arXiv subjects

Ilija Bogunovic

Publications and source records attributed to Ilija Bogunovic.

At least 19 recordsLinked to original sources

arXiv2026

Temporal Self-Distillation: Faster Inference in Discrete Diffusion Language Models

Diffusion language models (dLLMs) promise fast inference by generating multiple tokens in parallel, but suffer severe performance degradation when parallel decoding is pushed too aggressively. We introduce Temporal Self-Distillation (TSD), a simple on-policy method that trains dLLMs for fast inference by distilling predictions across time. Specifically, TSD distills the model's denoising distribution at earlier timesteps toward its distribution at the final timestep at which a token is committed. This encourages earlier predictions to better anticipate the model's eventual output, enabling much more aggressive parallel decoding. Because its teacher signal comes from the model itself, TSD requires no offline teacher generation and applies seamlessly to both base and post-trained policies. Across seven benchmarks in mathematics, planning, and code, TSD substantially shifts the speed--quality frontier toward the low-compute regime. TSD thus provides a simple, single-stage approach to accelerating dLLMs, achieving speedups competitive with offline distillation while avoiding a complex two-stage pipeline.

cs.LG↗

arXiv2026

RSPO: Regularized Self-Play Alignment of Large Language Models

Self-play-based policy optimization 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 Regularized Self-Play Policy Optimization (RSPO), a novel framework that unifies prior methods and enables simple plug-and-play regularizers, meanwhile preserving convergence to Nash equilibrium of the corresponding regularized game. We empirically show that RSPO with appropriate regularizers can substantially improve the length-controlled win rate (LCWR) on AlpacaEval-2 across a range of base models, while also achieving consistently superior performance on Arena-Hard, MT-Bench, ArmoRM, and response diversity. In particular, RSPO improves unregularized self-play baseline (SPPO) on AlpacaEval-2 LCWR from $28.5\%$ to $ 35.4\%$ with base model Mistral-7B, from $38.77\%$ to $43.66\%$ with LLaMA-8B, and from $50.54\%$ to $51.83\%$ with Gemma-2B. Combining simplicity, convergence guarantees, and significant empirical gains, RSPO offers a strong foundation for exploring regularized self-play in alignment. Code is available at https://github.com/xiaohangt/RSPO

cs.LG↗

arXiv2026

Multi-Task GRPO: Reliable LLM Reasoning Across Tasks

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.

cs.CL↗

arXiv2026

Meta-Learning Preferences for Multilingual LLM Alignment

Unequal availability of human preference data across languages poses a significant challenge for aligning large language models in multilingual settings. To address the lack of sufficient data in low-resource language alignment, we propose a meta-learning framework for Reinforcement Learning from Human Feedback and Direct Preference Optimization. By leveraging preference data from other languages, our framework learns a transferable initialization that enables effective adaptation to a target language with minimal data. We provide theoretical guarantees for both the meta-reward modeling and meta-policy optimization settings, and empirically demonstrate the effectiveness of our approach on multilingual benchmarks. In an extremely low-resource setting with only 100 target-language preference samples, our approach achieves up to $28\%$ win-rate improvements over baseline methods, and consistently outperforms baselines across multiple target languages and model scales. Our approaches retain these advantages across different combinations of meta-training languages and varying linguistic distances from the target languages.

cs.CL↗

arXiv2026

SWE-Router: Routing in Multi-turn Agentic Software Engineering Tasks

Large language models (LLMs) embedded in multi-turn agentic harnesses are reshaping software engineering (SWE), but routing every task to a frontier model is wasteful when many issues admit cheap fixes. Existing LLM routers operate on the task description alone, which inherits an information-theoretic Bayes-error floor in agentic settings: a similar issue can hide either a localized typo or a multi-module refactor, and the prompt does not separate the two. We introduce SWE-Router, a value-based temporal approach that lets a cheap model run for a few exploratory turns and reads the resulting partial trajectory before deciding whether to continue cheaply or to escalate to an expensive model. We provide a Bayes-optimality theorem showing that conditioning on the partial trajectory never harms routing and is strictly better whenever exploration is informative. Across the LLM pairs of weak and strong models spanning the contemporary cost--capability frontier, we show that SWE-Router greatly improves the cost efficiency of SWE tasks, while maintaining the majority of the performances of the stronger model. We additionally release a multi-LLM trajectory dataset which allows reproduction of our trajectory-level routing.

cs.SE↗

arXiv2026

LLM-WikiRace Benchmark: How Far Can LLMs Plan over Real-World Knowledge Graphs?

