arXiv subjects

Jungsoo Park

Publications and source records attributed to Jungsoo Park.

15 recordsLinked to original sources

arXiv2026

Who Pays More for Safety? Measuring the Disparate Cost of Safety Alignment across Languages

Safety alignment helps models adhere to human values, but it often reduces response utility. We ask a critical but understudied question: Does safety alignment impose the cost equally across language groups? To answer this, we introduce a rigorous protocol to measure the utility loss imposed solely by safety alignment, which we term Safety Cost. Through direct pairwise comparisons between safety-aligned models and their unaligned counterparts, we find a systematic inequity: non-English users consistently bear a higher Safety Cost than English users. We further identify three underlying patterns. First, multiple languages lie in a double-penalty zone, experiencing both weaker safety protection and larger utility loss. Second, certain languages exhibit apparent utility gains that are in fact a consequence of safety filters failing to engage. Third, even high-resource languages pay a larger Safety Cost than English to reach the same level of safety. We show that these disparities arise from both explicit refusals and implicit qualitative differences across multiple dimensions. By accurately measuring the disparate effects of safety alignment, our findings expose a systematic disparity in current safety alignment practices.

cs.CL↗

arXiv2026

Soft Token Alignment for Cross-Lingual Reasoning

Multilingual large language models often produce inconsistent reasoning and answers for semantically equivalent prompts in different languages. Prior work suggests that intermediate representations can be relatively language-agnostic, but generation becomes increasingly language-specific as models commit to discrete output tokens. This is problematic because language-specific lexical choices can cause semantically equivalent reasoning paths to diverge across languages. These divergences motivate searching for a cross-lingual alignment signal that is less tied to any single vocabulary item or script. We propose SOLAR, an auxiliary objective for supervised fine-tuning that aligns soft-token representations across languages, using English as a pivot. Soft tokens are probability-weighted mixtures over the vocabulary embeddings, yielding continuous representations that can aggregate information from semantically related tokens across languages. We then align each non-English soft-token summary to its English counterpart in the shared embedding space. Across four multilingual reasoning benchmarks, SOLAR improves accuracy by up to +17.7 points over the base model and +3.8 over standard supervised fine-tuning, with the largest gains on low-resource languages. SOLAR also strengthens final-layer cross-lingual similarity and substantially reduces language-cluster separability, suggesting that aligning soft-token representations helps preserve shared semantic structure during multilingual reasoning.

cs.CL↗

arXiv2026

Distribution-Aware Reward: Reinforcement Learning over Predictive Distributions for LLM Regression

Large language models can predict real-valued quantities from heterogeneous inputs such as text, code, and molecular strings, but most training objectives score each decoded floating-point number independently, improving point estimates without ensuring calibrated predictive distributions. This limits applications requiring candidate ranking or uncertainty estimation. We introduce Distribution-Aware Reward, an on-policy reinforcement learning objective whose main contribution is to train language models to produce better predictive distributions for regression tasks, rather than only optimizing individual decoded outputs against scalar targets. Our method treats multiple decoded samples as an empirical predictive distribution, evaluates it with the Continuous Ranked Probability Score, and assigns leave-one-out credit based on each rollout's marginal contribution to distribution quality, rewarding predictions that are both accurate and appropriately dispersed. We evaluate our method on a controlled Gaussian-mixture task, code performance prediction, and molecular property prediction from SMILES strings. Across tasks, our method improves over supervised fine-tuning and pointwise reinforcement learning baselines, with strong rank-correlation gains, including a 6-point Spearman improvement on KBSS. On MoleculeNet, it uses only SMILES strings yet remains competitive with strong graph-based and 3D molecular models. Further analyses show that our method mitigates rollout diversity collapse and improves uncertainty diagnostics, suggesting that directly optimizing predictive distributions makes language model regression more robust and better calibrated.

