ICLR 2026 Papers

Federated Learning (FL) enables clients to collaboratively train a global model without sharing their private data. However, the presence of malicious (Byzantine) clients poses significant challenges to the robustness of FL, particularly when data distributions across clients are heterogeneous. In this paper, we propose a novel Byzantine-robust FL optimization problem that incorporates adaptive weighting into the aggregation process. Unlike conventional approaches, our formulation treats aggregation weights as learnable parameters, jointly optimizing them alongside the global model parameters. To solve this optimization problem, we develop an alternating minimization algorithm with strong convergence guarantees under adversarial attack. We analyze the Byzantine resilience of the proposed objective. We evaluate the performance of our algorithm against state-of-the-art Byzantine-robust FL approaches across various datasets and attack scenarios. Experimental results demonstrate that our method consistently outperforms existing approaches, particularly in settings with highly heterogeneous data and a large proportion of malicious clients.

Recent neural solvers have achieved strong performance on vehicle routing problems (VRPs), yet they mainly assume symmetric Euclidean distances, restricting applicability to real-world scenarios. A core challenge is encoding the relational features in asymmetric distance matrices of VRPs. Early attempts directly encoded these matrices but often failed to produce compact embeddings and generalized poorly at scale. In this paper, we propose RADAR, a scalable neural framework that augments existing neural VRP solvers with the ability to handle asymmetric inputs. RADAR addresses asymmetry from both static and dynamic perspectives. It leverages Singular Value Decomposition (SVD) on the asymmetric distance matrix to initialize compact and generalizable embeddings that inherently encode the *static asymmetry* in the inbound and outbound costs of each node. To further model *dynamic asymmetry* in embedding interactions during encoding, it replaces the standard softmax with Sinkhorn normalization that imposes joint row and column distance awareness in attention weights. Extensive experiments on synthetic and real-world benchmarks across various VRPs show that RADAR outperforms strong baselines on both in-distribution and out-of-distribution instances, demonstrating robust generalization and superior performance in solving asymmetric VRPs.

Diffusion generative models have demonstrated remarkable empirical success in the recent years and are now considered the state-of-the-art generative models in modern AI. These models consist of a forward process, which gradually diffuses the data distribution to a noise distribution spanning the whole space, and a backward process, which inverts this transformation to recover the data distribution from noise. Most of the existing literature assumes that the underlying space is Euclidean. However, in many practical applications, the data are constrained to lie on a submanifold of Euclidean space. Addressing this setting, de Bortoli et al. (2022) introduced Riemannian diffusion models and proved that using an exponentially small step size yields small sampling error in Wasserstein distance, provided the data distribution is smooth and strictly positive. In this paper, we prove that a *polynomially small stepsize* suffices to guarantee small sampling error in total variation distance, without any assumption on the smoothness or positivity of the data distribution. Our analysis only requires mild and standard curvature assumptions on the underlying manifold. The main ingredients in our proof are Li-Yau estimate for log-gradient of heat kernel, and Minakshisundaram-Pleijel parametrix expansion for perturbed heat equation. Our approach opens the door to a sharper analysis of diffusion models on non-Euclidean spaces.

Zeroth-order (ZO) optimization enables dimension-free communication in federated learning (FL), making it attractive for fine-tuning of large language models (LLMs) due to significant communication savings. However, existing ZO-FL methods largely overlook curvature information, despite its well-established benefits for convergence acceleration. To address this, we propose **HiSo**, a Hessian-informed ZO federated optimization method that accelerates convergence by leveraging global diagonal Hessian approximations, while strictly preserving scalar-only communication **without transmitting any second-order information**. Theoretically, for non-convex functions, we show that HiSo can achieve an accelerated convergence rate that is independent of the Lipschitz constant $L$ and model dimension $d$ under some Hessian approximation assumptions, offering a plausible explanation for the observed phenomenon of ZO convergence being much faster than its worst-case $O(d)$-bound. Empirically, across diverse LLM fine-tuning benchmarks, HiSo delivers a 1$\sim$5× speedup in communication rounds over existing state-of-the-art ZO-FL baselines. This superior convergence not only cuts communication costs but also provides strong empirical evidence that Hessian information acts as an effective accelerator in federated ZO optimization settings.