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.

cs.AI↗

arXiv2026

Re-evaluating Confidence Remasking in Masked Diffusion Language Models

Masked diffusion language models (dLLMs) have recently emerged as a competitive alternative to autoregressive language models, with the promise of faster inference via parallel token generation. A notable limitation of the masked formulation, however, is that once a token has been unmasked it can no longer be revised, leaving dLLMs vulnerable to early sampling mistakes. To address this, a growing body of work has sought to extend masked dLLMs with self-correcting (remasking) capabilities. One appealing subset of these methods does so in a training-free, post-hoc manner based on token confidences, with encouraging early reported results. In this work, we revisit the empirical evaluation of a representative post-hoc remasking method, WINO [Hong et al., 2026], and find that under standard decoding settings (shorter block lengths) it brings little-to-no benefit over confidence-based unmasking alone [Wu et al., 2025]. Extending the evaluation to non-greedy decoding, we find that while confidence-based remasking can mitigate errors introduced by increased stochasticity to some extent, it also exacerbates the diversity collapse previously reported for confidence-based unmasking. Overall, our results show that the benefits of post-hoc confidence-based remasking are highly setting-dependent, underscoring the need for a more comprehensive evaluation framework.

cs.LG↗

arXiv2026

PROWL: Prioritized Regret-Driven Optimization for World Model Learning

Modern action-conditioned video world models achieve strong short-horizon visual realism, yet remain unreliable on rare, interaction-critical transitions that dominate downstream planning and policy performance. Because passive demonstration data systematically under-samples these high-impact regimes, improving robustness requires actively eliciting model failures rather than relying on their natural occurrence. We introduce a KL-constrained adversarial curriculum in which a policy is trained to expose high-error trajectories of a diffusion-based world model while remaining close to the behavior distribution. The world model is continuously fine-tuned on these adversarially discovered trajectories, yielding an adversarial training loop that converts rare failures into a stable, near-distribution training signal without drifting into out-of-distribution exploitation. To maintain pressure on unresolved weaknesses as the model improves, we propose a Prioritized Adversarial Trajectory (PAT) buffer that re-ranks trajectories based on prediction error, action fidelity, and learning progress, focusing training on unresolved failure modes rather than repeatedly revisiting solved cases. We implement our approach in the MineRL framework and evaluate it on held-out out-of-distribution trajectories; PROWL improves robustness over models trained on passive data alone, reveals reward-hacking behaviors under weak behavioral constraints, and demonstrates that effective adversarial world-model training critically depends on balancing exploratory failure discovery with explicit behavioral regularization. Our results suggest that scalable world models benefit not only from larger datasets, but also from selectively generating informative training data.

cs.LG↗

arXiv2026

GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models

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.

cs.LG↗

arXiv2026

Overton Pluralistic Reinforcement Learning for Large Language Models

Existing alignment paradigms remain limited in capturing the pluralistic nature of human values. Overton Pluralism addresses this gap by generating responses with diverse perspectives from a single query. This paper introduces OP-GRPO (Overton Pluralistic Group Relative Policy Optimization), a reinforcement learning framework for implicit Overton Pluralism that enables a single large language model to produce pluralistic responses without explicit prompting or modular orchestration. Our workflow consists of two main steps. First, similarity estimator training fine-tunes a Sentence Transformer for Overton Pluralism tasks to provide more accurate coverage evaluation of generated responses. Second, OP-GRPO training incorporates this similarity estimator into a dual-reward system designed to ensure both broad coverage of genuine human perspectives and the uniqueness of each perspective, thereby promoting diversity. Empirical results demonstrate a "small models, big perspective coverage" effect. The trained Qwen2.5-3B-Instruct model surpasses a 20B GPT-OSS baseline with a 37.4 percent relative accuracy gain on a Natural Language Inference benchmark, and also outperforms a modular architecture baseline with a 19.1 percent relative improvement. Additional evaluations using GPT-4.1 as a large language model judge further confirm the robustness of the approach.

