[smlee] 학생자율연구 기획

*정보 정리성 게시글이며, 계속해서 추가 및 정리 예정입니다.

 

학부생연구지원프로그램: https://snuc.snu.ac.kr/%ea%b3%b5%ec%a7%80%ec%82%ac%ed%95%ad/?mod=document&uid=631 -> 4/13까지 신청 예정

[신청 시 필요 정보]
- 연구자: 성함, 소속 (단과대, 학과), 전화번호, 이메일, 등록학기, 직전학기 평점 평균, 전자서명 필요
- 연구지원조교: 학번, 박사 n학기/수료 정보, 전자서명 필요 
- 지도교수: 연구 과제 추천 사유, 전자서명 필요

- 연구계획서, 연구비 집행 계획서 작성 예정 @이수민
- 연구과제 추천 및 지도교수 승인서 -> 승인 요청 후 작성 예정

 

연구 방향: Biosignal * LLM 통합

경훈님 피드백: https://team-sudal.tistory.com/145

 

연구자 IRB 교육 수강 -> SNU IRB 정회원 등업 필요

---

Archived Topic 1-4 by smlee, Topic 5 by tien, Topic 6&7 by sblee

Topic Candidates Ranking (by Claude, prompted to see novelty based on related works)

 

#1 — Topic 4: Zombie Clicks → STRONG GAP ✅
#2 — Topic 6: Gaze × AI Hallucinations → MODERATE GAP ✅
#3 — Topic 1: Physiological Transparency → MODERATE GAP ⚠️
#4 — Topic 2: BodyDraft → MODERATE GAP, leaning weak ⚠️
#5 — Topic 7: Biosignal → Uncanny Valley → MODERATE GAP, leaning weak ⚠️
#6 — Topic 3: CogLoad Interruption → WEAK GAP ⛔
#7 — Topic 5: Biosignal → Music → DEAD ☠️

---

Topic 4 — Zombie Clicks: Automation Bias vs. Cognitive Offloading via Pupillometry (동공측정기)

Verdict: STRONG GAP ✅ — Submit

RQ: Can pupillometry distinguish unconscious automation bias (zombie click) from intentional cognitive offloading (strategic click) within the same observable behavior of accepting an AI recommendation?

Verified Prior Works

Buçinca, Malaya & Gajos — "To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making"
CSCW 2021, Proc. ACM Hum.-Comput. Interact. Vol. 5, No. CSCW1, Article 188
✅ Verified · https://dl.acm.org/doi/10.1145/3449287
Establishes dual-process theory (System 1 vs. System 2) as the framework for AI overreliance. Designs and tests cognitive forcing interventions. Behavioral measures only — no physiological measurement. This is the theoretical predecessor your paper operationalizes.

Deniz, Liu & Hassib — "Measuring the Effect of Mental Workload and Explanations on Appropriate AI Reliance Using EEG"
Behaviour & Information Technology, published online November 2024
✅ Verified · https://www.tandfonline.com/doi/full/10.1080/0144929X.2024.2431055
Closest physiological approach. Uses EEG to measure workload's effect on AI reliance patterns. Treats reliance as a single construct — no dissociation between automation bias and cognitive offloading.

Lee & Gutzwiller — "Do the Eyes Have It? A Review of Using Eye Tracking for Automation Trust Measurement"
Human Factors, 2025
✅ Verified · https://journals.sagepub.com/doi/10.1177/00187208251348395
Comprehensive review concluding that the trust-eye-tracking relationship is "inconsistent and unreliable." This is useful for you: the inconsistency exists precisely because prior work did not distinguish automation bias from intentional offloading. Your paper explains the inconsistency.

Gaube et al. — "Eye Tracking Insights into Physician Behaviour with Safe and Unsafe AI Recommendations"
npj Digital Medicine, 2024
✅ Verified · https://pmc.ncbi.nlm.nih.gov/articles/PMC11297294/
Uses fixation duration and blink rate to study physician engagement with AI recommendations. Shows gaze metrics differentiate engagement levels but does not attempt within-behavior cognitive process dissociation.

Romeo — "Exploring Automation Bias in Human-AI Collaboration"
AI & Society, 2025
✅ Verified · https://link.springer.com/article/10.1007/s00146-025-02422-7
Conceptual review of automation bias in HAI. No physiological measures. Sets the problem space.

The Gap

No paper dissociates automation bias from cognitive offloading within the same surface behavior using pupillometry. The contribution is the first empirical operationalization of a theoretical distinction Buçinca et al. (2021) call for but never deliver.

