{"id":"8e0fcc9c-ab6c-4ab2-98bc-1e00da0d6e7f","arxiv_id":"2501.16240","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A proactive smart-glasses AI that infers learning desires from gaze and context can deliver personalized, low-disruption knowledge during everyday activities.","lead":"AiGet is a prototype AI assistant on AR glasses that watches where you look and proactively tells you surprising facts about things around you, from plants to products. It aims to turn casual walks and shopping into low-effort learning moments, and a small user study suggests people enjoy and remember the knowledge it delivers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gaze-to-desire mapping is unvalidated: the LLM's classification of Saccade/Quick Browse/Focused has no accuracy check, and the only supporting ablation reuses the same 12 participants who defined the mapping.","rationale":"The paper is a solid systems/HCI feasibility study. The in-lab ablation is a well-designed within-subject comparison, and the real-world deployment with 40 sessions is a reasonable initial probe. I credit the open-sourced code and detailed prompts. My concern is not that the gaze-mapping idea is wrong, but that the paper provides no direct evidence for its accuracy. The reader identified this same assumption, and I agree it is the most load-bearing, because the proactive gaze-and-context-driven mechanism is the paper's core novelty and determines both timing and content. The circularity of using the same participants as formative informants and raters is a real threat: they may rate knowledge as desirable simply because it matches what they previously said they wanted. A fresh-rater ablation would help. The most decisive check, however, is a direct accuracy test of the gaze-mode classifier against ground-truth cognitive states, because without it the system could be delivering the right kind of knowledge only by luck. I therefore recommend keeping the reader's CONDITIONAL verdict: the claim is promising but conditional on this mapping being reliable, which is currently unshown.","tokens_in":49063,"tokens_out":4368,"duration_ms":42812,"concrete_test":"Run an independent validation study with N≥20 participants wearing the same Pupil Core + XReal setup during scripted daily activities (walking, shopping, museum). After each 30-60s episode, participants self-report their cognitive state (e.g., focused on one object, scanning shelves, thinking about unrelated matters, open to info) via a cued-recall prompt. Run the AiGet context-analysis LLM on the same FPV+gaze frames to output a gaze-mode label. Compute Cohen's kappa / confusion matrix between LLM labels and self-reports, and have two independent coders label the videos as a check. If kappa < 0.4 or per-mode F1 < 0.6, the mapping is too noisy to support the claim that AiGet infers learning desire reliably. Additionally, re-run the in-lab ablation with a fresh rater pool (n≥12) blind to the design goals to rule out demand characteristics from formative-study participants.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The system's proactive behavior hinges on classifying gaze into three modes (Saccade, Quick Browse, Focused) and mapping each mode to a learning-desire type and content rule (Sec 4.5.2, Sec 5.1.2, App C.2). This mapping is never independently validated. The in-lab ablation (Sec 5.3) shows that a pipeline including these rules outperforms baselines, but the raters are the very same 12 participants from the formative study (Sec 5.3.2) who articulated the mapping, creating circularity: they rate knowledge as desirable when it matches their own stated preferences. Moreover, the gaze-mode classification is performed by Gemini-1.5-Flash on FPV frames with gaze overlays; no accuracy, confusion matrix, or inter-rater agreement for this classification is reported. If the LLM misclassifies, the content-type rules (e.g., comparative info for Focused, fun facts for Saccade) fire in the wrong situations, undermining the low-disruption and seamless-embedding claims even if the delivered knowledge is accurate. The real-world study (Sec 6) does not log ground-truth cognitive states, so in situ classification accuracy remains unknown. Thus the central effectiveness claim depends on an unverified perceptual-cognitive model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AiGet, a proactive AI assistant on AR smart glasses that aims to embed informal learning into low-demand daily activities (e.g., casual walking, shopping, museum visits). The system analyzes real-time gaze patterns, environmental context, and user profiles via multimodal LLMs, then delivers personalized knowledge through audio, text keywords, and images with bounding boxes, with minimal disruption to primary tasks. The authors present a formative study with 12 participants, an in-lab ablation study with the same 12 participants comparing AiGet against two baseline pipelines, and a real-world feasibility study with 18 participants (6 of whom completed additional sessions over up to 7 days). They report that AiGet generated desirable, novel, personalized, and surprising knowledge, was not annoying, and enhanced users' connection with the environment, and they derive design guidelines for future wearable informal-learning assistants.","tokens_in":49294,"tokens_out":5243,"duration_ms":47446,"significance":"If the findings hold, AiGet addresses a relatively underexplored niche: proactive, gaze-context-aware incidental