ICLR 2026 Papers

Efficient deployment of deep neural networks increasingly relies on Post-Training Quantization (PTQ). Logarithmic PTQ, in particular, promises multiplier-free hardware efficiency, but its performance is often limited by the nonlinear and symmetric quantization grid and standard rounding-to-nearest (RTN) approach. While learnable rounding has significantly advanced linear PTQ, its application to the non-linear and often discrete nature of logarithmic domain remains unexplored. This paper introduces learnable Logarithmic Adaptive Rounding Techniques (LogART) that pioneer task-aware learnable rounding specifically for the logarithmic domain. LogART further extends the learnable rounding strategy to flexibly support outlier-aware, asymmetric, and hardware-friendly dynamic logarithmic bases, determined in a distribution-aware manner using an efficient search strategy. Extensive experiments demonstrate that LogART achieves state-of-the-art accuracy while maintaining efficiency in quantizing models across various architectures and ultra-low bitwidths, outperforming existing logarithmic PTQ methods and paving the way for more effective hardware deployment. The code is available at https://github.com/logart-lab/logart.

Diffusion-based language models (dLLMs) have emerged as a promising alternative to traditional autoregressive LLMs by enabling parallel token generation and significantly reducing inference latency. However, existing sampling strategies for dLLMs, such as confidence-based or semi-autoregressive decoding, often suffer from static behavior, leading to suboptimal efficiency and limited flexibility. In this paper, we propose SlowFast Sampling, a novel dynamic sampling strategy that adaptively alternates between exploratory and accelerated decoding stages. Our method is guided by three golden principles: certainty principle, convergence principle, and positional principle, which govern when and where tokens can be confidently and efficiently decoded. We further integrate our strategy with dLLM-Cache to reduce redundant computation. Extensive experiments across benchmarks and models show that SlowFast Sampling achieves up to 15.63× speedup on LLaDA with minimal accuracy drop, and up to 34.22× when combined with caching. Notably, our approach outperforms strong autoregressive baselines like LLaMA3 8B in throughput, demonstrating that well-designed sampling can unlock the full potential of dLLMs for fast and high-quality generation.

Current automated agent design frameworks produce either static workflows that lack adaptability or per-query optimizers that prevent the accumulation of deep, agent-level task expertise. We propose a new direction that reconciles these paradigms: creating stateful teams of specialist agents that accumulate knowledge over time and can be reconfigured for novel tasks entirely without human intervention. To this end, we introduce \textsc{ASpec}, a framework that manages this full agent lifecycle by first autonomously $\textbf{discovering}$ specialist archetypes via evolutionary search and then $\textbf{cultivating}$ their expertise through experience, mirroring how human experts learn through practice and reflection. We further introduce a lightweight hierarchical control policy, "retain-then-escalate," which governs when to leverage the established agent system versus when to adapt its structure. Through comprehensive experiments, we demonstrate that this approach leads to significant performance gains on expert-level scientific benchmarks like GPQA while matching the state-of-the-art on broader domain tasks, demonstrating a promising path toward agent systems that are simultaneously expert, adaptive, and efficient.

Diffusion-based large language models (dLLMs), despite their promising performance, still suffer from inferior inference efficiency. This is because dLLMs rely on bidirectional attention and cannot directly benefit from the standard key-value (KV) cache as autoregressive models (ARMs) do. To tackle this issue, we introduce \textit{Dual aDaptive Cache} (d$^2$Cache), which is a training-free approximate KV cache framework for accelerating dLLM inference. d$^2$Cache features a two-stage fine-grained selection strategy to identify tokens and adaptively update their KV states at each decoding step, while caching the KV states of the remaining tokens for reuse. Furthermore, d$^2$Cache naturally offers a more reliable decoding alternative, which can enable quasi left-to-right generation and mitigate premature overconfidence in tokens at the end of the sequence. Extensive experimental results on two representative dLLMs (\ie, LLaDA and Dream) demonstrate that d$^2$Cache not only achieves substantial inference speedups, but also yields consistent improvements in generation quality. The anonymous evaluation codes are available at \url{https://anonymous.4open.science/r/d2Cache-5538}.

