arXiv subjects

Guijin Son

Publications and source records attributed to Guijin Son.

At least 19 recordsLinked to original sources

arXiv2026

LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training

Reinforcement learning (RL) post-training has shown to improve reasoning in large language models (LLMs). However, there has been little exploration on the problem of data contamination in RL post-training, potentially undermining generalization and evaluation reliability of the training process itself. Existing detection methods primarily rely on output-level signals such as likelihood or entropy, which become unreliable for RL-trained models since RL shapes behavior through trajectory-level rewards rather than token likelihoods. We propose LaRA, a layer-wise representation analysis framework for detecting contamination in RL post-trained LLMs. LaRA introduces three complementary metrics, measuring perturbation sensitivity, directional collapse, and local representation rigidity under controlled perturbations. We find that contamination produces progressive geometric deviations across layers, including amplified perturbation sensitivity, stronger directional collapse, and enhanced local rigidity. Based on our findings, we also develop a contamination detection protocol that aggregates representation-level deviations across layers and metrics. Experiments on RL-trained reasoning models show that our protocol outperforms existing output-level baselines for contamination detection.

cs.LG↗

arXiv2026

Small Models Scout Bottleneck Order for Large-Model Data Control

Small proxy models are commonly used to identify data mixtures for larger-scale training. We ask whether their training trajectories reveal another transferable structure: the order in which larger models should resolve skill bottlenecks. We formulate first-passage skill training, where each monitored skill has a target floor and the objective is to minimize the tokens required to reach all floors. We introduce LogFloor, a closed-loop controller that directs each round toward current bottlenecks, producing phase-ordered resolution trajectories. Across five bAbI skill slices on Qwen2.5-1.5B, LogFloor reduces token cost by 56.2% on average. In 70M-to-12B transfer, three-round replay of a 70M scout path reaches every floor in all eight target runs, saving 30.9% by pair mean, 39.4% in pooled training tokens, and 37.6% under source-cost accounting. On MMLU-control, a frozen scout path succeeds across all eight 12B runs. Collapsing a path to its static marginal mixture or reversing its phase order removes most benefits, while bottleneck labels alone remain partially useful. These results identify phase-ordered bottleneck resolution as a transferable curriculum structure for monitored skill-targeted training.

cs.AI↗

arXiv2026

Source-Grounded Data Generation for Text-to-JSON Learning

From financial filings to clinical records, legacy industries rely heavily on long, unstructured documents to store high-value information. Reliably extracting this information into structured, machine-readable representations is a key prerequisite to making the contents accessible to automated systems. JSON is a natural target for such structured extraction, yet constructing reliable and scalable text-to-JSON training data remains challenging. To address this gap, we propose STAGE (Spreadsheet-grounded Text-to-JSON Artifact GEneration), a source-grounded data generation pipeline that constructs reports and JSON schema by using LLMs for scalable synthesis while validating ground-truth values against the underlying spreadsheet. Evaluations on STAGE-Eval, our source-grounded benchmark with an 851-example test set, show that STAGE produces stronger training data than existing approaches. This improves Qwen3-4B exact match from 31.37% to 74.27% and value accuracy from 45.46% to 90.69%.

cs.CL↗

arXiv2026

K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts

Frontier model evaluations are shifting from foundational capabilities (e.g., instruction following and reasoning) toward compositional, agentic ones, but Korean agentic benchmarks remain scarce. We introduce K-BrowseComp, a web-browsing agent benchmark grounded in Korean contexts, consisting of 400 problems. The 300-problem K-BrowseComp-Verified subset is manually constructed and validated by native Korean speakers. On this subset, frontier LLMs, including GPT-5.5, DeepSeek-V4-Pro, and GLM-5.1, reach only 30.00--45.67\%, a substantial drop from BrowseComp, while Korean LLMs released through Korea's Proprietary AI Foundation Model program obtain only 0.00--10.33\%. We further construct a 100-problem synthetic split using hard few-shot exemplars and failure-mode-targeted generation to exploit the asymmetry between solving and creating web browsing problems. On the adversarially filtered synthetic diagnostic split, the strongest model reaches only 26.00\%, and we report this split separately as a targeted stress test. We publicly release our data and code.

