arXiv subjects

Hyunwoo Kim

Publications and source records attributed to Hyunwoo Kim.

At least 55 records · Page 3Linked to original sources

arXiv2025

HAICOSYSTEM: An Ecosystem for Sandboxing Safety Risks in Human-AI Interactions

AI agents are increasingly autonomous in their interactions with human users and tools, leading to increased interactional safety risks. We present HAICOSYSTEM, a framework examining AI agent safety within diverse and complex social interactions. HAICOSYSTEM features a modular sandbox environment that simulates multi-turn interactions between human users and AI agents, where the AI agents are equipped with a variety of tools (e.g., patient management platforms) to navigate diverse scenarios (e.g., a user attempting to access other patients' profiles). To examine the safety of AI agents in these interactions, we develop a comprehensive multi-dimensional evaluation framework that uses metrics covering operational, content-related, societal, and legal risks. Through running 1840 simulations based on 92 scenarios across seven domains (e.g., healthcare, finance, education), we demonstrate that HAICOSYSTEM can emulate realistic user-AI interactions and complex tool use by AI agents. Our experiments show that state-of-the-art LLMs, both proprietary and open-sourced, exhibit safety risks in over 50\% cases, with models generally showing higher risks when interacting with simulated malicious users. Our findings highlight the ongoing challenge of building agents that can safely navigate complex interactions, particularly when faced with malicious users. To foster the AI agent safety ecosystem, we release a code platform that allows practitioners to create custom scenarios, simulate interactions, and evaluate the safety and performance of their agents.

cs.AI↗

arXiv2025

Toward Automated Hypervisor Scenario Generation Based on VM Workload Profiling for Resource-Constrained Environments

In the automotive industry, the rise of software-defined vehicles (SDVs) has driven a shift toward virtualization-based architectures that consolidate diverse automotive workloads on a shared hardware platform. To support this evolution, chipset vendors provide board support packages (BSPs), hypervisor setups, and resource allocation guidelines. However, adapting these static configurations to varying system requirements and workloads remain a significant challenge for Tier 1 integrators. This paper presents an automated scenario generation framework, which helps automotive vendors to allocate hardware resources efficiently across multiple VMs. By profiling runtime behavior and integrating both theoretical models and vendor heuristics, the proposed tool generates optimized hypervisor configurations tailored to system constraints. We compare two main approaches for modeling target QoS based on profiled data and resource allocation: domain-guided parametric modeling and deep learning-based modeling. We further describe our optimization strategy using the selected QoS model to derive efficient resource allocations. Finally, we report on real-world deployments to demonstrate the effectiveness of our framework in improving integration efficiency and reducing development time in resource-constrained environments.

cs.SE↗

arXiv2025

Hypothesis-Driven Theory-of-Mind Reasoning for Large Language Models

Existing LLM reasoning methods have shown impressive capabilities across various tasks, such as solving math and coding problems. However, applying these methods to scenarios without ground-truth answers or rule-based verification methods - such as tracking the mental states of an agent - remains challenging. Inspired by the sequential Monte Carlo algorithm, we introduce thought-tracing, an inference-time reasoning algorithm designed to trace the mental states of specific agents by generating hypotheses and weighting them based on observations without relying on ground-truth solutions to questions in datasets. Our algorithm is modeled after the Bayesian theory-of-mind framework, using LLMs to approximate probabilistic inference over agents' evolving mental states based on their perceptions and actions. We evaluate thought-tracing on diverse theory-of-mind benchmarks, demonstrating significant performance improvements compared to baseline LLMs. Our experiments also reveal interesting behaviors of the recent reasoning models - e.g., o3 and R1 - on theory-of-mind, highlighting the difference of social reasoning compared to other domains.