Diffusion Large Language Models (dLLMs) offer a promising alternative to autoregressive models, excelling in text generation tasks due to their bidirectional attention mechanisms. However, their computational complexity, scaling as $\mathcal{O}(L^3)$ with sequence length $L$, poses significant challenges for long-sequence and real-time applications, primarily due to the lack of compatibility with key-value caching and the non-autoregressive nature of denoising steps. Existing acceleration methods rely on static caching or parallel decoding strategies, which fail to account for the dynamic behavior of token properties across layers and decoding steps. We propose \textbf{Dynamic-dLLM}, a training-free framework that enhances dLLM inference efficiency through two components: Dynamic Cache Updating (DCU), which adaptively allocates cache-update budgets based on layer-wise token dynamics, and Adaptive Parallel Decoding (APD), which dynamically calibrates decoding thresholds to balance generation quality and efficiency. Extensive experiments on models like LLaDA-8B-Instruct, LLaDA-1.5, and Dream-v0-7B-Instruct across benchmarks such as MMLU, GSM8K, and HumanEval demonstrate that Dynamic-dLLM significantly improves inference speed, attaining an average speedup of exceeding 3$\times$ while maintaining performance. Dynamic-dLLM outperforms state-of-the-art acceleration methods and provides a plug-and-play solution for efficient dLLM deployment without compromising performance. Code and models will be made publicly available.

Post-hoc explanations aim at understanding which input features (or groups thereof) are the most impactful toward certain model decisions. Many such methods have been proposed (ArchAttribute, Occlusion, SHAP, RISE, LIME, Integrated Gradient) and it is hard for practitioners to understand the differences between them. Even worse, faithfulness metrics, often used to quantitatively compare explanation methods, also exhibit inconsistencies. To address these issues, recent work has unified explanation methods through the lens of Functional Decomposition. We extend such work to scenarios where input features are partitioned into groups (e.g. pixel patches) and prove that disagreements between explanation methods and faithfulness metrics are caused by between-group interactions. Crucially, getting rid of between-group interactions would lead to a single explanation that is optimal according to all faithfulness metrics. We finally show how to reduce the disagreements by adaptively grouping features/pixels on tabular/image data.

The one-epoch overfitting problem has drawn widespread attention, especially in CTR and CVR estimation models in search, advertising, and recommendation domains. These models which rely heavily on large-scale sparse categorical features, often suffer a significant decline in performance when trained for multiple epochs. Although recent studies have proposed heuristic solutions, the fundamental cause of this phenomenon remains unclear. In this work, we present a theoretical explanation grounded in Rademacher complexity, supported by empirical experiments, to explain why overfitting occurs in models with large-scale sparse categorical features. Based on this analysis, we propose a regularization method that constrains the norm budget of embedding layers adaptively. Our approach not only prevents the severe performance degradation observed during multi-epoch training, but also improves model performance within a single epoch. This method has already been deployed in online production systems.

Analyzing dynamic tensor streams is fundamentally challenged by complex, evolving temporal dynamics and the need to identify informative data from high-velocity streams. Existing methods often lack the expressiveness to model multi-scale temporal dependencies, limiting their ability to capture evolving patterns. We propose SONATA, a novel framework that unifies expressive dynamic embedding modeling with adaptive coreset selection. SONATA leverages principled machine learning techniques for efficient evaluation of each observation for uncertainty, novelty, influence, and information gain, and dynamically prioritizes learning from the most valuable data using Bellman-inspired optimization. Entity dynamics are modeled with Linear Dynamical Systems and expressive temporal kernels for fine-grained temporal representation. Experiments on synthetic and real-world datasets show that SONATA consistently outperforms state-of-the-art methods in modeling complex temporal patterns and improving predictive accuracy for dynamic tensor streams.

Accurately solving partial differential equations (PDEs) on arbitrary geometries and a variety of meshes is an important task in science and engineering applications. In this paper, we propose Adaptive Fourier Mamba Operators (AFMO), which integrates reproducing kernels for state-space models (SSMs) rather than the kernel integral formulation of SSMs. This is achieved by constructing Takenaka-Malmquist systems for the PDEs. AFMO offers new representations that align well with the adaptive Fourier decomposition (AFD) theory and can approximate the solution manifold of PDEs on a wide range of geometries and meshes. In several challenging benchmark PDE problems in the field of fluid physics, solid physics, and finance on point clouds, structured meshes, regular grids, and irregular domains, AFMO consistently outperforms state-of-the-art solvers in terms of relative $L^2$ error. Overall, this work presents a new paradigm for designing explainable neural operator frameworks.

