ICML 2026 Papers

Direct Preference Optimization (DPO) is often tuned by selecting checkpoints with high preference margins. We show that this selection rule can fail. Across dense $\beta$ sweeps in three 7B open-weight families under controlled DPO recipes, the DPO preference margin can strongly anticorrelate with capability probes, with the strongest observed case in Llama logic probes (Pearson $r=-0.91$, $n=13$). In Mistral, matched-duration controls show that transient exposure to higher $\beta$ produces persistent arithmetic ($d_z = 3.73$, $p = 0.002$) and format ($d_z = 3.38$, $p = 0.003$) degradation relative to a duration-matched constant-$\beta$ run, while logic and sycophancy show no significant effect. An expanded 14-probe logic suite further shows that aggregate capability scores can hide stable opposition between probe clusters: direct-inference probes are consistently negative while fallacy-detection probes are consistently positive, with 13 of 14 probes sign-stable across 3 seeds and 3 $\beta$ values. Sensitivity profiles differ across architectures: Mistral shows elevated seed variance near $\beta \approx 10^{-2}$, Llama shows capability-specific rigidity, and Qwen trades off smoothly. These findings motivate probe-resolved $\beta$ sweeps rather than margin-based checkpoint selection.

Diffusion models for super-resolution (SR) produce high-quality visual results but require expensive computational costs. Despite the development of several methods to accelerate diffusion-based SR models, some (e.g., SinSR) fail to produce realistic perceptual details, while others (e.g., OSEDiff) may hallucinate non-existent structures. To overcome these issues, we present **RSD**, a new distillation method for ResShift. Our method is based on training the student network to produce images such that a new fake ResShift model trained on them will coincide with the teacher model. RSD achieves single-step restoration and outperforms the teacher by a noticeable margin in various perceptual metrics (LPIPS, CLIPIQA, MUSIQ). We show that our distillation method can surpass SinSR, the other distillation-based method for ResShift, making it on par with state-of-the-art diffusion SR distillation methods with limited computational costs in terms of perceptual quality. Compared to SR methods based on pre-trained text-to-image models, RSD produces competitive perceptual quality and requires fewer parameters, GPU memory, and training cost. We provide experimental results on various real-world and synthetic datasets, including RealSR, RealSet65, DRealSR, ImageNet, and DIV2K. We provide the code at https://github.com/Daniil-Selikhanovych/RSD.

LLM-as-a-judge ensembles are the standard paradigm for scalable evaluation, but their aggregation mechanisms suffer from a fundamental flaw: they implicitly assume that judges provide independent estimates of true quality. However, in practice, LLM judges exhibit correlated errors caused by shared latent confounders---such as verbosity, stylistic preferences, or training artifacts---causing standard aggregation rules like majority vote or averaging to provide little gain or even amplify systematic mistakes. To address this, we introduce CARE, a confounder-aware aggregation framework that explicitly models LLM judge scores as arising from both a latent true-quality signal and shared confounding factors. Rather than heuristically re-weighting judges, CARE separates quality from confounders without access to ground-truth labels. We provide theoretical guarantees for identifiability and finite-sample recovery under shared confounders, and we quantify the systematic bias incurred when aggregation models omit confounding latent factors. Across 12 public benchmarks spanning continuous scoring, binary classification, and pairwise preference settings, CARE improves aggregation accuracy, reducing error by up to 26.8%.

Large Language Models (LLMs) often produce hallucinated outputs, which limit their reliability in high-stakes applications. Conformal prediction can provide guarantees on the correctness and factuality of LLM outputs, but existing approaches rely on the exchangeability assumption, which rarely holds in online settings where user queries and interests change over time. To solve this problem, in this paper, we propose **PACE** (**P**roactive **A**daptive **C**onformal Inferenc**E**), a novel framework that sequentially updates the time-varying target miscoverage parameter with a dynamic step size to maintain valid coverage under online distribution shifts. PACE is motivated by the theoretical connections between expected miscoverage error and key factors such as distribution shifts and instantaneous parameter error. It integrates two complementary signals: (1) a proactive shift detection to estimate the magnitude of distribution shifts, and (2) a reactive error that scales updates according to the local coverage gap. Extensive experiments on synthetic and real-world datasets demonstrate that PACE consistently outperforms advanced adaptive baselines. It reduces the deviation from the target error rate by up to 60\% in QA tasks and accelerates coverage recovery by over 2.5x during abrupt shifts, ensuring stable factuality guarantees without compromising utility and stability.

