ICLR 2026 论文列表

A hallmark of modern large-scale machine learning techniques is the use of training objectives that provide dense supervision to intermediate computations, such as teacher forcing the next token in language models or denoising step-by-step in diffusion models. This enables models to learn complex functions in a generalizable manner. Motivated by this observation, we investigate the benefits of iterative computation for temporal difference (TD) methods in reinforcement learning (RL). Typically, they represent value functions in a monolithic fashion, without iterative compute. We introduce floq (flow-matching Q-functions), an approach that parameterizes the Q-function using a velocity field and trains it with techniques from flow-matching, typically used in generative modeling. This velocity field underneath the flow is trained using a TD-learning objective, which bootstraps from values produced by a target velocity field, computed by running multiple steps of numerical integration. Crucially, floq allows for more fine-grained control and scaling of the Q-function capacity than monolithic architectures, by appropriately setting the number of integration steps. Across a suite of challenging offline RL benchmarks and online fine-tuning tasks, floq improves performance by nearly 1.8x. floq scales capacity far better than standard TD-learning architectures, highlighting the potential of iterative computation for value learning.

Hybrid model architectures that combine computational primitives (e.g., Attention, MLP) in different ratios have shown promising performance beyond Transformers. Some studies have shown that different interleavings of primitives can affect model quality as well. However, prior works explore the hybrid model architecture design space manually. Due to the large design space and training costs, discovering hybrid models that combine key computational primitives for pre-training is challenging. In this work, we take a principled approach in designing a modular hybrid model architecture search framework — Composer. Composer explores model architectures at a small scale and extrapolates the top-performing model architectures to a larger scale using our proposed scaling strategies. Using Composer, we discover new hybrid LLM architectures that outperform Llama 3.2. Compared to Llama 3.2 and previous state-of-the-art baselines, the new model architectures consistently reduce validation loss at parameter scales of 350M-3B and improve evaluation accuracy on the downstream tasks by up to 2.8-8.3% (1.1-3.1% on average) while improving both training and inference efficiency.

Large reasoning models (LRMs) trained with *Reinforcement Learning with Verifiable Rewards* (RLVR) have achieved remarkable progress on complex reasoning tasks. However, RLVR heavily relies on on-policy rollout generation, whose cost grows rapidly with rollout length and model size, eventually becoming the training bottleneck. Our empirical analysis reveals that independent rollouts for the same query often share similar early steps, indicating substantial redundancy. To address this, we propose **Pros** (**P**refix **R**euse for **O**n-policy **S**ampling), a paradigm that reuses promising prefixes of historical rollouts in RLVR training. **Pros** appends these self-generated partial rollouts to the original queries to form *Augmented Queries*, which are then used as regular training inputs in subsequent iterations, thereby reducing redundant computation. To select training batch from augmented queries, **Pros** adopts a hierarchical Bayesian model to estimate their pass rates and prioritize those with the highest reward uncertainty. Experiments across diverse settings show that **Pros** consistently improves training efficiency and achieves higher accuracy than strong baselines. These results highlight **Pros** as a practical path toward scalable and compute-efficient RLVR.

Deep generative models have made significant advances in generating complex content, yet conditional generation remains a fundamental challenge. Existing conditional generative adversarial networks often struggle to balance the dual objectives of assessing authenticity and conditional alignment of input samples within their conditional discriminators. To address this, we propose a novel discriminator design that integrates three key capabilities: unconditional discrimination, matching-aware supervision to enhance alignment sensitivity, and adaptive weighting to dynamically balance all objectives. Specifically, we introduce Sum of Naturalness and Alignment (SONA), which employs separate projections for naturalness (authenticity) and alignment in the final layer with an inductive bias, supported by dedicated objective functions and an adaptive weighting mechanism. Extensive experiments on class-conditional generation tasks show that SONA achieves superior sample quality and conditional alignment compared to state-of-the-art methods. Furthermore, we demonstrate its effectiveness in text-to-image generation, confirming the versatility and robustness of our approach.

Many offline reinforcement learning algorithms are underpinned by $f$-divergence regularization, but their sample complexity *defined with respect to regularized objectives* still lacks tight analyses, especially in terms of concrete data coverage conditions. In this paper, we study the exact concentrability requirements to achieve the $\tilde{\Theta}(\epsilon^{-1})$ sample complexity for offline $f$-divergence-regularized contextual bandits. For reverse Kullback–Leibler (KL) divergence, arguably the most commonly used one, we achieve an $\tilde{O}(\epsilon^{-1})$ sample complexity under single-policy concentrability for the first time via a novel pessimism-based analysis, surpassing existing $\tilde{O}(\epsilon^{-1})$ bound under all-policy concentrability and $\tilde{O}(\epsilon^{-2})$ bound under single-policy concentrability. We also propose a near-matching lower bound, demonstrating that a multiplicative dependency on single-policy concentrability is necessary to maximally exploit the curvature property of reverse KL. Moreover, for $f$-divergences with strongly convex $f$, to which reverse KL *does not* belong, we show that the sharp sample complexity $\tilde{\Theta}(\epsilon^{-1})$ is achievable even without pessimistic estimation or single-policy concentrability. We further corroborate our theoretical insights with numerical experiments and extend our analysis to contextual dueling bandits. We believe these results take a significant step towards a comprehensive understanding of objectives with $f$-divergence regularization.

