Search arXiv⌕ Search

arXiv subjects

Hyunjun Kim

Publications and source records attributed to Hyunjun Kim.

At least 19 recordsLinked to original sources
arXiv2026

EquiGQNet: Fast Grasp Quality Evaluation via Shared Equivariant Point Cloud Encoding

Planning six-degree-of-freedom (6-DoF) grasps for unseen objects in cluttered tabletop scenes from a single-view depth image requires accurate and efficient evaluation of diverse grasp candidates. Existing early-fusion methods capture local object geometry relative to each grasp candidate but repeatedly encode the scene, whereas late-fusion methods reuse a shared scene representation but may lose this grasp-relative local geometry. We propose EquiGQNet, an efficient 6-DoF grasp quality evaluator that combines the strengths of both approaches. For grasp orientation, EquiGQNet replaces the early-fusion operation of rotating and re-encoding the point cloud for each grasp candidate with an SO(3)-equivariant encode-once-then-rotate scheme, yielding grasp-aligned geometric features from a shared scene encoding. For grasp translation, Mid-level Action Fusion (MAF) injects the grasp position into intermediate features before global aggregation, retaining local geometry relative to each candidate. We evaluate EquiGQNet in two grasp planning pipelines: Cross-Entropy Method (CEM)-based continuous grasp search and candidate ranking with a pretrained generative planner. In simulation, EquiGQNet achieves grasping performance comparable to the early-fusion baseline and substantially outperforms late fusion on objects with complex geometry and limited graspable regions, while reducing CEM planning time from 3.31s to 0.48s, a 6.9x speedup over early fusion. In real-world household-object decluttering, EquiGQNet achieves a 95.2% grasp success rate and 230 picks per hour, versus 153 and 170 for early- and late-fusion baselines. Code is available at https://equigqnet.github.io/.

cs.RO↗
arXiv2026

URF: A Unified Robot Control-Policy Framework for Stable Contact Aware Manipulation

Learning-based manipulation policies usually predict robot actions from sensory observations and leave their execution to a separate low-level controller. In rigid contact, this separation can be problematic: the same motion to a virtual target or compliant motion command can lead to unstable contact, tracking error, excessive loading, or tool damage, depending on the low-level controller. In this paper, we propose a \textit{Unified Robot Control-Policy Framework} (URF), which connects compliant action prediction with unified impedance-admittance control. Given multimodal observations, URF predicts a virtual target, a stiffness matrix, and an impedance-admittance switch ratio. The switch ratio determines when the controller should behave more like admittance control for accurate motion tracking and when it should move toward impedance control for safer rigid contact. Because demonstration data do not provide ground-truth environment stiffness, we construct switch-ratio labels from measured contact forces and use them to supervise controller-mode prediction. Across box-flipping and line-pressing tasks, URF achieves higher task success rates while reducing failure modes observed with admittance-only execution, including rapid force buildup, large force oscillations, tool breakage, and robot safety stops. These results suggest that contact-aware policies benefit from predicting not only compliant actions but also the controller behavior used to execute them. Project page: https://jiyou384.github.io/urf_project_page/

cs.RO↗
arXiv2026

Technical Report: AI-Assisted Gated DeltaNet Optimization on NVIDIA Blackwell

AI-assisted GPU programming is often framed as a kernel-generation loop: ask a model to produce faster CUDA code, benchmark the result, and repeat. This case study argues that contest-grade optimization involves more than improving the kernel body. We examine the Agent-Assisted submission by our team, MSInfer, to the MLSys 2026 FlashInfer Contest. The submission optimized Gated DeltaNet decode and prefill on NVIDIA B200/Blackwell and achieved an official $1.58\times$ speedup, with approximate average latencies of $9.315\,\mu\mathrm{s}$ for decode and $239.48\,\mu\mathrm{s}$ for prefill. Our experience shows that even effective local kernel improvements can plateau when a workload requires structural reformulation and evaluator-aligned measurement. We therefore characterize AI-assisted kernel optimization as an end-to-end systems problem that encompasses algorithm design, workload specialization, measurement tooling, build and evaluation surfaces, evaluator alignment, and human interpretation.

