Searcharxiv⌕ Search

arXiv subjects

Hoejoon Kwon

Publications and source records attributed to Hoejoon Kwon.

3 recordsLinked to original sources

AnachroBench: Evaluating Historical Consistency in Large Vision-Language Models

Existing Large Vision-Language Model benchmarks primarily evaluate visual understanding in general everyday scenarios, leaving historical consistency with a specified era and country largely underexplored. To address this gap, we propose AnachroBench, a controlled discriminative benchmark for evaluating historical consistency in visual scenes. It contains 1,600 image pairs from eight countries spanning the late 19th to the early 20th century, each pairing an original historical image with an anachronistic version containing a seamlessly inserted modern object. AnachroBench evaluates three aspects: Inserted Object Perception, which tests whether models detect the inserted modern object; Era-Country Consistency, which assesses whether the scene is consistent with the specified era and country; and Conditional Era-Country Consistency, which measures consistency judgments only when the inserted object is correctly perceived. A false consistency judgment occurs when a model perceives the modern object but still judges the scene as historically consistent. Overall, AnachroBench reveals a gap between inserted object perception and era-country consistency judgments, providing a targeted evaluation of this behavior.

cs.CV↗

Beyond Safe Answers: Segment-Aware Listwise Alignment for Reasoning Safety in Large Reasoning Models

Large Reasoning Models (LRMs) pose a dual-surface safety challenge: both intermediate reasoning traces and final answers can contain harmful content. Existing alignment methods often operate at the whole-response level, allowing unsafe reasoning to be masked by a safe-looking final answer. We propose Segment-aware Listwise Target DPO (SaLT-DPO), which addresses this gap through three mechanisms: (1) segment-aware listwise alignment that decomposes responses into reasoning and answer segments, independently scores each segment's safety, and aligns length-normalized segment rewards with soft target distributions over multiple candidates; (2) joint safety coherence regularization that applies a weakest-link principle to promote safety consistency across both segments; and (3) utility anchoring on benign prompts to mitigate over-refusal and reasoning degradation. Experiments on three LRMs show that SaLT-DPO consistently reduces unsafe rates for both reasoning and answer segments while mitigating degradation in benign compliance and preserving general reasoning performance. Ablation studies demonstrate the complementary contributions of its components.

cs.AI↗

ALTSTEER: Selective Safety Steering for Moving Beyond Hard Refusals to Constructive Alternatives

Safety alignment is essential for deploying large language models, requiring systems to prevent harmful compliance while preserving helpfulness on benign requests. Activation steering offers a training-free inference-time approach to safety control, but effective safety steering requires addressing two coupled questions: when to intervene and how generation should be shaped after intervention. However, existing safety steering methods remain limited along both dimensions, as their triggering mechanisms can be unstable across domains and refusal-oriented steering often yields rigid refusals rather than constructive safe guidance. To address these limitations, we propose ALTSTEER, an inference-time framework that couples selective intervention with refusal-anchored constructive redirection within a single inference pass. ALTSTEER uses an internal refusal-relevant signal to decide when to steer, and applies staged steering to shift generation from refusal-oriented control toward constructive alternatives. Evaluations on Llama-3.1 and Qwen2.5 show that ALTSTEER preserves benign utility while improving constructive safe-completion behavior, especially on models that otherwise tend to produce short refusals for harmful requests.

cs.CL↗