Academic papers
Open research notes — cite abstracts, link originals.
Papers brief: Korean journal abstracts show a late-2024 LLM register shift
arXiv study of 398,296 KCI abstracts finds morphology-aware excess vocabulary rising from late 2024 — sisahada-style hedging up, plain verbs down, with lower-bound LLM prevalence estimates through 2026.
Papers brief: PetQA tests whether AI can answer Korean vet questions like a clinic
arXiv PetQA benchmarks 18 LLMs and LVLMs on 18,827 Korean long-form dog-and-cat QA pairs from real pet-owner queries — text and multimodal — with expert veterinarian answers.
Papers brief: KBO's ABS rollout maps where human strike zones bend under count pressure
arXiv audit of Korean Baseball Organization pitch calls before and after automated ball-strike: count pressure shows the clearest human context bias; home-field suspicions mostly fail.
Papers brief: Online Bayesian node classification when graphs keep shifting
arXiv graph-learning study on inductive nodes under distribution shift: variational Bayesian last layers plus online Laplace updates claim best accuracy and calibration across five benchmarks — with a Korea fraud-and-recommendation angle.
Papers brief: Frozen transformers as in-context samplers — diffusion without retraining
arXiv theory on transformers simulating generative samplers from prompt examples alone: closed-form diffusion, U-shaped layer geometry, and what Korean LLM builders should ask before betting on prompt-only generation.
Papers brief: Korean style alignment moves abstention and disclosure the objective never named
arXiv study post-trains Qwen3.8-27B for Korean response form — lists, register, verbosity — and finds off-target shifts in KoBBQ abstention and unprompted securities disclosure driven mainly by answer propensity.
Papers brief: Korean census personas expose limits of simulated public deliberation
arXiv study with census-grounded Korean LLM agents on real policy questions: personas miss survey demographics, and debate-style interaction may not drive the opinion shifts it appears to.
Papers brief: AI training load surges as grid stability tool — TILS on Korea's network
arXiv 2608.30901 proposes training-induced load surge (TILS): after a fault clears, ramp flexible AI training workloads to pull more power locally and limit generator first-swing acceleration — tested on IEEE 39-bus and a large-scale Korean system.
Papers brief: A.X K2 — SK Telecom’s 688B MoE bets on sparse long context
arXiv A.X K2 technical report: from-scratch agentic MoE, Sparse Gated Attention to 256K, Think-Fusion modes — what Korea open-weight buyers should price before the parameter headline.
Papers brief: Korean science-ed research tracks policy — until it doesn't
arXiv 2608.26675: BERTopic on 2008–2025 Korean science education papers finds STEAM, gifted, and teacher-policy topics move with government waves; digital and disciplinary cores hold steady.
Papers brief: NOLLI maps where Korean LLM gaps actually break
arXiv NOLLI’s 7,500 procedurally generated English–Korean puzzles show presentation language is cheap; Hangul jamo multi-step work and kinship notation are not.
Papers brief: a scorecard for national AI rules — and Korea is on the board
arXiv framework scores AI policy portfolios across China, India, Japan, Singapore, South Korea, UK, US, and the EU — what on-paper design readiness means for Korea-touching watchers.
Papers brief: KoVRE — efficient Korean visual document search without giant models
arXiv KoVRE trains a 2B single-vector retriever on 708,729 Korean–English query-page pairs, beating larger English-centric VDR baselines on Korean benchmarks.
Papers brief: Korean Go TV shows AI winrates — then stops saying “AI”
arXiv decade study of ~1,900 hours of Korean Go YouTube finds AI graphs everywhere while explicit AI talk shrinks — domestication of machine judgment.
Papers brief: AI pension advice moves Korean DC portfolios — partly
arXiv 2608.11371: among 400 Korean DC workers, ~37% of aggressive-vs-conservative AI advice difference passes through; rationales add little; not investment advice.
Papers brief: EdgeXpert — fixing MoE + speculative decoding on the edge
KAIST’s arXiv 2608.05303 co-designs software and silicon so mixture-of-experts and speculative decoding stop fighting over memory.
Papers brief: learning pitch-contour tokens for Korean traditional music
arXiv 2608.10979: a VQ-VAE that learns unlabeled pitch-contour tokens recovers sigimsae categories and pansori modes — what MIR teams should watch.
Papers brief: Korean charts lag Billboard eras — then the gap halves
ISMIR 2026 accepted arXiv study trains Billboard-era CNNs on Melon charts and finds a 4–5 year lag that shrinks to 2–3 years from the 1990s.
Papers brief: Umulsai digitizes Jeongganbo without a Western staff hack
arXiv report on Umulsai — a browser-native plain-text format and engraver for Jeongganbo, Korea's grid mensural notation still used in traditional music.
Papers brief: ARI — when Joseon archives need more than the page in front of you
arXiv ARI pairs LLMs with retrieval-augmented generation to restore damaged Korean historical documents, especially proper nouns masked language models miss.
Papers brief: TIDES tracks bilingual Korea team meetings for LLMs
KAIST’s TIDES logs 75,971 EN/KO utterances across a semester — why next-speaker gains still fail the naturalness test for Korea-touching AI products.
Papers brief: leveraged ETF rebalance loop and Korea’s 2026 vol spike
arXiv preprint argues closing LETF rebalance + arbitrage added ~47 ppt annualized vol and moved KRW 4T from retail in eight weeks — and why dispersing the order may not fix it.
Papers brief: K-EXAONE 2.0 scales Korea’s open MoE to 750B
LG AI Research’s arXiv report: upcycled MoE, 10 languages, Apache 2.0 — what overseas buyers should ask beyond the parameter headline.
Papers brief: Korea's gender wage gap mostly opens after job entry
An August 2026 arXiv paper finds Korea's OECD-leading gender pay gap is small at college entry and widens across the career — what overseas HR and expats should watch.
Papers brief: Semiconductor AI meets CBAM — a Safe and Sustainable by Design gap
arXiv scoping review of 1,465 papers maps a structural hole between AI metrology on the fab floor and downstream ESG/CBAM compliance — why Korea-touching chip buyers should care.
Papers brief: Korea sits in the Syntax Ceiling of AI literacy policy
arXiv 15-nation study frames secondary programming language policy as an AI-literacy equity problem — South Korea as assessment-driven washback.
Papers brief: Pluralis puts Korea inside a multicultural AI safety test
arXiv Pluralis v0.1 spans 6,448 prompts across six Asia-Pacific countries including Korea — why Western-default AI safety scores miss locale failures.
Papers brief: Solar Open 2 — 250B MoE built for long agent trajectories
arXiv Solar Open 2 technical report: 250B-A15B MoE, 1M-token hybrid attention, strong Korean benchmark average among compared models.
Papers brief: GenAI is absorbing the junior-to-senior path in Korean SE
arXiv interview study in South Korea argues generative AI redirects entry-level work into senior-AI workflows and erodes how the next seniors form.
Papers brief: Korean traffic sentencing — LLM extracts, solver does the math
arXiv neuro-symbolic pipeline for Korea summary criminal cases keeps statutes in an SMT solver so LLMs cannot hallucinate fine arithmetic.
Papers brief: Canada vs Korea — same AI rules, different student ethics
arXiv survey finds Canadian computing students judge GenAI coding help as unethical more often than South Korean peers under near-identical policies.
Papers brief: some LLMs show consistent risk attitudes
An arXiv study finds stable belief-to-decision risk postures across navigation, triage, and finance tasks — what non-lab readers should take from it.