cs.DC↗
arXiv2026

CyBiasBench: Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios

Large language models (LLMs) are increasingly deployed as autonomous agents in offensive cybersecurity. In this paper, we reveal an interesting phenomenon: different agents exhibit distinct attack patterns. Specifically, each agent exhibits an attack-selection bias, disproportionately concentrating its efforts on a narrow subset of attack families regardless of prompt variations. To systematically quantify this behavior, we introduce CyBiasBench, a comprehensive 630-session benchmark that evaluates five agents on three targets and four prompt conditions with ten attack families. We identify explicit bias across agents, with different dominant attack families and varying entropy levels in their attack-family allocation distributions. Such bias is better characterized as a trait of the agents, rather than a factor associated with the attack success rate. Furthermore, our experiments reveal a bias momentum effect, where agents resist explicit steering toward attack families that conflict with their bias. This forced distribution shift does not yield measurable improvements in attack performance. To ensure reproducibility and facilitate future research, we release an interactive result dashboard at https://trustworthyai.co.kr/CyBiasBench/ and a reproducibility artifact with aggregated session-level statistics and full evaluation scripts at https://github.com/Harry24k/CyBiasBench.

cs.CR↗
arXiv2026

Robust Grounding with MLLMs Against Occlusion and Small Objects via Language-Guided Semantic Cues

While Multimodal Large Language Models (MLLMs) have enhanced grounding capabilities in general scenes, their robustness in crowded scenes remains underexplored. Crowded scenes entail visual challenges (i.e., occlusion and small objects), which impair object semantics and degrade grounding performance. In contrast, language expressions are immune to such degradation and preserve object semantics. In light of these observations, we propose a novel method that overcomes such constraints by leveraging Language-Guided Semantic Cues (LGSCs). Specifically, our approach introduces a Semantic Cue Extractor (SCE) to derive semantic cues of objects from the visual pipeline of an MLLM. We then guide these cues using corresponding text embeddings to produce LGSCs as linguistic semantic priors. Subsequently, they are reintegrated into the original visual pipeline to refine object semantics. Extensive experiments and analyses demonstrate that incorporating LGSCs into an MLLM effectively improves grounding accuracy in crowded scenes.

cs.CV↗
arXiv2026

LION-DG: Layer-Informed Initialization with Deep Gradient Protocols for Accelerated Neural Network Training

Weight initialization remains decisive for neural network optimization, yet existing methods are largely layer-agnostic. We study initialization for deeply-supervised architectures with auxiliary classifiers, where untrained auxiliary heads can destabilize early training through gradient interference. We propose LION-DG, a layer-informed initialization that zero-initializes auxiliary classifier heads while applying standard He-initialization to the backbone. We prove that this implements Gradient Awakening: auxiliary gradients are exactly zero at initialization, then phase in naturally as weights grow -- providing an implicit warmup without hyperparameters. Experiments on CIFAR-10 and CIFAR-100 with DenseNet-DS and ResNet-DS architectures demonstrate: (1) DenseNet-DS: +8.3% faster convergence on CIFAR-10 with comparable accuracy, (2) Hybrid approach: Combining LSUV with LION-DG achieves best accuracy (81.92% on CIFAR-10), (3) ResNet-DS: Positive speedup on CIFAR-100 (+11.3%) with side-tap auxiliary design. We identify architecture-specific trade-offs and provide clear guidelines for practitioners. LION-DG is simple, requires zero hyperparameters, and adds no computational overhead.

