arXiv subjects

Chanjun Park

Publications and source records attributed to Chanjun Park.

At least 19 recordsLinked to original sources

arXiv2026

The Blindness of Document-Level Translation Evaluation

Document-level machine translation (MT) evaluation extends segment-level protocols by presenting full documents to annotators, on the assumption that such presentation elicits document-level judgments. We test this assumption with a counterfactual condition (MIX) in which each document combines segments drawn from different systems, preserving document-level presentation while breaking cross-segment consistency. Across 18,420 expert Englis-to-Korean annotations and 14 automatic metrics, scores, system rankings, and error annotations are statistically equivalent between coherent and incoherent documents. Perception does not explain this: shown matched passages, raters identify the coherent one as the work of a single translator in 87.3% of trials. Document presentation does change how annotators work, but that change does not reach the recorded output. What is blind is the protocol, not the annotator. The concern is not that scores fall short, but that the resources invested in document-level systems, metrics, and annotation may not be measuring what they are intended to measure.

cs.CL↗

arXiv2026

Discourse Dependency: A Continuous Criterion for Translation Difficulty

Recent calls for harder machine translation benchmarks have not clarified what difficulty should mean. We argue that one meaningful and currently unmeasured axis is referential reach, the distance a segment must look back into its document to resolve the entities and pronouns it contains. We formalize this as discourse dependency (DDP), a metric-free, source-side measure computed from named entity re-mentions and pronominal coreference. Validated against gold coreference, DDP errs one-sidedly in 99.2% of segments, so a high-DDP segment is certified to require long-range context. Applying DDP to WMT24++ and WMT25 shows that both are heavily skewed toward low-DDP segments, which domain labels do not distinguish. Building on DDP, we compare five context injection strategies in an English-Korean post-editing setup, varying context size and selection. As DDP grows, no strategy keeps pace with human post-editing. On segments with DDP >= 15 raters prefer human translations, while automatic metrics register no difference. As frontier systems saturate aggregate scores, DDP shifts evaluation from how well models score to how far they can reach.

cs.CL↗

arXiv2026

Skill Following: Evaluating Actual Skill Use in Retrieval-Enabled LLM Agents

Large Language Model (LLM) agents increasingly rely on external skills, yet standard evaluations obscure whether retrieving these skills actually helps. Aggregate metrics often compare retrieved versus non-retrieved tasks, introducing severe selection bias and failing to isolate the true effect of skill use. To measure this actual-use capability-which we formalize as Skill Following (SF)-we introduce the Retrieval-Invoked Actual-Use Effect (RAE). RAE computes the same-task outcome difference between matched skill-enabled and skill-disabled executions, conditioned exclusively on tasks where the agent actively retrieved a skill. Evaluating 17 LLMs across coding and mathematical domains, we uncover a stark evaluation paradox: models frequently show positive aggregate retrieval lift but negative RAE. On MBPP+, multiple models that appear to benefit system-wide actually harm their own performance on the exact tasks where retrieval occurred. These findings demonstrate that aggregate averages can create a misleading illusion of tool-use proficiency, whereas RAE directly measures whether the retrieval-to-answer pipeline genuinely rescues more outcomes than it harms.

cs.CL↗

arXiv2026

Auditing MCQA Benchmarks through Probability Landscapes

As Large Language Models rapidly advance, performance on standard multiple-choice question answering (MCQA) benchmarks is reaching saturation. While the community has responded by developing increasingly difficult datasets, validating question quality and filtering flawed items remains a labor-intensive process. To provide a scalable diagnostic approach, we propose a two-component probabilistic framework for auditing MCQA benchmarks using model output distributions. First, for benchmark-level analysis, we characterize the probability landscape using the top prediction probability ($P_{top1}$) and normalized residual entropy ($H_{norm}$), summarized globally by Mean Pairwise Distance (MPD). Second, for item-level diagnostics, we introduce noise injection to reduce meaningful distractor competition, enabling us to flag candidate items for targeted human review and categorize residual failure patterns. Across four MCQA benchmarks, our landscape analysis reveals benchmark-level differences in model confidence and residual option competition. Concurrently, our noise-injection method flags potentially actionable item-level issues, showing alignment with expert error annotations from MMLU-Redux. These results suggest that our probability-based framework provides a lightweight audit lens for comparing macro-level benchmark structure and prioritizing individual items for targeted human review.