cs.AI↗

arXiv2025

Socratic-MCTS: Test-Time Visual Reasoning by Asking the Right Questions

Recent research in vision-language models (VLMs) has centered around the possibility of equipping them with implicit long-form chain-of-thought reasoning -- akin to the success observed in language models -- via distillation and reinforcement learning. But what about the non-reasoning models already trained and deployed across the internet? Should we simply abandon them, or is there hope for a search mechanism that can elicit hidden knowledge and induce long reasoning traces -- without any additional training or supervision? In this paper, we explore this possibility using a Monte Carlo Tree Search (MCTS)-inspired algorithm, which injects subquestion-subanswer pairs into the model's output stream. We show that framing reasoning as a search process -- where subquestions act as latent decisions within a broader inference trajectory -- helps the model "connect the dots" between fragmented knowledge and produce extended reasoning traces in non-reasoning models. We evaluate our method across three benchmarks and observe consistent improvements. Notably, our approach yields a 2% overall improvement on MMMU-PRO, including a significant 9% gain in Liberal Arts.

cs.CV↗

arXiv2025

Machine-Facing English: Defining a Hybrid Register Shaped by Human-AI Discourse

Machine-Facing English (MFE) is an emergent register shaped by the adaptation of everyday language to the expanding presence of AI interlocutors. Drawing on register theory (Halliday 1985, 2006), enregisterment (Agha 2003), audience design (Bell 1984), and interactional pragmatics (Giles & Ogay 2007), this study traces how sustained human-AI interaction normalizes syntactic rigidity, pragmatic simplification, and hyper-explicit phrasing - features that enhance machine parseability at the expense of natural fluency. Our analysis is grounded in qualitative observations from bilingual (Korean/English) voice- and text-based product testing sessions, with reflexive drafting conducted using Natural Language Declarative Prompting (NLD-P) under human curation. Thematic analysis identifies five recurrent traits - redundant clarity, directive syntax, controlled vocabulary, flattened prosody, and single-intent structuring - that improve execution accuracy but compress expressive range. MFE's evolution highlights a persistent tension between communicative efficiency and linguistic richness, raising design challenges for conversational interfaces and pedagogical considerations for multilingual users. We conclude by underscoring the need for comprehensive methodological exposition and future empirical validation.

cs.CL↗

arXiv2025

Adaptive Episode Length Adjustment for Multi-agent Reinforcement Learning

In standard reinforcement learning, an episode is defined as a sequence of interactions between agents and the environment, which terminates upon reaching a terminal state or a pre-defined episode length. Setting a shorter episode length enables the generation of multiple episodes with the same number of data samples, thereby facilitating an exploration of diverse states. While shorter episodes may limit the collection of long-term interactions, they may offer significant advantages when properly managed. For example, trajectory truncation in single-agent reinforcement learning has shown how the benefits of shorter episodes can be leveraged despite the trade-off of reduced long-term interaction experiences. However, this approach remains underexplored in MARL. This paper proposes a novel MARL approach, Adaptive Episode Length Adjustment (AELA), where the episode length is initially limited and gradually increased based on an entropy-based assessment of learning progress. By starting with shorter episodes, agents can focus on learning effective strategies for initial states and minimize time spent in dead-end states. The use of entropy as an assessment metric prevents premature convergence to suboptimal policies and ensures balanced training over varying episode lengths. We validate our approach using the StarCraft Multi-agent Challenge (SMAC) and a modified predator-prey environment, demonstrating significant improvements in both convergence speed and overall performance compared to existing methods. To the best of our knowledge, this is the first study to adaptively adjust episode length in MARL based on learning progress.

cs.MA↗

arXiv2025

Retro-Search: Exploring Untaken Paths for Deeper and Efficient Reasoning

Large reasoning models exhibit remarkable reasoning capabilities via long, elaborate reasoning trajectories. Supervised fine-tuning on such reasoning traces, also known as distillation, can be a cost-effective way to boost reasoning capabilities of student models. However, empirical observations reveal that these reasoning trajectories are often suboptimal, switching excessively between different lines of thought, resulting in under-thinking, over-thinking, and even degenerate responses. We introduce Retro-Search, an MCTS-inspired search algorithm, for distilling higher quality reasoning paths from large reasoning models. Retro-Search retrospectively revises reasoning paths to discover better, yet shorter traces, which can then lead to student models with enhanced reasoning capabilities with shorter, thus faster inference. Our approach can enable two use cases: self-improvement, where models are fine-tuned on their own Retro-Search-ed thought traces, and weak-to-strong improvement, where a weaker model revises stronger model's thought traces via Retro-Search. For self-improving, R1-distill-7B, fine-tuned on its own Retro-Search-ed traces, reduces the average reasoning length by 31.2% while improving performance by 7.7% across seven math benchmarks. For weak-to-strong improvement, we retrospectively revise R1-671B's traces from the OpenThoughts dataset using R1-distill-32B as the Retro-Search-er, a model 20x smaller. Qwen2.5-32B, fine-tuned on this refined data, achieves performance comparable to R1-distill-32B, yielding an 11.3% reduction in reasoning length and a 2.4% performance improvement compared to fine-tuning on the original OpenThoughts data. Our work counters recently emergent viewpoints that question the relevance of search algorithms in the era of large reasoning models, by demonstrating that there are still opportunities for algorithmic advancements, even for frontier models.

