Gemini frequently corrects board-written math errors with correction logs, but still shows case-consistency and notation-substitution failures¶
Claim · All claims
Evidence · 1 study (2 entries) · designr2–r3·q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i? Gemini often corrects obvious board-written physics or math errors on its own, with correction logs appended as comments, yet in the v6m transcription of video 1 it consistently used uppercase V rather than lowercase v for velocity despite an explicit contextual case-consistency directive. → Looney 2026
q2 i? Gemini occasionally makes systematic notation substitutions, such as transcribing the volume element dτ as dV in every transcription of video 10, presumably due to dominance in its training data. → Looney 2026 (2)
Evidence¶
Looney 2026¶
Looney, C. W., & Duston, C. L. (2026). Using Gemini and LuaLaTeX to transcribe physics videos into PDF/UA-2 and ISO 32005 math-accessible PDFs. https://arxiv.org/abs/2608.20733 doi:10.48550/arxiv.2608.20733
q2 · i? · design · r3
Qualitative evaluation of transcribed mathematics in Section 5.4: the authors report that "Gemini will often correct obvious board-written physics or math errors on its own", with correction logs preserved in the transcripts of videos 1, 3, 6, 8, 9 (test 2), 10, and 13. No effect size is reported.
"We have found that Gemini will often correct obvious board-written physics or math errors on its own."
Looney 2026 (2)¶
Looney, C. W., & Duston, C. L. (2026). Using Gemini and LuaLaTeX to transcribe physics videos into PDF/UA-2 and ISO 32005 math-accessible PDFs. https://arxiv.org/abs/2608.20733. doi:10.48550/arxiv.2608.20733
q2 · i? · design · r2
Qualitative evaluation in Section 5.4 of recurring notation substitution: the authors report Gemini "has made this exact same substitution every single time" for video 10, attributing it to training-data dominance. No effect size is reported.
"in the v6m transcription of video 10, Gemini transcribed the board-written volume element 𝑑𝜏 as 𝑑𝑉 . Versions of this video have been transcribed several times by multiple iterations of the system instructions, and Gemini has made this exact same substitution every single time, presumably due to the dominance of 𝑑𝑉 in Gemini’s training data."
Discussion¶
Related Claims¶
- The v6m subscript framework robustly transcribed a wide range of board-written subscript configurations, with stacked-subscript unstacking less reliable — related
- Successful transcriptions of videos using Dirac notation, eigenstate derivations, and vector calculus indicate generalized capability beyond introductory physics — related
- The v6m workflow produced compilations passing PDF/UA-2 and ISO 32005 validation in 16 of 17 successive transcription tests across 16 videos — related
- Most AI grading errors (87% for the best model) were attributable to transcription failures rather than rubric misapplication — related
- Gold text descriptions improve VLM correctness-and-error assessment, but performance still lags other question types — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
gemini-error-correction-with-case-consistency-failures |
title |
Gemini frequently corrects board-written math errors with correction logs, but still shows case-consistency and notation-substitution failures |
description |
Gemini frequently corrects board-written math errors with correction logs, but still shows case-consistency and notation-substitution failures |
status |
review |
evidence_strength |
mixed |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-10 |
trust |
unverified — no one has confirmed this page's content |
sources |
2 — looney-2026, looney-2026-2 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.