learning in everyday settings, as opposed to reactive query-based systems such as GazePointAR or G-VOILA. The paper contributes a complete system description with detailed prompts, an open-source implementation, and design guidelines that could inform future wearable AI assistants. The real-world deployment, despite its limitations, provides useful feasibility data and user-experience insights for this class of systems.","major_comments":[{"comment":"The in-lab ablation study reuses the same 12 participants from the formative study (Sec 4.1) to rate knowledge generated from their own recorded activities. Since these participants' stated preferences directly shaped the design goals and the gaze-pattern-to-content rules (Sec 4.5.2, Sec 5.1.3), the evaluation is not independent of the system's design. The observed advantage of AiGet over the two baselines (Sec 5.3.5, Figure 5) could be inflated by participants rating knowledge that matches their own previously articulated desires. To support the central effectiveness claim, the authors should either re-run the ablation with independent raters or provide a post-hoc analysis demonstrating that the results hold for participants who did not contribute to specific rules.","section":"Sec 5.3.2, Sec 4.5.2-4.5.3"},{"comment":"The system's trigger timing and content selection depend on an LLM classification of gaze patterns into Saccade, Quick Browse, and Focused modes. The paper reports no validation of this classification: no accuracy, no confusion matrix, no inter-rater agreement between the LLM and human coders, and no ground-truth recording of user cognitive states in the real-world study. If the classification is noisy or biased, the rules for what knowledge to deliver and when will fire inappropriately, which would undermine the low-disruption and seamless-embedding claims even though the delivered content is individually accurate. The authors should provide at least a small-scale evaluation of the gaze-pattern classifier (e.g., comparing LLM labels against human annotations on a sample of the recorded FPV frames) or explicitly acknowledge and discuss the impact of this uncertainty on the design guidelines.","section":"Sec 5.1.2, Sec 6.6, App C.1"},{"comment":"The abstract and discussion claim that AiGet 'demonstrate[s] effectiveness' in enhancing enjoyment, curiosity, and connection with the environment. The supporting real-world evidence is an uncontrolled, self-reported study with 18 participants (6 of whom had repeated sessions), all in low-demand activities, and the authors themselves state in Sec 6 that 'fully addressing these questions with a high degree of quantitative evidence requires a larger user base and longer periods of daily usage.' The absence of a comparison condition, objective knowledge-retention measures, or behavior-change indicators means the evidence supports feasibility and positive user experience but not causal effectiveness. The claims should be tempered accordingly, or the study design should be extended to include a control group and pre/post measures.","section":"Abstract, Sec 6, Sec 7"}],"minor_comments":[{"comment":"Multiple Wilcoxon signed-rank tests are conducted across many outcome measures without multiple-comparison correction; the authors should report adjusted p-values or explicitly treat these analyses as exploratory.","section":"Sec 5.3.5"},{"comment":"The cancellation rate is computed as the ratio of means (2.4/54.2), which is not necessarily the mean of per-session ratios; reporting the distribution of cancellation rates across sessions would be more informative.","section":"Sec 6.6.1"},{"comment":"The operational definitions of the three gaze modes are qualitative; providing quantitative thresholds (e.g., fixation duration, scanpath velocity) would improve reproducibility.","section":"App C.1"},{"comment":"Figure 9 shows an upward trend in subjective ratings across days, but no statistical test is reported; please report the test statistics or describe the trend as descriptive.","section":"Sec 6.6.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be the camera-ready version of a CHI 2025 paper. If the authors are seeking journal extension, the two main risks are the circularity in the in-lab evaluation and the unvalidated gaze-pattern classifier. I recommend that they address these directly, perhaps by adding an independent-rater evaluation and a classifier validation study. The paper's contribution as a system and design-guideline paper is solid; the current version overstates the evidence in the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate systems paper with a new interaction pattern—query-free, gaze-triggered knowledge delivery on smart glasses—and an honest feasibility evaluation. Read it as a feasibility study, not as a demonstration of effectiveness, and you'll get real value from the design insights and the open-sourced implementation.\n\nWhat's new: prior gaze-based assistants like GazePointAR and G-VOILA require explicit verbal queries or focus only on foveated objects. AiGet infers Saccade/Quick Browse/Focused gaze modes, expands to peripheral entities, and uses a novelty × (alignment + utility + unexpectedness) score to select and withhold content. That is a concrete, useful design contribution. The in-lab ablation comparing the full pipeline against two baselines is well-structured, and the qualitative examples are convincing. The real-world study, including repeated sessions across up to 7 days, is a reasonable first deployment, and the paper's own limitations section is candid about scope. The GitHub link and detailed prompts are a plus.