cs.AI↗

arXiv2025

Geometry in Style: 3D Stylization via Surface Normal Deformation

We present Geometry in Style, a new method for identity-preserving mesh stylization. Existing techniques either adhere to the original shape through overly restrictive deformations such as bump maps or significantly modify the input shape using expressive deformations that may introduce artifacts or alter the identity of the source shape. In contrast, we represent a deformation of a triangle mesh as a target normal vector for each vertex neighborhood. The deformations we recover from target normals are expressive enough to enable detailed stylizations yet restrictive enough to preserve the shape's identity. We achieve such deformations using our novel differentiable As-Rigid-As-Possible (dARAP) layer, a neural-network-ready adaptation of the classical ARAP algorithm which we use to solve for per-vertex rotations and deformed vertices. As a differentiable layer, dARAP is paired with a visual loss from a text-to-image model to drive deformations toward style prompts, altogether giving us Geometry in Style. Our project page is at https://threedle.github.io/geometry-in-style.

cs.GR↗

arXiv2025

MeshUp: Multi-Target Mesh Deformation via Blended Score Distillation

We propose MeshUp, a technique that deforms a 3D mesh towards multiple target concepts, and intuitively controls the region where each concept is expressed. Conveniently, the concepts can be defined as either text queries, e.g., "a dog" and "a turtle," or inspirational images, and the local regions can be selected as any number of vertices on the mesh. We can effectively control the influence of the concepts and mix them together using a novel score distillation approach, referred to as the Blended Score Distillation (BSD). BSD operates on each attention layer of the denoising U-Net of a diffusion model as it extracts and injects the per-objective activations into a unified denoising pipeline from which the deformation gradients are calculated. To localize the expression of these activations, we create a probabilistic Region of Interest (ROI) map on the surface of the mesh, and turn it into 3D-consistent masks that we use to control the expression of these activations. We demonstrate the effectiveness of BSD empirically and show that it can deform various meshes towards multiple objectives. Our project page is at https://threedle.github.io/MeshUp.

cs.CV↗

arXiv2025

Alpaca against Vicuna: Using LLMs to Uncover Memorization of LLMs

In this paper, we introduce a black-box prompt optimization method that uses an attacker LLM agent to uncover higher levels of memorization in a victim agent, compared to what is revealed by prompting the target model with the training data directly, which is the dominant approach of quantifying memorization in LLMs. We use an iterative rejection-sampling optimization process to find instruction-based prompts with two main characteristics: (1) minimal overlap with the training data to avoid presenting the solution directly to the model, and (2) maximal overlap between the victim model's output and the training data, aiming to induce the victim to spit out training data. We observe that our instruction-based prompts generate outputs with 23.7% higher overlap with training data compared to the baseline prefix-suffix measurements. Our findings show that (1) instruction-tuned models can expose pre-training data as much as their base-models, if not more so, (2) contexts other than the original training data can lead to leakage, and (3) using instructions proposed by other LLMs can open a new avenue of automated attacks that we should further study and explore. The code can be found at https://github.com/Alymostafa/Instruction_based_attack .

cs.CL↗

arXiv2024

Perceptions to Beliefs: Exploring Precursory Inferences for Theory of Mind in Large Language Models