Fine-tuning large language models (LLMs) often faces GPU memory bottlenecks: the backward pass of first-order optimizers like Adam increases memory usage to more than 10 times the inference level (e.g., 633~GB for OPT-30B). Zeroth-order (ZO) optimizers avoid this cost by estimating gradients only from forward passes, yet existing methods like MeZO usually need tens of times more steps to converge. Can this trade-off between speed and memory in ZO be fundamentally improved? Normalized-SGD, for instance, demonstrates strong empirical performance with greater memory efficiency than Adam. In light of this, we introduce FZOO, a Fast Zeroth-Order Optimizer towards Adam-Scale Speed. On the one hand, FZOO reduces the total forward passes needed for convergence by employing batched one-sided estimates that adapt step-sizes based on the standard deviation of batch losses. On the other hand, it accelerates per-batch computation through the use of Rademacher random vector (±1) perturbations, which also enables further speedups through batched evaluation. Extensive experiments on diverse models (including RoBERTa-large, the OPT family (350M-66B), Phi-2, and Llama3) across 11 varied downstream tasks validate FZOO's effectiveness. On average, FZOO outperforms MeZO by +3% in accuracy while requiring 3$\times$fewer forward passes. Notably, for the RoBERTa-large model, FZOO achieves average improvements of +5.6% in accuracy and 18$\times$reduction in forward passes compared to MeZO, achieving convergence speeds comparable to Adam. We also provide theoretical analysis proving FZOO’s formal equivalence to a normalized-SGD update rule and establishing its convergence guarantees. Beyond full-parameter tuning, FZOO plugs smoothly into PEFT techniques, unlocking even larger memory savings. Taken together, our results make single-GPU, high-speed, full-parameter fine-tuning realistic today and point toward future work on memory-efficient pre-training. Code: https://github.com/DKmiyan/FZOO

The long-output context generation of large reasoning models enables extended chain of thought (CoT) but also drives rapid growth of the key–value (KV) cache, quickly overwhelming GPU memory. To address this challenge, we propose ThinKV, a thought-adaptive KV cache compression framework. ThinKV is based on the observation that attention sparsity reveals distinct thought types with varying importance within the CoT. It applies a hybrid quantization–eviction strategy, assigning token precision by thought importance and progressively evicting tokens from less critical thoughts as reasoning trajectories evolve. Furthermore, to implement ThinKV, we design a kernel that extends PagedAttention to enable efficient reuse of evicted tokens' memory slots, eliminating compaction overheads. Extensive experiments on DeepSeek-R1-Distill, GPT-OSS, and NVIDIA AceReason across mathematics and coding benchmarks show that ThinKV achieves near-lossless accuracy with less than 5% of the original KV cache, while improving performance with up to 5.8x higher inference throughput over SoTA baselines.

Adam [Kingma & Ba, 2015] is the de facto optimizer in deep learning, yet its theoretical understanding remains limited. Prior analyses show that Adam favors solutions aligned with $\ell_\infty$-geometry, but these results are restricted to the full-batch regime. In this work, we study the implicit bias of incremental Adam (using one sample per step) for logistic regression on linearly separable data, and show that its bias can deviate from the full-batch behavior. As an extreme example, we construct datasets on which incremental Adam provably converges to the $\ell_2$-max-margin classifier, in contrast to the $\ell_\infty$-max-margin bias of full-batch Adam. For general datasets, we characterize its bias using a proxy algorithm for the $\beta_2 \to 1$ limit. This proxy maximizes a data-adaptive Mahalanobis-norm margin, whose associated covariance matrix is determined by a data-dependent dual fixed-point formulation. We further present concrete datasets where this bias reduces to the standard $\ell_2$- and $\ell_\infty$-max-margin classifiers. As a counterpoint, we prove that Signum [Bernstein et al., 2018] converges to the $\ell_\infty$-max-margin classifier for any batch size. Overall, our results highlight that the implicit bias of Adam crucially depends on both the batching scheme and the dataset, while Signum remains invariant.