Reinforcement Learning from Verifiable Rewards (RLVR) significantly enhances large language models (LLMs) reasoning but severely suffers from calibration degeneration, where models become excessively over-confident in incorrect answers. Previous studies devote to directly incorporating calibration objective into existing optimization target. However, our theoretical analysis demonstrates that there exists a fundamental gradient conflict between the optimization for maximizing policy accuracy and minimizing calibration error. Building on this insight, we propose DCPO, a simple yet effective framework that systematically decouples reasoning and calibration objectives. Extensive experiments demonstrate that our DCPO not only preserves accuracy on par with GRPO but also achieves the best calibration performance and substantially mitigates the over-confidence issue. Our study provides valuable insights and practical solution for more reliable LLM deployment.

Pruning is a practical approach to compress large language models (LLMs), but it can amplify text degeneration, especially repetition loops, even when perplexity and task accuracy remain largely unchanged. In this work, we present a token-level analysis of this failure mode by viewing decoding as a dynamical process that enters and persists in a small set of recurrent contexts. Our analysis decomposes degeneration into loop entry risk and loop persistence, and shows that persistence is controlled by the escape mass assigned to plausible alternatives within the token sampling set. Motivated by these findings, we propose two token-level guidance objectives for post-pruning fine-tuning. FOCUS reweights distillation toward high-confidence teacher regions to suppress leakage, while RePAIR uses onset-centered positive/negative continuation pairs with a margin loss to promote plausible alternatives and prevent early commitment to repetition loops. Experiments on open-ended continuation and instruction-based generation show that both methods consistently reduce repetition and improve generation quality.

Standard reward models typically predict scalar scores that fail to capture the multifaceted nature of response quality in non-verifiable domains, such as creative writing or open-ended instruction following. To address this limitation, we propose Rubric-ARM, a framework that jointly optimizes a rubric generator and a judge using reinforcement learning from preference feedback. Unlike existing methods that rely on static rubrics or disjoint training pipelines, our approach treats rubric generation as a latent action learned to maximize judgment accuracy. We introduce an alternating optimization strategy to mitigate the non-stationarity of simultaneous updates, providing theoretical analysis that demonstrates how this schedule reduces gradient variance during training. Extensive experiments show that Rubric-ARM achieves state-of-the-art performance among baselines on multiple benchmarks and significantly improves downstream policy alignment in both offline and online reinforcement learning settings.

Large language models frequently exhibit hallucinations: fluent and confident outputs that are factually incorrect or unsupported by the input context. While recent hallucination detection methods have explored various features derived from attention maps, the underlying mechanisms they exploit remain poorly understood. In this work, we propose SinkProbe, a hallucination detection method grounded in the observation that hallucinations are deeply entangled with attention sinks - tokens that accumulate disproportionate attention mass during generation - indicating a transition from distributed, input-grounded attention to compressed, prior-dominated computation. Importantly, although sink scores are computed solely from attention maps, we find that the classifier preferentially relies on sinks whose associated value vectors have large norms. Moreover, we show that previous methods implicitly depend on attention sinks by establishing their mathematical relationship to sink scores. Our findings yield a novel hallucination detection method grounded in theory that produces state-of-the-art results across popular datasets and LLMs.

Stochastic Variance Reduced Gradient (SVRG) and its variants aim to speed-up training by using gradient corrections. Originally proposed over a decade ago, these methods have never been connected to any Bayesian method at a fundamental level. Here, we fill this gap and derive surprising new connections of SVRG to a recently proposed Bayesian method called `posterior correction'. Our main contribution is to show that SVRG can be recovered as a special case of posterior-correction over isotropic-Gaussian posteriors. Novel extensions of SVRG are automatically obtained by using more flexible exponential-family posteriors. We derive two new such extensions by using Gaussian families: a Newton-like variant with novel Hessian corrections, and an Adam-like extension that scales to large problems. Our work is the first to connect SVRG to Bayes and use it to speed-up training.

Even for the gradient descent (GD) method applied to neural network training, understanding its optimization dynamics, including convergence rate, iterate trajectories, function value oscillations, and especially its implicit acceleration, remains a challenging problem. We analyze nonlinear models with the logistic loss and show that the steps of GD reduce to those of generalized perceptron algorithms (Rosenblatt, 1958), providing a new perspective on the dynamics. This reduction yields significantly simpler algorithmic steps, which we analyze using classical linear algebra tools. Using these tools, we demonstrate on a minimalistic example that the nonlinearity in a two-layer model can provably yield a faster iteration complexity $\tilde{\mathcal{O}}(\sqrt{d})$ compared to $\Omega(d)$ achieved by linear models, where $d$ is the number of features. This helps explain the optimization dynamics and the implicit acceleration phenomenon observed in neural networks. The theoretical results are supported by extensive numerical experiments. We believe that this alternative view will further advance research on the optimization of neural networks.

