Xiaohang Tang

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.

Juliusz Ziomek, William Bankes, Lorenz Wolf et al. · oxford

We introduce LLM-Wikirace, a benchmark for evaluating planning, reasoning, and world knowledge in large language models (LLMs). In LLM-Wikirace, models must efficiently navigate Wikipedia hyperlinks step by step to reach a target page from a given source, requiring look-ahead planning and the ability to reason about how concepts are connected in the real world. We evaluate a broad set of open- and closed-source models, including Gemini-3, GPT-5, and Claude Opus 4.5, which achieve the strongest results on the easy level of the task and demonstrate superhuman performance. Despite this, performance drops sharply on hard difficulty: the best-performing model, Gemini-3, succeeds in only 23\% of hard games, highlighting substantial remaining challenges for frontier models. Our analysis shows that world knowledge is a necessary ingredient for success, but only up to a point, beyond this threshold, planning and long-horizon reasoning capabilities become the dominant factors. Trajectory-level analysis further reveals that even the strongest models struggle to replan after failure, frequently entering loops rather than recovering. LLM-Wikirace is a simple benchmark that reveals clear limitations in current reasoning systems, offering an open arena where planning-capable LLMs still have much to prove. Our code and leaderboard available at https:/llmwikirace.github.io.

Yang Luo, Shengju Qian, Xiaohang Tang et al.

Autoregressive video generators are attractive for streaming, long-horizon, and interactive applications, but distilling strong black-box teachers into causal students remains difficult. The student must learn under its own rollout distribution, whereas practical teachers may expose only prompt-conditioned completed videos and may differ in architecture, capacity, temporal design, and sampling schedule. This interface makes supervised fine-tuning off-policy, score-based distillation inapplicable, and direct adversarial imitation too sparse for denoising-time credit assignment. We propose Adversarial Flow Distillation (AFD), an on-policy framework for heterogeneous black-box video distillation. AFD queries the teacher and rolls out the current student on the same prompts, trains a prompt-paired Bradley-Terry discriminator to estimate clean-sample teacher-student discrepancy, and converts the resulting on-policy advantage into forward-process flow-matching updates on the student's own noised states. Thus, AFD provides dense velocity-field supervision while requiring no teacher scores, latents, denoising trajectories, step alignment, or reverse-chain reinforcement learning. Experiments across two causal AR student families show that AFD consistently improves motion- and physics-sensitive generation while preserving general video quality, and ablations validate the importance of adaptive on-policy feedback and forward-process credit assignment. The method requires only clean teacher videos and student rollouts, providing a practical route for distilling proprietary or heterogeneous video generators into efficient autoregressive students.

Xiaohang Tang, Afonso Marques, Parameswaran Kamalaruban et al.

Decision Transformer (DT), as one of the representative Reinforcement Learning via Supervised Learning (RvS) methods, has achieved strong performance in offline learning tasks by leveraging the powerful Transformer architecture for sequential decision-making. However, in adversarial environments, these methods can be non-robust, since the return is dependent on the strategies of both the decision-maker and adversary. Training a probabilistic model conditioned on observed return to predict action can fail to generalize, as the trajectories that achieve a return in the dataset might have done so due to a suboptimal behavior adversary. To address this, we propose a worst-case-aware RvS algorithm, the Adversarially Robust Decision Transformer (ARDT), which learns and conditions the policy on in-sample minimax returns-to-go. ARDT aligns the target return with the worst-case return learned through minimax expectile regression, thereby enhancing robustness against powerful test-time adversaries. In experiments conducted on sequential games with full data coverage, ARDT can generate a maximin (Nash Equilibrium) strategy, the solution with the largest adversarial robustness. In large-scale sequential games and continuous adversarial RL environments with partial data coverage, ARDT demonstrates significantly superior robustness to powerful test-time adversaries and attains higher worst-case returns compared to contemporary DT methods.

Xiaohang Tang, Yi Zhou, Taichi Aida et al.

Semantic Change Detection (SCD) of words is an important task for various NLP applications that must make time-sensitive predictions. Some words are used over time in novel ways to express new meanings, and these new meanings establish themselves as novel senses of existing words. On the other hand, Word Sense Disambiguation (WSD) methods associate ambiguous words with sense ids, depending on the context in which they occur. Given this relationship between WSD and SCD, we explore the possibility of predicting whether a target word has its meaning changed between two corpora collected at different time steps, by comparing the distributions of senses of that word in each corpora. For this purpose, we use pretrained static sense embeddings to automatically annotate each occurrence of the target word in a corpus with a sense id. Next, we compute the distribution of sense ids of a target word in a given corpus. Finally, we use different divergence or distance measures to quantify the semantic change of the target word across the two given corpora. Our experimental results on SemEval 2020 Task 1 dataset show that word sense distributions can be accurately used to predict semantic changes of words in English, German, Swedish and Latin.

Xiaohang Tang, Yi Zhou, Danushka Bollegala

Dynamic contextualised word embeddings (DCWEs) represent the temporal semantic variations of words. We propose a method for learning DCWEs by time-adapting a pretrained Masked Language Model (MLM) using time-sensitive templates. Given two snapshots $C_1$ and $C_2$ of a corpus taken respectively at two distinct timestamps $T_1$ and $T_2$, we first propose an unsupervised method to select (a) \emph{pivot} terms related to both $C_1$ and $C_2$, and (b) \emph{anchor} terms that are associated with a specific pivot term in each individual snapshot. We then generate prompts by filling manually compiled templates using the extracted pivot and anchor terms. Moreover, we propose an automatic method to learn time-sensitive templates from $C_1$ and $C_2$, without requiring any human supervision. Next, we use the generated prompts to adapt a pretrained MLM to $T_2$ by fine-tuning using those prompts. Multiple experiments show that our proposed method reduces the perplexity of test sentences in $C_2$, outperforming the current state-of-the-art.