cs.CL↗

arXiv2026

Self-Improving CAD Generation Agents with Finite Element Analysis as Feedback

Computer-aided design (CAD) is the backbone of modern industrial design, yet learned CAD generators still fall short of real engineering pipelines: they neither iterate like engineers nor evaluate what engineering requires. Prior work has treated CAD generation as two disjoint steps, part synthesis and assembly, where the former is graded by proximity to a gold reference and the latter, when handled at all, is reduced to a separate constraint solving step. In this work, we introduce a more industry-native task formulation that requires a model to produce a fully assembled multi-part STEP file from a free-form engineering brief, which is then validated via finite element analysis (FEA). FEA validation reveals that Codex (GPT-5.5) and Claude Code (Opus-4.7) agents do not produce a single strict-passing artifact in the main first-attempt sweep, with the best configuration meeting only about 20% of typed requirements on average. Moreover, we introduce two additional supervision signals, a novel text-only blueprint schema and a 21-view image renderer that aids the agent's visual inspection, that better align the generation loop with how engineers iterate in practice. On S2O and Fusion360, the same feedback tools improve geometric reconstruction, with GPT-5.5/xhigh rising from 0.444 to 0.592 Box-IoU on S2O and from 0.397 to 0.505 on Fusion360. Together these signals move CAD programs toward artifacts that are not only visually plausible but also checked against physical and structural requirements.

cs.GR↗

arXiv2026

ResearchMath-14K: Scaling Research-Level Mathematics via Agents

The frontier of mathematics is defined by problems whose solutions are not yet known, yet it remains unclear whether language models can meaningfully engage with such problems without human intervention. A major obstacle is the lack of large-scale research-level math datasets. To this end, we introduce ResearchMath-14k, a set of $14{,}056$ problems curated from academic sources via a multi-agent pipeline, making it the largest collection of research-level mathematical problems to date. We further generate ResearchMath-Reasoning, $220$K teacher trajectories from two open models, where we observe recurring avoidance behaviors such as non-attempts and fabricated references. Interestingly, across eight open-weight models, newer generations produce $5.6\times$ more references and $5.0\times$ more fake references per trace. After agentic filtering of ResearchMath-Reasoning, fine-tuning Qwen3 models from 4B to 30B parameters improves over base models by $9.2$ points on average. This shows that filtered open-problem attempts can provide useful supervision even without fully correct reasoning traces. We make ResearchMath-14k publicly available for future works on research-level mathematical reasoning.

cs.CL↗

arXiv2026

Soohak: A Mathematician-Curated Benchmark for Evaluating Research-level Math Capabilities of LLMs

Following the recent achievement of gold-medal performance on the IMO by frontier LLMs, the community is searching for the next meaningful and challenging target for measuring LLM reasoning. Whereas olympiad-style problems measure step-by-step reasoning alone, research-level problems use such reasoning to advance the frontier of mathematical knowledge itself, emerging as a compelling alternative. Yet research-level math benchmarks remain scarce because such problems are difficult to source (e.g., Riemann Bench and FrontierMath-Tier 4 contain 25 and 50 problems, respectively). To support reliable evaluation of next-generation frontier models, we introduce Soohak, a 439-problem benchmark newly authored from scratch by 64 mathematicians. Soohak comprises two subsets. On the Challenge subset, frontier models including Gemini-3-Pro, GPT-5, and Claude-Opus-4.5 reach 30.4%, 26.4%, and 10.4% respectively, leaving substantial headroom, while leading open-weight models such as Qwen3-235B, GPT-OSS-120B, and Kimi-2.5 remain below 15%. Notably, beyond standard problem solving, Soohak introduces a refusal subset that probes a capability intrinsic to research mathematics: recognizing ill-posed problems and pausing rather than producing confident but unjustified answers. On this subset, no model exceeds 50%, identifying refusal as a new optimization target that current models do not directly address. To prevent contamination, the dataset will be publicly released in late 2026, with model evaluations available upon request in the interim.