We analyze Muon as originally proposed and used in practice---using the momentum orthogonalization with a few Newton-Schulz steps. The prior theoretical results replace this key step in Muon with an exact SVD-based polar factor. We prove that Muon with Newton-Schulz converges to a stationary point with the same rate as the SVD-polar idealization, up to a constant factor for given the number of Newton-Schulz steps $q$. We further analyze this constant factor, and prove that it converges to 1 doubly exponentially in $q$ and improves with $\kappa$, which is the degree of a polynomial used in Newton-Schulz required when approximating the orthogonalization direction. We also prove that Muon removes the typical square-root-of-rank loss compared to its vector-based counterpart, SGD with momentum. Our results explain why Muon with a few low-degree Newton-Schulz steps matches exact-polar (SVD) behavior at much faster wall-clock time, and explain how much momentum matrix orthogonalization via Newton-Schulz benefits over the vector-based optimizer. Overall, our theory justifies the practical Newton-Schulz design of Muon, narrowing its practice–theory gap.

We introduce fast algorithms for solving $\ell_{p}$ regression problems using the iteratively reweighted least squares (IRLS) method. Our approach achieves state-of-the-art iteration complexity, outperforming the IRLS algorithm by Adil-Peng-Sachdeva (NeurIPS 2019) and matching the theoretical bounds established by the complex algorithm of Adil-Kyng-Peng-Sachdeva (SODA 2019, J. ACM 2024) via a simpler lightweight iterative scheme. This bridges the existing gap between theoretical and practical algorithms for $\ell_{p}$ regression. Our algorithms depart from prior approaches, using a primal-dual framework, in which the update rule can be naturally derived from an invariant maintained for the dual objective. Empirically, we show that our algorithms significantly outperform both the IRLS algorithm by Adil-Peng-Sachdeva and MATLAB/CVX implementations.

Previous work has explored the computational complexity of deriving two fundamental types of explanations for ML model predictions: (1) *sufficient reasons*, which are subsets of input features that, when fixed, determine a prediction, and (2) *contrastive reasons*, which are subsets of input features that, when modified, alter a prediction. Prior studies have examined these explanations in different contexts, such as non-probabilistic versus probabilistic frameworks and local versus global settings. In this study, we introduce a unified framework for analyzing these explanations, demonstrating that they can all be characterized through the minimization of a unified probabilistic value function. We then prove that the complexity of these computations is influenced by three key properties of the value function: (1) *monotonicity*, (2) *submodularity*, and (3) *supermodularity* - which are three fundamental properties in *combinatorial optimization*. Our findings uncover some counterintuitive results regarding the nature of these properties within the explanation settings examined. For instance, although the *local* value functions do not exhibit monotonicity or submodularity/supermodularity whatsoever, we demonstrate that the *global* value functions do possess these properties. This distinction enables us to prove a series of novel polynomial-time results for computing various explanations with provable guarantees in the global explainability setting, across a range of ML models that span the interpretability spectrum, such as neural networks, decision trees, and tree ensembles. In contrast, we show that even highly simplified versions of these explanations become NP-hard to compute in the corresponding local explainability setting.

Deep neural networks rely on backpropagation (BP) for optimization, but its strictly sequential backward pass hinders parallelism and scalability. Direct Feedback Alignment (DFA) has been proposed as a promising approach for parallel learning of deep neural networks, relying on fixed random projections to enable layer-wise parallel updates, but fails on deep convolutional networks, and performs poorly on modern transformer architectures. We introduce GrAPE (Gradient-Aligned Projected Error), a hybrid feedback-alignment method that (i) estimates rank-1 Jacobians via forward-mode JVPs and (ii) aligns each layer’s feedback matrix by minimizing a local cosine-alignment loss. To curb drift in very deep models, GrAPE performs infrequent BP anchor steps on a single mini-batch, preserving mostly parallel updates. We show that the forward-gradient estimator has strictly positive expected cosine with the true Jacobian and, inspired by Zoutendijk-style arguments, derive a convergence-in-expectation result under a positive expected-cosine condition. Empirically, GrAPE consistently outperforms prior alternatives to BP, enabling the training of modern architectures, closing a large fraction of the gap to BP while retaining layer-parallel updates for the vast majority of steps.

