Junseok Kim

Dohyeong Kim, Hyeokjin Kwon, Junseok Kim et al.

As the complexity of tasks addressed through reinforcement learning (RL) increases, the definition of reward functions also has become highly complicated. We introduce an RL method aimed at simplifying the reward-shaping process through intuitive strategies. Initially, instead of a single reward function composed of various terms, we define multiple reward and cost functions within a constrained multi-objective RL (CMORL) framework. For tasks involving sequential complex movements, we segment the task into distinct stages and define multiple rewards and costs for each stage. Finally, we introduce a practical CMORL algorithm that maximizes objectives based on these rewards while satisfying constraints defined by the costs. The proposed method has been successfully demonstrated across a variety of acrobatic tasks in both simulation and real-world environments. Additionally, it has been shown to successfully perform tasks compared to existing RL and constrained RL algorithms. Our code is available at https://github.com/rllab-snu/Stage-Wise-CMORL.

Junseok Kim, Nakyeong Yang, Kyomin Jung

Recent studies demonstrate that prompting a role-playing persona to an LLM improves reasoning capability. However, assigning an adequate persona is difficult since LLMs are extremely sensitive to assigned prompts; thus, inaccurately defined personas sometimes hinder LLMs and degrade their reasoning capabilities. In this paper, we first investigate the potential negative impact of injecting persona into language models. Furthermore, we propose a novel framework, Jekyll \& Hyde, which ensembles the outcomes of both role-playing and neutral prompts to enhance the robustness of reasoning ability. Specifically, Jekyll \& Hyde predicts an appropriate persona using an LLM when defining the role-playing prompt. Then, Jekyll \& Hyde collects two potential solutions from role-playing and neutral prompts and selects a better solution using the LLM evaluator. The experimental analysis demonstrates that role-playing prompts sometimes distract LLMs, degrading their reasoning abilities in 7 out of 12 datasets in llama3. Meanwhile, Jekyll \& Hyde improve reasoning capabilities by selecting better choices among the potential solutions on twelve widely-used natural language reasoning datasets. In addition, we reveal that assigning LLM-generated personas obtains more stable results than handcrafted personas.

Junseok Kim, Dohyeong Kim, Mineui Hong et al.

Compositional generalization is essential for reaching unseen goals under novel contextual variations in offline goal-conditioned reinforcement learning (GCRL), where a generalist goal-reaching agent must be learned from limited data. Most prior approaches pursue this via trajectory stitching over temporally contiguous segments, which limits composing behaviors across varying contexts. To overcome this limitation, we formalize analogy transduction as synthesizing new plans by composing task-endogenous analogies with given contexts and propose a novel analogy representation tailored for it. Grounded in our theory, this analogy representation captures what changes under optimal task execution, remains invariant to contextual variations, and is sufficient for optimal goal reaching. We further contend that generalization to unseen analogy-context pairs is a practical obstacle in analogy transduction, and introduce a new approach for offline GCRL that enables analogy transduction beyond seen pairs to unseen combinations. We empirically demonstrate the effectiveness of our approach on OGBench manipulation environments, substantially outperforming prior methods that do not perform analogy transduction. Project page: https://rllab-snu.github.io/projects/CTA/

Junseok Kim, Nakyeong Yang, Kyungmin Min et al.

Self-Consistency improves reasoning reliability through multi-sample aggregation, but incurs substantial inference cost. Adaptive self-consistency methods mitigate this issue by adjusting the sampling budget; however, they rely on count-based stopping rules that treat all responses equally, often leading to unnecessary sampling. We propose Reliability-Aware Adaptive Self-Consistency (ReASC), which addresses this limitation by reframing adaptive sampling from response counting to evidence sufficiency, leveraging response-level confidence for principled information aggregation. ReASC operates in two stages: a single-sample decision stage that resolves instances confidently answerable from a single response, and a reliability-aware accumulation stage that aggregates responses by jointly leveraging their frequency and confidence. Across five models and four datasets, ReASC consistently achieves the best accuracy-cost trade-off compared to existing baselines, yielding improved inference efficiency across model scales from 3B to 27B parameters. As a concrete example, ReASC reduces inference cost by up to 70\% relative to self-consistency while preserving accuracy on GSM8K using Gemma-3-4B-it.

Junseok Kim, Nakyeong Yang, Kyomin Jung

Role-play prompting is known to steer the behavior of language models by injecting a persona into the prompt, improving their zero-shot reasoning capabilities. However, such improvements are inconsistent across different tasks or instances. This inconsistency suggests that zero-shot and role-play prompting may offer complementary strengths rather than one being universally superior. Building on this insight, we propose Persona Switch, a novel decoding method that dynamically combines the benefits of both prompting strategies. Our method proceeds step-by-step, selecting the better output between zero-shot and role-play prompting at each step by comparing their output confidence, as measured by the logit gap. Experiments with widely-used LLMs demonstrate that Persona Switch consistently outperforms competitive baselines, achieving up to 5.13% accuracy improvement. Furthermore, we show that output confidence serves as an informative measure for selecting the more reliable output.

Nakyeong Yang, Jiwon Moon, Junseok Kim et al.

Enabled by large-scale text corpora with huge parameters, pre-trained language models operate as multi-task experts using a single model architecture. However, recent studies have revealed that certain neurons play disproportionately important roles in solving specific tasks, suggesting that task-relevant substructures can be isolated and selectively activated for each task. Therefore, we introduce Decomposition of Experts (DoE), a novel framework that dynamically identifies and activates task-specific experts within a language model to reduce inference cost without sacrificing accuracy. We first define a task expert as a set of parameters that significantly influence the performance of a specific task and propose a four-step unplug-and-play process: (1) receiving a user request, (2) identifying the corresponding task expert, (3) performing inference using the expert-localized model, and (4) restoring the original model and waiting for the next task. Using attribution methods and prompt tuning, DoE isolates task-relevant neurons, minimizing computational overhead while maintaining task performance. We assume a setting where a language model receives user requests from five widely used natural language understanding benchmarks, processing one task at a time. In this setup, we demonstrate that DoE achieves up to a x1.73 inference speed-up with a 65% pruning rate, without compromising accuracy. Comparisons with various task expert localization methods reveal that DoE effectively identifies task experts, while ablation studies validate the importance of its components. Additionally, we analyze the effects of batch size, token count, and layer types on inference speed-up, providing practical insights for adopting DoE. The proposed framework is both practical and scalable, applicable to any transformer-based architecture, offering a robust solution for efficient task-specific inference.

Chankyu Choi, Youngmin Yoon, Junsu Lee et al.

Recent state-of-the-art scene text recognition methods have primarily focused on horizontal text in images. However, in several Asian countries, including China, large amounts of text in signs, books, and TV commercials are vertically directed. Because the horizontal and vertical texts exhibit different characteristics, developing an algorithm that can simultaneously recognize both types of text in real environments is necessary. To address this problem, we adopted the direction encoding mask (DEM) and selective attention network (SAN) methods based on supervised learning. DEM contains directional information to compensate in cases that lack text direction; therefore, our network is trained using this information to handle the vertical text. The SAN method is designed to work individually for both types of text. To train the network to recognize both types of text and to evaluate the effectiveness of the designed model, we prepared a new synthetic vertical text dataset and collected an actual vertical text dataset (VTD142) from the Web. Using these datasets, we proved that our proposed model can accurately recognize both vertical and horizontal text and can achieve state-of-the-art results in experiments using benchmark datasets, including the street view test (SVT), IIIT-5k, and ICDAR. Although our model is relatively simple as compared to its predecessors, it maintains the accuracy and is trained in an end-to-end manner.