Searcharxiv⌕ Search

arXiv subjects

Seunghyun Lee

Publications and source records attributed to Seunghyun Lee.

At least 19 recordsLinked to original sources

"Don't Leave the Task Unfinished": Persistence-Driven Violations under Quiet Oversight

LLM agents are increasingly given long, multi-step tasks to carry out, with the user reviewing little of what happens in between. Persistence, the tendency to keep working until a task is complete, is installed on purpose to improve performance on long tasks. Long-horizon tasks are typically complex, and the user's instruction can fall short of what the work actually requires, or conflict with the constraints the agent encounters in the task. We study what persistence does on long-horizon operations tasks that cannot be solved without breaking a documented constraint. We find that the agent pursues completion to the point of violating the constraint, and that a critically engaged user lowers the violation rate. Persistence takes violations from 29.2% to 58.3% across four models. Violations stay high wherever oversight goes quiet, whether the user is silent, assenting, or delegating with expressed trust. Only a questioning user, who asks for the basis of each action before allowing the next, brings the rate back to roughly the level measured without persistence. Practitioners should treat high persistence as a setting that carries a safety cost, and should recognize that the cost appears when oversight goes quiet. Safety here is better cast as a property of the human-AI pair than of the model alone. Project page: https://francis21c.github.io/quiet-oversight/

cs.LG↗

Unfinished-Bench: Measuring Rule Violations by Persistent Agents in Realistic Work Environments

AI agents are increasingly equipped with persistence toward task completion to run delegated work autonomously for long periods, under many unstated constraints from internal documents and other people's authority. When completion requires violating a rule, the agent must decide alone, yet benchmarks of instruction-rule conflicts state the rule in the prompt or keep it hidden and do not evaluate persistence. We present Unfinished-Bench, a benchmark that gives models persistence in autonomous runs without a user, in realistic work environments with embedded rules. In each of its 97 scenarios across 7 organizations and 7 risk categories, the goal and the rule cannot both hold. We find that persistence raised violations of a rule in an internal document from 32.2% to 43.6%, five times the 8.5% for the same rule stated by the user. Of the document-rule violations, 98.3% occurred after the rule was displayed, and in 89% of a judged sample the agent first wrote down the rule. Violations of rules placed in the environment without user oversight can lead to serious incidents, so persistence-driven risks need to be evaluated. Benchmark documentation: https://francis21c.github.io/unfinished-bench/

cs.LG↗

DM-Call: Decision-Model-Guided Function Calling with Cumulative Verification Feedback

Reliable function calling requires Large Language Models (LLMs) to select appropriate functions and generate arguments consistent with user intent. However, syntactically valid calls may still be semantically incorrect due to inappropriate function selection or mismatched arguments. In this paper, we propose Decision-Model-Guided Function Calling (DM-Call), a framework that combines local schema validation with separate decision-model verification of function selection and argument correctness. For rejected candidates, DM-Call accumulates verification probabilities, binary judgments, and schema errors across attempts. A hybrid feedback representation further combines numerical probabilities with verbal correctness labels to guide iterative refinement without executing external tools. Extensive evaluation across BFCL and API-Bank (first-call) demonstrates DM-Call boosts function-calling accuracy versus direct generation, at modest additional cost.

cs.IR↗

Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs

In a dual-use field, hazardous knowledge and the benign expertise around it share vocabulary, concepts, and causal structure. That overlap makes selective refusal hard, and existing methods such as unlearning and refusal training overgeneralize their suppression into the benign side. We argue for conditioning, not suppression. In this work, we introduce Token Inoculation, which marks hazardous sentences with a special token, the inoculation token, during continued pre-training. We then fine-tune the model to refuse hazardous questions and, on the same questions under the inoculation token, to predict the hazardous answers. At test time the inoculation token is absent, so the model does not express the hazardous answers it learned to predict. Across 1.5B to 14B parameters and three architecture families it holds hazardous accuracy near 18% while retaining 93% of the instruct baseline's MMLU performance, the best safety-utility balance among the baselines we compare. The same advantage reproduces under open-ended generation and in a cybersecurity transfer. We test this with a causal do(X) analysis that randomizes the accuracy of the conditioning signal, and find that refusal becomes more selective the more of the hazardous domain the model still knows. Our findings underscore the importance of training to predict the responses that must be refused, and outline a direction for safety alignment that does not degrade models' growing frontier knowledge. Project page: https://token-inoculation.github.io/

