[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83331-en":3,"doc-seo-83331-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},83331,687197207919,"Theodora","https://ap-avatar.wpscdn.com/avatar/a000253d6f5f7c60be?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779446848396160552",8,"Research & Report","LEXIC: Lightweight Eye-tracking eXtension via Injected Complexity","On the EyeBench benchmark, predicting reading comprehension from eye movements reveals a large performance gap: text-aware models using pretrained language models reach 56–63% AUROC, while gaze-only models remain at chance. LEXIC studies how far gaze-only inference can be improved through lightweight, language-model-free conditioning. Using the EyeBench AhnCNN baseline, LEXIC injects three precomputed word-level difficulty signals via concatenation or a residual mechanism. Results on OneStop show consistent AUROC gains on Unseen Text and select improvements on Unseen Reader, with an identified boundary for residual transfer.","LEXIC: Lightweight Eye-tracking eXtension via  \nInjected Complexity  \narXiv :2607 .08 152v 1 [ cs .CL] 9 Jul 2026  \nSumin Lee Kyeonghun Kim Subeen Lee Jiwon Yang  \nSeoul National University OUTTA Seoul National University Seoul National University  \n[cirtuare@snu.ac.kr](cirtuare@snu.ac.kr) [kyeonghun.kim@outta.ai](kyeonghun.kim@outta.ai) [5584sb@snu.ac.kr](5584sb@snu.ac.kr) [jwyang29@snu.ac.kr](jwyang29@snu.ac.kr)  \nTien Nguyen Ken Ying-Kai Liao Nam-Joon Kim†  \nSeoul National University NVIDIA Seoul National University  \ntien [noob@snu.ac.kr](noob@snu.ac.kr) [kenyingkail@nvidia.com](kenyingkail@nvidia.com) [knj01@snu.ac.kr](knj01@snu.ac.kr)  \nAbstract—On the recent EyeBench benchmark, predicting reading comprehension from eye movements exposes a stark gap: text-aware models using pretrained language models reach 56– 63% AUROC, while gaze-only models operate at chance. We ask how far a gaze-only model can be pushed by lightweight, language-model-free conditioning. Building on the EyeBench AhnCNN baseline (LEXIC-Base), we propose two mechanisms to inject three precomputed word-level difficulty signals (GPT-2 surprisal, word frequency, and word length) into the per-fixation input: direct concatenation (LEXIC-Concat) and a residual mechanism (LEXIC-Res) where a small head predicts typicalreader gaze response and the encoder is conditioned on the deviation. On the OneStop reading comprehension task, with K=5 seed-ensemble training across ten folds, both mechanisms produce statistically consistent AUROC gains on Unseen Text (+1 .8 to +2 .2pp, Wilcoxon p ≤ 0.065); LEXIC-Concat additionally lifts Unseen Reader by +2 .9pp (p=0 .010). We trace an architectural boundary in LEXIC-Res on Unseen Reader (+1 .8pp, p=0 . 19) to the prediction head being calibrated to training readers, transferring imperfectly to out-of-distribution readers.  \nIndex Terms—eye tracking, reading comprehension, lightweight feature injection, CNN  \nI. INTRODUCTION  \nEye movements collected during reading reflect cognitive processes ranging from lexical access to discourse comprehension [1] . Predicting reader-level cognitive states or reader–text interactions from these movements has both scientific value for refining cognitive theories and practical value for adaptive interfaces. The EyeBench benchmark [2] standardizes these prediction tasks over six datasets, evaluating models across out-of-distribution generalization regimes on corpora such as OneStop [3] and MECO [4] .  \nOn the reader–text tasks of EyeBench—specifically binary comprehension classification—predicting cognitive processing has become an active interdisciplinary target [5], [6] . The leaderboard exposes a sharp dichotomy: models that ingest text content via large pretrained language models (PLMs) reach AUROC in the 56–63% range [7], [8], whereas pure gazeonly models operate at or near chance (49 .6–50.0%) [2], [9] . Models that augment gaze with linguistic features without a language model, like BEyeLSTM [10], sit between the two regimes (52 .5%) . Comprehension depends largely on the  \n† Corresponding author  \ncontent being read; thus, gaze-only models lack text-side signal.  \nWe ask: can a gaze-only model be pushed above chance by injecting lightweight text signals, rather than the full text? We use three precomputed word-level difficulty signals: GPT- 2 [11] surprisal, English word frequency [12], and word length. All three are computed offline once per dataset and merged onto fixation rows; no language-model forward pass is invoked at inference. The architectural changes to the gazeonly AhnCNN backbone are minimal: three additional input channels (concatenation), or a small auxiliary head predicting typical-reader gaze, whose residual against the observed gaze augments the input (residual injection) .  \nOur contributions are:  \n• We propose two lightweight mechanisms for injecting precomputed word-level difficulty features into the gaze-only AhnCNN baseline of EyeBench, requiring no language-mo","cbCainvzuTCqUNzZ","https://ap.wps.com/l/cbCainvzuTCqUNzZ","pdf",539669,1,4,"English","en",105,"# Introduction\n# Related Work\n## Predictive Modeling on Eye Movements\n## Lightweight Linguistic Features","[{\"question\":\"What problem does LEXIC address on the EyeBench benchmark?\",\"answer\":\"LEXIC addresses the performance gap between text-aware models that use pretrained language models and gaze-only models that operate near chance when predicting reading comprehension from eye movements.\"},{\"question\":\"How does LEXIC inject linguistic information without language-model inference at deployment?\",\"answer\":\"LEXIC precomputes three word-level difficulty signals—GPT-2 surprisal, word frequency, and word length offline—then merges them onto fixation rows for model input.\"},{\"question\":\"What are the two proposed LEXIC mechanisms and what differs between them?\",\"answer\":\"LEXIC-Concat directly concatenates the difficulty signals into the per-fixation input, while LEXIC-Res adds a residual mechanism using an auxiliary head to predict typical-reader gaze and conditions the encoder on the deviation.\"}]",1784186780,10,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"lexic-lightweight-eye-tracking-extension-via-injected-complexity","",{"@graph":35,"@context":84},[36,52,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":21},"https://docshare.wps.com/document/lexic-lightweight-eye-tracking-extension-via-injected-complexity/83331/",{"url":51,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":23,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":40,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-23","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What problem does LEXIC address on the EyeBench benchmark?","Question",{"text":74,"@type":75},"LEXIC addresses the performance gap between text-aware models that use pretrained language models and gaze-only models that operate near chance when predicting reading comprehension from eye movements.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does LEXIC inject linguistic information without language-model inference at deployment?",{"text":79,"@type":75},"LEXIC precomputes three word-level difficulty signals—GPT-2 surprisal, word frequency, and word length offline—then merges them onto fixation rows for model input.",{"name":81,"@type":72,"acceptedAnswer":82},"What are the two proposed LEXIC mechanisms and what differs between them?",{"text":83,"@type":75},"LEXIC-Concat directly concatenates the difficulty signals into the per-fixation input, while LEXIC-Res adds a residual mechanism using an auxiliary head to predict typical-reader gaze and conditions the encoder on the deviation.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,133],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":28,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":28,"slug":132},"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]