In modern applications such as ECG monitoring, neuroimaging, wearable sensing, and industrial equipment diagnostics, complex and continuously structured data are ubiquitous, presenting both challenges and opportunities for functional data analysis. However, existing methods face a critical trade-off: conventional functional models are limited by linearity, whereas deep learning approaches lack interpretable region selection for sparse effects. To bridge these gaps, we propose a sparse Bayesian functional deep neural network (sBayFDNN). It learns adaptive functional embeddings through a deep Bayesian architecture to capture complex nonlinear relationships, while a structured prior enables interpretable, region-wise selection of influential domains with quantified uncertainty. Theoretically, we establish rigorous approximation error bounds, posterior consistency, and region selection consistency. These results provide the first theoretical guarantees for a Bayesian deep functional model, ensuring its reliability and statistical rigor. Empirically, comprehensive simulations and real‑world studies confirm the effectiveness and superiority of sBayFDNN. Crucially, sBayFDNN excels in recognizing intricate dependencies for accurate predictions and more precisely identifies functionally meaningful regions, capabilities fundamentally beyond existing approaches.

Kleinberg and Mullainathan showed that language generation in the limit is always possible at the level of computability: given enough positive examples, a learner can eventually generate data indistinguishable from a target language. However, such existence results do not address feasibility. We study the sample complexity of language generation in the limit for several canonical classes of formal languages. Our results show that infeasibility already appears for context-free and regular languages, and persists even for strict subclasses such as locally threshold testable languages, as well as for incomparable classes such as non-erasing pattern languages, a well-studied class in the theory of language identification. Overall, our results establish a clear gap between the theoretical possibility of language generation in the limit and its computational feasibility.

Federated Graph Learning (FGL) enables distributed clients to collaboratively train graph neural networks while strictly preserving data privacy. However, existing FGL methods implicitly assume the reliability of local graph structures and lack elastic awareness of parameter importance during model aggregation, leading to representation degradation under topological noise and catastrophic forgetting caused by model drift. To address these challenges, we propose GraphP-FL, a general personalized FGL framework. (1) Specifically, we design a self-supervised dynamic topology reconstruction mechanism on the client side. This mechanism mines implicit dependencies to adaptively rectify noisy topologies, effectively suppressing topological noise propagation and capturing precise structural relationships for high-quality representations. (2) Additionally, we introduce a Fisher-based Elastic Parameter Alignment (FRPA) algorithm. FRPA imposes anisotropic regularization constraints in the parameter space to precisely quantify parameter importance, enabling the model to strictly preserve critical local knowledge while flexibly aligning with the global model, thus effectively overcoming catastrophic forgetting. Extensive experiments on seven benchmarks (including biochemical molecules, social networks, and large-scale encrypted traffic) demonstrate that GraphP-FL significantly outperforms state-of-the-art methods, improving accuracy by up to 8.6% while exhibiting superior generalization and robustness.

With the rapid development of deep learning, the issue of data scarcity has become increasingly prominent, inspiring emerging interests towards research on data augmentation techniques over recent years. However, our literature survey indicates that existing efforts often suffer from two issues of semantic infidelity, including: (i) visual semantics infidelity, such as visual artifacts, manifold intrusion, and unnatural blending boundaries etc, and (ii) label semantic infidelity, where augmented images do not match the original labels, creating extra label noises. To address these issues, we propose a Semantics Calibrated and Diffusion-Enhanced Augmentation (SCalDA) scheme to achieve accurate semantics calibration across image, label and feature domains. Compared with the existing approaches, our proposed features in precise guidance in label domain, semantics driven synthesis across three domains (image, label and feature), and semantics-aware metric learning. Extensive experiments on multiple datasets demonstrate that SCalDA yields consistent and significant performance improvements for both fine-grained and general classification tasks, validating the effectiveness and broad applicability of the proposed.