\n\nSoft spots, in proportion: the biggest is the unvalidated gaze-to-desire mapping. The system treats three LLM-classified gaze modes as reliable signals of learning desire, but no accuracy, confusion matrix, or inter-rater agreement is reported for that classification. The in-lab ablation uses the same 12 participants from the formative study who helped articulate the mapping, so their ratings can partly reflect self-confirmation. That is a real circularity, though not fatal for a proof-of-concept. The real-world evidence is small, self-reported, and uncontrolled; the 'revived curiosity' and 'reduced phone use' claims rest on interviews, not logs. The paper also only covers low-demand activities, and several thresholds (12-second interval, 0.6 similarity, 4.91 degrees, retention score 2) are tuned through pilot testing without sensitivity analysis. Minor: no commit hash or dataset, which weakens reproducibility.\n\nOverall, the central idea holds up as design research. The evidence supports feasibility and user acceptance; it does not support the stronger wording in the abstract about demonstrating effectiveness. That mismatch is the main thing a referee or reader should push on.\n\nWho it's for: HCI, learning technology, and wearable assistance researchers. I'd bring it to a reading group focused on proactive AI or smart glasses. I'd cite it as an example of gaze-conditioned proactive assistance. It deserves serious peer review; I would send it out rather than desk reject.","headline":"A credible feasibility study of proactive, gaze-driven incidental learning on smart glasses; read the effectiveness claims as promising, not proven.","tokens_in":49849,"tokens_out":2228,"would_cite":true,"duration_ms":23260,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Wearable AI that reads gaze can turn routine walks into learning opportunities.","keywords":["smart glasses","informal learning","incidental learning","large language models","gaze pattern analysis","proactive AI assistance","wearable AI","knowledge discovery"],"falsifier":"Run a field study where participants self-report their cognitive state at random moments while wearing the glasses and compare those reports with the system's Saccade/Quick Browse/Focused classifications; near-chance agreement, or a sharp rise in cancellation when users report being deep in thought, would show the trigger mapping is too weak to support the effectiveness claim.","tokens_in":48852,"feed_emoji":"🕶️","tokens_out":8499,"duration_ms":78758,"temperature":0.7,"pith_summary":"Adults often walk past plants, buildings, and products they would genuinely like to know about, but time pressure, inattentional blindness, an illusion of explanatory depth, and discomfort with the unfamiliar keep that knowledge out of reach. AiGet is a smart-glasses assistant built to close that gap: it reads the wearer's gaze pattern, senses the environment, consults a personal profile, and asks a large language model to compose a short, personalized knowledge snippet. The paper reports in-lab ablation results and 40 real-world sessions, including repeated use over several days, in which proactive suggestions were mostly accepted, novel facts were recalled, and users felt more connected to their surroundings. The central claim is that proactive, gaze-aware knowledge delivery can embed informal learning into low-demand daily activities without significantly disrupting the primary task.","feed_headline":"AI glasses uncover knowledge hiding in everyday places","feed_subtitle":"Prototype on AR glasses delivered 54 knowledge items per session and users canceled only about 4 percent.","key_machinery":"The load-bearing component is the gaze-pattern classifier with three modes: Saccade (random gaze movement, treated as relaxed openness to fun facts), Quick Browse (rapid scanning of familiar related items, treated as medium interest), and Focused (sustained fixation, treated as high interest or decision-making). These modes decide both when the system asks the model for knowledge and what kind of knowledge it should produce, for example comparisons for Focused users and serendipitous facts for Saccade users. A second central mechanism is the knowledge-selection gate, which scores each candidate with Novelty times the sum of alignment, utility, and unexpectedness, keeps items scoring at least 2, filters out content similar to recent history, and caps each delivery at two items.","core_discovery":"The core discovery is that a wearable eye tracker plus an LLM reasoning pipeline can make incidental learning happen during casual walking, shopping, and museum visits. AiGet classifies gaze into three modes—Saccade, Quick Browse, and Focused—and treats those modes as proxies for what the user is open to learning. It overlays gaze on the first-person view, adds location, time, and a user profile, then prompts the model to identify primary and peripheral entities, infer the user's intention, and generate knowledge scored by $\\text{Novelty} \\times (\\text{AlignWithUserPreference} + \\text{Utility} + \\text{Unexpectedness})$. Only items scoring at least 2 are retained, at most two are shown at once, and output is split across audio, keyword-and-emoji text, and a bounding-box image so the user can keep attending to the environment. In the real-world sessions, AI-initiated