cs.LG↗
arXiv2026

Defensive M2S: Training Guardrail Models on Compressed Multi-turn Conversations

Guardrail models are essential for ensuring the safety of Large Language Model (LLM) deployments, but processing full multi-turn conversation histories incurs significant computational cost. We propose Defensive M2S, a training paradigm that fine-tunes guardrail models on Multi-turn to Single-turn (M2S) compressed conversations rather than complete dialogue histories. We provide a formal complexity analysis showing that M2S reduces training cost from $O(n^2)$ to $O(n)$ for $n$-turn conversations. Empirically, on our training dataset (779 samples, avg. 10.6 turns), M2S requires only 169K tokens compared to 15.7M tokens for the multi-turn baseline -- a 93$\times$ reduction. We evaluate Defensive M2S across three guardrail model families (LlamaGuard, Nemotron, Qwen3Guard) and three compression templates (hyphenize, numberize, pythonize) on SafeDialBench, a comprehensive multi-turn jailbreak benchmark. Our best configuration, Qwen3Guard with hyphenize compression, achieves 93.8% attack detection recall while reducing inference tokens by 94.6% (from 3,231 to 173 tokens per conversation). This represents a 38.9 percentage point improvement over the baseline while dramatically reducing both training and inference costs. Our findings demonstrate that M2S compression can serve as an effective efficiency technique for guardrail deployment, enabling scalable safety screening of long multi-turn conversations.

cs.CL↗
arXiv2026

Geometric Regularization in Mixture-of-Experts: The Disconnect Between Weights and Activations

Mixture-of-Experts (MoE) models achieve efficiency through sparse activation, but the role of geometric regularization in expert specialization remains unclear. We apply orthogonality loss to enforce expert diversity and find it fails on multiple fronts: it does not reduce weight-space overlap (MSO actually increases by up to 114%), activation-space overlap remains high (~0.6) regardless of regularization, and effects on performance are inconsistent -- marginal improvement on WikiText-103 (-0.9%), slight degradation on TinyStories (+0.9%), and highly variable results on PTB (std > 1.0). Our analysis across 7 regularization strengths reveals no significant correlation (r = -0.293, p = 0.523) between weight and activation orthogonality. These findings demonstrate that weight-space regularization neither achieves its geometric goal nor reliably improves performance, making it unsuitable for MoE diversity.

cs.LG↗
arXiv2026

Entropic Context Shaping: Information-Theoretic Filtering for Context-Aware LLM Agents

Context engineering for large language model (LLM) agents requires distinguishing pragmatically useful information from misleading distractors. We introduce Entropic Context Shaping (ECS), an information-theoretic framework that measures context utility via the shift in the model's answer distribution toward the correct answer. Unlike lexical similarity methods that rely on word overlap, ECS captures pragmatic utility -- whether a passage actually helps answer the question. We formalize utility as the signed change in answer probability and provide theoretical analysis showing that task-irrelevant updates yield near-zero distribution shift. We evaluate on multi-turn context selection tasks using LongMemEval (session-level) and LoCoMo (turn-level) benchmarks. On fine-grained turn selection, ECS with Llama-3.1-8B achieves F1=0.265, a 71.83% relative improvement over TF-IDF (F1=0.154), demonstrating that pragmatic utility outperforms lexical similarity when precise context selection matters. Code and data are available in the supplementary materials.

cs.CL↗
arXiv2025

HOLOGRAPH: Active Causal Discovery via Sheaf-Theoretic Alignment of Large Language Model Priors

Causal discovery from observational data remains fundamentally limited by identifiability constraints. Recent work has explored leveraging Large Language Models (LLMs) as sources of prior causal knowledge, but existing approaches rely on heuristic integration that lacks theoretical grounding. We introduce HOLOGRAPH, a framework that formalizes LLM-guided causal discovery through sheaf theory--representing local causal beliefs as sections of a presheaf over variable subsets. Our key insight is that coherent global causal structure corresponds to the existence of a global section, while topological obstructions manifest as non-vanishing sheaf cohomology. We propose the Algebraic Latent Projection to handle hidden confounders and Natural Gradient Descent on the belief manifold for principled optimization. Experiments on synthetic and real-world benchmarks demonstrate that HOLOGRAPH provides rigorous mathematical foundations while achieving competitive performance on causal discovery tasks with 50-100 variables. Our sheaf-theoretic analysis reveals that while Identity, Transitivity, and Gluing axioms are satisfied to numerical precision (<10^{-6}), the Locality axiom fails for larger graphs, suggesting fundamental non-local coupling in latent variable projections. Code is available at [https://github.com/hyunjun1121/holograph](https://github.com/hyunjun1121/holograph).