cs.LG↗

Scaling Limits of Multichannel Spectral Routers for Snapshot Imaging

Inverse-designed spectral routers can enable compact snapshot spectral imaging by directing different wavelengths to designated detector sub-pixels without mechanical scanning or absorptive filters. Here, we examine how routing performance varies as the number of spectral channels increases while the visible bandwidth remains fixed. A delay-bandwidth analysis relates channel count, worst-channel efficiency, and device thickness, explaining the growing optical capacity required to generate many spatially distinct spectral outputs. We then use adjoint-based topology optimization and three-dimensional finite-difference time-domain simulations to design TiO2/SiO2 routers with 9, 16, 25, and 36 channels under matched material, bandwidth, and thickness conditions. The average routing efficiency increases with sub-pixel size and approaches a plateau, while the maximum near-plateau efficiency decreases monotonically from 97.0% for 9 channels to 82.3% for 36 channels. At a fixed channel count, compressing the wavelength spacing to 8 nm or rearranging the wavelength-to-sub-pixel assignment changes the efficiency by less than one percentage point. These results indicate that the observed efficiency penalty is governed mainly by the number of wavelength-dependent routing constraints rather than by wavelength spacing or local wavelength arrangement. The findings quantify the trade-off between spectral sampling and optical throughput in compact snapshot spectral imagers.

physics.optics↗

Real-time Appearance-based Gaze Estimation for Open Domains

Appearance-based gaze estimation (AGE) has achieved remarkable performance in constrained settings, yet we reveal a significant generalization gap where existing AGE models often fail in practical, unconstrained scenarios, particularly those involving facial wearables and poor lighting conditions. We attribute this failure to two core factors: limited image diversity and inconsistent label fidelity across different datasets, especially along the pitch axis. To address these, we propose a robust AGE framework that enhances generalization without requiring additional human-annotated data. First, we expand the image manifold via an ensemble of augmentation techniques, including synthesis of eyeglasses, masks, and varied lighting. Second, to mitigate the impact of anisotropic inter-dataset label deviation, we reformulate gaze regression as a multi-task learning problem, incorporating multi-view supervised contrastive (SupCon) learning, discretized label classification, and eye-region segmentation as auxiliary objectives. To rigorously validate our approach, we curate new benchmark datasets designed to evaluate gaze robustness under challenging conditions, a dimension largely overlooked by existing evaluation protocols. Our MobileNet-based lightweight model achieves generalization performance competitive with the state-of-the-art (SOTA) UniGaze-H, while utilizing less than 1\% of its parameters, enabling high-fidelity, real-time gaze tracking on mobile devices.

cs.CV↗

BV Loss: Block Verification-Aware Loss for Block Diffusion Speculative Decoding

Diffusion drafters accelerate speculative decoding by proposing multiple tokens in parallel. Despite recent advances in speculative decoding through sequence-level drafting and verification, existing training objectives remain largely designed around token-level verification. To address this mismatch, we introduce Block Verification-aware loss (BV loss), a training objective designed to maximize the expected acceptance length of a drafted sequence. BV loss is directly derived from the block verification acceptance rule, providing a principled connection between the drafter training objective and the inference-time verification mechanism at the sequence level. Across math, code, and chat benchmarks, BV loss increases the mean number of tokens accepted per verification call under block verification by 13.0--21.0\% over cross-entropy loss training for DFlash and DSpark with Qwen3-4B and Qwen3-8B without changing the inference procedure. BV loss also outperforms tokenwise acceptance objectives such as TV loss and LK loss, and its gains extend to token verification and greedy decoding. These results demonstrate the benefit of training block diffusion drafters with an objective aligned with sequence-level verification, rather than optimizing each token independently.

cs.AI↗

Cultural Divergence Preservation: Diagnosing Flattening and Caricature in LLM-Simulated Survey Populations