Direct Preference Optimization (DPO) has become the de facto standard for offline preference alignment of large language models, but its reliance on a reference policy introduces a critical tension. DPO weighs each update relative to a reference, which stabilizes the training by regularizing the updates within a trusted region. This reliance becomes problematic for pessimistic pairs, where the reference model prefers the rejected response. For these pairs, DPO prematurely attenuates the gradient as soon as the policy margin ($\Delta_\theta$) merely beats the reference margin ($\Delta_{\mathrm{ref}}$) even if the policy is still wrong ($\Delta_{\theta}<0$). We name this failure premature satisfaction, which is a concrete form of the training–inference mismatch. Reference-free objectives remove this mismatch by optimizing the absolute margin, but at the cost of discarding the stabilizing signal of the reference. We mitigate this tension with Hybrid-DPO (HyPO), a drop-in modification to DPO that applies reference conditionally: HyPO behaves exactly like DPO when the reference is optimistic or neutral, and it treats the reference as neutral when it is pessimistic by replacing $\Delta_\theta-\Delta_{\mathrm{ref}}$ with $\Delta_\theta-\max\{0,\Delta_{\mathrm{ref}}\}$. This one-line change strictly strengthens per-example learning signals on pessimistic pairs while preserving DPO’s objective form and computational cost. By conditionally debiasing the pessimistic reference signal, HyPO mitigates premature satisfaction; empirically, across preference alignment, HyPO improves inference-aligned metrics and achieves higher pairwise win rates. Our results provide evidence that direct preference alignment could be enhanced by conditionally debiasing the reference signal, rather than discarding it.

Binary Neural Networks (BNNs), which constrain both weights and activations to binary values, offer substantial reductions in computational complexity, memory footprint, and energy consumption. These advantages make them particularly well suited for deployment on resource-constrained devices. However, training BNNs via gradient-based optimization remains challenging due to the discrete nature of their variables. The dominant approach, quantization-aware training, circumvents this issue by employing surrogate gradients. Yet, this method requires maintaining latent full-precision parameters and performing the backward pass with floating-point arithmetic, thereby forfeiting the efficiency of binary operations during training. While alternative approaches based on local learning rules exist, they are unsuitable for global credit assignment and for back-propagating errors in multi-layer architectures. This paper introduces Binary Error Propagation (BEP), the first learning algorithm to establish a principled, discrete analog of the backpropagation chain rule. This mechanism enables error signals, represented as binary vectors, to be propagated backward through multiple layers of a neural network. BEP operates entirely on binary variables, with all forward and backward computations performed using only bitwise operations. Crucially, this makes BEP the first solution to enable end-to-end binary training for recurrent neural network architectures. We validate the effectiveness of BEP on both multi-layer perceptrons and recurrent neural networks, demonstrating gains of up to $+6.89$% and $+10.57$% in test accuracy, respectively. The proposed algorithm is released as an open-source repository.

Standard human preference-based alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), are a cornerstone technology for aligning Large Language Models (LLMs) with human values. However, these methods are all underpinned by a strong assumption that the collected preference data is clean and that all observed labels are equally reliable. In reality, large-scale preference datasets contain substantial label noise due to annotator errors, inconsistent instructions, varying expertise, and even adversarial or low-effort feedback. This creates a discrepancy between the recorded data and the ground-truth preferences, which can misguide the model and degrade its performance. To address this challenge, we introduce Robust Preference Optimization (RPO). RPO employs an Expectation-Maximization algorithm to infer the posterior probability of each label’s correctness, which is used to adaptively re-weigh each data point in the training loss to mitigate noise. We further generalize this approach by establishing a theoretical link between arbitrary preference losses and their corresponding probabilistic models. This generalization enables the systematic transformation of existing alignment algorithms into their robust counterparts, elevating RPO from a specific algorithm to a general framework for robust preference alignment. Theoretically, we prove that under the condition of a perfectly calibrated model, RPO is guaranteed to converge to the true noise level of the dataset. Our experiments demonstrate RPO’s effectiveness as a general framework, consistently enhancing four state-of-the-art alignment algorithms (DPO, IPO, SimPO, and CPO). When applied to Mistral and Llama 3 models, the RPO-enhanced methods improve AlpacaEval 2 win rates by up to 7.0 percentage points over their respective baselines.