cs.LG↗
arXiv2025

DrawingBench: Evaluating Spatial Reasoning and UI Interaction Capabilities of Large Language Models through Mouse-Based Drawing Tasks

As agentic AI systems increasingly operate autonomously, establishing trust through verifiable evaluation becomes critical. Yet existing benchmarks lack the transparency and auditability needed to assess whether agents behave reliably. We present DrawingBench, a verification framework for evaluating the trustworthiness of agentic LLMs through spatial reasoning tasks that require generating sequences of low-level GUI actions. Unlike opaque evaluations, DrawingBench provides transparent, rule-based assessment: 8 objective criteria enable reproducible scoring, while action-level inspection allows stakeholders to audit agent behavior. Our framework comprises 250 diverse prompts across 20 categories and 4 difficulty levels, deterministic evaluation metrics, and an external oversight mechanism through multi-turn feedback that enables human control over agent refinement. Evaluating four state-of-the-art LLMs (Claude-4 Sonnet, GPT-4.1, GPT-4.1-mini, Gemini-2.5 Flash) across 1,000 tests, we establish both capabilities and limitations: models achieved 92.8% perfect performance with structured external feedback driving significant improvements (average +3.2%, up to +32.8% for complex scenes), but systematic error patterns emerged in tool state management and long-horizon planning. Notably, specification clarity proved more important than task complexity -- models achieved 100% perfect performance when given explicit, verifiable criteria. These findings demonstrate that transparent evaluation frameworks can establish trust in agentic systems, with external oversight proving more reliable than self-correction for guiding agent behavior. Our open-source framework provides a template for trustworthy agent assessment. Code and data: https://github.com/hyunjun1121/DrawingBench

cs.CL↗
arXiv2025

Continual Error Correction on Low-Resource Devices

The proliferation of AI models in everyday devices has highlighted a critical challenge: prediction errors that degrade user experience. While existing solutions focus on error detection, they rarely provide efficient correction mechanisms, especially for resource-constrained devices. We present a novel system enabling users to correct AI misclassifications through few-shot learning, requiring minimal computational resources and storage. Our approach combines server-side foundation model training with on-device prototype-based classification, enabling efficient error correction through prototype updates rather than model retraining. The system consists of two key components: (1) a server-side pipeline that leverages knowledge distillation to transfer robust feature representations from foundation models to device-compatible architectures, and (2) a device-side mechanism that enables ultra-efficient error correction through prototype adaptation. We demonstrate our system's effectiveness on both image classification and object detection tasks, achieving over 50% error correction in one-shot scenarios on Food-101 and Flowers-102 datasets while maintaining minimal forgetting (less than 0.02%) and negligible computational overhead. Our implementation, validated through an Android demonstration app, proves the system's practicality in real-world scenarios.

cs.CV↗
arXiv2025

MacroBench: A Novel Testbed for Web Automation Scripts via Large Language Models

We introduce MacroBench, a code-first benchmark that evaluates whether LLMs can synthesize reusable browser-automation programs (macros) from natural-language goals by reading HTML/DOM and emitting Selenium. MacroBench instantiates seven self-hosted sites covering 681 tasks across interaction complexity and targeting difficulty. Our end-to-end protocol validates generated code via static checks, sandboxed execution, and outcome verification (DOM assertions, database snapshots), and includes a safety suite for scraping, spam/abuse, and credential/privacy prompts. Across 2,636 model-task runs, we observe stratified success: GPT-4o-mini (96.8%), GPT-4o (95.3%), Gemini (89.0%), DeepSeek (83.4%). Models handle simple tasks reliably (91.7%) but fail on complex workflows (0.0%), and none meet production-quality coding practices despite functional completion. We release our complete benchmark pipeline, evaluation framework, and experimental results at https://github.com/hyunjun1121/MacroBench to enable reproducible assessment of macro synthesis for web automation.