Large language models (LLMs) are increasingly used as synthetic survey respondents to estimate population response distributions. In cross-cultural survey simulation, evaluations should assess not only distributional fidelity within countries but also whether differences across countries are preserved. However, existing distance-based metrics such as Jensen--Shannon divergence (JSD) do not directly capture such cross-country differences. To address this limitation, we introduce Cultural Divergence Preservation (CDP), a reference-light diagnostic based on a one-time human calibration. CDP identifies reduced cross-country divergence as cultural flattening and increased divergence as cultural caricature. To evaluate CDP, we conduct experiments across four LLM backbones, three persona-based prompting methods, and two survey domains, the World Values Survey (WVS) and the Big Five Personality Test. The results reveal a systematic discrepancy between conventional fidelity metrics and CDP. Controlled experiments show that CDP changes monotonically as cross-country divergence is attenuated or amplified, while the corresponding changes in JSD remain relatively small. In our audit of real LLM generations, DeepPersona-Inspired prompting is frequently favored by conventional fidelity metrics but exhibits the strongest flattening in every model--domain block. CDP thus complements fidelity metrics by directly quantifying the attenuation or amplification of cross-country divergence.

cs.CL↗

Automated Visualization Code Synthesis via Multi-Path Reasoning and Feedback-Driven Optimization

Large Language Models (LLMs) have become a cornerstone for automated visualization code generation, enabling users to create charts through natural language instructions. Despite improvements from techniques like few-shot prompting and query expansion, existing methods often struggle when requests are underspecified in actionable details (e.g., data preprocessing assumptions, solver or library choices, etc.), frequently necessitating manual intervention. To overcome these limitations, we propose VisPath: a Multi-Path Reasoning and Feedback-Driven Optimization Framework for Visualization Code Generation. VisPath handles underspecified queries through structured, multi-stage processing. It begins by using Chain-of-Thought (CoT) prompting to reformulate the initial user input, generating multiple extended queries in parallel to surface alternative plausible concretizations of the request. These queries then generate candidate visualization scripts, which are executed to produce diverse images. By assessing the visual quality and correctness of each output, VisPath generates targeted feedback that is aggregated to synthesize an optimal final result. Extensive experiments on MatPlotBench and Qwen-Agent Code Interpreter Benchmark show that VisPath outperforms state-of-the-art methods, providing a more reliable framework for AI-driven visualization generation.

cs.SE↗

Solar Open 2 Technical Report

We present Solar Open 2, a 250B-A15B Mixture-of-Experts language model built for long-horizon agentic tasks, scaled up from Solar Open 1 (Solar Open 100B). To hold entire agent trajectories in a single context, Solar Open 2 reaches a 1M-token window through a hybrid attention stack that interleaves one softmax layer among every three linear-attention layers, using no positional encoding and a gated delta rule extended to negative eigenvalues. To train at this scale under a fixed compute budget, we make training efficient in two ways: a stronger starting point, and higher-value data. For the starting point, we initialize Solar Open 2 from Solar Open 1, transferring the 5.69B-parameter shared skeleton that survives the architectural change and learning everything else through full pre-training. For the data, we curate for value per token: quality- and rarity-aware data curation and mixture-ratio optimization refine a 20T pool into a 10T mixture that, at equal token budget, outperforms the Solar Open 1 recipe. To build its agent skills, we train twelve domain specialists across purpose-built scenarios, then consolidate them into a single model by Multi-teacher On-Policy Distillation (MOPD). Against comparably sized open-weight models on English benchmarks, Solar Open 2 leads on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic suite, and stays competitive with the strongest (DeepSeek-V4-Flash and MiMo-V2.5) elsewhere. On Korean benchmarks, Solar Open 2 records the highest average of any model compared, including fast-tier closed APIs, and on Ko-GDPval, an in-house Korean officework-agent benchmark, it is competitive with DeepSeek-V4-Pro (1.6T) at less than a sixth of its size.

cs.CL↗

Beyond Local Independence: High-Dimensional Latent Class Graphical Models with Shared Block Structure