Xiaoyuan Cheng, Yiming Yang, Xiaohang Tang et al.

An efficient way to control systems with unknown nonlinear dynamics is to find an appropriate embedding or representation for simplified approximation (e.g. linearization), which facilitates system identification and control synthesis. Nevertheless, there has been a lack of embedding methods that can guarantee (i) embedding the dynamical system comprehensively, including the vector fields (ODE form) of the dynamics, and (ii) preserving the consistency of control effect between the original and latent space. To address these challenges, we propose Koopman Embedded Equivariant Control (KEEC) to learn an embedding of the states and vector fields such that a Koopman operator is approximated as the latent dynamics. Due to the Koopman operator's linearity, learning the latent vector fields of the dynamics becomes simply solving linear equations. Thus in KEEC, the analytical form of the greedy control policy, which is dependent on the learned differential information of the dynamics and value function, is also simplified. Meanwhile, KEEC preserves the effectiveness of the control policy in the latent space by preserving the metric in two spaces. Our algorithm achieves superior performances in the experiments conducted on various control domains, including the image-based Pendulum, Lorenz-63 and the wave equation. The code is available at https://github.com/yyimingucl/Koopman-Embedded-Equivariant-Control.

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.

Xiaohang Tang, Zhuowen Cheng, Satyabrat Kumar

The Transformer, a highly expressive architecture for sequence modeling, has recently been adapted to solve sequential decision-making, most notably through the Decision Transformer (DT), which learns policies by conditioning on desired returns. Yet, the adversarial robustness of reinforcement learning methods based on sequence modeling remains largely unexplored. Here we introduce the Conservative Adversarially Robust Decision Transformer (CART), to our knowledge the first framework designed to enhance the robustness of DT in adversarial stochastic games. We formulate the interaction between the protagonist and the adversary at each stage as a stage game, where the payoff is defined as the expected maximum value over subsequent states, thereby explicitly incorporating stochastic state transitions. By conditioning Transformer policies on the NashQ value derived from these stage games, CART generates policy that are simultaneously less exploitable (adversarially robust) and conservative to transition uncertainty. Empirically, CART achieves more accurate minimax value estimation and consistently attains superior worst-case returns across a range of adversarial stochastic games.

Xiang Li, Xiaohang Tang, Xin Tong et al.

Myopia is an eye condition that makes it difficult for people to focus on faraway objects. It has become one of the most serious eye conditions worldwide and negatively impacts the quality of life of those who suffer from it. Although myopia is prevalent, many non-myopic people have misconceptions about it and encounter challenges empathizing with myopia situations and those who suffer from it. In this research, we developed two virtual reality (VR) games, (1) Myopic Bike and (2) Say Hi, to provide a means for the non-myopic population to experience the frustration and difficulties of myopic people. Our two games simulate two inconvenient daily life scenarios (riding a bicycle and greeting someone on the street) that myopic people encounter when not wearing glasses. We evaluated four participants' game experiences through questionnaires and semi-structured interviews. Overall, our two VR games can create an engaging and non-judgmental experience for the non-myopic population to better understand and empathize with those who suffer from myopia.

Diego Monteiro, Hai-Ning Liang, Xiaohang Tang et al.

Identifying cybersickness in virtual reality (VR) applications such as games in a fast, precise, non-intrusive, and non-disruptive way remains challenging. Several factors can cause cybersickness, and their identification will help find its origins and prevent or minimize it. One such factor is virtual movement. Movement, whether physical or virtual, can be represented in different forms. One way to represent and store it is with a temporally annotated point sequence. Because a sequence is memory-consuming, it is often preferable to save it in a compressed form. Compression allows redundant data to be eliminated while still preserving changes in speed and direction. Since changes in direction and velocity in VR can be associated with cybersickness, changes in compression rate can likely indicate changes in cybersickness levels. In this research, we explore whether quantifying changes in virtual movement can be used to estimate variation in cybersickness levels of VR users. We investigate the correlation between changes in the compression rate of movement data in two VR games with changes in players' cybersickness levels captured during gameplay. Our results show (1) a clear correlation between changes in compression rate and cybersickness, and(2) that a machine learning approach can be used to identify these changes. Finally, results from a second experiment show that our approach is feasible for cybersickness inference in games and other VR applications that involve movement.

Xiaoteng Ma, Xiaohang Tang, Li Xia et al.

Most of reinforcement learning algorithms optimize the discounted criterion which is beneficial to accelerate the convergence and reduce the variance of estimates. Although the discounted criterion is appropriate for certain tasks such as financial related problems, many engineering problems treat future rewards equally and prefer a long-run average criterion. In this paper, we study the reinforcement learning problem with the long-run average criterion. Firstly, we develop a unified trust region theory with discounted and average criteria and derive a novel performance bound within the trust region with the Perturbation Analysis (PA) theory. Secondly, we propose a practical algorithm named Average Policy Optimization (APO), which improves the value estimation with a novel technique named Average Value Constraint. Finally, experiments are conducted in the continuous control environment MuJoCo. In most tasks, APO performs better than the discounted PPO, which demonstrates the effectiveness of our approach. Our work provides a unified framework of the trust region approach including both the discounted and average criteria, which may complement the framework of reinforcement learning beyond the discounted objectives.