While humans naturally develop theory of mind (ToM), the capability to understand other people's mental states and beliefs, state-of-the-art large language models (LLMs) underperform on simple ToM benchmarks. We posit that we can extend our understanding of LLMs' ToM abilities by evaluating key human ToM precursors$-$perception inference and perception-to-belief inference$-$in LLMs. We introduce two datasets, Percept-ToMi and Percept-FANToM, to evaluate these precursory inferences for ToM in LLMs by annotating characters' perceptions on ToMi and FANToM, respectively. Our evaluation of eight state-of-the-art LLMs reveals that the models generally perform well in perception inference while exhibiting limited capability in perception-to-belief inference (e.g., lack of inhibitory control). Based on these results, we present PercepToM, a novel ToM method leveraging LLMs' strong perception inference capability while supplementing their limited perception-to-belief inference. Experimental results demonstrate that PercepToM significantly enhances LLM's performance, especially in false belief scenarios.

cs.CL↗

arXiv2024

Is this the real life? Is this just fantasy? The Misleading Success of Simulating Social Interactions With LLMs

Recent advances in large language models (LLM) have enabled richer social simulations, allowing for the study of various social phenomena. However, most recent work has used a more omniscient perspective on these simulations (e.g., single LLM to generate all interlocutors), which is fundamentally at odds with the non-omniscient, information asymmetric interactions that involve humans and AI agents in the real world. To examine these differences, we develop an evaluation framework to simulate social interactions with LLMs in various settings (omniscient, non-omniscient). Our experiments show that LLMs perform better in unrealistic, omniscient simulation settings but struggle in ones that more accurately reflect real-world conditions with information asymmetry. Our findings indicate that addressing information asymmetry remains a fundamental challenge for LLM-based agents.

cs.CL↗

arXiv2024

CULTURE-GEN: Revealing Global Cultural Perception in Language Models through Natural Language Prompting

As the utilization of large language models (LLMs) has proliferated world-wide, it is crucial for them to have adequate knowledge and fair representation for diverse global cultures. In this work, we uncover culture perceptions of three SOTA models on 110 countries and regions on 8 culture-related topics through culture-conditioned generations, and extract symbols from these generations that are associated to each culture by the LLM. We discover that culture-conditioned generation consist of linguistic "markers" that distinguish marginalized cultures apart from default cultures. We also discover that LLMs have an uneven degree of diversity in the culture symbols, and that cultures from different geographic regions have different presence in LLMs' culture-agnostic generation. Our findings promote further research in studying the knowledge and fairness of global culture perception in LLMs. Code and Data can be found here: https://github.com/huihanlhh/Culture-Gen/

cs.CL↗

arXiv2024

Probabilistic Vision-Language Representation for Weakly Supervised Temporal Action Localization

Weakly supervised temporal action localization (WTAL) aims to detect action instances in untrimmed videos using only video-level annotations. Since many existing works optimize WTAL models based on action classification labels, they encounter the task discrepancy problem (i.e., localization-by-classification). To tackle this issue, recent studies have attempted to utilize action category names as auxiliary semantic knowledge through vision-language pre-training (VLP). However, there are still areas where existing research falls short. Previous approaches primarily focused on leveraging textual information from language models but overlooked the alignment of dynamic human action and VLP knowledge in a joint space. Furthermore, the deterministic representation employed in previous studies struggles to capture fine-grained human motions. To address these problems, we propose a novel framework that aligns human action knowledge and VLP knowledge in a probabilistic embedding space. Moreover, we propose intra- and inter-distribution contrastive learning to enhance the probabilistic embedding space based on statistical similarities. Extensive experiments and ablation studies reveal that our method significantly outperforms all previous state-of-the-art methods. Code is available at https://github.com/sejong-rcv/PVLR.

cs.CV↗

arXiv2024

Using LLMs to Investigate Correlations of Conversational Follow-up Queries with User Satisfaction