Latent class models are central tools for multivariate categorical data from heterogeneous populations, but their standard local-independence assumption is often unrealistic in modern high-dimensional applications. We propose a high-dimensional latent class graphical model for ordinal responses with block-structured local dependence. The model retains the interpretability and parsimony of classical latent class analysis by imposing a shared block partition of variables, while allowing class-specific graphical dependence within each block. We develop a scalable three-step estimator that first recovers latent classes by spectral clustering of a flattened response matrix, then estimates class-specific latent covariance matrices and aggregates them to recover the shared block partition, and finally estimates sparse within-block precision matrices. We establish finite-sample error bounds for clustering, covariance estimation, block recovery, and precision-matrix estimation, yielding end-to-end consistency of all model components under high-dimensional scaling. Simulations demonstrate accurate recovery of latent classes, the shared block partition, and class-specific dependence graphs with scalable computation. Applications to American National Election Studies survey data and HapMap3 genotype data show that the method uncovers interpretable local dependence structures while accounting for latent heterogeneity.

stat.ME↗

RoTRAG: Rule of Thumb Reasoning for Conversation Harm Detection with Retrieval-Augmented Generation

Detecting harmful content in multi turn dialogue requires reasoning over the full conversational context rather than isolated utterances. However, most existing methods rely mainly on models internal parametric knowledge, without explicit grounding in external normative principles. This often leads to inconsistent judgments in socially nuanced contexts, limited interpretability, and redundant reasoning across turns. To address this, we propose RoTRAG, a retrieval augmented framework that incorporates concise human written moral norms, called Rules of Thumb (RoTs), into LLM based harm assessment. For each turn, RoTRAG retrieves relevant RoTs from an external corpus and uses them as explicit normative evidence for turn level reasoning and final severity classification. To improve efficiency, we further introduce a lightweight binary routing classifier that decides whether a new turn requires retrieval grounded reasoning or can reuse existing context. Experiments on ProsocialDialog and Safety Reasoning Multi Turn Dialogue show that RoTRAG consistently improves both harm classification and severity estimation over competitive baselines, with an average relative gain of around 40% in F1 across benchmark datasets and an average relative reduction of 8.4% in distributional error, while reducing redundant computation without sacrificing performance.

cs.CL↗

MARIC: Multi-Agent Reasoning for Image Classification

Image classification has traditionally relied on parameter-intensive model training, requiring large-scale annotated datasets and extensive fine tuning to achieve competitive performance. While recent vision language models (VLMs) alleviate some of these constraints, they remain limited by their reliance on single pass representations, often failing to capture complementary aspects of visual content. In this paper, we introduce Multi Agent based Reasoning for Image Classification (MARIC), a multi agent framework that reformulates image classification as a collaborative reasoning process. MARIC first utilizes an Outliner Agent to analyze the global theme of the image and generate targeted prompts. Based on these prompts, three Aspect Agents extract fine grained descriptions along distinct visual dimensions. Finally, a Reasoning Agent synthesizes these complementary outputs through integrated reflection step, producing a unified representation for classification. By explicitly decomposing the task into multiple perspectives and encouraging reflective synthesis, MARIC mitigates the shortcomings of both parameter-heavy training and monolithic VLM reasoning. Experiments on 4 diverse image classification benchmark datasets demonstrate that MARIC significantly outperforms baselines, highlighting the effectiveness of multi-agent visual reasoning for robust and interpretable image classification.

cs.CV↗

SSAFE: Simple and Strong AI-Generated Image Detection via Frozen Vision Encoders

The rapid advancement of generative models has blurred the boundary between synthetic and real imagery, creating an urgent need for reliable deepfake detection. Yet most existing approaches rely on massive real--fake datasets, which are increasingly difficult to maintain as new generators continue to emerge. In this work, we investigate how much information about image authenticity is already encoded in modern multimodal vision representations. We find that frozen multimodal encoders naturally separate real and synthetic images in their embedding space, enabling a simple linear classifier to achieve strong performance without task-specific fine-tuning. Motivated by this observation, we develop a representation-aware data curation strategy that selects a compact set of representative generators for training. The resulting training set contains only 10K images, compared to 288K in AIGIBench and 4M in OpenFake, while improving robustness to unseen generators and distribution shifts. We additionally introduce RealWorldBench, a benchmark consisting of modern camera photographs, contemporary stock images, and outputs from recent commercial generators. Experiments across multiple benchmarks show that combining frozen multimodal representations with carefully curated training data provides a simple and effective approach to AI-generated image detection.

