Sangwoong Yoon

Xiaohang Tang, Keyue Jiang, Che Liu et al.

Reinforcement learning (RL) can be used to improve the policy (denoiser) of diffusion large language models (dLLMs), while being hindered by the intractability of the policy likelihood. A dominant and efficient family of methods replaces the likelihood in standard RL with its evidence lower bound (ELBO), estimated from randomly masked sequences. Despite being well aligned with pre-training, these approaches introduce bias through training--inference mismatch by using the ELBO as a likelihood surrogate, which can degrade performance. In this work, we propose Guided Denoiser Self-Distillation (GDSD) to directly distill the denoiser of dLLMs from an advantage-guided self-teacher, derived from the closed-form optimum of reverse-KL regularized RL. GDSD matches the dLLM's denoiser logits to the teacher's via a normalization-free objective, which reduces RL to likelihood-free self-distillation and thus bypasses the TIM biases. Recent ELBO-based methods emerge as instances of applying different distillation divergences, but with diagnosable pathologies that GDSD avoids. On planning, math, and coding benchmarks with LLaDA-8B and Dream-7B, GDSD consistently outperforms prior state-of-the-art ELBO-based methods with a more stable training reward dynamics, achieving test-accuracy improvements of up to $+19.6\%$. These results suggest that direct denoiser self-distillation, without relying on an ELBO likelihood surrogate, can provide a more stable and effective RL procedure for dLLMs. Code is available at https://github.com/GaryBall/GDSD.

Shyam Sundhar Ramesh, Xiaotong Ji, Matthieu Zimmer et al.

RL-based post-training with GRPO is widely used to improve large language models on individual reasoning tasks. However, real-world deployment requires reliable performance across diverse tasks. A straightforward multi-task adaptation of GRPO often leads to imbalanced outcomes, with some tasks dominating optimization while others stagnate. Moreover, tasks can vary widely in how frequently prompts yield zero advantages (and thus zero gradients), which further distorts their effective contribution to the optimization signal. To address these issues, we propose a novel Multi-Task GRPO (MT-GRPO) algorithm that (i) dynamically adapts task weights to explicitly optimize worst-task performance and promote balanced progress across tasks, and (ii) introduces a ratio-preserving sampler to ensure task-wise policy gradients reflect the adapted weights. Experiments on both 3-task and 9-task settings show that MT-GRPO consistently outperforms baselines in worst-task accuracy. In particular, MT-GRPO achieves 16-28% and 6% absolute improvement on worst-task performance over standard GRPO and DAPO, respectively, while maintaining competitive average accuracy. Moreover, MT-GRPO requires 50% fewer training steps to reach 50% worst-task accuracy in the 3-task setting, demonstrating substantially improved efficiency in achieving reliable performance across tasks.

Sangwoong Yoon, Young-Uk Jin, Yung-Kyun Noh et al.

We present a new method of training energy-based models (EBMs) for anomaly detection that leverages low-dimensional structures within data. The proposed algorithm, Manifold Projection-Diffusion Recovery (MPDR), first perturbs a data point along a low-dimensional manifold that approximates the training dataset. Then, EBM is trained to maximize the probability of recovering the original data. The training involves the generation of negative samples via MCMC, as in conventional EBM training, but from a different distribution concentrated near the manifold. The resulting near-manifold negative samples are highly informative, reflecting relevant modes of variation in data. An energy function of MPDR effectively learns accurate boundaries of the training data distribution and excels at detecting out-of-distribution samples. Experimental results show that MPDR exhibits strong performance across various anomaly detection tasks involving diverse data types, such as images, vectors, and acoustic signals.

Sangwoong Yoon, Jinwon Choi, Yonghyeon Lee et al.

A reliable evaluation method is essential for building a robust out-of-distribution (OOD) detector. Current robustness evaluation protocols for OOD detectors rely on injecting perturbations to outlier data. However, the perturbations are unlikely to occur naturally or not relevant to the content of data, providing a limited assessment of robustness. In this paper, we propose Evaluation-via-Generation for OOD detectors (EvG), a new protocol for investigating the robustness of OOD detectors under more realistic modes of variation in outliers. EvG utilizes a generative model to synthesize plausible outliers, and employs MCMC sampling to find outliers misclassified as in-distribution with the highest confidence by a detector. We perform a comprehensive benchmark comparison of the performance of state-of-the-art OOD detectors using EvG, uncovering previously overlooked weaknesses.

Sangwoong Yoon, Frank C. Park, Gunsu S Yun et al.

Kernel density estimation (KDE) is integral to a range of generative and discriminative tasks in machine learning. Drawing upon tools from the multidimensional calculus of variations, we derive an optimal weight function that reduces bias in standard kernel density estimates for density ratios, leading to improved estimates of prediction posteriors and information-theoretic measures. In the process, we shed light on some fundamental aspects of density estimation, particularly from the perspective of algorithms that employ KDEs as their main building blocks.

Himchan Hwang, Hyeokju Jeong, Gene Chung et al.

