arXiv subjects

Sundong Kim

Publications and source records attributed to Sundong Kim.

At least 19 recordsLinked to original sources

arXiv2026

Cost-efficient Active Learning for Referring Image Segmentation and Grounding

Collecting natural-language referring expressions along with region annotations, such as masks or boxes, is a major bottleneck in visual grounding (VG), as annotators must write descriptions that distinguish target regions from visually similar ones. We tackle this by formulating active learning (AL) for VG under the realistic setting where only raw images are available without accompanying text. Since ground-truth text is unavailable, sample selection must estimate which images contain ambiguous regions that would require discriminative referring expressions. To address this, we generate auxiliary region-text pairs using foundation models, and introduce Referred Region Ambiguity, a new acquisition function that measures whether the model's confidence collapses onto a single region or disperses across multiple candidates. It allows our method to prioritize images with strong cross-region competition, which are more informative due to their visual ambiguity. We also design a referring-expression annotation interface that helps annotators quickly focus on writing discriminative language with a few clicks. Experiments on RIS and REC benchmarks show that our AL framework consistently outperforms several AL baselines, while a user study shows up to 1.6X faster description labeling of ours.

cs.CV↗

arXiv2026

From Noise to Diversity: Random Embedding Injection in LLM Reasoning

Recent soft prompt research has tried to improve reasoning by inserting trained vectors into LLM inputs, yet whether the gain comes from the learned content or from the act of injection itself has not been carefully separated. We study Random Soft Prompts (RSPs), which drop the training step entirely and append a freshly drawn sequence of random embedding vectors to the input. Each RSP vector is sampled from an isotropic Gaussian fitted to the entrywise mean and variance of the pretrained embedding table; the sequence carries no learned content, and yet reaches accuracy comparable to optimized soft prompts on math reasoning benchmarks in several settings. The mechanism unfolds in two stages: because attention has to absorb a never-seen-before random position, the distribution over the first few generated tokens flattens and reasoning trajectories branch, and as generation continues this influence dilutes naturally so the response commits to a single completion. We show that during inference RSPs lift early-stage token diversity and, combined with temperature sampling, widen Pass@N, the probability that at least one out of N attempts is correct. Beyond inference, we carry the same effect into DAPO training and demonstrate practical gains. Our contributions are: (i) RSP isolates the simplest form of soft prompt -- training-free, freshly resampled -- providing a unified lens for the structural effect of injection that variants otherwise differing in training and form all share; (ii) a theoretical and empirical validation of the underlying mechanism; and (iii) an extension from inference to training.

cs.AI↗

arXiv2025

ARCTraj: A Dataset and Benchmark of Human Reasoning Trajectories for Abstract Problem Solving

We present ARCTraj, a dataset and methodological framework for modeling human reasoning through complex visual tasks in the Abstraction and Reasoning Corpus (ARC). While ARC has inspired extensive research on abstract reasoning, most existing approaches rely on static input-output supervision, which limits insight into how reasoning unfolds over time. ARCTraj addresses this gap by recording temporally ordered, object-level actions that capture how humans iteratively transform inputs into outputs, revealing intermediate reasoning steps that conventional datasets overlook. Collected via the O2ARC web interface, it contains around 10,000 trajectories annotated with task identifiers, timestamps, and success labels across 400 training tasks from the ARC-AGI-1 benchmark. It further defines a unified reasoning pipeline encompassing data collection, action abstraction, Markov decision process (MDP) formulation, and downstream learning, enabling integration with reinforcement learning, generative modeling, and sequence modeling methods such as PPO, World Models, GFlowNets, Diffusion agents, and Decision Transformers. Analyses of spatial selection, color attribution, and strategic convergence highlight the structure and diversity of human reasoning. Together, these contributions position ARCTraj as a structured and interpretable foundation for studying human-like reasoning, advancing explainability, alignment, and generalizable intelligence.

cs.AI↗

arXiv2025

Can Large Language Models Develop Gambling Addiction?

This study identifies the specific conditions under which large language models exhibit human-like gambling addiction patterns, providing critical insights into their decision-making mechanisms and AI safety. We analyze LLM decision-making at cognitive-behavioral and neural levels based on human addiction research. In slot machine experiments, we identified cognitive features such as illusion of control and loss chasing, observing that greater autonomy in betting parameters substantially amplified irrational behavior and bankruptcy rates. Neural circuit analysis using a Sparse Autoencoder confirmed that model behavior is controlled by abstract decision-making features related to risk, not merely by prompts. These findings suggest LLMs internalize human-like cognitive biases beyond simply mimicking training data.

cs.AI↗

arXiv2025

The Othello AI Arena: Evaluating Intelligent Systems Through Limited-Time Adaptation to Unseen Boards