Learning compact and reliable convex hulls from data is a fundamental yet challenging problem with broad applications in classification, constraint learning, and decision optimization. We propose Projection Convex Hull (PCH), a scalable framework for learning polyhedral trust regions in high-dimensional spaces. Starting from an exact MINLP formulation, we derive an unconstrained surrogate objective and show that, under suitable weight assignments, the optimal hyperplanes of the MINLP are recovered as stationary points of the surrogate. Building on this theoretical foundation, PCH adaptively constructs and refines hyperplanes by subregion partition, strategic weight assignment, and gradient-based updates, yielding convex hulls that tightly enclose the positive class while excluding negatives. The learned polyhedra can serve as geometric trust regions to enhance selective classification and constraint learning. Extensive experiments on synthetic and real-world datasets demonstrate that PCH achieves strong performance in accuracy, scalability, and model compactness, outperforming classical geometric algorithms and recent optimization-based approaches, especially in high-dimensional and large-scale settings. These results confirm the value of PCH as a theoretically grounded and practically effective framework for trust-region learning.

Multimodal large language models (MLLMs) have achieved remarkable progress in vision–language reasoning, yet they remain vulnerable to hallucination, where generated content deviates from the visual evidence. Existing mitigation strategies either demand costly supervision during training or introduce additional latency at inference. Recent vision-enhancement methods attempt to address this by reinforcing visual tokens during decoding, but they typically inject all tokens indiscriminately, leading to interference from background regions and distracting the model from critical cues. To overcome this challenge, we propose an **A**daptive v**I**sual **R**einforcement framework for MLLMs, dubbed as **AIR**. AIR consists of two main components: prototype-based token reduction, which condenses the large pool of visual tokens into a compact subset to suppress redundancy, and OT-guided patch reinforcement, which quantifies the alignment between hidden state and patch embeddings to selectively integrate the most consistent patches into the feed-forward layers. As a result, AIR enhances the model’s reliance on salient visual information and effectively mitigates hallucination. Extensive experiments across representative MLLMs demonstrate that AIR substantially reduces hallucination while preserving general capabilities, establishing it as an effective and independent solution for building reliable MLLMs.

The rapid growth in the parameter scale of large language models (LLMs) has created a high demand for efficient compression techniques. As a hardware-agnostic and highly compatible technique, low-rank compression has been widely adopted. However, existing methods typically compress each layer independently by minimizing per-layer reconstruction error, overlooking a critical limitation: the reconstruction error propagates and accumulates through the network, which leads to amplified global deviations from the full-precision baseline. To address this, we propose **Self-Adaptive Error Suppression SVD (SAES-SVD)**, a LLMs compression framework that jointly optimizes intra-layer reconstruction and inter-layer error compensation. SAES-SVD is composed of two novel components: **Cumulative Error-Aware Layer Compression (CEALC),** which formulates the compression objective as a combination of local reconstruction and weighted cumulative error compensation. Based on it, we derive a closed-form low-rank solution relied on second-order activation statistics, which explicitly aligns each layer's output with its full-precision counterpart to compensate for accumulated errors. \ding{183} Adaptive Collaborative Error Suppression (ACES), which automatically adjusts the weighting coefficient to enhance the low-rank structure of the compression objective in CELAC. Specifically, the coefficient is optimized to maximize the ratio between the Frobenius norm of the compressed layer's output and that of the compression objective under a fixed rank, thus ensuring that the rank budget is utilized effectively. Extensive experiments across multiple LLM architectures and tasks show that, without fine-tuning or additional tricks, SAES-SVD consistently improves post-compression performance. For example, at a 0.2 compression ratio on LLaMA-7B, existing methods exhibit an average accuracy drop exceeding 0.05, whereas SAES-SVD restricts the drop to only 0.02. These improvements underscore the potential of SAES-SVD to effectively narrow the gap between compressed models and their full-precision counterparts, paving the way for more reliable compression of LLMs.