cs.CL↗

arXiv2026

Revisiting the Uniform Information Density Hypothesis in LLM Reasoning

The Uniform Information Density (UID) hypothesis proposes that effective communication is achieved by maintaining a stable flow of information. In this work, we revisit this principle in the context of Large Language Model (LLM) reasoning, asking whether step-level uniformity reflects reasoning quality. To this end, we introduce a novel framework to quantify uniformity of information flow at both local and global levels, using an entropy-based stepwise density metric. Across experiments on seven reasoning benchmarks, we see a counter-intuitive pattern: while high-quality reasoning exhibit smooth step-by-step transitions local uniformity and structured, non-uniform information flow at the trajectory level global non-uniformity. The results demonstrate that these uniformities outperform alternative internal signals as predictors of reasoning quality, and such divergence with human communication is not a model deficiency, but a byproduct of distinct objectives between human communication and LLM reasoning.

cs.AI↗

arXiv2026

KMMMU: Evaluation of Massive Multi-discipline Multimodal Understanding in Korean Language and Context

We introduce KMMMU, a native Korean benchmark for evaluating multimodal understanding in Korean cultural and institutional settings. KMMMU contains 3,466 questions from exams natively written in Korean, covering nine disciplines and nine visual modality categories, along with a 300-item Korean-specific subset and a hard subset of 627 questions. Unlike translated or English-centric benchmarks, KMMMU targets information-dense problems shaped by local conventions, official standards, and discipline-specific visual formats. Experiments show that the strongest open-source model reaches only 42.05% accuracy on the full set, while the best proprietary model achieves 52.42% on the hard subset. Performance varies across disciplines, with some disciplines emerging as bottlenecks, and Korean-specific questions showing gaps of up to 13.43%. Error analysis suggests that these failures stem less from insufficient reasoning depth than from weak convention-to-label mapping, few-shot symbolic induction, localized knowledge recall, and domain-specific standards understanding. KMMMU provides a testbed for multimodal evaluation beyond English-centric benchmarks and for developing more reliable systems for expert real-world tasks.

cs.CL↗

arXiv2026

Pushing the Boundaries of Multiple Choice Evaluation to One Hundred Options

Multiple choice evaluation is widely used for benchmarking large language models, yet near ceiling accuracy in low option settings can be sustained by shortcut strategies that obscure true competence. Therefore, we propose a massive option evaluation protocol that scales the candidate set to one hundred options and sharply reduces the impact of chance performance. We apply this framework to a Korean orthography error detection task where models must pick the single incorrect sentence from a large candidate set. With fixed targets and repeated resampling and shuffling, we obtain stable estimates while separating content driven failures from positional artifacts. Across experiments, results indicate that strong performance in low option settings can overstate model competence. This apparent advantage often weakens under dense interference at high $N$, revealing gaps that conventional benchmarks tend to obscure. We identify two failure modes, semantic confusion and position bias toward early options under uncertainty. To isolate the effect of context length, we run padding controlled and length matched tests, which suggest that the main bottleneck is candidate ranking rather than context length. Together, these findings support massive option evaluation as a general framework for stress testing model reliability under extreme distractor density, beyond what low option benchmarks can reveal.

cs.CL↗

arXiv2026

What Users Leave Unsaid: Under-Specified Queries Limit Vision-Language Models

Current vision-language benchmarks predominantly feature well-structured questions with clear, explicit prompts. However, real user queries are often informal and underspecified. Users naturally leave much unsaid, relying on images to convey context. We introduce HAERAE-Vision, a benchmark of 653 real-world visual questions from Korean online communities (0.76% survival from 86K candidates), each paired with an explicit rewrite, yielding 1,306 query variants in total. Evaluating 39 VLMs, we find that even state-of-the-art models (GPT-5, Gemini 2.5 Pro) achieve under 50% on the original queries. Crucially, query explicitation alone yields 8 to 22 point improvements, with smaller models benefiting most. We further show that even with web search, under-specified queries underperform explicit queries without search, revealing that current retrieval cannot compensate for what users leave unsaid. Our findings demonstrate that a substantial portion of VLM difficulty stem from natural query under-specification instead of model capability, highlighting a critical gap between benchmark evaluation and real-world deployment.