cs.CL↗

arXiv2026

Robust Multi-Objective Controlled Decoding of Large Language Models

We introduce Robust Multi-Objective Decoding (RMOD), a novel inference-time algorithm that robustly aligns Large Language Models (LLMs) to multiple human objectives (e.g., instruction-following, helpfulness, safety) by maximizing the worst-case rewards. RMOD formulates the robust decoding problem as a maximin two-player game between adversarially computed reward weights and the sampling policy, solvable through a Nash equilibrium. We demonstrate that this game reduces to a convex optimization problem to identify the worst-case reward weights, with the optimal sampling policy analytically derived. For practical applications, we propose an efficient algorithm of RMOD tailored for contemporary LLMs, introducing minimal computational overhead compared to standard non-robust Controlled Decoding methods. Experimental results across a range of popular alignment datasets with up to 10 objectives show the effectiveness of RMOD and its distilled version, consistently outperforming baselines in worst-case rewards and win rates.

cs.LG↗

arXiv2026

Robust Bayesian Optimisation with Unbounded Corruptions

Bayesian Optimization is critically vulnerable to extreme outliers. Existing provably robust methods typically assume a bounded cumulative corruption budget, which makes them defenseless against even a single corruption of sufficient magnitude. To address this, we introduce a new adversary whose budget is only bounded in the frequency of corruptions, not in their magnitude. We then derive RCGP-UCB, an algorithm coupling the famous upper confidence bound (UCB) approach with a Robust Conjugate Gaussian Process (RCGP). We present stable and adaptive versions of RCGP-UCB, and prove that they achieve sublinear regret in the presence of up to $O(T^{1/4})$ and $O(T^{1/7})$ corruptions with possibly infinite magnitude. This robustness comes at near zero cost: without outliers, RCGP-UCB's regret bounds match those of the standard GP-UCB algorithm.

stat.ML↗

arXiv2026

wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Models

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 can lead to large variance and estimation error in RL objective -- particularly in computing the policy ratio for importance sampling. To mitigate these issues, we introduce wd1, a novel ratio-free policy optimization approach that reformulates the RL objective as a weighted log-likelihood, requiring only a single approximation for the current parametrized policy likelihood. We formally show that our proposed method can be interpreted as energy-guided discrete diffusion training combined with negative sample unlearning, thereby confirming its theoretical soundness. In experiments on LLaDA-8B model, wd1 outperforms diffusion-based GRPO (d1) while requiring lower computational cost, achieving up to a $+59\%$ improvement in accuracy. Furthermore, we extend wd1 to denoising-stepwise weighted policy optimization (wd1++), achieving state-of-the-art math performance of $44.2\%$ on MATH500 and $84.5\%$ on GSM8K with only 20 RL training steps.

cs.LG↗

arXiv2026

Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift

Current Large Language Model (LLM) preference optimization algorithms do not account for temporal preference drift, which can lead to severe misalignment. To address this limitation, we propose Non-Stationary Direct Preference Optimisation (NS-DPO) that models time-dependent reward functions with a Dynamic Bradley-Terry model. NS-DPO proposes a computationally efficient solution by introducing only a single discount parameter in the loss function, which is used for exponential weighting that proportionally focuses learning on more time-relevant datapoints. We theoretically analyze the convergence of NS-DPO in a general setting where the exact nature of the preference drift is not known, providing upper bounds on the estimation error and regret caused by non-stationary preferences. Finally, we demonstrate the effectiveness of NS-DPO for fine-tuning LLMs under drifting preferences. Using scenarios where various levels of preference drift is introduced, with popular LLM reward models and datasets, we show that NS-DPO fine-tuned LLMs remain robust under non-stationarity, significantly outperforming baseline algorithms that ignore temporal preference changes, without sacrificing performance in stationary cases.