Adaptive optimizers can reduce to normalized steepest descent (NSD) when only adapting to the current gradient, suggesting a close connection between the two algorithmic families. A key distinction in their analyses, however, lies in the smoothness assumptions they rely on. In the convex setting, adaptive optimizers are governed by a stronger adaptive smoothness condition, while NSD relies on the standard notion of smoothness. We extend the theory of adaptive smoothness to the nonconvex setting and show that it precisely characterizes the convergence of adaptive optimizers. Moreover, we establish that adaptive smoothness enables acceleration of adaptive optimizers with Nesterov momentum, a guarantee unattainable under standard smoothness. We further develop an analogous comparison for stochastic optimization by introducing adaptive variance, which parallels adaptive smoothness and leads to qualitatively stronger guarantees than the standard variance.

We study the alternating gradient descent-ascent (AltGDA) algorithm in two-player zero-sum games. Alternating methods, where players take turns to update their strategies, have long been recognized as simple and practical approaches for learning in games, exhibiting much better numerical performance than their simultaneous counterparts. However, our theoretical understanding of alternating algorithms remains limited, and results are mostly restricted to the unconstrained setting. We show that for two-player zero-sum games that admit an interior Nash equilibrium, AltGDA converges at an $O(1/T)$ ergodic convergence rate when employing a small constant stepsize. This is the first result showing that alternation improves over the simultaneous counterpart of GDA in the constrained setting. For games without an interior equilibrium, we show an $O(1/T)$ local convergence rate with a constant stepsize that is independent of any game-specific constants. In a more general setting, we develop a performance estimation programming (PEP) framework to jointly optimize the AltGDA stepsize along with its worst-case convergence rate. The PEP results indicate that AltGDA may achieve an $O(1/T)$ convergence rate for a finite horizon $T$, whereas its simultaneous counterpart appears limited to an $O(1/\sqrt{T})$ rate.

Learned optimizers are powerful alternatives to hand-designed rules like Adam, yet they have seen limited practical adoption since they often fail to meta-generalize beyond their training distribution and incur high meta-training cost. For instance, prior work, VeLO, scaled meta-training to 4,000 TPU months ($\sim$10$\times$ GPT-3 compute) to meta-train a general-purpose optimizer but it failed to generalize beyond 600M parameters tasks. In this work, we present a surprising finding: by crafting a simple normalized optimizer architecture, and augmenting meta-training, it becomes feasible to meta-train a performant general-purpose learned update rule on a tiny fraction of VeLO compute, 4.5 GPU hours to be precise. Our learned update rule scales stably to a billion-scale pretraining task (GPT-3 XL 1.3B) which is six orders of magnitude larger than its meta-training distribution. Furthermore, it shows strong performance across diverse out-of-distribution ML workloads and is compatible with modern optimization harness that includes orthogonalization, distinct update rules for input-output and hidden weights, and decoupled weight decay. In all, this work paves the way for practically applicable _learnable_ optimization algorithms, unlocking exploration of richer meta-training and data curation recipes to further improve performance.

Attention-based regression models are often trained by jointly optimizing Mean Squared Error (MSE) loss and Pearson correlation coefficient (PCC) loss, emphasizing the magnitude of errors and the order or shape of targets, respectively. A common but poorly understood phenomenon during training is the *PCC plateau*: PCC stops improving early in training, even as MSE continues to decrease. We provide the first rigorous theoretical analysis of this behavior, revealing fundamental limitations in both optimization dynamics and model capacity. First, in regard to the flattened PCC curve, we uncover a critical conflict where lowering MSE (magnitude matching) can *paradoxically* suppress the PCC gradient (shape matching). This issue is exacerbated by the softmax attention mechanism, particularly when the data to be aggregated is highly homogeneous. Second, we identify a limitation in the model capacity: we derived a PCC improvement limit for *any* convex aggregator (including the softmax attention), showing that the convex hull of the inputs strictly bounds the achievable PCC gain. We demonstrate that data homogeneity intensifies both limitations. Motivated by these insights, we propose the Extrapolative Correlation Attention (ECA), which incorporates novel, theoretically-motivated mechanisms to improve the PCC optimization and extrapolate beyond the convex hull. Across diverse benchmarks, including challenging homogeneous data setting, ECA consistently breaks the PCC plateau, achieving significant improvements in correlation without compromising MSE performance.