suggestions averaged 54.2 per session against 2.3 user-initiated queries, users canceled roughly 4% of the proactive items, and subjective ratings for usefulness, interest, surprise, and relevance all averaged above 5.9 out of 7.","pith_inferences":["The paper leaves implicit that the same gaze-triggered, context-scored pipeline could serve other proactive wearables, from accessibility aids that narrate overlooked hazards to retail assistants that explain unfamiliar products rather than only recommending them.","A direct test not reported in the paper would compare the classifier's Saccade/Quick Browse/Focused output with experience-sampling self-reports of mental state in real time, pinpointing where the trigger policy needs per-user calibration.","The paper's own numbers also leave room for a caution: with one prototype, one-week usage, and experimenter-present sessions, the 4% cancellation rate could shift under fully unsupervised, long-term deployment.","An untested extension follows from the environment-connection result: placing serendipitous knowledge in peripheral vision while keeping requested content central might be a general design principle for attention-maintaining displays."],"forward_implications":["If the claim holds, query-free delivery creates learning opportunities that user-initiated visual question answering cannot, because users often do not know what to ask about in familiar places.","The 4% cancellation rate suggests proactive knowledge is acceptable during low-demand activities, but the same design may need interruptibility prediction before it can be used in time-sensitive or cognitively demanding tasks.","Repeated visits to the same place need not exhaust novelty when the system filters semantically similar content; first- and second-visit novelty ratings stayed close (5.7 vs 5.8 out of 7).","Gaze-conditional rules give later systems a concrete recipe: deliver fun facts during Saccade, interest-raising knowledge during Quick Browse, and decision-support comparisons during Focused engagement.","Multi-day usage reports indicate the effect can outlast the session, with users noticing environmental details such as a nesting bird even when not wearing the glasses."],"supporting_citations":[{"why":"Supplies the gaze-fixation trigger, ring-mouse control, and wearable-assistant interaction design that AiGet adapts for proactive learning.","marker":"[25]"},{"why":"Documents gaze and motion patterns during mobile information seeking, grounding the Saccade/Quick Browse/Focused classification.","marker":"[50]"},{"why":"Provides the gaze-assisted voice-query baseline that AiGet extends from explicit user queries to proactive delivery.","marker":"[61]"},{"why":"Gives the gaze-facilitated querying framework and the in-lab pipeline-ablation method used to compare AiGet against baselines.","marker":"[100]"},{"why":"Motivates using user profiles and conversation history in LLM-based eyewear systems, which AiGet adapts for personalized knowledge selection.","marker":"[104]"},{"why":"Supplies the retrieval-augmented-generation idea behind the semantic-history filter that prevents repeated knowledge on revisits.","marker":"[109]"}],"fun_headline_variants":["Gaze-tracking glasses turn walks into curiosity boosts","AI on glasses finds surprise facts during casual strolls","Proactive AI glasses surface hidden knowledge as you wander","Smart glasses spark curiosity with gaze-aware knowledge drops","Wearable AI converts everyday errands into learning moments"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a person's gaze mode reliably signals their current mental state and openness to learning, so the system can choose the right moment and the right fact; if that mapping is noisy, proactive prompts will arrive at the wrong times and the reported acceptance and enjoyment would not generalize.","fun_headline_variants_meta":{"raw":{"variants":["Gaze-tracking glasses turn walks into curiosity boosts","AI on glasses finds surprise facts during casual strolls","Proactive AI glasses surface hidden knowledge as you wander","Smart glasses spark curiosity with gaze-aware knowledge drops","Wearable AI converts everyday errands into learning moments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1323,"prompt_tokens":944,"completion_tokens":379,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":303}},"tokens_in":560,"tokens_out":379,"duration_ms":3995,"temperature":1.0,"reasoning_tokens":303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:36:00.783374+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a field study where participants self-report their cognitive state at random moments while wearing the glasses and compare those reports with the system's Saccade/Quick Browse/Focused classifications; near-chance agreement, or a sharp rise in cancellation when users report being deep in thought, would show the trigger mapping is too weak to support the effectiveness claim.","supporting_citations":[{"cited_title":"Visual Behaviors and Mobile Information Acquisition","cited_arxiv_id":"2202.02748","evidence_quote":"Documents gaze and motion patterns during mobile information seeking, grounding the Saccade/Quick Browse/Focused classification."},{"cited_title":"Dick, Fan Yang, Tun Lu, Ning Gu, and Li Shang","cited_arxiv_id":null,"evidence_quote":"Motivates using user profiles and conversation history in LLM-based eyewear systems, which AiGet adapts for personalized knowledge selection."}],"review_version":1}