cs.CV↗

arXiv2026

Redefining Evaluation Standards: A Unified Framework for Evaluating the Korean Capabilities of Language Models

Recent advancements in Korean large language models (LLMs) have driven numerous benchmarks and evaluation methods, yet inconsistent protocols cause up to 10 p.p performance gaps across institutions. Overcoming these reproducibility gaps does not mean enforcing a one-size-fits-all evaluation. Rather, effective benchmarking requires diverse experimental approaches and a framework robust enough to support them. To this end, we introduce HRET (Haerae Evaluation Toolkit), an open-source, registry-based framework that unifies Korean LLM assessment. HRET integrates major Korean benchmarks, multiple inference backends, and multi-method evaluation, with language consistency enforcement to ensure genuine Korean outputs. Its modular registry design also enables rapid incorporation of new datasets, methods, and backends, ensuring the toolkit adapts to evolving research needs. Beyond standard accuracy metrics, HRET incorporates Korean-focused output analyses-morphology-aware Type-Token Ratio (TTR) for evaluating lexical diversity and systematic keyword-omission detection for identifying missing concepts-to provide diagnostic insights into language-specific behaviors. These targeted analyses help researchers pinpoint morphological and semantic shortcomings in model outputs, guiding focused improvements in Korean LLM development.

cs.CE↗

arXiv2026

Judging What We Cannot Solve: A Consequence-Based Approach for Oracle-Free Evaluation of Research-Level Math

Recent progress in reasoning models suggests that generating plausible attempts for research-level mathematics may be within reach, but verification remains a bottleneck, consuming scarce expert time. We hypothesize that a meaningful solution should contain enough method-level information that, when applied to a neighborhood of related questions, it should yield better downstream performance than incorrect solutions. Building on this idea, we propose \textbf{Consequence-Based Utility}, an oracle-free evaluator that scores each candidate by testing its value as an in-context exemplar in solving related yet verifiable questions. Our approach is evaluated on an original set of research-level math problems, each paired with one expert-written solution and nine LLM-generated solutions. Notably, Consequence-Based Utility consistently outperforms reward models, generative reward models, and LLM judges on ranking quality. Specifically, for GPT-OSS-120B, it improves Acc@1 from 67.2 to 76.3 and AUC from 71.4 to 79.6, with similarly large AUC gains on GPT-OSS-20B (69.0 to 79.2). Furthermore, compared to LLM-Judges, it also exhibits a larger solver-evaluator gap, maintaining a stronger correct-wrong separation even on instances where the underlying solver often fails to solve.

cs.CL↗

arXiv2026

Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought

Recent frontier models employ long chain-of-thought reasoning to explore solution spaces in context and achieve stonger performance. While many works study distillation to build smaller yet capable models, most focus on English and little is known about language-specific reasoning. To bridge this gap, we first introduct **Language-Mixed CoT**, a reasoning schema that switches between English and a target language, using English as an anchor to excel in reasoning while minimizing translation artificats. As a Korean case study, we curate **Yi-Sang**: 5.79M native-Korean prompts from web Q&A, exams, STEM, and code; 3.7M long reasoning traces generated from Qwen3-32B; and a targeted 260k high-yield subset. We train ninve models (4B-35B) across six families (Qwen2.5, Llama-3.1, Gemma-3, etc). Our best model, **KO-REAson-35B**, achieves state-of-the-art performance, with the highest overall average score (64.0 \pm 25), ranking first on 5/9 benchmarks and second on the remainder. Samller and mid-sized models also benefit substantially, with an average improvement of +18.6 points across teh evaluated nine benchmarks. Ablations show **Language-Mixed CoT** is more effective than monolingual CoT, also resulting in cross-lingual and mult-modal performance gains. We release our data-curation pipeline, evaluation system, datasets, and models to advance research on language-specific reasoning. Data and model collection: https://huggingface.co/KOREAson.