Masked diffusion large language models (dLLMs) are emerging as promising alternatives to autoregressive LLMs, offering competitive performance while supporting unique generation capabilities such as inpainting. We explore how inpainting can inform RL algorithm design for dLLMs. Aligning LLMs with reinforcement learning faces an exploration challenge: sparse reward signals and sample waste when models fail to discover correct solutions. While this inefficiency affects LLMs broadly, dLLMs offer a distinctive opportunity—their inpainting ability can guide exploration. We introduce IGPO (Inpainting Guided Policy Optimization), an RL framework that strategically inserts partial ground-truth reasoning traces during online sampling. Unlike providing full solutions, inpainting steers exploration toward promising trajectory spaces while preserving self-generated reasoning, bridging supervised fine-tuning and reinforcement learning. We apply IGPO to group-based optimization methods such as GRPO, where exploration failures cause zero advantages and gradients. IGPO restores meaningful gradients while improving sample efficiency. We also propose supervised fine-tuning on synthetically rewritten concise traces that better align with dLLM generation patterns. With additional techniques including entropy-based filtering, our training recipe yields substantial gains across four mathematical benchmarks—GSM8K, Math500, AMC and Minerva—achieving new state-of-the-art results for full-attention masked dLLMs.

Discrete diffusion models have demonstrated great promise in modeling various sequence data, ranging from human language to biological sequences. Inspired by the success of RL in language models, there is growing interest in further improving the models by alignment with a certain reward. In this work, we propose an offline preference optimization method to approach trajectory alignment for discrete diffusion models. Instead of applying the reward on the final output and backpropagating the gradient to the entire denoising process, we decompose the problem into a set of stepwise alignment objectives by matching the per-step posterior. This framework enables efficient diffusion optimization, is compatible with arbitrary reward functions, and importantly, yields an equivalent optimal solution under additive factorization of the trajectory reward. Experiments across multiple domains including DNA sequence design, protein inverse folding, and language modeling consistently demonstrate the superiority of our approach. Notably, it achieves an up to 12\% improvement over the most competitive RL-based baseline in terms of predicted activity on DNA sequence design, and further improves the GSM8K score from 78.6 to 81.2 on LLaDA-8B-Instruct for language modeling.

Speculative sampling reduces the latency of autoregressive decoding for target model LLMs without sacrificing inference quality, by using a cheap draft model to suggest a candidate token and a verification criterion to accept or resample this token. To improve acceptance and decoding efficiency, recent work has explored the multi-draft extension, where at each step $n$ draft tokens are generated, and the verification criterion is a distribution conditioned on these. When this criterion maximizes the probability of accepting some draft token, it is called the optimal transport (OT). However, finding the OT is difficult, as it is the solution of a linear program (OTLP) in over $V^n$ variables, with $V$ being the vocabulary size. Two recent theoretical works have reframed the OTLP in terms of importance sampling or subset selection. In this work, we prove that these formulations are equivalent to an exponentially large relaxed OTLP, so it remains infeasible to solve. Then, we reverse engineer subset selection to formulate the OTLP as a max-flow problem. With a novel application of polymatroid theory, we reduce the exponentially large OTLP to a convex optimization problem in at most $V$ variables. This allows us to devise an algorithm for optimal $n$-draft speculative sampling when the $n$ tokens are chosen i.i.d. from a single draft model, which can be tuned to arbitrary accuracy. Finally, we measure acceptance rates and algorithm runtimes for various $n$ and top-$k$ draft sampling settings. Our findings give the first multi-draft algorithm with 90\% acceptance and under 100 ms of overhead per generated token with negligible deviation from the target model distribution.

The field of preference optimization has made outstanding contributions to the alignment of language models with human preferences. Despite these advancements, recent methods still rely heavily on substantial paired (labeled) feedback data, leading to substantial resource expenditures. To address these challenges, we study the problem of Semi-Supervised Preference Optimization in which the idea is to learn from both a small number of pairwise preference labels and a large pool of unpaired samples simultaneously. Our key theoretical contribution proves the existence of an optimal reward threshold capable of separating winning and losing responses with high probability, which enables a principled pseudo-labeling of unpaired data. By leveraging these pseudo-labels, SSPO effectively distills latent preferences from large-scale unpaired data, thus maintaining human alignment while drastically reducing acquisition costs. Extensive experiments across datasets validate this remarkable data efficiency; for instance, SSPO trained with Mistral-7B-Instruct on just 1% of UltraFeedback consistently surpasses strong baselines trained on 10% of UltraFeedback.