cs.CL↗

arXiv2026

Beyond Consensus: Downward Bias and Role Asymmetry in Multi-Agent LLM Judges for Subjective Evaluation

Multi-Agent Debate (MAD) has been widely adopted to improve LLM-based evaluation by prompting multiple agents to negotiate and reach a consensus. However, for subjective rubric-based scoring, inter-agent agreement does not guarantee alignment with human judgments. In this paper, we compare a single-judge baseline against a consensus-based MAD protocol on subjective evaluation tasks and design three ablations to isolate the impact of role prompting, multi-round interaction, and explicit score sharing. Evaluations across six LLMs show that the single-judge baseline achieves the strongest human alignment on average across six judge models, whereas MAD shows degradation in human alignment on both tasks. Our ablations demonstrate that this performance drop stems primarily from asymmetric role prompting rather than the interaction itself. Specifically, assigning a strict judge role introduces a systematic downward bias that the consensus process fails to correct. The central finding is that this bias reflects strict-stance dominance beyond averaging: the consensus score falls well beyond the arithmetic midpoint of the standalone strict and lenient conditions, rather than averaging them out. Removing role asymmetry (Symmetric MAD) largely recovers baseline performance, while masking peer scores widens inter-agent disagreement on average and worsens average human alignment. These findings demonstrate that multi-agent consensus can enforce artificial agreement at the expense of true human alignment, revealing a structural limitation in consensus-style, role-specialized MAD protocols for subjective scoring.

cs.CL↗

arXiv2026

EvoSkill Injection: Red-Teaming Autonomous Skill Generation and Evolution in Self-Evolving Agents

LLM-based agent systems increasingly adopt skill-based architectures to reduce repetitive reasoning costs and improve stable, efficient task execution. Recent studies propose self-evolving agents that autonomously generate, refine, and reuse skills from past experiences to enable continuous capability evolution. However, autonomous skill evolution introduces a new attack surface in which malicious capabilities are generated, stored, and reused as legitimate skills. In this paper, we define EvoSkill Injection as a threat model targeting the autonomous skill generation and evolution pipeline of self-evolving agents. We further propose SARGE (Red-teaming Autonomous Skill Generation and Evolution in self-evolving agents), a red-teaming framework for evaluating this threat model through iterative generation, escalation, and reinforcement interactions. To support our framework, we construct EvoSkillBench, a benchmark dataset of malicious interaction trajectories for inducing malicious skill formation in self-evolving agents, and introduce EvoSkillSafetyBench, a post-attack benchmark for evaluating whether injected malicious skills are subsequently retrieved and activated as harmful behaviors. Our evaluation shows that SARGE induces malicious skill formation and that injected skills are persistently stored and repeatedly activated, highlighting the risk of persistent capability corruption.

cs.AI↗

arXiv2026

Vision Is Not Overhead: One-Pass Block Drafting for Lossless Speculative Decoding in Vision-Language Models

Speculative decoding accelerates generation without changing its output, yet on vision-language models (VLMs) it has been caught in a self-defeating cycle. The drafter stays autoregressive, so it must stay small. A small drafter cannot afford the image at every step, so vision is compressed, pruned, or hidden. A drafter cut off from the image is then least reliable exactly where the image makes text predictable. We present GLANCE, the first one-pass block drafter that is lossless on an unmodified VLM target, and it breaks the cycle at both ends. A block-diffusion head reads the target's already-fused vision-language state, so vision costs the drafter nothing, and fills a whole block in one forward pass, so depth costs no sequential steps. A wide candidate tree is verified in one target pass, and every audited prompt reproduces greedy decoding exactly. Grounded workloads reward this most, entering a verbatim-copy regime whose long runs cost an autoregressive drafter a pass for every token and a block drafter one in total. Under one engine and one round budget, GLANCE decodes up to 2.93x faster than autoregression, from one draft pass a round where the production EAGLE3-VL head takes eight, and accepts 2.7x longer blocks than an EAGLE-3 head trained on the same corpus. One law organizes these results. Accepted length is set by the target's next-token entropy, with a fitted slope that steepens with grounding across all five tasks. The law transfers across targets and modalities and names its own boundary, since free-running text still favors a chain. Our code is available at https://github.com/js-lee-AI/GLANCE.

