arXiv subjects

Su-Hyeon Kim

Publications and source records attributed to Su-Hyeon Kim.

8 recordsLinked to original sources

arXiv2026

One Adapter Pair per Model: A Universal Activation Interface for Language Models

Activation-based tools are usually tied to one model's native hidden space, requiring probes, sparse autoencoders, and natural-language interpreters to be rebuilt or rediscovered for each new language model. We present a Universal Activation Bus, a framework that provides a common activation interface across compatible language models. Using a small set of source models, we learn a shared dense space together with one lightweight linear encoder--decoder adapter pair per model. After source training, the interface is frozen; a new model joins by fitting only its adapter pair on unlabeled matched text. The resulting interface allows activation-based tools to be shared across connected models, including common probes and SAE features as well as access to an NLA originally trained for a different model. Across five models, semantically related texts form consistent neighborhoods in the shared space, and an onboarded model reuses these tools effectively without retraining them. We further show that an intermediate activation from one model can be used by another model's frozen upper layers to produce predictions. These results establish a stable, model-wise activation contract for reusable tools across compatible language models.

cs.AI↗

arXiv2026

Obfuscation Rules for Detecting and Detoxifying Korean Toxicity

As language models become increasingly deployed in online environments, toxicity detection and detoxification have received growing attention. Existing studies primarily focus on non-obfuscated text, which limits robustness when users intentionally disguise toxic expressions. In particular, Korean toxic expressions can be easily disguised through agglutinative morphology and Hangeul-specific orthographic variation. However, obfuscation in Korean remains largely unexplored, which motivates us to introduce a KOTOX: Korean toxic dataset for deobfuscation and detoxification. We categorize Korean obfuscation patterns into linguistically grounded classes, define transformation rules derived from real-world examples, and provide the resulting obfuscation framework as an open transformation package. Using these rules, we provide paired neutral and toxic sentences alongside their obfuscated counterparts. Models trained on our dataset better handle obfuscated text without sacrificing performance on non-obfuscated text. This is the first dataset that simultaneously supports deobfuscation and detoxification for the Korean language. We expect the dataset to facilitate better understanding and mitigation of obfuscated toxic content in LLM for Korean. Our code and data are available at https://github.com/leeyejin1231/KOTOX.

cs.CL↗

arXiv2026

ReSyn: A Generalized Recursive Regular Expression Synthesis Framework

Existing Programming-By-Example (PBE) systems often rely on simplified benchmarks that fail to capture the high structural complexity of real-world regexes, such as deeper nesting and frequent use of union operations. To overcome the resulting performance drop, we propose ReSyn, a synthesizer-agnostic divide-and-conquer framework that decomposes complex synthesis problem into manageable sub-problems. We also introduce Set2Regex, a parameter-efficient synthesizer capturing the permutation invariance of examples. Experimental results demonstrate that ReSyn significantly boosts accuracy across various synthesizers, and its combination with Set2Regex establishes a new state-of-the-art on challenging real-world benchmark. The complete source code, datasets, and pre-trained model checkpoints are publicly available at https://github.com/mrseongminkim/ReSyn.

cs.PL↗

arXiv2026

CRaFT: Circuit-Guided Refusal Feature Selection via Cross-Layer Transcoders

While modern LLMs are aligned to refuse harmful requests, it is essential to understand the underlying mechanistic basis of this refusal behavior for model safety analysis. For example, steering-based jailbreak attacks exploit this by identifying and manipulating sparse, neuron-like refusal features to bypass safety guardrails. Current feature selection methods primarily rely on how strongly features activate on harmful prompts. However, activation strength alone often captures superficial heuristics such as topic or lexical cues, rather than the true causal mechanisms. Thus, selecting refusal features requires measuring inter-feature relationships, rather than treating each feature as an isolated activation signal. Based on this insight, we propose CRaFT, a circuit-guided framework for identifying critical refusal features that directly govern the refusal decision. CRaFT leverages cross-layer transcoders to map the model's internal computations into a sparse feature circuit graph, where edges quantify inter-feature influences and their contributions to the final output logits. By aggregating the effects propagating along the paths to refusal, CRaFT effectively ranks the most influential features. Extensive evaluations across four jailbreak benchmarks show that CRaFT significantly improves average performance from 6.7% to 57.4% and generates more specific harmful completions compared to current SOTA methods.

cs.AI↗

arXiv2026

Cross-Family Universality of Behavioral Axes via Anchor-Projected Representations