---

Topic 6 — Gaze Patterns on AI Hallucinations × Critical Thinking

Verdict: MODERATE GAP ✅ — Submit with a strong theoretical model

RQ: Do gaze patterns while reading AI-generated hallucinations differ by critical thinking skills and AI usage duration?

Verified Prior Works

"Enhancing Students' Critical Thinking Literacy in a Generative AI Context: Eye Movement Patterns of Deepfake Detection"
Computers & Education, 2025
✅ Verified · https://www.sciencedirect.com/science/article/pii/S0360131525002982
Your biggest threat. Studies eye tracking + critical thinking + AI-generated content. Studies visual deepfakes, not textual LLM hallucinations. Uses pre/post intervention design, not individual differences as predictor. Must be cited and differentiated explicitly.

Sümer et al. — FakeNewsPerception Eye Movement Dataset
Data in Brief, 2021
✅ Verified · https://www.sciencedirect.com/science/article/pii/S2352340921001931
Collected eye movements during fake/real news reading alongside Cognitive Reflection Test scores — essentially your design applied to traditional misinformation rather than AI text.

"Misinformation Identification as a Digital Literacy Skill: An Eye Tracking Study"
Humanities and Social Sciences Communications (Nature), 2025
✅ Verified · https://www.nature.com/articles/s41599-025-04938-1
N=83 study linking digital literacy to gaze patterns during credibility assessment. Establishes the individual-differences → gaze-patterns → misinformation-detection pipeline.

"Reading the Readers Mind through Eye Tracking: Can AI Generated Texts Match Human Authors?"
ETRA 2025
✅ Verified · https://dl.acm.org/doi/10.1145/3715669.3726846
Eye tracking distinguishes AI-generated from human text through reading behavior. No hallucination focus, no critical thinking variable, no individual differences.

Nahar et al. — "Fakes of Varying Shades: How Warning Affects Human Perception Regarding LLM Hallucinations"
COLM 2024
✅ Verified · https://openreview.net/forum?id=c30qeMg8dv
N=419 study of human perception of LLM hallucinations. Self-report only, no gaze.

The Gap

No paper studies gaze during reading of LLM-generated textual hallucinations with critical thinking as moderator. The AI usage duration variable is entirely novel. The three-way intersection (CT × gaze × AI hallucination) is unstudied.

---

Topic 1 — Physiological Transparency in Adaptive AI

Verdict: MODERATE GAP ⚠️ — Viable, rename required

RQ: Is transparent biosignal surfacing better than silent AI adaptation for autonomy, metacognition, and task outcomes?

Verified Prior Works

Benke, Gnewuch & Maedche — "Understanding the Impact of Control Levels Over Emotion-Aware Chatbots"
Computers in Human Behavior, 2022
✅ Verified · https://www.sciencedirect.com/science/article/abs/pii/S0747563221004453
N=176 comparing control levels over an affect-aware chatbot. Same DVs and design logic as proposed work. Uses text-based emotion detection, not physiological biosignals.

Chiossi, Turgut, Welsch & Mayer — "Adapting Visual Complexity Based on Electrodermal Activity Improves Working Memory Performance in Virtual Reality"
Proc. ACM Hum.-Comput. Interact. 7(MHCI), Article 196, 2023
✅ Verified · https://dl.acm.org/doi/10.1145/3604243
State-of-the-art EDA-based silent adaptation in VR. Implements your Condition A. No transparency comparison.

Dongre et al. — "Integrating Physiological Data with Large Language Models for Empathic Human-AI Interaction"
PhysioCHI Workshop, CHI 2024 (also a Doctoral Consortium paper, CHI EA 2024)
✅ Verified · arXiv: https://arxiv.org/abs/2404.15351
LLM + EDA/BVP/ST for adaptive chatbot. Pilot study. Implements silent physiological LLM adaptation. No transparency comparison.

"PhysioLLM: Supporting Personalized Health Insights with Wearables and Large Language Models"
arXiv 2024
✅ Verified · https://arxiv.org/abs/2406.19283
Surfaces physiological data to users via LLM — essentially your Condition B in health context. N=24. No silent adaptation comparator.

Pu et al. — "Assistance or Disruption?" (Codellaborator)
CHI 2025
✅ Verified · https://dl.acm.org/doi/10.1145/3706598.3713357
The "presence and context" condition of Codellaborator surfaces AI process transparency to users — adjacent to your transparency concept, but this is transparency of the AI's actions, not of the user's own physiological state. Must be cited and distinguished.

The Gap

No study directly compares biosignal-triggered silent AI adaptation vs. biosignal-surfacing user transparency in a controlled experiment. ⚠️ Rename required: "Cognitive Mirror" is taken by Frontiers in Education 2025.