cs.SE↗
arXiv2025

X-Teaming Evolutionary M2S: Automated Discovery of Multi-turn to Single-turn Jailbreak Templates

Multi-turn-to-single-turn (M2S) compresses iterative red-teaming into one structured prompt, but prior work relied on a handful of manually written templates. We present X-Teaming Evolutionary M2S, an automated framework that discovers and optimizes M2S templates through language-model-guided evolution. The system pairs smart sampling from 12 sources with an LLM-as-judge inspired by StrongREJECT and records fully auditable logs. Maintaining selection pressure by setting the success threshold to $\theta = 0.70$, we obtain five evolutionary generations, two new template families, and 44.8% overall success (103/230) on GPT-4.1. A balanced cross-model panel of 2,500 trials (judge fixed) shows that structural gains transfer but vary by target; two models score zero at the same threshold. We also find a positive coupling between prompt length and score, motivating length-aware judging. Our results demonstrate that structure-level search is a reproducible route to stronger single-turn probes and underscore the importance of threshold calibration and cross-model evaluation. Code, configurations, and artifacts are available at https://github.com/hyunjun1121/M2S-x-teaming.

cs.CL↗
arXiv2025

ObjexMT: Objective Extraction and Metacognitive Calibration for LLM-as-a-Judge under Multi-Turn Jailbreaks

LLM-as-a-Judge (LLMaaJ) enables scalable evaluation, yet we lack a decisive test of a judge's qualification: can it recover the hidden objective of a conversation and know when that inference is reliable? Large language models degrade with irrelevant or lengthy context, and multi-turn jailbreaks can scatter goals across turns. We present ObjexMT, a benchmark for objective extraction and metacognition. Given a multi-turn transcript, a model must output a one-sentence base objective and a self-reported confidence. Accuracy is scored by semantic similarity to gold objectives, then thresholded once on 300 calibration items ($\tau^\star = 0.66$; $F_1@\tau^\star = 0.891$). Metacognition is assessed with expected calibration error, Brier score, Wrong@High-Confidence (0.80 / 0.90 / 0.95), and risk--coverage curves. Across six models (gpt-4.1, claude-sonnet-4, Qwen3-235B-A22B-FP8, kimi-k2, deepseek-v3.1, gemini-2.5-flash) evaluated on SafeMTData\_Attack600, SafeMTData\_1K, and MHJ, kimi-k2 achieves the highest objective-extraction accuracy (0.612; 95\% CI [0.594, 0.630]), while claude-sonnet-4 (0.603) and deepseek-v3.1 (0.599) are statistically tied. claude-sonnet-4 offers the best selective risk and calibration (AURC 0.242; ECE 0.206; Brier 0.254). Performance varies sharply across datasets (16--82\% accuracy), showing that automated obfuscation imposes challenges beyond model choice. High-confidence errors remain: Wrong@0.90 ranges from 14.9\% (claude-sonnet-4) to 47.7\% (Qwen3-235B-A22B-FP8). ObjexMT therefore supplies an actionable test for LLM judges: when objectives are implicit, judges often misinfer them; exposing objectives or gating decisions by confidence is advisable. All experimental data are in the Supplementary Material and at https://github.com/hyunjun1121/ObjexMT_dataset.

cs.CL↗
arXiv2025

Experimental Analysis of Productive Interaction Strategy with ChatGPT: User Study on Function and Project-level Code Generation Tasks