The ability to rapidly adapt to novel and unforeseen environmental changes is a cornerstone of artificial general intelligence (AGI), yet it remains a critical blind spot in most existing AI benchmarks. Traditional evaluation largely focuses on optimizing performance within fixed environments, failing to assess systems' flexibility and generalization capabilities when faced with even subtle rule or structural modifications. Addressing this gap, I introduce the Othello AI Arena, a novel benchmark framework designed to evaluate intelligent systems based on their capacity for limited-time adaptation to unseen environments. Our platform poses a meta-learning challenge: participants must develop systems that can analyze the specific configuration and rules of a novel Othello board within a strict time limit (60 seconds) and generate a tailored, high-performing strategy for that unique environment. With this, evaluation of the meta-level intelligence can be separated from the task-level strategy performance. The Arena features a diverse set of game stages, including public stages for development and private stages with structural and rule variations designed to test genuine adaptive and generalization capabilities. Implemented as an accessible web-based platform, the Arena provides real-time visualization, automated evaluation using multi-dimensional metrics, and comprehensive logging for post-hoc analysis. Initial observations from pilot tests and preliminary student engagements highlight fascinating patterns in adaptation approaches, ranging from rapid parameter tuning to rudimentary environmental model learning through simulation. The Othello AI Arena offers a unique educational tool and a valuable research benchmark for fostering and evaluating the crucial skill of rapid, intelligent adaptation in AI systems.

cs.AI↗

arXiv2025

TRACED: Transition-aware Regret Approximation with Co-learnability for Environment Design

Generalizing deep reinforcement learning agents to unseen environments remains a significant challenge. One promising solution is Unsupervised Environment Design (UED), a co-evolutionary framework in which a teacher adaptively generates tasks with high learning potential, while a student learns a robust policy from this evolving curriculum. Existing UED methods typically measure learning potential via regret, the gap between optimal and current performance, approximated solely by value-function loss. Building on these approaches, we introduce the transition-prediction error as an additional term in our regret approximation. To capture how training on one task affects performance on others, we further propose a lightweight metric called Co-Learnability. By combining these two measures, we present Transition-aware Regret Approximation with Co-learnability for Environment Design (TRACED). Empirical evaluations show that TRACED produces curricula that improve zero-shot generalization over strong baselines across multiple benchmarks. Ablation studies confirm that the transition-prediction error drives rapid complexity ramp-up and that Co-Learnability delivers additional gains when paired with the transition-prediction error. These results demonstrate how refined regret approximation and explicit modeling of task relationships can be leveraged for sample-efficient curriculum design in UED. Project Page: https://geonwoo.me/traced/

cs.LG↗

arXiv2025

Causal-Paced Deep Reinforcement Learning

Designing effective task sequences is crucial for curriculum reinforcement learning (CRL), where agents must gradually acquire skills by training on intermediate tasks. A key challenge in CRL is to identify tasks that promote exploration, yet are similar enough to support effective transfer. While recent approach suggests comparing tasks via their Structural Causal Models (SCMs), the method requires access to ground-truth causal structures, an unrealistic assumption in most RL settings. In this work, we propose Causal-Paced Deep Reinforcement Learning (CP-DRL), a curriculum learning framework aware of SCM differences between tasks based on interaction data approximation. This signal captures task novelty, which we combine with the agent's learnability, measured by reward gain, to form a unified objective. Empirically, CP-DRL outperforms existing curriculum methods on the Point Mass benchmark, achieving faster convergence and higher returns. CP-DRL demonstrates reduced variance with comparable final returns in the Bipedal Walker-Trivial setting, and achieves the highest average performance in the Infeasible variant. These results indicate that leveraging causal relationships between tasks can improve the structure-awareness and sample efficiency of curriculum reinforcement learning. We provide the full implementation of CP-DRL to facilitate the reproduction of our main results at https://github.com/Cho-Geonwoo/CP-DRL.

cs.LG↗

arXiv2025

AMPED: Adaptive Multi-objective Projection for balancing Exploration and skill Diversification

Skill-based reinforcement learning (SBRL) enables rapid adaptation in environments with sparse rewards by pretraining a skill-conditioned policy. Effective skill learning requires jointly maximizing both exploration and skill diversity. However, existing methods often face challenges in simultaneously optimizing for these two conflicting objectives. In this work, we propose a new method, Adaptive Multi-objective Projection for balancing Exploration and skill Diversification (AMPED), which explicitly addresses both: during pre-training, a gradient-surgery projection balances the exploration and diversity gradients, and during fine-tuning, a skill selector exploits the learned diversity by choosing skills suited to downstream tasks. Our approach achieves performance that surpasses SBRL baselines across various benchmarks. Through an extensive ablation study, we identify the role of each component and demonstrate that each element in AMPED is contributing to performance. We further provide theoretical and empirical evidence that, with a greedy skill selector, greater skill diversity reduces fine-tuning sample complexity. These results highlight the importance of explicitly harmonizing exploration and diversity and demonstrate the effectiveness of AMPED in enabling robust and generalizable skill learning. Project Page: https://geonwoo.me/amped/

cs.LG↗

arXiv2025

GIFARC: Synthetic Dataset for Leveraging Human-Intuitive Analogies to Elevate AI Reasoning