cs.AI↗

arXiv2026

HealMed: Multilingual Evaluation of Large Language Models in Medicine

We present HealMed, an expert-reviewed benchmark for multilingual evaluation of large language models in medicine. HealMed contains 1,000 examples in each of nine languages, drawn from nine datasets and covering three task formats: MCQA, NLI and open-ended QA. The benchmark was developed over two years by 23 physicians and medical experts based across nine countries and regions. Each translation was evaluated and revised by two experts fluent in English and the corresponding target language. On HealMed, performance declined most in low-resource languages, although the size of the gap varied markedly across languages and models. The strongest proprietary models were the most stable across languages, whereas many open-source and medically specialized models showed larger and less consistent gaps. Medical specialization alone did not ensure multilingual robustness. Furthermore, expert revision could either raise or lower measured performance, indicating that translation quality materially affects cross-language evaluation results.

cs.CL↗

arXiv2026

To Isolate or to Score? Model-Adaptive Assessment for Cost-Efficient Multi-Agent RAG

Multi-agent document assessment for retrieval-augmented generation is computationally expensive, driving practitioners toward smaller, deployable models whose assessment mechanisms remain poorly understood. We conduct a controlled study of training-free interventions on 7B-9B instruction-tuned models across diverse QA benchmarks, revealing a sharp dichotomy in how models benefit from assessment. For weaker baselines, the dominant mechanism is per-document isolation. Astoundingly, assessment-free isolation matches full multi-agent assessment, demonstrating that resolving multi-document context confusion, rather than scoring quality, drives outsized gains of up to 50 percentage points. Conversely, for strong baselines where scoring quality matters, we introduce Reasoning-Score Coupling, a label-free perturbation probe that classifies scoring behavior. Integrating these findings, we propose MADARA, a model-adaptive routing architecture. Crucially, MADARA's diagnostic thresholds derived from a single pilot model generalize zero-shot to four unseen model families, providing a robust, lightweight pipeline to eliminate computational overhead.

cs.AI↗

arXiv2026

DART: Draft-Agreement Routing for Training-Free Adaptive Thinking Budgets in Hybrid Reasoning Models

Hybrid reasoning models can answer directly or spend extra tokens on extended thinking. A practical router should choose between these modes for each query, so easy problems avoid unnecessary reasoning and hard problems receive enough budget to finish the answer. Existing routers move in this direction, but they typically require labeled training data or fix thinking budgets up front, ignoring answer-level evidence from the model itself. We introduce DART, a training-free routing framework that samples two cheap no-think drafts, accepts direct answering when the drafts agree, and predicts a thinking budget from draft entropy when they disagree. Across the main comparisons, DART preserves or improves always-thinking accuracy in most settings while reducing thinking-token use. Accuracy improves by up to +9.0 points on Olympiad-level math and by up to +22.5 points on code under execution-based equivalence, while thinking-token use drops by 32-73%. The Stage~1 signal extends across model scales (0.6B--32B), model families, and API-only hosted settings, with no labeled data and no gradient updates required. Our code is available at https://github.com/js-lee-AI/DART.

cs.AI↗

arXiv2026

Beyond Penalizing Mistakes: Stabilizing Efficiency Training in Large Reasoning Models via Adaptive Correct-Only Rewards

Training large language models to reason efficiently is a critical challenge. While integrating length-penalizing rewards into Group Relative Policy Optimization (GRPO) aims to reduce verbosity, it frequently triggers reward collapse, severely degrading reasoning capabilities. Through a systematic evaluation of various reward configurations, we identify the root mechanism: GRPO's group normalization creates divergent advantages when incorrect answers receive continuous length penalties. Consequently, methods penalizing the length of incorrect answers are structurally prone to collapse under sustained optimization. Furthermore, restricting penalties exclusively to correct answers avoids this primary failure, but leaves the model susceptible to a stochastic collapse driven by response over-compression. To robustly prevent both failure modes, we propose ACOER (Adaptive Correct-Only Efficiency Reward). ACOER eliminates the structural penalty loop by isolating brevity bonuses to correct completions and prevents stochastic compression via dynamic budget normalization and control-loop penalty adjustments. Evaluated across diverse mathematical reasoning benchmarks, ACOER improves overall accuracy compared to the base model while reducing token generation by over 60%, establishing a fundamentally stable approach for efficiency-aware optimization.