---

Topic 2 — BodyDraft: EDA/HRV Biofeedback During Creative Writing

Verdict: MODERATE GAP, leaning weak ⚠️ — Risky without domain-specific findings

Verified Prior Works

Yang, Feng & Kalantari — "Design with Myself: A BCI Tool that Predicts Live Emotion to Enhance Metacognitive Monitoring of Designers"
CHI EA 2023 version: DOI 10.1145/3544549.3585701; full IJHCS 2024 version
✅ Verified · https://www.sciencedirect.com/science/article/abs/pii/S1071581924000132
Uses EEG emotional detection during architectural design for metacognitive biofeedback during a creative process. Shares three core pillars with BodyDraft. A reviewer will immediately ask how this differs from Multi-Self applied to writing instead of design.

Frey, Ostrin, Grabli & Cauchard — "Physiologically Driven Storytelling: Concept and Software Tool"
CHI 2020, Best Paper Honorable Mention
✅ Verified · https://dl.acm.org/doi/10.1145/3313831.3376643
Note: this paper received a Best Paper Honorable Mention, which increases its visibility. Uses EDA + breathing + eye tracking with text. Focused on reading/consumption, not writing/production.

Mitrevska & Mayer — "Physiological Signals as Implicit Multimodal Input in Human-AI Interactions During Creative Tasks"
GenAICHI Workshop, CHI 2024
✅ Verified · https://sven-mayer.com/wp-content/uploads/2024/07/mitrevska2024physiological.pdf
Short position paper (3 pages) that explicitly proposes EDA during creative writing. No implementation but establishes conceptual priority.

Haghighi & Satyanarayan — "Self-Interfaces: Utilizing Real-Time Biofeedback in the Wild to Elicit Subconscious Behavior Change"
TEI 2020 (Work-in-Progress paper, not a full paper)
✅ Verified · https://dl.acm.org/doi/10.1145/3374920.3374979
Correction from previous drafts: this is a work-in-progress paper (6 pages), not a full TEI paper. Presents design and early development of a wearable haptic EDA biofeedback interface for self-awareness. Relevant as a mechanism predecessor but lower weight than a full paper.

The Gap

No fully implemented system provides real-time EDA/HRV biofeedback integrated into a creative writing interface. The contribution must rest on writing-domain-specific findings that could not have come from any other creative domain.

---

Topic 7 — Biosignal → Fixing Uncanny Valley in AI-Generated Images

Verdict: MODERATE GAP, leaning weak ⚠️ — Narrow gap, wrong venue

Verified Prior Works

De la Torre-Ortiz et al. — "Brain Relevance Feedback for Interactive Image Generation"
UIST 2020
✅ Verified · https://dl.acm.org/doi/abs/10.1145/3379337.3415821
Closed-loop EEG → GAN for face generation. N=31, 86.26% accuracy. Same pipeline, different signal (EEG vs. EDA) and target (relevance/attractiveness vs. uncanny valley).

Spapé et al. — "Brain-Computer Interface for Generating Personally Attractive Images"
IEEE Transactions on Affective Computing, 2023
✅ Verified · https://doi.org/10.1109/TAFFC.2021.3059043
EEG captures attractiveness reactions to GAN faces. Same paradigm applied to a different quality dimension (attractiveness vs. uncanny valley).

Ratajczyk et al. — "Evaluation of the Uncanny Valley Hypothesis Based on EDA"
Bio-Algorithms and Med-Systems, 2019 (DOI: 10.1515/bams-2019-0008)
⚠️ Partially verified — journal exists, paper is cited in literature but not directly confirmed via ACM DL. Finding: EDA correlates with uncanny valley ratings (Spearman -0.61) but no significant relationship between declared comfort and EDA was found, which could undermine the premise of using EDA as a clean proxy.

Chen et al. — "Realness of Face Images Can Be Decoded from Non-linear EEG Responses"
Scientific Reports, 2024
✅ Verified · https://doi.org/10.1038/s41598-024-56130-1
Neural signals decode face realness consistent with uncanny valley. Measurement only, no generative feedback.

---

Topic 3 — Cognitive Load-Gated Interruption for Agentic AI

Verdict: WEAK GAP ⛔ — Core mechanism is 20+ years old

Verified Prior Works

Iqbal, Zheng & Bailey — "Task-Evoked Pupillary Response to Mental Workload in Human-Computer Interaction"
CHI EA 2004 (Extended Abstracts, 4 pages)
✅ Verified · https://dl.acm.org/doi/10.1145/985921.986094
Correction from previous drafts: this is in CHI Extended Abstracts, not a full CHI paper. However, it directly links pupillometry to interruption management in HCI.