The application of Large Language Models (LLMs) is growing in the productive completion of Software Engineering tasks. Yet, studies investigating productive prompting techniques often employed a limited problem space, focusing primarily on well-known prompting patterns and targeting function-level SE practices. We identify significant gaps in real-world workflows that involve complexities beyond class-level (e.g., multi-class dependencies) and different features that can impact Human-LLM Interaction (HLI) processes in code generation. To address these issues, we designed an experiment to comprehensively analyze HLI features related to code generation productivity. Our study presents two project-level benchmark tasks that extend beyond function-level evaluations. We conducted a user study with 36 participants from diverse backgrounds, asking them to solve the assigned tasks by interacting with the GPT assistant using specific prompting patterns. We also examined participants' experiences and behavioral features during interactions by analyzing screen recordings and GPT chat logs. Our empirical investigation, based on statistical guidance, revealed (1) that three out of 15 HLI features emerged as consistently supported factors for productivity; (2) five primary guidelines for enhancing productivity for HLI processes; and (3) a taxonomy of 29 runtime and logic errors that can occur during HLI processes, along with suggested mitigation plans.

cs.SE↗
arXiv2025

ReFoCUS: Reinforcement-guided Frame Optimization for Contextual Understanding

Recent progress in Large Multi-modal Models (LMMs) has enabled effective vision-language reasoning, yet the ability to video understanding remains constrained by suboptimal frame selection strategies, albeit with the rapid development of video-specialized LMMs. Prior works attempted to solve this with static heuristics or external retrieval modules to feed frame-level information, but these approaches often fail to capture visual cues grounded to the given user queries conflating raw visual dynamics with true semantic relevance. In this paper, we introduce ReFoCUS (Reinforcement-guided Frame Optimization for Contextual UnderStanding), the first framework to integrate online policy-gradient reinforcement learning into frame-level optimization for video-LLMs. ReFoCUS aims to learn a frame selection policy, leveraging reward signals derived from reference models to capture their underlying scoring behavior over frame combinations that best support temporally grounded responses. To efficiently explore the large combinatorial frame space, we employ an autoregressive and query-conditional selection architecture that ensures contextual consistency while reducing complexity. Our policy learning removes the need for explicit frame-level supervision, as it implicitly discovers optimal and semantically consistent frame compositions. ReFoCUS consistently improves reasoning accuracy across multiple video QA benchmarks, demonstrating the advantage of aligning frame selection with model-internal utility.

cs.CV↗
arXiv2025

DIP-R1: Deep Inspection and Perception with RL Looking Through and Understanding Complex Scenes

MLLMs have demonstrated significant visual understanding capabilities, yet their fine-grained visual perception in complex real-world scenarios, such as densely crowded public areas, remains limited. Inspired by the recent success of RL in both LLMs and MLLMs, in this paper, we explore how RL can enhance visual perception ability of MLLMs. Then we develop a novel RL-based framework, Deep Inspection and Perception with RL (DIP-R1) designed to enhance the visual perception capabilities of MLLMs, by comprehending complex scenes and looking through visual instances closely. DIP-R1 guides MLLMs through detailed inspection of visual scene via three simply designed rule-based reward modeling. First, we adopt a standard reasoning reward encouraging the model to include three-step reasoning process: 1) comprehending entire visual scene, 2) observing for looking through interested but ambiguous regions, and 3) decision-making for predicting answer. Second, a variance-guided looking reward is designed to encourage MLLM to examine uncertain regions during the observing process, guiding it to inspect ambiguous areas and mitigate perceptual uncertainty. This reward promotes variance-driven visual exploration, enabling MLLM to reason about region-level uncertainty and explicitly indicate interpretable uncertain regions. Third, we model a weighted precision-recall accuracy reward enhancing accurate decision-making. We verify its effectiveness across diverse fine-grained object detection data consisting of challenging real-world scenes, such as densely crowded scenes. Built upon existing MLLMs, DIP-R1 achieves consistent and significant improvement across various in-domain and out-of-domain scenarios, outperforming various existing baselines and SFT method. Our findings highlight the substantial potential of integrating RL into MLLMs for enhancing capabilities in complex real-world perception tasks.

cs.CV↗