cs.AI↗

arXiv2026

MultiDocFusion: Hierarchical and Multimodal Chunking Pipeline for Enhanced RAG on Long Industrial Documents

RAG-based QA has emerged as a powerful method for processing long industrial documents. However, conventional text chunking approaches often neglect complex and long industrial document structures, causing information loss and reduced answer quality. To address this, we introduce MultiDocFusion, a multimodal chunking pipeline that integrates: (i) detection of document regions using vision-based document parsing, (ii) text extraction from these regions via OCR, (iii) reconstruction of document structure into a hierarchical tree using large language model (LLM)-based document section hierarchical parsing (DSHP-LLM), and (iv) construction of hierarchical chunks through DFS-based grouping. Extensive experiments across industrial benchmarks demonstrate that MultiDocFusion improves retrieval precision by 8-15% and ANLS QA scores by 2-3% compared to baselines, emphasizing the critical role of explicitly leveraging document hierarchy for multimodal document-based QA. These significant performance gains underscore the necessity of structure-aware chunking in enhancing the fidelity of RAG-based QA systems.

cs.AI↗

arXiv2026

Omanic: Towards Step-wise Evaluation of Multi-hop Reasoning in Large Language Models

Evaluating the reasoning abilities of large language models (LLMs) solely from final answers can obscure failures in intermediate steps, especially in multi-hop QA benchmarks without step-level annotations. To address this gap, we introduce Omanic, an open-domain 4-hop QA benchmark designed not only to measure final-answer accuracy but also to diagnose where reasoning breaks down. Omanic contains 10,296 machine-generated training examples (OmanicSynth) and 967 expert-reviewed human-annotated evaluation examples (OmanicBench), with each evaluation question decomposed into single-hop sub-questions, intermediate answers, and structured graph topologies. Experiments with proprietary and open-source LLMs show that Omanic is challenging, while step-wise analysis reveals a later-hop bottleneck, factual knowledge floor, and error propagation along reasoning chains. Fine-tuning on OmanicSynth transfers to six reasoning and mathematics benchmarks, yielding a 7.41-point average gain and validating its effectiveness as supervision for reasoning-capability transfer. We release the data at https://huggingface.co/datasets/li-lab/Omanic and the code at https://github.com/XiaojieGu/Omanic.

cs.CL↗

arXiv2026

Sensory-Aware Sequential Recommendation via Review-Distilled Representations

We propose a novel framework for sensory-aware sequential recommendation that enriches item representations with linguistically extracted sensory attributes from product reviews. Our approach, ASER (Attribute-based Sensory-Enhanced Representation), introduces an offline extraction-and-distillation pipeline in which a large language model is first fine-tuned as a teacher to extract structured sensory attribute-value pairs, such as color: matte black and scent: vanilla, from unstructured review text. The extracted structures are then distilled into a compact student transformer that produces fixed-dimensional sensory embeddings for each item. These embeddings encode experiential semantics in a reusable form and are incorporated into standard sequential recommender architectures as additional item-level representations. We evaluate our method on five Amazon domains and integrate the learned sensory embeddings into SASRec, BERT4Rec, BSARec, and DIFF. Across 20 domain-backbone combinations, sensory-enhanced models improve over matched non-sensory counterparts in 19 cases for both HR@10 and NDCG@10, with average relative gains of 7.9% in HR@10 and 11.2% in NDCG@10. Qualitative analysis further shows that the extracted attributes align closely with human perceptions of products, enabling interpretable connections between natural language descriptions and recommendation behavior. Overall, this work demonstrates that sensory attribute distillation offers a principled and scalable way to bridge information extraction and sequential recommendation through structured semantic representation learning.

cs.CL↗

arXiv2026

LANGSAE EDITING: Improving Multilingual Information Retrieval via Post-hoc Language Identity Removal