We propose MATE, a simple yet effective memory architecture for solving Contextual Markov Decision Processes (CMDPs), a family of MDPs parameterized by an unobserved context. In CMDPs, an optimal agent can adapt online by maintaining the posterior belief over contexts. MATE replaces this intractable posterior with a sum-aggregated memory, leveraging the posterior's permutation invariance to retain provably sufficient expressiveness. Compared to prior memory architectures, MATE avoids the growing per-step rollout cost of Transformers and the gradient issues commonly associated with Recurrent Neural Networks (RNNs). Extensive evaluations across diverse benchmarks demonstrate that MATE provides clear computational advantages while achieving performance comparable to standard sequence-model baselines.

Xiaohang Tang, Rares Dolga, Sangwoong Yoon et al.

Improving the reasoning capabilities of diffusion-based large language models (dLLMs) through reinforcement learning (RL) remains an open problem. The intractability of dLLMs likelihood function necessitates approximating the current, old, and reference policy likelihoods at each policy optimization step. This reliance introduces additional computational overhead and lead to potentially large bias -- particularly when approximation errors occur in the denominator of policy ratios used for importance sampling. To mitigate these issues, we introduce $\mathtt{wd1}$, a novel policy optimization approach that reformulates the objective as a weighted likelihood, requiring only a single approximation for the current parametrized policy likelihood. Experiments on widely used reasoning benchmarks demonstrate that $\mathtt{wd1}$, without supervised fine-tuning (SFT) or any supervised data, outperforms existing RL methods for dLLMs, achieving up to 16% higher accuracy. $\mathtt{wd1}$ delivers additional computational gains, including reduced training time and fewer function evaluations (NFEs) per gradient step. These findings, combined with the simplicity of method's implementation and R1-Zero-like training (no SFT), position $\mathtt{wd1}$ as a more effective and efficient method for applying RL to dLLMs reasoning.

Seongho Son, William Bankes, Sangwoong Yoon et al.

Test-time alignment of Large Language Models (LLMs) to human preferences offers a flexible way to generate responses aligned to diverse objectives without extensive retraining of LLMs. Existing methods achieve alignment to multiple objectives simultaneously (e.g., instruction-following, helpfulness, conciseness) by optimizing their corresponding reward functions. However, they often rely on predefined weights or optimize for averages, sacrificing one objective for another and leading to unbalanced outcomes. To address this, we introduce Robust Multi-Objective Decoding (RMOD), a novel inference-time algorithm that optimizes for improving worst-case rewards. RMOD formalizes the robust decoding problem as a maximin two-player game between reward weights and the sampling policy, solving for the Nash equilibrium. We show that the game reduces to a convex optimization problem to find the worst-case weights, while the best response policy can be computed analytically. We also introduce a practical RMOD variant designed for efficient decoding with contemporary LLMs, incurring minimal computational overhead compared to non-robust Multi-Objective Decoding (MOD) methods. Our experimental results showcase the effectiveness of RMOD in generating responses equitably aligned with diverse objectives, outperforming baselines up to 20%.

Xiaohang Tang, Sangwoong Yoon, Seongho Son et al.

Self-play alignment has emerged as an effective approach for fine-tuning large language models (LLMs), formulating preference optimization as a two-player game. However, the regularization with respect to the reference policy, which is crucial for mitigating over-optimization, has been insufficiently investigated in self-play alignment. To study the impact of different regularization strategies, we propose \textbf{Regularized Self-Play Policy Optimization (RSPO)}, a general and modular framework that unifies prior methods and enables simple plug-and-play integration of various regularizers, meanwhile preserving convergence to Nash equilibrium of the corresponding regularized game.Our empirical study involving over $120$ fine-tuned Mistral-7B-Instruct models reveals that forward KL divergence regularization reduces response length, whereas reverse KL divergence markedly improves raw win rates. Crucially, RSPO regularized with a linear combination of forward and reverse KL divergence significantly boosts the length-controlled win rate on AlpacaEval-2 from $28.5\%$ (unregularized self-play, SPPO) to $35.4\%$, and consistently demonstrates superior performance on Arena-Hard, MT-Bench, ArmoRM scores, and response diversity. Combining simplicity, convergence guarantees, and significant empirical gains, RSPO offers a strong foundation for exploring regularized self-play in language model alignment.

Sangwoong Yoon, Himchan Hwang, Hyeokju Jeong et al.

We propose the Value Gradient Sampler (VGS), a trainable sampler based on the interpretation of sampling as discrete-time sequential decision-making. VGS generates samples from a given unnormalized density (i.e., energy) by drifting and diffusing randomly initialized particles. In VGS, finding the optimal drift is equivalent to solving an optimal control problem where the cost is the upper bound of the KL divergence between the target density and the samples. We employ value-based dynamic programming to solve this optimal control problem, which gives the gradient of the value function as the optimal drift vector. The connection to sequential decision making allows VGS to leverage extensively studied techniques in reinforcement learning, making VGS a fast, adaptive, and accurate sampler that achieves competitive results in various sampling benchmarks. Furthermore, VGS can replace MCMC in contrastive divergence training of energy-based models. We demonstrate the effectiveness of VGS in training accurate energy-based models in industrial anomaly detection applications.

