Accessibility Concept Emergence in the Pythia Suite: Thresholds,  Binding, and the Declarative-Evaluative Gap

preprint OA: closed CC-BY-4.0

Abstract

Sustained deep-network binding of accessibility compounds appears to be a necessary structural condition for behavioral capability — present in every model that correctly defines core concepts, absent in every model that fails. We use Web Content Accessibility Guidelines (WCAG) as our test domain because accessibility represents a specialized, low-frequency domain in web-scale training data — a small, well-defined vocabulary with unambiguous answers and direct relevance to real-world tooling decisions, making it an unusually clean lens for studying emergence: concepts are concrete enough to evaluate and rare enough to show scale sensitivity. Using the Pythia suite (160M–12B) and GPT-2 (small–XL) with TransformerLens, we investigate not just what models know but how that knowledge is encoded internally. All models show compound binding in early layers; the differentiating factor is whether that binding persists to late network layers. Screen reader, skip link, and alt text emerge behaviorally at ~2.8B; WCAG first appears at 6.9B; Accessible Rich Internet Applications (ARIA) exhibits fluent wrongness at every scale tested — producing confident, plausible expansions that are consistently incorrect. Models prefer correct definitions before they can produce them, and a declarative-evaluative gap persists even at maximum scale: models that correctly define accessibility concepts cannot reliably identify violations in code. The gap is robust across 15 prompts spanning three elicitation strategies and is not an artifact of prompt design. Entropy analysis reveals that the gap has internal structure — the model enters distinct failure states depending on how it is asked, from high-entropy stalling to low-entropy confident parroting. Extending the binding analysis to Pythia 12B introduces a late-layer resurgence pattern: a cluster of binding heads re-engaging near the output layers that scales monotonically with model size.
Full text 3,027 characters · extracted from oa-doi-fallback · 2 sections · click to expand

Abstract

Sustained deep-network binding of accessibility compounds appears to be a necessary structural condition for behavioral capability — present in every model that correctly defines core concepts, absent in every model that fails. We use Web Content Accessibility Guidelines (WCAG) as our test domain because accessibility represents a specialized, low-frequency domain in web-scale training data — a small, well-defined vocabulary with unambiguous answers and direct relevance to real-world tooling decisions, making it an unusually clean lens for studying emergence: concepts are concrete enough to evaluate and rare enough to show scale sensitivity. Using the Pythia suite (160M–12B) and GPT-2 (small–XL) with TransformerLens, we investigate not just what models know but how that knowledge is encoded internally. All models show compound binding in early layers; the differentiating factor is whether that binding persists to late network layers. Screen reader, skip link, and alt text emerge behaviorally at ~2.8B; WCAG first appears at 6.9B; Accessible Rich Internet Applications (ARIA) exhibits fluent wrongness at every scale tested — producing confident, plausible expansions that are consistently incorrect. Models prefer correct definitions before they can produce them, and a declarative-evaluative gap persists even at maximum scale: models that correctly define accessibility concepts cannot reliably identify violations in code. The gap is robust across 15 prompts spanning three elicitation strategies and is not an artifact of prompt design. Entropy analysis reveals that the gap has internal structure — the model enters distinct failure states depending on how it is asked, from high-entropy stalling to low-entropy confident parroting. Extending the binding analysis to Pythia 12B introduces a late-layer resurgence pattern: a cluster of binding heads re-engaging near the output layers that scales monotonically with model size. Supplementary Material File (accessibility-knowledge-emergence-v3.pdf) - Download - 549.53 KB Information & Authors Information Copyright This work is licensed under a Creative Commons Attribution 4.0 International License

Keywords

Authors Metrics & Citations Metrics Article Usage 205views 90downloads Citations Download citation Trisha Salas. Accessibility Concept Emergence in the Pythia Suite: Thresholds, Binding, and the Declarative-Evaluative Gap. Authorea. 25 March 2026. DOI: https://doi.org/10.22541/au.177282002.24340653/v2 DOI: https://doi.org/10.22541/au.177282002.24340653/v2 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu.

Text is read by the "Ask this paper" AI Q&A widget below. Extraction quality varies by source — PMC NXML preserves structure cleanly, OA-HTML may include some navigation residue, and OA-PDF can have broken hyphenation. The publisher copy (via DOI) is the canonical version.

My notes (saved in your browser only)

Ask this paper AI returns verbatim quotes from the full text · source: oa-doi-fallback

Answers must be backed by verbatim quotes from this paper's full text. Hallucinated quotes are dropped automatically; if no verbatim passage answers the question, we say so. How this works

Citation neighborhood (no data yet)

We don't have any in-corpus citations linked to this paper yet. This is a recent paper (2026) — citers typically take a year or two to land, and the OpenAlex reference graph may still be filling in.

Source provenance

europepmc
last seen: 2026-05-20T01:45:00.602351+00:00
unpaywall
last seen: 2026-05-24T02:00:01.246996+00:00
License: CC-BY-4.0