The Abstraction and Reasoning Corpus (ARC) poses a stringent test of general AI capabilities, requiring solvers to infer abstract patterns from only a handful of examples. Despite substantial progress in deep learning, state-of-the-art models still achieve accuracy rates of merely 40-55% on 2024 ARC Competition, indicative of a significant gap between their performance and human-level reasoning. In this work, we seek to bridge that gap by introducing an analogy-inspired ARC dataset, GIFARC. Leveraging large language models (LLMs) and vision-language models (VLMs), we synthesize new ARC-style tasks from a variety of GIF images that include analogies. Each new task is paired with ground-truth analogy, providing an explicit mapping between visual transformations and everyday concepts. By embedding robust human-intuitive analogies into ARC-style tasks, GIFARC guides AI agents to evaluate the task analogically before engaging in brute-force pattern search, thus efficiently reducing problem complexity and build a more concise and human-understandable solution. We empirically validate that guiding LLM with analogic approach with GIFARC affects task-solving approaches of LLMs to align with analogic approach of human.

cs.AI↗

arXiv2024

Abductive Symbolic Solver on Abstraction and Reasoning Corpus

This paper addresses the challenge of enhancing artificial intelligence reasoning capabilities, focusing on logicality within the Abstraction and Reasoning Corpus (ARC). Humans solve such visual reasoning tasks based on their observations and hypotheses, and they can explain their solutions with a proper reason. However, many previous approaches focused only on the grid transition and it is not enough for AI to provide reasonable and human-like solutions. By considering the human process of solving visual reasoning tasks, we have concluded that the thinking process is likely the abductive reasoning process. Thus, we propose a novel framework that symbolically represents the observed data into a knowledge graph and extracts core knowledge that can be used for solution generation. This information limits the solution search space and helps provide a reasonable mid-process. Our approach holds promise for improving AI performance on ARC tasks by effectively narrowing the solution space and providing logical solutions grounded in core knowledge extraction.

cs.AI↗

arXiv2024

Diffusion-Based Offline RL for Improved Decision-Making in Augmented ARC Task

Effective long-term strategies enable AI systems to navigate complex environments by making sequential decisions over extended horizons. Similarly, reinforcement learning (RL) agents optimize decisions across sequences to maximize rewards, even without immediate feedback. To verify that Latent Diffusion-Constrained Q-learning (LDCQ), a prominent diffusion-based offline RL method, demonstrates strong reasoning abilities in multi-step decision-making, we aimed to evaluate its performance on the Abstraction and Reasoning Corpus (ARC). However, applying offline RL methodologies to enhance strategic reasoning in AI for solving tasks in ARC is challenging due to the lack of sufficient experience data in the ARC training set. To address this limitation, we introduce an augmented offline RL dataset for ARC, called Synthesized Offline Learning Data for Abstraction and Reasoning (SOLAR), along with the SOLAR-Generator, which generates diverse trajectory data based on predefined rules. SOLAR enables the application of offline RL methods by offering sufficient experience data. We synthesized SOLAR for a simple task and used it to train an agent with the LDCQ method. Our experiments demonstrate the effectiveness of the offline RL approach on a simple ARC task, showing the agent's ability to make multi-step sequential decisions and correctly identify answer states. These results highlight the potential of the offline RL approach to enhance AI's strategic reasoning capabilities.

cs.AI↗

arXiv2024

DIAR: Diffusion-model-guided Implicit Q-learning with Adaptive Revaluation

We propose a novel offline reinforcement learning (offline RL) approach, introducing the Diffusion-model-guided Implicit Q-learning with Adaptive Revaluation (DIAR) framework. We address two key challenges in offline RL: out-of-distribution samples and long-horizon problems. We leverage diffusion models to learn state-action sequence distributions and incorporate value functions for more balanced and adaptive decision-making. DIAR introduces an Adaptive Revaluation mechanism that dynamically adjusts decision lengths by comparing current and future state values, enabling flexible long-term decision-making. Furthermore, we address Q-value overestimation by combining Q-network learning with a value function guided by a diffusion model. The diffusion model generates diverse latent trajectories, enhancing policy robustness and generalization. As demonstrated in tasks like Maze2D, AntMaze, and Kitchen, DIAR consistently outperforms state-of-the-art algorithms in long-horizon, sparse-reward environments.