cs.LG↗

arXiv2026

Safe and Scalable Web Agent Learning via Recreated Websites

Training autonomous web agents is fundamentally limited by the environments they learn from: real-world websites are unsafe to explore, hard to reset, and rarely provide verifiable feedback. We propose VeriEnv, a framework that treats language models as environment creators, automatically cloning real-world websites into fully executable, verifiable synthetic environments. By exposing controlled internal access via a Python SDK, VeriEnv enables agents to self-generate tasks with deterministic, programmatically verifiable rewards, eliminating reliance on heuristic or LLM-based judges. This design decouples agent learning from unsafe real-world interaction while enabling scalable self-evolution through environment expansion. Through experiments on web agent benchmarks, we show that agents trained with VeriEnv generalize to unseen websites, achieve site-specific mastery through self-evolving training, and benefit from scaling the number of training environments. Code and resources will be released at https://github.com/kyle8581/VeriEnv upon acceptance.

cs.CL↗

arXiv2026

Anticipatory Evaluation of Language Models

Progress in large language models is increasingly constrained by an evaluation bottleneck: benchmarks must be built and models run before iteration can begin. We investigate whether evaluation outcomes can be forecast before any experiments are conducted. Specifically, we study text-only performance prediction, where models estimate performance from task descriptions and experimental configurations alone, without access to dataset instances. To support systematic study, we curate PRECOG, a corpus of description-performance pairs spanning diverse tasks, domains, and metrics. We scrape task and configuration descriptions from arXiv, yielding 2,290 instances covering 1,519 papers, and construct a test split using papers published after the evaluated models' knowledge cutoff. Experiments show the task is challenging but feasible: reasoning models achieve a non-trivial forecasting skill reaching mean absolute error as low as 9.9 at high-confidence thresholds. Overall, our corpus and analyses offer an initial step toward open-ended anticipatory evaluation, supporting difficulty estimation and smarter resource allocation.

cs.CL↗

arXiv2026

Making Expert Reasoning Learnable with Self-Distillation

Improving the reasoning capabilities of large language models (LLMs) typically relies either on the model's ability to sample a correct solution to be reinforced or the existence of a stronger model able to solve the problem. However, many difficult problems remain intractable for even current frontier models, preventing the extraction of valid training signals. A promising alternative is to leverage high-quality expert human solutions, yet naive imitation of this data fails because it is fundamentally out-of-distribution: expert solutions are typically didactic, containing implicit reasoning gaps intended for human readers rather than computational models. Furthermore, high-quality expert solutions are expensive, necessitating generalizable, sample-efficient training methods. We propose Distribution Aligned Imitation Learning (DAIL), a two-step self-distillation method that bridges the distributional gap by first transforming expert solutions into detailed, in-distribution reasoning traces and then applying a contrastive objective to focus learning on expert insights and methodologies. We find that DAIL can leverage fewer than 1000 high-quality expert solutions to achieve up to 31% pass@128 gains on Qwen2.5-Instruct and Qwen3, double reasoning efficiency, and enable out-of-domain generalization.

cs.LG↗

arXiv2025

Data Transformation Strategies to Remove Heterogeneity

Data heterogeneity is a prevalent issue, stemming from various conflicting factors, making its utilization complex. This uncertainty, particularly resulting from disparities in data formats, frequently necessitates the involvement of experts to find resolutions. Current methodologies primarily address conflicts related to data structures and schemas, often overlooking the pivotal role played by data transformation. As the utilization of artificial intelligence (AI) continues to expand, there is a growing demand for a more streamlined data preparation process, and data transformation becomes paramount. It customizes training data to enhance AI learning efficiency and adapts input formats to suit diverse AI models. Selecting an appropriate transformation technique is paramount in preserving crucial data details. Despite the widespread integration of AI across various industries, comprehensive reviews concerning contemporary data transformation approaches are scarce. This survey explores the intricacies of data heterogeneity and its underlying sources. It systematically categorizes and presents strategies to address heterogeneity stemming from differences in data formats, shedding light on the inherent challenges associated with each strategy.

