested fields, prompts, and execution configurations. - [Supporting source](https://arxiv.org/abs/2611.05798) ## Source and contributor context Cha, Sungguk, Kim, Mintae, Han, Youngsub, Jeon, Byoung-Ki, Lee, Sangyeob · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Neural network A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data. - [Definition reference](https://developers.google.com/machine-learning/glossary#neural-network) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 8ce9055a-86a5-4996-81af-6b875c5c57a2. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T12:19:55.897Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 1749edcd6352f6d939512d2ddc7f222afd064d3efee427babb8c4c4227a769db. Source text hash: 28318f65936b72c71a7eda14ad4d07002b333eb083b3802b10fca61d0d82e4f3. ### uncertain — 2026-11-09T12:19:55.897Z — anthropic/claude-opus-5.5 Publication: 8ce9055a-86a5-4996-81af-6b875c5c57a2 attribution — pass: The reading attributes the argument and results to the paper's authors, using the arXiv title and abstract. It does not present the authors' claims as established fact. > Authors:Sungguk Cha, Mintae Kim, Youngsub Han, Byoung-Ki Jeon, Sangyeob Lee summary — uncertain: The summary is cut off mid-sentence, ending with 'cannot be read from'. Smallest repair: complete it, for example by adding 'the supplied abstract.' Central terms are not explained for a general reader, including neural field, weight reader, ReLU, hidden-neuron permutations and canonicalization. Without them, a non-specialist cannot follow what was tested or why it matters. Keep the exact terms but explain them through term-library tooltips, adding entries where missing. Apart from these gaps, the faithful account and the authors' caution are preserved. date — pass: The originalPublished date of 2026-11-04 matches the v1 submission date shown on the arXiv page. The 2026-11-05 feed date is correctly not used. > \[Submitted on 4 Nov 2026\] support — pass: Each claim matches the abstract and is flagged as source-stated. The reading correctly notes that the numeric results are unreplaced placeholders. The question about generalization follows from the authors' own restriction of scope. No benefits, costs or outcomes are invented. > Canonicalization makes equivalent prompts identical, but this neither establishes recognition nor guarantees identical generated responses. media — pass: The selected media is the verified arXiv abstract page, which is the direct original source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/48b07091b45342847bc9f86b8b4db215aa4ddad626c892d6a4ff1f2de0626fdf/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/consistency-is-not-understanding-orbit-audits-for-neural-field-48b07091.json)