cs.LG↗

arXiv2024

System 2 Reasoning for Human-AI Alignment: Generality and Adaptivity via ARC-AGI

Despite their broad applicability, transformer-based models still fall short in System~2 reasoning, lacking the generality and adaptivity needed for human--AI alignment. We examine weaknesses on ARC-AGI tasks, revealing gaps in compositional generalization and novel-rule adaptation, and argue that closing these gaps requires overhauling the reasoning pipeline and its evaluation. We propose three research axes: (1) Symbolic representation pipeline for compositional generality, (2) Interactive feedback-driven reasoning loop for adaptivity, and (3) Test-time task augmentation balancing both qualities. Finally, we demonstrate how ARC-AGI's evaluation suite can be adapted to track progress in symbolic generality, feedback-driven adaptivity, and task-level robustness, thereby guiding future work on robust human--AI alignment.

cs.AI↗

arXiv2024

Addressing and Visualizing Misalignments in Human Task-Solving Trajectories

Understanding misalignments in human task-solving trajectories is crucial for enhancing AI models trained to closely mimic human reasoning. This study categorizes such misalignments into three types: (1) lack of functions to express intent, (2) inefficient action sequences, and (3) incorrect intentions that cannot solve the task. To address these issues, we first formalize and define these three misalignment types in a unified framework. We then propose a heuristic algorithm to detect misalignments in ARCTraj trajectories and analyze their impact hierarchically and quantitatively. We also present an intention estimation method based on our formalism that infers missing alignment between user actions and intentions. Through trajectory alignment, we experimentally demonstrate that AI models trained on human task-solving trajectories improve performance in mimicking human reasoning. Based on hierarchical analysis and experiments, we highlight the importance of trajectory-intention alignment and demonstrate the effectiveness of intention-aligned training.

cs.AI↗

arXiv2024

Enhancing Analogical Reasoning in the Abstraction and Reasoning Corpus via Model-Based RL

This paper demonstrates that model-based reinforcement learning (model-based RL) is a suitable approach for the task of analogical reasoning. We hypothesize that model-based RL can solve analogical reasoning tasks more efficiently through the creation of internal models. To test this, we compared DreamerV3, a model-based RL method, with Proximal Policy Optimization, a model-free RL method, on the Abstraction and Reasoning Corpus (ARC) tasks. Our results indicate that model-based RL not only outperforms model-free RL in learning and generalizing from single tasks but also shows significant advantages in reasoning across similar tasks.

cs.AI↗

arXiv2024

ARCLE: The Abstraction and Reasoning Corpus Learning Environment for Reinforcement Learning

This paper introduces ARCLE, an environment designed to facilitate reinforcement learning research on the Abstraction and Reasoning Corpus (ARC). Addressing this inductive reasoning benchmark with reinforcement learning presents these challenges: a vast action space, a hard-to-reach goal, and a variety of tasks. We demonstrate that an agent with proximal policy optimization can learn individual tasks through ARCLE. The adoption of non-factorial policies and auxiliary losses led to performance enhancements, effectively mitigating issues associated with action spaces and goal attainment. Based on these insights, we propose several research directions and motivations for using ARCLE, including MAML, GFlowNets, and World Models.

cs.AI↗

arXiv2024

Reasoning Abilities of Large Language Models: In-Depth Analysis on the Abstraction and Reasoning Corpus

The existing methods for evaluating the inference abilities of Large Language Models (LLMs) have been predominantly results-centric, making it challenging to assess the inference process comprehensively. We introduce a novel approach using the Abstraction and Reasoning Corpus (ARC) benchmark to evaluate the inference and contextual understanding abilities of LLMs in a process-centric manner, focusing on three key components from the Language of Thought Hypothesis (LoTH): Logical Coherence, Compositionality, and Productivity. Our carefully designed experiments reveal that while LLMs demonstrate some inference capabilities, they still significantly lag behind human-level reasoning in these three aspects. The main contribution of this paper lies in introducing the LoTH perspective, which provides a method for evaluating the reasoning process that conventional results-oriented approaches fail to capture, thereby offering new insights into the development of human-level reasoning in artificial intelligence systems.

cs.CL↗

arXiv2023

Explainable Product Classification for Customs

The task of assigning internationally accepted commodity codes (aka HS codes) to traded goods is a critical function of customs offices. Like court decisions made by judges, this task follows the doctrine of precedent and can be nontrivial even for experienced officers. Together with the Korea Customs Service (KCS), we propose a first-ever explainable decision supporting model that suggests the most likely subheadings (i.e., the first six digits) of the HS code. The model also provides reasoning for its suggestion in the form of a document that is interpretable by customs officers. We evaluated the model using 5,000 cases that recently received a classification request. The results showed that the top-3 suggestions made by our model had an accuracy of 93.9\% when classifying 925 challenging subheadings. A user study with 32 customs experts further confirmed that our algorithmic suggestions accompanied by explainable reasonings, can substantially reduce the time and effort taken by customs officers for classification reviews.

cs.AI↗