With large language models (LLMs), conversational search engines shift how users retrieve information from the web by enabling natural conversations to express their search intents over multiple turns. Users' natural conversation embodies rich but implicit signals of users' search intents and evaluation of search results to understand user experience with the system. However, it is underexplored how and why users ask follow-up queries to continue conversations with conversational search engines and how the follow-up queries signal users' satisfaction. From qualitative analysis of 250 conversational turns from an in-lab user evaluation of Naver Cue:, a commercial conversational search engine, we propose a taxonomy of 18 users' follow-up query patterns from conversational search, comprising two major axes: (1) users' motivations behind continuing conversations (N = 7) and (2) actions of follow-up queries (N = 11). Compared to the existing literature on query reformulations, we uncovered a new set of motivations and actions behind follow-up queries, including asking for subjective opinions or providing natural language feedback on the engine's responses. To analyze conversational search logs with our taxonomy in a scalable and efficient manner, we built an LLM-powered classifier (73% accuracy). With our classifier, we analyzed 2,061 conversational tuples collected from real-world usage logs of Cue: and examined how the conversation patterns from our taxonomy correlates with satisfaction. Our initial findings suggest some signals of dissatisfactions, such as Clarifying Queries, Excluding Condition, and Substituting Condition with follow-up queries. We envision our approach could contribute to automated evaluation of conversation search experience by providing satisfaction signals and grounds for realistic user simulations.

cs.HC↗

arXiv2024

Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory

The interactive use of large language models (LLMs) in AI assistants (at work, home, etc.) introduces a new set of inference-time privacy risks: LLMs are fed different types of information from multiple sources in their inputs and are expected to reason about what to share in their outputs, for what purpose and with whom, within a given context. In this work, we draw attention to the highly critical yet overlooked notion of contextual privacy by proposing ConfAIde, a benchmark designed to identify critical weaknesses in the privacy reasoning capabilities of instruction-tuned LLMs. Our experiments show that even the most capable models such as GPT-4 and ChatGPT reveal private information in contexts that humans would not, 39% and 57% of the time, respectively. This leakage persists even when we employ privacy-inducing prompts or chain-of-thought reasoning. Our work underscores the immediate need to explore novel inference-time privacy-preserving approaches, based on reasoning and theory of mind.

cs.AI↗

arXiv2024

Towards Diverse Evaluation of Class Incremental Learning: A Representation Learning Perspective

Class incremental learning (CIL) algorithms aim to continually learn new object classes from incrementally arriving data while not forgetting past learned classes. The common evaluation protocol for CIL algorithms is to measure the average test accuracy across all classes learned so far -- however, we argue that solely focusing on maximizing the test accuracy may not necessarily lead to developing a CIL algorithm that also continually learns and updates the representations, which may be transferred to the downstream tasks. To that end, we experimentally analyze neural network models trained by CIL algorithms using various evaluation protocols in representation learning and propose new analysis methods. Our experiments show that most state-of-the-art algorithms prioritize high stability and do not significantly change the learned representation, and sometimes even learn a representation of lower quality than a naive baseline. However, we observe that these algorithms can still achieve high test accuracy because they enable a model to learn a classifier that closely resembles an estimated linear classifier trained for linear probing. Furthermore, the base model learned in the first task, which involves single-task learning, exhibits varying levels of representation quality across different algorithms, and this variance impacts the final performance of CIL algorithms. Therefore, we suggest that the representation-level evaluation should be considered as an additional recipe for more diverse evaluation for CIL algorithms.

cs.LG↗

arXiv2024

HPC resources for CMS offline computing: An integration and scalability challenge for the Submission Infrastructure

The computing resource needs of LHC experiments are expected to continue growing significantly during the Run 3 and into the HL-LHC era. The landscape of available resources will also evolve, as High Performance Computing (HPC) and Cloud resources will provide a comparable, or even dominant, fraction of the total compute capacity. The future years present a challenge for the experiments' resource provisioning models, both in terms of scalability and increasing complexity. The CMS Submission Infrastructure (SI) provisions computing resources for CMS workflows. This infrastructure is built on a set of federated HTCondor pools, currently aggregating 400k CPU cores distributed worldwide and supporting the simultaneous execution of over 200k computing tasks. Incorporating HPC resources into CMS computing represents firstly an integration challenge, as HPC centers are much more diverse compared to Grid sites. Secondly, evolving the present SI, dimensioned to harness the current CMS computing capacity, to reach the resource scales required for the HLLHC phase, while maintaining global flexibility and efficiency, will represent an additional challenge for the SI. To preventively address future potential scalability limits, the SI team regularly runs tests to explore the maximum reach of our infrastructure. In this note, the integration of HPC resources into CMS offline computing is summarized, the potential concerns for the SI derived from the increased scale of operations are described, and the most recent results of scalability test on the CMS SI are reported.

cs.DC↗