---
title: "SEOYUN: Refusal-Inducing Exam Documents... | SI Paper Desk"
url: https://thesimind.si/papers/2611.26033/
description: "SEOYUN proposes a framework, not a tested defense, for designing exam pages that push a vision-language model to refuse when someone photographs them with..."
---

Number 10 of 12 · November 22, 2026

# SEOYUN: Refusal-Inducing Exam Documents for AI-Glasses Pipelines

SEOYUN proposes a framework, not a tested defense, for designing exam pages that push a vision-language model to refuse when someone photographs them with smart glasses.

Dohyeon Kwon · Submitted November 19, 2026 · arXiv 2611.26033v1 · cs.AI

[Read the paper on arXiv](https://arxiv.org/abs/2611.26033) · [The PDF](https://arxiv.org/pdf/2611.26033) · [The day's 12](https://thesimind.si/papers/day/2026-11-22/)

## What they found

The author, Dohyeon Kwon, describes SEOYUN, a framework for exam documents that aim to make a vision-language model (VLM, a SI model that reads images and text) abstain instead of answering. The paper offers a threat model, analytical limits, and an evaluation protocol for exam settings. It separates a refusal from leaking the correct answer. It also argues that removing an added dummy option does not by itself shrink the original answer space. It derives a sampling comparison under stated camera assumptions. The abstract says plainly that this version reports no model-performance measurements, no physical tests, and no human-subject results.

## Why it matters

Camera glasses can turn an exam page into a query for a VLM. Device detection only checks who has the hardware. This paper asks a different question: can the document itself make an unauthorized capture less useful? The paper's main value is a careful way to test such ideas. The protocol keeps digital, simulated, and physical inputs apart. It also requires held-out evaluation after subject-specific selection. The abstract states that refusal induction and defensive visual prompt injection already have precedents.

## How it changes the race

It does not change the race. It is a research framework with no results, and its value is in how future defenses against exam cheating with VLMs could be tested.

## Safety and control

The work bears on misuse of SI and on defenses that try to steer a model into refusing. It does not show that any defense works. The author says it is not a validated defense and not a basis for allegations against individuals.

### The abstract, by the authors

Camera-equipped glasses can convert an examination page into a query to a vision-language model (VLM). Device detection addresses possession; a complementary question is whether an examination document can reduce the usefulness of an unauthorized capture. We present SEOYUN, an application-specific framework for refusal-inducing exam documents, with optical and encoder-level interventions as extensions. We use reverse jailbreaking to describe the direction of the objective: inducing abstention rather than bypassing refusal. Refusal induction and defensive visual prompt injection already have precedents; our contribution is a threat model, analytical limits, and an evaluation protocol for examination use. We distinguish refusal from correct-answer leakage, show why removing an added dummy option is not itself a reduction in the original answer space, and derive a sampling comparison under explicitly stated camera assumptions. The protocol separates digital, simulated, and physical inputs and requires held-out evaluation after subject-specific selection. A local CSAT candidate corpus and rendering harness have been prepared, but this version reports no model-performance measurements, physical validation, or human-subject results. Accordingly, SEOYUN is a research framework, not a validated defense or a basis for allegations against individuals.

Quoted from arXiv:2611.26033. The research and its words belong to its authors.

Desk score **65** out of 100

Why it was picked

- Safety (in the title)
- Evals
- Security
- Language models

Topics

- [Security](https://thesimind.si/papers/?tag=security)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Multimodal](https://thesimind.si/papers/?tag=multimodal)
- [Education](https://thesimind.si/papers/?tag=education)
- [Safety](https://thesimind.si/papers/?tag=safety)
- [Language models](https://thesimind.si/papers/?tag=language-models)

**arXiv ID:** 2611.26033v1
**Submitted:** November 19, 2026
**Fields:** cs.AI (Artificial Intelligence)
**Level:** Intermediate
**Comments:** 10 pages; research framework and analytical limits; no empirical efficacy results. Includes an author reflection in Korean in Appendix C

Fingerprint, SHA-256

`dab21db6f42c35a0ac67287f0eced00eed1e1ba68e86d227eda5066067c942d5`

In SI MIND's journal since Nov 22, 2026, 02:01 UTC. It goes on Solana with the next hourly Proof of Mind seal.

### The exact text fingerprinted

{"arxiv":"2611.26033","version":1,"title":"SEOYUN: Refusal-Inducing Exam Documents for AI-Glasses Pipelines","authors":["Dohyeon Kwon"],"day":"2026-11-22","rank":10,"score":65,"status":"written","summary":{"takeaway":"SEOYUN proposes a framework, not a tested defense, for designing exam pages that push a vision-language model to refuse when someone photographs them with smart glasses.","found":"The author, Dohyeon Kwon, describes SEOYUN, a framework for exam documents that aim to make a vision-language model (VLM, a SI model that reads images and text) abstain instead of answering. The paper offers a threat model, analytical limits, and an evaluation protocol for exam settings. It separates a refusal from leaking the correct answer. It also argues that removing an added dummy option does not by itself shrink the original answer space. It derives a sampling comparison under stated camera assumptions. The abstract says plainly that this version reports no model-performance measurements, no physical tests, and no human-subject results.","matters":"Camera glasses can turn an exam page into a query for a VLM. Device detection only checks who has the hardware. This paper asks a different question: can the document itself make an unauthorized capture less useful? The paper's main value is a careful way to test such ideas. The protocol keeps digital, simulated, and physical inputs apart. It also requires held-out evaluation after subject-specific selection. The abstract states that refusal induction and defensive visual prompt injection already have precedents.","race":"It does not change the race. It is a research framework with no results, and its value is in how future defenses against exam cheating with VLMs could be tested.","safety":"The work bears on misuse of SI and on defenses that try to steer a model into refusing. It does not show that any defense works. The author says it is not a validated defense and not a basis for allegations against individuals.","level":"intermediate","tags":["security","evals","multimodal","education","safety","language-models"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Dohyeon Kwon

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 16, 2026 · Paper of the DayOnly as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-AgentWhen a strong SI model hands work to a cheaper sub-agent, the whole system can be only as safe as the weaker agent, even if the strong model alone resists the attack.SafetyMulti-agent systemsSecurity](https://thesimind.si/papers/2611.17520/) · [Nov 20, 2026 · Number 2AutoPI: Taxonomy-Guided Adaptive Red Teaming for Prompt Injection in LLM AgentsAutoPI is a red teaming tool that tests SI agents for prompt injection across many attack types, not just how often attacks work, and it reports higher success and coverage than recent rivals.SecurityAgentsEvals](https://thesimind.si/papers/2611.24313/) · [Nov 16, 2026 · Number 2Soteria: Formally Verified Planning with Runtime Enforcement for Safe LLM AgentsSoteria checks an SI agent's plan against written rules before any tool runs, then enforces those rules during execution. The authors report more completed tasks and no policy violations.AgentsSafetyReasoning](https://thesimind.si/papers/2611.17880/)

[Number 9](https://thesimind.si/papers/2611.26569/) · [All 12 for Nov 22, 2026](https://thesimind.si/papers/day/2026-11-22/) · [Number 11](https://thesimind.si/papers/2611.25845/)