cs.CV↗

Toward Culturally Aligned LLMs through Ontology-Guided Multi-Agent Reasoning

Large Language Models (LLMs) increasingly support culturally sensitive decision making, yet often exhibit misalignment due to skewed pretraining data and the absence of structured value representations. Existing methods can steer outputs, but often lack demographic grounding and treat values as independent, unstructured signals, reducing consistency and interpretability. We propose OG-MAR, an Ontology-Guided Multi-Agent Reasoning framework. OG-MAR summarizes respondent-specific values from the World Values Survey (WVS) and constructs a global cultural ontology by eliciting relations over a fixed taxonomy via competency questions. At inference time, it retrieves ontology-consistent relations and demographically similar profiles to instantiate multiple value-persona agents, whose outputs are synthesized by a judgment agent that enforces ontology consistency and demographic proximity. Experiments on regional social-survey benchmarks across four LLM backbones show that OG-MAR improves cultural alignment and robustness over competitive baselines, while producing more transparent reasoning traces.

cs.CL↗

ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents

Exploitation is not a binary event. It is a ladder of acquiring progressive capabilities, from executing a single buggy line of code to taking full control of the target. However, existing LLM security benchmarks treat a crash as exploitation success. That single binary outcome collapses the hard parts of exploitation: the transition from triggering a bug to constructing reusable primitives and control. We present ExploitBench, a capability-graded benchmark that decomposes exploitation into 16 measurable flags, from coverage and crash through sandbox primitives, arbitrary read/write, control-flow hijack, and arbitrary code execution. Each capability is verified by a deterministic oracle that uses a per-run randomized challenge-response for primitives, differential execution against ground-truth binaries to measure progress, and a signal-handler proof for code execution. We instantiate ExploitBench on 41 V8 bugs because V8 is both widely deployed and exploitation-hardened. We report three arms: as the primary measurement of model-environment capability, as a secondary arm that adds adaptive coaching to test whether targeted feedback shifts outcomes, and as an ablation that swaps in the model's native CLI to check whether vendor-side optimizations increase exploitation capabilities. Our results show a sharp capability split between publicly deployed frontier models and the private frontier. Across the 8 publicly deployed models tested, reaching the vulnerable code and triggering a crash is routine, but arbitrary code execution is not. The private model shows arbitrary code execution on approximately half. Overall, results suggest that exploit construction against hardened targets is an emerging frontier capability.

cs.CR↗

Hot Days, Unsafe Schools? The Impact of Heat on School Shootings

Using data on shootings in U.S.\ K--12 schools from 1981 to 2022, we estimate the effect of temperature on school shootings and assess climate-change impacts. We find that days with maximum temperatures above 90$^{\circ}$F increase school shooting incidence by approximately 90\% relative to days with maximum temperatures below 70$^{\circ}$F. The response is concentrated in interpersonal incidents and in non-class periods, such as before school, dismissal, after school, and lunch: shootings during these periods more than triple on days with maximum temperatures above 90$^{\circ}$F, while shootings during class time show no detectable temperature response. The estimated effects are positive for both indoor and outdoor shootings and are larger for shootings involving fatalities or injuries than for shootings involving only minor or no injuries. Applying the estimated dose-response to future warming, we estimate that interpersonal school shootings increase by 6\% by mid-century (2051--2060) under moderate emissions (SSP2--4.5) and 8\% under high emissions (SSP5--8.5), or about 12 and 16 additional incidents per decade. The present discounted value of mid-century social costs is \$599 million under SSP2--4.5 and \$799 million under SSP5--8.5, driven primarily by lost lifetime earnings among exposed students. The results suggest that climate damages in schools may include rare but high-cost safety events, not only heat stress and learning losses.

econ.GN↗

Fluctuations in random field Ising models

This paper establishes a CLT for linear statistics of the form $\langle \mathbf{q},\boldsymbolσ \rangle$ with quantitative Berry-Esseen bounds, where $\boldsymbolσ$ is an observation from an exponential family with a quadratic form as its sufficient statistic, in the \enquote{high-temperature} regime. We apply our general result to random field Ising models with both discrete and continuous spins. To demonstrate the generality of our techniques, we apply our results to derive both quenched and annealed CLTs in various examples, which include Ising models on some graph ensembles of common interest (Erdős-Rényi, regular, dense bipartite), and the Hopfield spin glass model. Our proofs rely on a combination of Stein's method of exchangeable pairs and Chevet type concentration inequalities.

math.PR↗