cs.LG↗

arXiv2025

Can LLMs Help Uncover Insights about LLMs? A Large-Scale, Evolving Literature Analysis of Frontier LLMs

The surge of LLM studies makes synthesizing their findings challenging. Analysis of experimental results from literature can uncover important trends across studies, but the time-consuming nature of manual data extraction limits its use. Our study presents a semi-automated approach for literature analysis that accelerates data extraction using LLMs. It automatically identifies relevant arXiv papers, extracts experimental results and related attributes, and organizes them into a structured dataset, LLMEvalDB. We then conduct an automated literature analysis of frontier LLMs, reducing the effort of paper surveying and data extraction by more than 93% compared to manual approaches. We validate LLMEvalDB by showing that it reproduces key findings from a recent manual analysis of Chain-of-Thought (CoT) reasoning and also uncovers new insights that go beyond it, showing, for example, that in-context examples benefit coding & multimodal tasks but offer limited gains in math reasoning tasks compared to zero-shot CoT. Our automatically updatable dataset enables continuous tracking of target models by extracting evaluation studies as new data becomes available. Through LLMEvalDB and empirical analysis, we provide insights into LLMs while facilitating ongoing literature analyses of their behavior.

cs.CL↗

arXiv2025

Design and Development of a Locomotion Interface for Virtual Reality Lower-Body Haptic Interaction

This work presents the design, build, control, and preliminary user data of a locomotion interface called ForceBot. It delivers lower-body haptic interaction in virtual reality (VR), enabling users to walk in VR while interacting with various simulated terrains. It utilizes two planar gantries to give each foot two degrees of freedom and passive heel-lifting motion. The design used motion capture data with dynamic simulation for ergonomic human-robot workspace and hardware selection. Its system framework uses open-source robotic software and pairs with a custom-built power delivery system that offers EtherCAT communication with a 1,000 Hz soft real-time computation rate. This system features an admittance controller to regulate physical human-robot interaction (pHRI) alongside a walking algorithm to generate walking motion and simulate virtual terrains. The system's performance is explored through three measurements that evaluate the relationship between user input force and output pHRI motion. Overall, this platform presents a unique approach by utilizing planar gantries to realize VR terrain interaction with an extensive workspace, reasonably compact footprint, and preliminary user data.

cs.RO↗

arXiv2023

Optimizing Test-Time Query Representations for Dense Retrieval

Recent developments of dense retrieval rely on quality representations of queries and contexts from pre-trained query and context encoders. In this paper, we introduce TOUR (Test-Time Optimization of Query Representations), which further optimizes instance-level query representations guided by signals from test-time retrieval results. We leverage a cross-encoder re-ranker to provide fine-grained pseudo labels over retrieval results and iteratively optimize query representations with gradient descent. Our theoretical analysis reveals that TOUR can be viewed as a generalization of the classical Rocchio algorithm for pseudo relevance feedback, and we present two variants that leverage pseudo-labels as hard binary or soft continuous labels. We first apply TOUR on phrase retrieval with our proposed phrase re-ranker, and also evaluate its effectiveness on passage retrieval with an off-the-shelf re-ranker. TOUR greatly improves end-to-end open-domain question answering accuracy, as well as passage retrieval performance. TOUR also consistently improves direct re-ranking by up to 2.0% while running 1.3-2.4x faster with an efficient implementation.

cs.CL↗

arXiv2023

Empowering Sentence Encoders with Prompting and Label Retrieval for Zero-shot Text Classification