As large pre-trained models rapidly scale, Parameter-Efficient Fine-Tuning (PEFT) through methods like Low-Rank Adaptation (LoRA) becomes increasingly crucial. While LoRA has emerged as a cornerstone of PEFT, excelling at preserving performance with minimal additional parameters, exploring parameter-sharing mechanisms of LoRA remains critical to pushing efficiency boundaries. However, existing naive LoRA sharing methods often degrade performance due to sacrificed representational diversity and weakened model expressiveness. To overcome this issue, we conduct an in-depth analysis of pre-trained models using gradient-based proxy entropy, and uncover two critical, previously overlooked properties: Local Similarity and Layer-wise Information Heterogeneity. Building on these insights, we propose E²LoRA, a novel dual-adaptive sharing framework. It enables adaptive sharing interval partitioning, guided by inter-layer proxy entropy similarity, and adaptive rank allocation, informed by layer-wise absolute proxy entropy. This unique design leverages inherently informative properties of pre-trained models to significantly reduce parameter redundancy while maintaining or enhancing expressiveness. Comprehensive evaluations across diverse tasks, modalities, and models consistently demonstrate that E²LoRA achieves an excellent balance of efficiency and effectiveness, consistently matching or surpassing baselines with approximately 50% fewer trainable parameters.

Speculative decoding accelerates large language model (LLM) inference by using a small draft model to generate candidate tokens for a larger target model to verify. The efficacy of this technique hinges on the trade-off between the time spent on drafting candidates and verifying them. However, current state-of-the-art methods rely on a static time allocation, while recent dynamic approaches optimize for proxy metrics like acceptance length, often neglecting the true time cost and treating the drafting and verification phases in isolation. To address these limitations, we introduce Learning to Draft (LTD), a novel method that directly optimizes for throughput of each draft-and-verify cycle. We formulate the problem as a reinforcement learning environment and train two co-adaptive policies to dynamically coordinate the draft and verification phases. This encourages the policies to adapt to each other and explicitly maximize decoding efficiency. We conducted extensive evaluations on five diverse LLMs and four distinct tasks. Our results show that LTD achieves speedup ratios ranging from 2.24x to 4.32x, outperforming the state-of-the-art method Eagle3 up to 36.4\%.

The conditional average treatment effect (CATE) is widely used in personalized medicine to inform therapeutic decisions. However, state-of-the-art methods for CATE estimation (so-called meta-learners) often perform poorly in the presence of low overlap. In this work, we introduce a new approach to tackle this issue and improve the performance of existing meta-learners in the low-overlap regions. Specifically, we introduce Overlap-Adaptive Regularization (OAR) that regularizes target models proportionally to overlap weights so that, informally, the regularization is higher in regions with low overlap. To the best of our knowledge, our OAR is the first approach to leverage overlap weights in the regularization terms of the meta-learners. Our OAR approach is flexible and works with any existing CATE meta-learner: we demonstrate how OAR can be applied to both parametric and non-parametric second-stage models. Furthermore, we propose debiased versions of our OAR that preserve the Neyman-orthogonality of existing meta-learners and thus ensure more robust inference. Through a series of (semi-)synthetic experiments, we demonstrate that our OAR significantly improves CATE estimation in low-overlap settings in comparison to constant regularization.

Associative memory models are content-addressable memory systems fundamental to biological intelligence and are notable for their high interpretability. However, existing models evaluate the quality of retrieval based on proximity, which cannot guarantee that the retrieved pattern has the strongest association with the query, failing correctness. We reframe this problem by proposing that a query is a generative variant of a stored memory pattern, and define a variant distribution to model this subtle context-dependent generative process. Consequently, correct retrieval should return the memory pattern with the maximum a posteriori probability of being the query's origin. This perspective reveals that an ideal similarity measure should approximate the likelihood of each stored pattern generating the query in accordance with variant distribution, which is impossible for fixed and pre-defined similarities used by existing associative memories. To this end, we develop adaptive similarity, a novel mechanism that learns to approximate this insightful but unknown likelihood from samples drawn from context, aiming for correct retrieval. We theoretically prove that our proposed adaptive similarity achieves optimal correct retrieval under three canonical and widely applicable types of variants: noisy, masked, and biased. We integrate this mechanism into a novel adaptive Hopfield network (`A-Hop`), and empirical results show that it achieves state-of-the-art performance across diverse tasks, including memory retrieval, tabular classification, image classification, and multiple instance learning. Our code is publicly available at https://anonymous.4open.science/r/Adaptive-Hopfield-Network-C137/.