cs.LG↗

arXiv2025

Imagined Autocurricula

Training agents to act in embodied environments typically requires vast training data or access to accurate simulation, neither of which exists for many cases in the real world. Instead, world models are emerging as an alternative leveraging offline, passively collected data, they make it possible to generate diverse worlds for training agents in simulation. In this work, we harness world models to generate imagined environments to train robust agents capable of generalizing to novel task variations. One of the challenges in doing this is ensuring the agent trains on useful generated data. We thus propose a novel approach, IMAC (Imagined Autocurricula), leveraging Unsupervised Environment Design (UED), which induces an automatic curriculum over generated worlds. In a series of challenging, procedurally generated environments, we show it is possible to achieve strong transfer performance on held-out environments, having trained only inside a world model learned from a narrower dataset. We believe this opens the path to utilizing larger-scale, foundation world models for generally capable agents.

cs.AI↗

arXiv2025

Synthetic Data is Sufficient for Zero-Shot Visual Generalization from Offline Data

Offline reinforcement learning (RL) offers a promising framework for training agents using pre-collected datasets without the need for further environment interaction. However, policies trained on offline data often struggle to generalise due to limited exposure to diverse states. The complexity of visual data introduces additional challenges such as noise, distractions, and spurious correlations, which can misguide the policy and increase the risk of overfitting if the training data is not sufficiently diverse. Indeed, this makes it challenging to leverage vision-based offline data in training robust agents that can generalize to unseen environments. To solve this problem, we propose a simple approach generating additional synthetic training data. We propose a two-step process, first augmenting the originally collected offline data to improve zero-shot generalization by introducing diversity, then using a diffusion model to generate additional data in latent space. We test our method across both continuous action spaces (Visual D4RL) and discrete action spaces (Procgen), demonstrating that it significantly improves generalization without requiring any algorithmic changes to existing model-free offline RL methods. We show that our method not only increases the diversity of the training data but also significantly reduces the generalization gap at test time while maintaining computational efficiency. We believe this approach could fuel additional progress in generating synthetic data to train more general agents in the future.

cs.CV↗

arXiv2025

On Almost Surely Safe Alignment of Large Language Models at Inference-Time

We introduce a novel inference-time alignment approach for LLMs that aims to generate safe responses almost surely, i.e., with probability approaching one. Our approach models the generation of safe responses as a constrained Markov Decision Process (MDP) within the LLM's latent space. We augment a safety state that tracks the evolution of safety constraints and dynamically penalize unsafe generations to ensure the generation of safe responses. Consequently, we demonstrate formal safety guarantees w.r.t. the given cost model upon solving the MDP in the latent space with sufficiently large penalties. Building on this foundation, we propose InferenceGuard, a practical implementation that safely aligns LLMs without modifying the model weights. Empirically, we demonstrate that InferenceGuard effectively balances safety and task performance, outperforming existing inference-time alignment methods in generating safe and aligned responses. Our findings contribute to the advancement of safer LLM deployment through alignment at inference-time, thus presenting a promising alternative to resource-intensive, overfitting-prone alignment techniques like RLHF.

cs.LG↗

arXiv2025

Sample Efficient Preference Alignment in LLMs via Active Exploration

Preference-based feedback is important for many applications in machine learning where evaluation of a reward function is not feasible. Notable recent examples arise in preference alignment for large language models, including in reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO). For many applications of preference alignment, the cost of acquiring human feedback can be substantial. In this work, we take advantage of the fact that one can often choose contexts at which to obtain human feedback to most efficiently identify a good policy, and formalize the setting as an active contextual dueling bandit problem. We propose an active exploration algorithm to efficiently select the data and provide theoretical proof that it has a polynomial worst-case regret bound. We extend the setting and methodology for practical use in preference alignment of large language models. We provide two extensions, an online and an offline approach. Our method outperforms the baselines with limited samples of human preferences on several language models and four real-world datasets including two new datasets that we contribute to the literature.

cs.LG↗