arXiv subjects

Suhwan Kim

Publications and source records attributed to Suhwan Kim.

4 recordsLinked to original sources

arXiv2026

Towards Understanding Pause Token Fine-Tuning Dynamics: A Mode Retention Perspective

Pause-token methods improve LLM reasoning by inserting special tokens into sequences. Prior work explains these gains through computational expressivity. However, there is relatively little investigation into the training dynamics of pause tokens. We explore how pause tokens reshape the training dynamics of fine-tuning. Two controlled pilots expose distinct asymmetries. On a synthetic continual-learning task, masked pauses overwrite a previously-learned distribution roughly 4x less at matched final adaptation (H1, mode retention); on a synthetic math-reasoning probe, the boundary-adjacent token comes to encode substantially more downstream-step information (H2, non-myopic compression). We formalize a training rule consistent with both - Masked Boundary Pause (MBP), pause tokens placed at reasoning-step boundaries with their loss masked. Across 1B-8B Qwen and Llama models, MBP consistently improves reasoning, achieving gains of up to 6 points on math and 2.5 points on code, while preserving general language understanding abilities. We further demonstrate that this mode-preserving strategy extend gains to GRPO. These results recast pause tokens as a training-dynamics intervention on the retention-adaptation trade-off, rather than merely an inference-time computation device.

cs.CL↗

arXiv2026

A Regret Minimization Framework on Preference Learning in Large Language Models

Reinforcement learning with verifiable rewards (RLVR) has enabled progress on reasoning-intensive tasks by relying on task-specific verifiers that provide automated correctness signals. However, many realistic language tasks are difficult to equip with reliable verifiers, motivating a growing reliance on reinforcement learning from human feedback (RLHF). In this setting, we argue that a closer examination of how human feedback should be interpreted is essential. We introduce Regret-based Preference Optimization $(\textbf{RePO})$, which reframes RLHF through $\textit{regret minimization}$ rather than reward maximization. Human preferences are often shaped by $\textit{prospective}$ anticipation of outcomes and $\textit{counterfactual}$ comparisons to alternative behaviors, rather than by immediate, outcome-independent utility. $\textbf{RePO}$ captures this structure by modeling preferences as behavior-conditioned assessments of relative suboptimality. Experiments on mathematical reasoning benchmarks and human preference datasets demonstrate consistent performance gains, indicating that $\textbf{RePO}$ is an effective and human-aligned approach for training large language models.

cs.AI↗

arXiv2025

Mitigating Attention Localization in Small Scale: Self-Attention Refinement via One-step Belief Propagation

Transformer-based self-attention mechanism serves as the core of modern language models, yet it often suffers from localization, where attentions collapse onto a limited subset of tokens and fail to capture long-range dependencies. To address this issue, we propose Self-Attention One-step Belief Propagation (SAOBP), a refinement framework that injects multi-hop relationships through a belief propagation process. To interpret and quantify these interactions, we introduce Global Token Dependency (GTD) that captures the relative contribution of multihop connections within the attention graph. Empirical results indicate that SAOBP helps prevent entropy collapse in deeper layers and adaptively maintains GTD at task-appropriate levels, thereby supporting improvements in model performance. Importantly, we observe competitive gains in small-scale models, highlighting its potential for improving inference quality in resource-constrained scenarios.

cs.CL↗

arXiv2025

Experience-based Knowledge Correction for Robust Planning in Minecraft

Large Language Model (LLM)-based planning has advanced embodied agents in long-horizon environments such as Minecraft, where acquiring latent knowledge of goal (or item) dependencies and feasible actions is critical. However, LLMs often begin with flawed priors and fail to correct them through prompting, even with feedback. We present XENON (eXpErience-based kNOwledge correctioN), an agent that algorithmically revises knowledge from experience, enabling robustness to flawed priors and sparse binary feedback. XENON integrates two mechanisms: Adaptive Dependency Graph, which corrects item dependencies using past successes, and Failure-aware Action Memory, which corrects action knowledge using past failures. Together, these components allow XENON to acquire complex dependencies despite limited guidance. Experiments across multiple Minecraft benchmarks show that XENON outperforms prior agents in both knowledge learning and long-horizon planning. Remarkably, with only a 7B open-weight LLM, XENON surpasses agents that rely on much larger proprietary models. Project page: https://sjlee-me.github.io/XENON

cs.LG↗