Large language models from different families use different hidden dimensions, tokenizers, and training procedures, making behavioral directions difficult to compare or transfer across models. We introduce an anchor-projection framework that maps hidden representations from each model into a shared anchor coordinate space (ACS). Behavioral directions extracted from source models are projected into ACS and averaged into a canonical direction. For a new model, the canonical direction is reconstructed into its native hidden space using only anchor activations, without fine-tuning or target-specific direction extraction. We evaluate five instruction-tuned model families and ten behavioral axes. We find that same-axis directions align tightly across the Llama-Qwen-Mistral-Phi (LQMP) cluster in ACS. This shared structure transfers to downstream tasks. For the aligned LQMP cluster, held-out targets achieve (0.83) ten-way detection accuracy and (0.95) mean binary AUROC, while canonical steering induces refusal-rate shifts of up to +0.46% under distribution shift. Sensitivity analyses show that two source models and small anchor pools already suffice to approximate transferable directions. Overall, ACS provides a novel perspective on cross-family interpretability, revealing that representation-level transfer remains robust across model families.

cs.AI↗

arXiv2026

How Does the Thinking Step Influence Model Safety? An Entropy-based Safety Reminder for LRMs

Large Reasoning Models (LRMs) achieve remarkable success through explicit thinking steps, yet the thinking steps introduce a novel risk by potentially amplifying unsafe behaviors. Despite this vulnerability, conventional defense mechanisms remain ineffective as they overlook the unique reasoning dynamics of LRMs. In this work, we find that the emergence of safe-reminding phrases within thinking steps plays a pivotal role in ensuring LRM safety. Motivated by this finding, we propose SafeRemind, a decoding-time defense method that dynamically injects safe-reminding phrases into thinking steps. By leveraging entropy triggers to intervene at decision-locking points, SafeRemind redirects potentially harmful trajectories toward safer outcomes without requiring any parameter updates. Extensive evaluations across five LRMs and six benchmarks demonstrate that SafeRemind substantially enhances safety, achieving improvements of up to 45.5%p while preserving core reasoning utility.

cs.AI↗

arXiv2025

ECO: Enhanced Code Optimization via Performance-Aware Prompting for Code-LLMs

Code runtime optimization-the task of rewriting a given code to a faster one-remains challenging, as it requires reasoning about performance trade-offs involving algorithmic and structural choices. Recent approaches employ code-LLMs with slow-fast code pairs provided as optimization guidance, but such pair-based methods obscure the causal factors of performance gains and often lead to superficial pattern imitation rather than genuine performance reasoning. We introduce ECO, a performance-aware prompting framework for code optimization. ECO first distills runtime optimization instructions (ROIs) from reference slow-fast code pairs; Each ROI describes root causes of inefficiency and the rationales that drive performance improvements. For a given input code, ECO in parallel employs (i) a symbolic advisor to produce a bottleneck diagnosis tailored to the code, and (ii) an ROI retriever to return related ROIs. These two outputs are then composed into a performance-aware prompt, providing actionable guidance for code-LLMs. ECO's prompts are model-agnostic, require no fine-tuning, and can be easily prepended to any code-LLM prompt. Our empirical studies highlight that ECO prompting significantly improves code-LLMs' ability to generate efficient code, achieving speedups of up to 7.81x while minimizing correctness loss.

cs.PL↗

arXiv2022

Neuro-Symbolic Regex Synthesis Framework via Neural Example Splitting

Due to the practical importance of regular expressions (regexes, for short), there has been a lot of research to automatically generate regexes from positive and negative string examples. We tackle the problem of learning regexes faster from positive and negative strings by relying on a novel approach called `neural example splitting'. Our approach essentially split up each example string into multiple parts using a neural network trained to group similar substrings from positive strings. This helps to learn a regex faster and, thus, more accurately since we now learn from several short-length strings. We propose an effective regex synthesis framework called `SplitRegex' that synthesizes subregexes from `split' positive substrings and produces the final regex by concatenating the synthesized subregexes. For the negative sample, we exploit pre-generated subregexes during the subregex synthesis process and perform the matching against negative strings. Then the final regex becomes consistent with all negative strings. SplitRegex is a divided-and-conquer framework for learning target regexes; split (=divide) positive strings and infer partial regexes for multiple parts, which is much more accurate than the whole string inferring, and concatenate (=conquer) inferred regexes while satisfying negative strings. We empirically demonstrate that the proposed SplitRegex framework substantially improves the previous regex synthesis approaches over four benchmark datasets.

cs.LG↗