Multi-adapter serving systems route entire sequences to a single adapter, forcing a choice when requests span multiple domains. This assumption fails in two important settings: (1) multimodal generation, where text and image tokens require different adapters within the same sequence, and (2) mixed-capability requests like ``write code to solve this equation,'' which need expertise from multiple specialized adapters. We introduce \emph{per-token routing}, which routes individual tokens to adapters based on either vocabulary structure (for multimodal models) or learned gating (for semantic specialization). Per-token routing is provably optimal for mixed-adapter requests: $N$ work for $N$ tokens, versus $K \cdot N$ for per-sequence systems that must replay one adapter per pass. Our key contribution is MoLoRA (Mixture of LoRA), which enables \emph{composable specialization}: load multiple domain-specific adapters and let a learned router select the appropriate adapter per-token. We demonstrate that specialization dramatically beats scale: MoLoRA enables Qwen3-1.7B to exceed Qwen3-8B across four reasoning benchmarks while being 4.7$\times$ smaller. This enables modular expertise at inference time: train focused LoRAs independently, combine them without retraining, and add new capabilities by simply loading new adapters.

We present adversarial flow models, a class of generative models that belongs to both the adversarial and flow families. Our method supports native one-step and multi-step generation and is trained with an adversarial objective. Unlike traditional GANs, in which the generator learns an arbitrary transport map between the noise and data distributions, our generator is encouraged to learn a deterministic noise-to-data mapping. This significantly stabilizes adversarial training. Unlike consistency-based methods, our model directly learns one-step or few-step generation without having to learn the intermediate timesteps of the probability flow for propagation. This preserves model capacity and avoids error accumulation. Under the same 1NFE setting on ImageNet-256px, our B/2 model approaches the performance of consistency-based XL/2 models, while our XL/2 model achieves a new best FID of 2.38. We additionally demonstrate end-to-end training of 56-layer and 112-layer models without any intermediate supervision, achieving FIDs of 2.08 and 1.94 with a single forward pass and surpassing the corresponding 28-layer 2NFE and 4NFE counterparts with equal compute and parameters.

State Space Models (SSMs) provide linear-time alternatives to attention for vision, but require serializing 2D images into 1D sequences using a predefined scan order. We identify scan order as a previously underexplored inductive bias that fundamentally shapes spatial dependency modeling in Vision SSMs. Fixed scan paths distort local adjacency, fragment object structure, and induce anisotropic representations that are brittle under geometric transformations such as rotation. We propose Partial RIng Scan Mamba (PRIS-Mamba), a rotation-robust traversal that decomposes images into concentric rings, performs permutation-invariant aggregation within each ring, and models cross-ring dependencies via short radial SSMs. This design induces a structured factorization of spatial dependencies that preserves isotropy while maintaining linear complexity. To improve efficiency without sacrificing expressivity, we introduce partial channel filtering, selectively applying recurrent modeling to informative channels while routing others through a residual pathway. Empirically, PRIS-Mamba improves accuracy, efficiency, and rotation robustness over prior Vision SSMs on ImageNet-1K. Our results position scan-order design as a core representational choice in Vision SSMs, with implications for robustness and generalization beyond architectural scaling. The code will be released upon paper acceptance.

Relational prediction tasks are fundamental in many real-world applications, where data are naturally stored in relational databases (RDBs). Relational Deep Learning (RDL) addresses this problem by modeling RDBs as graphs and applying graph neural networks (GNNs) for end-to-end learning. However, the full-resolution property is commonly adopted as a design principle in graph construction for RDBs to preserve relational semantics, which leads most existing methods to rely on fixed graph structures. In this paper, we propose **FROG**, a *F*ull-*R*esolution and *O*ptimizable *G*raph Structure Learning framework for RDL that formulates relational structure learning as a learnable table role modeling problem, allowing tables to contribute as nodes and edges in message passing. We further design role-driven message passing mechanisms to capture relational semantics, enabling joint optimization of graph structure and GNN representations. To ensure semantic consistency, we introduce functional dependency constraints that regularize representations across table and entity levels. Extensive experiments demonstrate that our method outperforms existing approaches and reveal how table roles impact downstream tasks, offering new insights into graph construction for RDL.

Research on continual learning for LLMs seeks to acquire new skills without catastrophic forgetting of established prior knowledge. However, domain-specific fine-tuning still triggers severe, long-tailed forgetting issues even under narrow updates, particularly when the pre-training data is inaccessible. To tackle this challenge, we propose **ALER**, a data-free continual learning framework that adversarially searches for a small set of latent prompt embeddings to maximize logit divergence from a frozen reference model, proactively exposing high-risk forgetting modes at each step. It then performs online distillation from the frozen reference using the discovered embeddings to retain prior behaviors while preserving target-domain adaptation. We provide theoretical guarantees on the efficiency of our targeted repair, and extensive experiments demonstrate consistent improvements in the retention–adaptation frontier over representative baselines across $2$ domain-specific fine-tuning datasets and $6$ general-purpose benchmarks, suggesting a more proactive approach for LLM continual learning.