cs.CL↗

arXiv2025

Revisiting the UID Hypothesis in LLM Reasoning Traces

Large language models (LLMs) often solve problems using step-by-step Chain-of-Thought (CoT) reasoning, yet these intermediate steps are frequently unfaithful or hard to interpret. Inspired by the Uniform Information Density (UID) hypothesis in psycholinguistics -- which posits that humans communicate by maintaining a stable flow of information -- we introduce entropy-based metrics to analyze the information flow within reasoning traces. Surprisingly, across three challenging mathematical benchmarks, we find that successful reasoning in LLMs is globally non-uniform: correct solutions are characterized by uneven swings in information density, in stark contrast to human communication patterns. This result challenges assumptions about machine reasoning and suggests new directions for designing interpretable and adaptive reasoning models.

cs.CL↗

arXiv2025

Ko-PIQA: A Korean Physical Commonsense Reasoning Dataset with Cultural Context

Physical commonsense reasoning datasets like PIQA are predominantly English-centric and lack cultural diversity. We introduce Ko-PIQA, a Korean physical commonsense reasoning dataset that incorporates cultural context. Starting from 3.01 million web-crawled questions, we employed a multi-stage filtering approach using three language models to identify 11,553 PIQA-style questions. Through GPT-4o refinement and human validation, we obtained 441 high-quality question-answer pairs. A key feature of Ko-PIQA is its cultural grounding: 19.7% of questions contain culturally specific elements like traditional Korean foods (kimchi), clothing (hanbok), and specialized appliances (kimchi refrigerators) that require culturally-aware reasoning beyond direct translation. We evaluate seven language models on Ko-PIQA, with the best model achieving 83.22% accuracy while the weakest reaches only 59.86%, demonstrating significant room for improvement. Models particularly struggle with culturally specific scenarios, highlighting the importance of culturally diverse datasets. Ko-PIQA serves as both a benchmark for Korean language models and a foundation for more inclusive commonsense reasoning research. The dataset and code will be publicly available.

cs.CL↗

arXiv2025

KAIO: A Collection of More Challenging Korean Questions

With the advancement of mid/post-training techniques, LLMs are pushing their boundaries at an accelerated pace. Legacy benchmarks saturate quickly (e.g., broad suites like MMLU over the years, newer ones like GPQA-D even faster), which makes frontier progress hard to track. The problem is especially acute in Korean: widely used benchmarks are fewer, often translated or narrow in scope, and updated more slowly, so saturation and contamination arrive sooner. Accordingly, at this moment, there is no Korean benchmark capable of evaluating and ranking frontier models. To bridge this gap, we introduce KAIO, a Korean, math-centric benchmark that stresses long-chain reasoning. Unlike recent Korean suites that are at or near saturation, KAIO remains far from saturated: the best-performing model, GPT-5, attains 62.8, followed by Gemini-2.5-Pro (52.3). Open models such as Qwen3-235B and DeepSeek-R1 cluster falls below 30, demonstrating substantial headroom, enabling robust tracking of frontier progress in Korean. To reduce contamination, KAIO will remain private and be served via a held-out evaluator until the best publicly known model reaches at least 80% accuracy, after which we will release the set and iterate to a harder version.

cs.CL↗

arXiv2025

Linguistic Generalizability of Test-Time Scaling in Mathematical Reasoning

Scaling pre-training compute has proven effective for achieving mulitlinguality, but does the same hold for test-time scaling? In this work, we introduce MCLM, a multilingual math benchmark featuring competition-level problems in 55 languages. We test three test-time scaling methods-Outcome Reward Modeling (ORM), Process Reward Modeling (ORM), and Budget Forcing (BF)-on both Qwen2.5-1.5B Math and MR1-1.5B, a multilingual LLM we trained for extended reasoning. Our experiments show that using Qwen2.5-1.5B Math with ORM achieves a score of 35.8 on MCLM, while BF on MR1-1.5B attains 35.2. Although "thinking LLMs" have recently garnered significant attention, we find that their performance is comparable to traditional scaling methods like best-of-N once constrained to similar levels of inference FLOPs. Moreover, while BF yields a 20-point improvement on English AIME, it provides only a 1.94-point average gain across other languages-a pattern consistent across the other test-time scaling methods we studied-higlighting that test-time scaling may not generalize as effectively to multilingual tasks. To foster further research, we release MCLM, MR1-1.5B, and evaluation results.

cs.CL↗