Dense retrieval in multilingual settings often searches over mixed-language collections, yet multilingual embeddings encode language identity alongside semantics. This language signal can inflate similarity for same-language pairs and crowd out relevant evidence written in other languages. We propose LANGSAE EDITING, a post-hoc sparse autoencoder trained on pooled embeddings that enables controllable removal of language-identity signal directly in vector space. The method identifies language-associated latent units using cross-language activation statistics, suppresses these units at inference time, and reconstructs embeddings in the original dimensionality, making it compatible with existing vector databases without retraining the base encoder or re-encoding raw text. Experiments across multiple languages show consistent improvements in ranking quality and cross-language coverage, with especially strong gains for script-distinct languages.

cs.CL↗

arXiv2025

KITE: A Benchmark for Evaluating Korean Instruction-Following Abilities in Large Language Models

The instruction-following capabilities of large language models (LLMs) are pivotal for numerous applications, from conversational agents to complex reasoning systems. However, current evaluations predominantly focus on English models, neglecting the linguistic and cultural nuances of other languages. Specifically, Korean, with its distinct syntax, rich morphological features, honorific system, and dual numbering systems, lacks a dedicated benchmark for assessing open-ended instruction-following capabilities. To address this gap, we introduce the Korean Instruction-following Task Evaluation (KITE), a comprehensive benchmark designed to evaluate both general and Korean-specific instructions. Unlike existing Korean benchmarks that focus mainly on factual knowledge or multiple-choice testing, KITE directly targets diverse, open-ended instruction-following tasks. Our evaluation pipeline combines automated metrics with human assessments, revealing performance disparities across models and providing deeper insights into their strengths and weaknesses. By publicly releasing the KITE dataset and code, we aim to foster further research on culturally and linguistically inclusive LLM development and inspire similar endeavors for other underrepresented languages.

cs.CL↗

arXiv2025

Benchmark Profiling: Mechanistic Diagnosis of LLM Benchmarks

Large Language Models are commonly judged by their scores on standard benchmarks, yet such scores often overstate real capability since they mask the mix of skills a task actually demands. For example, ARC is assumed to test reasoning, while HellaSwag is designed to evaluate commonsense. However, we lack a systematic way to verify if these benchmarks actually measure these labels. We introduce Benchmark Profiling, a diagnostic framework that decomposes benchmark performance into ten cognitively grounded abilities. The method combines gradient-based importance scoring with targeted parameter ablation to compute an Ability Impact Score (AIS) that quantifies how much each ability contributes to a model's success on a given benchmark. Profiling three instruction-tuned models across ten widely used benchmarks yields four key findings: (i) most benchmarks draw on several abilities rather than one, (ii) datasets with similar labels rely on distinct ability mixtures, (iii) code-generation benchmarks reward broad, multi-skill improvement and thus show only modest gains from narrow domain-specific fine-tuning, and (iv) abilities irrelevant to the task could negatively affect performance. Benchmark Profiling therefore explains why performance gains do not always translate into user-perceived competence and offers a transparent tool for benchmark audit and model interpretability.

cs.CL↗

arXiv2025

Mixture-of-Clustered-Experts: Advancing Expert Specialization and Generalization in Instruction Tuning

A sparse Mixture-of-Experts (MoE) architecture has emerged as a highly scalable solution by conditionally activating sub-modules without a proportional increase in computational costs. However, improving expert specialization to enhance performance and generalization remains a challenge for MoE, especially in instruction tuning scenarios characterized by significant input heterogeneity. In this work, we propose the Mixture-of-Clustered-Experts (MoCE) to address this limitation through a dual-stage routing mechanism. The first stage in the mechanism performs expert group routing based on sequence-level features, while the second stage activates the top-$k$ experts within the group at the token level. This approach enables the effective partitioning of heterogeneous inputs based on their knowledge requirements, encouraging expert group specialization while maintaining the advantages of token-level routing. We evaluate MoCE across a comprehensive set of benchmarks, demonstrating its consistent superiority over strong baselines and its enhanced generalization capabilities. Detailed analysis further highlights the robustness and effectiveness of MoCE.

cs.LG↗