Iqbal, Adamczyk, Zheng & Bailey — "Towards an Index of Opportunity: Understanding Changes in Mental Workload During Task Execution"
CHI 2005 (full paper)
✅ Verified (cited in Semantic Scholar, ResearchGate, and ACM proceedings)
This is the full paper. Builds computational "Index of Opportunity" from pupillary data to predict best moments for interruption. The exact concept proposed for agentic AI, published in 2005.

Chen & Vertegaal — "Using Mental Load for Managing Interruptions in Physiologically Attentive User Interfaces"
CHI EA 2004 (Extended Abstracts, 4 pages, Late-Breaking Work)
✅ Verified · https://interruptions.net/literature/Chen-CHI04-2p1513.pdf
Correction from previous drafts: also in CHI Extended Abstracts, not a full paper. Presents the PAUI prototype concept using HRV + EEG to defer notifications. The r²=0.95 result is from the follow-up paper below.

Chen, Hart & Vertegaal — "Towards a Physiological Model of User Interruptability"
INTERACT 2007
✅ Verified · https://link.springer.com/chapter/10.1007/978-3-540-74800-7_39
The full empirical paper. r²=0.95 predicting interruptibility from combined HRV/EMG signals. This is the "expanded" version of the 2004 concept.

Züger & Fritz — "Interruptibility of Software Developers and its Prediction Using Psycho-Physiological Sensors"
CHI 2015
✅ Verified · https://dl.acm.org/doi/abs/10.1145/2702123.2702593
91.5% accuracy predicting interruptibility from psycho-physiological sensors in a developer context. Directly applies the 2004–2007 paradigm to the programming domain Codellaborator later addresses.

Pu et al. (Codellaborator) — "Assistance or Disruption?"
CHI 2025
✅ Verified · https://dl.acm.org/doi/10.1145/3706598.3713357
Proactive LLM agent with rule-based and LLM-predicted timing. N=18.

Chen et al. — "Need Help? Designing Proactive AI Assistants for Programming"
CHI 2025
✅ Verified · https://dl.acm.org/doi/10.1145/3706598.3714002
Second proactive LLM assistant with context-based timing. Both CHI 2025 papers establish the contemporary problem but deliberately avoid physiological sensing.

The Gap (and why it's weak)

The fundamental mechanism — use physiology to detect low cognitive load → interrupt at that moment — was published at CHI in 2004–2005 and implemented as a working system in 2007. The only novel angle is "agentic AI as the interruptor," but Codellaborator and Need Help already address proactive agentic AI interruption timing with behavioral heuristics. To be CHI 2027 submittable, this paper would need to show that physiological gating significantly outperforms behavioral heuristics from Codellaborator in a head-to-head comparison.

---

Topic 5 — Biosignal → Music for Emotion Regulation

Verdict: DEAD ☠️ — Abandon

Verified Prior Works (sample)

"Digital Music Interventions for Stress with Bio-Sensing: A Survey"
Frontiers in Computer Science, 2023
✅ Verified · https://www.frontiersin.org/journals/computer-science/articles/10.3389/fcomp.2023.1165355/full
Reviews 10 existing closed-loop biosignal → adaptive music systems. The field is already surveyed.

Ayata, Yaslan & Kamasak — "Emotion Based Music Recommendation System Using Wearable Physiological Sensors"
IEEE Transactions on Consumer Electronics, 2018 (DOI: 10.1109/TCE.2018.2844736)
✅ Verified (IEEE Xplore)
GSR + PPG → ML emotion → music recommendation. Same signals, same pipeline, same goal.

MusicalHeart — "A Hearty Way of Listening to Music"
SenSys 2012
✅ Verified · https://dl.acm.org/doi/10.1145/2426656.2426662
HRV-based music selection system.

EarTune — "Exploring the Physiology of Music Listening"
UbiComp/IMWUT 2024
✅ Verified · https://doi.org/10.1145/3675094.3680519
2024 UbiComp work on physiology + music.

The space has been saturated for over a decade and was formally surveyed in 2023.

---

Final Recommendation

Submit Topic 4 (Zombie Clicks). All prior works are confirmed. The gap is real. No paper exists that uses pupillometry to dissociate automation bias from cognitive offloading within the same behavioral event in HAI. Buçinca et al. (CSCW 2021, verified) provides the theoretical foundation. Deniz et al. (B&IT 2024, verified) provides the closest physiological precedent but does not attempt the dissociation. The construct is novel, falsifiable, and CHI-appropriate.

Do not submit Topic 3 without a head-to-head comparison against Codellaborator's behavioral heuristics. The core mechanism has 20 years of prior art.

Do not submit Topic 5 under any circumstances.