With contrastive pre-training, sentence encoders are generally optimized to locate semantically similar samples closer to each other in their embedding spaces. In this work, we focus on the potential of their embedding spaces to be readily adapted to zero-shot text classification, as semantically distinct samples are already well-separated. Our framework, RaLP (Retrieval augmented Label Prompts for sentence encoder), encodes prompted label candidates with a sentence encoder, then assigns the label whose prompt embedding has the highest similarity with the input text embedding. In order to compensate for the potentially poorly descriptive labels in their original format, RaLP retrieves sentences that are semantically similar to the original label prompt from external corpora and use them as additional pseudo-label prompts. RaLP achieves competitive or stronger performance than much larger baselines on various closed-set classification and multiple-choice QA datasets under zero-shot settings. We show that the retrieval component plays a pivotal role in RaLP's success, and its results are robustly attained regardless of verbalizer variations.

cs.CL↗

arXiv2022

Consistency Training with Virtual Adversarial Discrete Perturbation

Consistency training regularizes a model by enforcing predictions of original and perturbed inputs to be similar. Previous studies have proposed various augmentation methods for the perturbation but are limited in that they are agnostic to the training model. Thus, the perturbed samples may not aid in regularization due to their ease of classification from the model. In this context, we propose an augmentation method of adding a discrete noise that would incur the highest divergence between predictions. This virtual adversarial discrete noise obtained by replacing a small portion of tokens while keeping original semantics as much as possible efficiently pushes a training model's decision boundary. Experimental results show that our proposed method outperforms other consistency training baselines with text editing, paraphrasing, or a continuous noise on semi-supervised text classification tasks and a robustness benchmark

cs.CL↗

arXiv2022

FaVIQ: FAct Verification from Information-seeking Questions

Despite significant interest in developing general purpose fact checking models, it is challenging to construct a large-scale fact verification dataset with realistic real-world claims. Existing claims are either authored by crowdworkers, thereby introducing subtle biases that are difficult to control for, or manually verified by professional fact checkers, causing them to be expensive and limited in scale. In this paper, we construct a large-scale challenging fact verification dataset called FAVIQ, consisting of 188k claims derived from an existing corpus of ambiguous information-seeking questions. The ambiguities in the questions enable automatically constructing true and false claims that reflect user confusions (e.g., the year of the movie being filmed vs. being released). Claims in FAVIQ are verified to be natural, contain little lexical bias, and require a complete understanding of the evidence for verification. Our experiments show that the state-of-the-art models are far from solving our new task. Moreover, training on our data helps in professional fact-checking, outperforming models trained on the widely used dataset FEVER or in-domain data by up to 17% absolute. Altogether, our data will serve as a challenging benchmark for natural language understanding and support future progress in professional fact checking.

cs.CL↗

arXiv2021

Learn to Resolve Conversational Dependency: A Consistency Training Framework for Conversational Question Answering

One of the main challenges in conversational question answering (CQA) is to resolve the conversational dependency, such as anaphora and ellipsis. However, existing approaches do not explicitly train QA models on how to resolve the dependency, and thus these models are limited in understanding human dialogues. In this paper, we propose a novel framework, ExCorD (Explicit guidance on how to resolve Conversational Dependency) to enhance the abilities of QA models in comprehending conversational context. ExCorD first generates self-contained questions that can be understood without the conversation history, then trains a QA model with the pairs of original and self-contained questions using a consistency-based regularizer. In our experiments, we demonstrate that ExCorD significantly improves the QA models' performance by up to 1.2 F1 on QuAC, and 5.2 F1 on CANARD, while addressing the limitations of the existing approaches.

cs.CL↗

arXiv2020

Adversarial Subword Regularization for Robust Neural Machine Translation

Exposing diverse subword segmentations to neural machine translation (NMT) models often improves the robustness of machine translation as NMT models can experience various subword candidates. However, the diversification of subword segmentations mostly relies on the pre-trained subword language models from which erroneous segmentations of unseen words are less likely to be sampled. In this paper, we present adversarial subword regularization (ADVSR) to study whether gradient signals during training can be a substitute criterion for exposing diverse subword segmentations. We experimentally show that our model-based adversarial samples effectively encourage NMT models to be less sensitive to segmentation errors and improve the performance of NMT models in low-resource and out-domain datasets.

cs.CL↗