Keonwoo Kim

Author directory

2026

Large Language Models (LLMs) often memorize and regenerate sensitive or protected content that requires inference-time suppression. While machine unlearning has emerged as a highly active area of research, existing unlearning methods often degrade performance on retain dataset or require costly retraining. In this paper, we introduce Exponential-Tilted Unlearning (ETU), an inference-time method that formulates unlearning as the KL I-projection of the base next-token distribution under a user-specified upper bound 𝜀 on the forget-token mass, with a closed-form exponential tilt whose scalar multiplier 𝜆 is set analytically by 𝜀. Under a single shared protocol, (i) ETU satisfies the 𝜀 constraint across nine instruction-tuned dense and Mixture-of-Experts backbones on TOFU, RWKU, and WPU while preserving retain utility; (ii) the same calibrated family covers the full 𝜀 spectrum with the strict zero-mass case as a limiting operating point; and (iii) representative weight-update methods expose no such 𝜀-calibrated residual-mass control.
Mixture-of-Experts (MoE) pretraining relies on an auxiliary load-balancing loss (LBL) to drive per-expert utilization toward uniformity. Post-training inherits a different situation: the base router already encodes non-uniform expert co-activation structure, which a re-imposed uniformity objective flattens away. We show that downstream performance depends instead on holding this inherited routing softly, a principle we term soft router anchoring, and instantiate it as Router Prior Bias (RPB), a training-time bias that pulls the router logits toward a prior read off the frozen base router while leaving the router itself trainable. On math post-training of Moonlight-16B-A3B, RPB attains 45.77 in-domain accuracy against 31.91 under re-applied LBL and 29.44 under unanchored fine-tuning, and retains more out-of-domain capability than either. The ordering against LBL reproduces on a second model family (Qwen3-30B-A3B-Base), and the advantage over LBL is resolvable on an independently sourced corpus. Anchors defined on the router weights, on its logits, or on its output distribution perform comparably with no consistent ordering, which places the effect in the softness of the constraint rather than in the particular prior RPB supplies. Retained community structure in the expert co-activation graph tracks these gains wherever the base router is non-uniform enough for communities to form, yet enforcing the same prior as a hard assignment preserves that structure while performance falls sharply. Community structure is therefore a footprint of soft anchoring rather than its source, and the practical lesson is that inherited routing should be held softly during post-training, since both flattening it toward uniformity and enforcing it absolutely carry a downstream cost. Our code will be released at https://github.com/naver-ai/rpb.

2025

Recent advances in Large Language Models (LLMs) have enabled multi-agent systems that simulate real-world interactions with near-human reasoning. While previous studies have extensively examined biases related to protected attributes such as race, the emergence and propagation of biases on socially contentious issues in multi-agent LLM interactions remain underexplored. This study explores how LLM agents shape public opinion through debates on five contentious topics. By simulating over 2,500 debates, we analyze how initially neutral agents, assigned a centrist disposition, adopt specific stances over time. Statistical analyses reveal significant group conformity mirroring human behavior; LLM agents tend to align with numerically dominant groups or more intelligent agents, exerting a greater influence. These findings underscore the crucial role of agent intelligence in shaping discourse and highlight the risks of bias amplification in online interactions. Our results emphasize the need for policy measures that promote diversity and transparency in LLM-generated discussions to mitigate the risks of bias propagation within anonymous online environments.

2024

As natural language generation (NLG) models have become prevalent, systematically assessing the quality of machine-generated texts has become increasingly important. Recent studies introduce LLM-based evaluators that operate as reference-free metrics, demonstrating their capability to adeptly handle novel tasks. However, these models generally rely on a single-agent approach, which, we argue, introduces an inherent limit to their performance. This is because there exist biases in LLM agent’s responses, including preferences for certain text structure or content. In this work, we propose DEBATE, an NLG evaluation framework based on multi-agent scoring system augmented with a concept of Devil’s Advocate. Within the framework, one agent is instructed to criticize other agents’ arguments, potentially resolving the bias in LLM agent’s answers. DEBATE substantially outperforms the previous state-of-the-art methods in two meta-evaluation benchmarks in NLG evaluation, SummEval and TopicalChat. We also show that the extensiveness of debates among agents and the persona of an agent can influence the performance of evaluators.

2023

With the growing volume of diverse information, the demand for classifying arbitrary topics has become increasingly critical. To address this challenge, we introduce DRAFT, a simple framework designed to train a classifier for few-shot topic classification. DRAFT uses a few examples of a specific topic as queries to construct Customized dataset with a dense retriever model. Multi-query retrieval (MQR) algorithm, which effectively handles multiple queries related to a specific topic, is applied to construct the Customized dataset. Subsequently, we fine-tune a classifier using the Customized dataset to identify the topic. To demonstrate the efficacy of our proposed approach, we conduct evaluations on both widely used classification benchmark datasets and manually constructed datasets with 291 diverse topics, which simulate diverse contents encountered in real-world applications. DRAFT shows competitive or superior performance compared to baselines that use in-context learning, such as GPT-3 175B and InstructGPT 175B, on few-shot topic classification tasks despite having 177 times fewer parameters, demonstrating its effectiveness.