Lorenz Wolf, Sangwoong Yoon, Ilija Bogunovic

Mixture of large language model (LLMs) Agents (MoA) architectures achieve state-of-the-art performance on prominent benchmarks like AlpacaEval 2.0 by leveraging the collaboration of multiple LLMs at inference time. Despite these successes, an evaluation of the safety and reliability of MoA is missing. We present the first comprehensive study of MoA's robustness against deceptive LLM agents that deliberately provide misleading responses. We examine factors like the propagation of deceptive information, model size, and information availability, and uncover critical vulnerabilities. On AlpacaEval 2.0, the popular LLaMA 3.1-70B model achieves a length-controlled Win Rate (LC WR) of 49.2% when coupled with 3-layer MoA (6 LLM agents). However, we demonstrate that introducing only a $\textit{single}$ carefully-instructed deceptive agent into the MoA can reduce performance to 37.9%, effectively nullifying all MoA gains. On QuALITY, a multiple-choice comprehension task, the impact is also severe, with accuracy plummeting by a staggering 48.5%. Inspired in part by the historical Doge of Venice voting process, designed to minimize influence and deception, we propose a range of unsupervised defense mechanisms that recover most of the lost performance.

Sangwoong Yoon, Himchan Hwang, Dohyun Kwon et al.

We present a maximum entropy inverse reinforcement learning (IRL) approach for improving the sample quality of diffusion generative models, especially when the number of generation time steps is small. Similar to how IRL trains a policy based on the reward function learned from expert demonstrations, we train (or fine-tune) a diffusion model using the log probability density estimated from training data. Since we employ an energy-based model (EBM) to represent the log density, our approach boils down to the joint training of a diffusion model and an EBM. Our IRL formulation, named Diffusion by Maximum Entropy IRL (DxMI), is a minimax problem that reaches equilibrium when both models converge to the data distribution. The entropy maximization plays a key role in DxMI, facilitating the exploration of the diffusion model and ensuring the convergence of the EBM. We also propose Diffusion by Dynamic Programming (DxDP), a novel reinforcement learning algorithm for diffusion models, as a subroutine in DxMI. DxDP makes the diffusion model update in DxMI efficient by transforming the original problem into an optimal control formulation where value functions replace back-propagation in time. Our empirical studies show that diffusion models fine-tuned using DxMI can generate high-quality samples in as few as 4 and 10 steps. Additionally, DxMI enables the training of an EBM without MCMC, stabilizing EBM training dynamics and enhancing anomaly detection performance.

Sangwoong Yoon, Dohyun Kwon, Himchan Hwang et al.

We present Generalized Contrastive Divergence (GCD), a novel objective function for training an energy-based model (EBM) and a sampler simultaneously. GCD generalizes Contrastive Divergence (Hinton, 2002), a celebrated algorithm for training EBM, by replacing Markov Chain Monte Carlo (MCMC) distribution with a trainable sampler, such as a diffusion model. In GCD, the joint training of EBM and a diffusion model is formulated as a minimax problem, which reaches an equilibrium when both models converge to the data distribution. The minimax learning with GCD bears interesting equivalence to inverse reinforcement learning, where the energy corresponds to a negative reward, the diffusion model is a policy, and the real data is expert demonstrations. We present preliminary yet promising results showing that joint training is beneficial for both EBM and a diffusion model. GCD enables EBM training without MCMC while improving the sample quality of a diffusion model.

Sangwoong Yoon, Yung-Kyun Noh, Frank Chongwoo Park

Likelihood is a standard estimate for outlier detection. The specific role of the normalization constraint is to ensure that the out-of-distribution (OOD) regime has a small likelihood when samples are learned using maximum likelihood. Because autoencoders do not possess such a process of normalization, they often fail to recognize outliers even when they are obviously OOD. We propose the Normalized Autoencoder (NAE), a normalized probabilistic model constructed from an autoencoder. The probability density of NAE is defined using the reconstruction error of an autoencoder, which is differently defined in the conventional energy-based model. In our model, normalization is enforced by suppressing the reconstruction of negative samples, significantly improving the outlier detection performance. Our experimental results confirm the efficacy of NAE, both in detecting outliers and in generating in-distribution samples.

Sangwoong Yoon, Woo Young Kang, Sungwook Jeon et al.

As a scene graph compactly summarizes the high-level content of an image in a structured and symbolic manner, the similarity between scene graphs of two images reflects the relevance of their contents. Based on this idea, we propose a novel approach for image-to-image retrieval using scene graph similarity measured by graph neural networks. In our approach, graph neural networks are trained to predict the proxy image relevance measure, computed from human-annotated captions using a pre-trained sentence similarity model. We collect and publish the dataset for image relevance measured by human annotators to evaluate retrieval algorithms. The collected dataset shows that our method agrees well with the human perception of image similarity than other competitive baselines.