We introduce _String Seed of Thought (SSoT)_, a novel prompting method for LLMs that improves _Probabilistic Instruction Following (PIF)_. We define PIF as a task requiring an LLM to select its answer from a predefined set of options, each associated with a specific probability, such that the empirical distribution of the generated answers aligns with the target distribution when prompted multiple times. While LLMs excel at tasks with single, deterministic answers, they often fail at PIF, exhibiting biases problematic for applications requiring non-deterministic behaviors, such as human-behavior simulation, content diversification, and multiplayer games. It also harms the diversity of generated responses, a crucial factor in test-time scaling, by causing the outputs to collapse into a limited set of answers. To address this, we propose SSoT, a simple prompting method that instructs an LLM to first output a random string to generate sufficient entropy. SSoT also instructs the LLM to extract randomness by manipulating this string to derive a final answer, thereby preserving diversity while adhering to specific constraints. We demonstrate that SSoT significantly improves the PIF performance of LLMs, approaching the ideal performance of a pseudo-random number generator. Notably, our experiments on NoveltyBench show SSoT's benefits extend beyond closed-set tasks to open-ended tasks by enhancing response diversity.

The use of target networks in deep reinforcement learning is a widely popular solution to mitigate the brittleness of semi-gradient approaches and stabilize learning. However, target networks notoriously require additional memory and delay the propagation of Bellman updates compared to an ideal target-free approach. In this work, we step out of the binary choice between target-free and target-based algorithms. We introduce a new method that uses a copy of the last linear layer of the online network as a target network, while sharing the remaining parameters with the up-to-date online network. This simple modification enables us to keep the target-free's low-memory footprint while leveraging the target-based literature. We find that combining our approach with the concept of iterated $Q$-learning, which consists of learning consecutive Bellman updates in parallel, helps improve the sample-efficiency of target-free approaches. Our proposed method, iterated Shared $Q$-Learning (iS-QL), bridges the performance gap between target-free and target-based approaches across various problems while using a single $Q$-network, thus stepping towards resource-efficient reinforcement learning algorithms.

The quadratic compute and memory costs of global self-attention severely limit its use in high-resolution images. Local attention reduces complexity by restricting attention to neighborhoods. Block-sparse kernels can further improve the efficiency of local attention, but conventional local attention patterns often fail to deliver significant speedups because tokens within a window are not contiguous in the 1D sequence. This work proposes a novel method for constructing windows and neighborhoods based on the Hilbert curve. Image tokens are first reordered along a Hilbert curve, and windows and neighborhoods are then formed on the reordered 1D sequence. From a block-sparse perspective, this strategy significantly increases block sparsity and can be combined with existing block-sparse kernels to improve the efficiency of 2D local attention. Experiments show that the proposed Hilbert Window Attention and Hilbert Slide Attention can accelerate window attention and slide attention by about $4\times$ and $18\times$, respectively. To assess practicality, the strategy is instantiated as the Hilbert Window Transformer and the Hilbert Neighborhood Transformer, both of which achieve end-to-end speedups with minimal accuracy loss. Overall, combining Hilbert-guided local attention with block-sparse kernels offers a general and practical approach to enhancing the efficiency of 2D local attention for images.

Previous work shows that the chain of continuous thought (continuous CoT) improves the reasoning capability of large language models (LLMs) by enabling implicit parallel thinking, and a subsequent work provided theoretical insight by showing that a two-layer transformer equipped with continuous CoT can efficiently solve directed graph reachability by maintaining a superposition of multiple reasoning traces in the continuous thought. However, it remains unclear how the superposition mechanism is naturally learned from gradient-based training methods. To fill this gap, we theoretically analyze the training dynamics of a simplified two-layer transformer on the directed graph reachability problem to unveil how the superposition mechanism emerges during training in two training stages -- (i) a *thought-generation* stage that autoregressively expands the continuous thought, and (ii) a *prediction* stage that converts the thought into the final answer. Our analysis reveals that during training using continuous thought, the index-matching logit, an important quantity which reflects the strength of the model's local search ability, will first increase and then remain bounded under mild assumptions. The bounded index-matching logit effectively balances exploration and exploitation during the reasoning process: the model will exploit local problem structures to identify plausible search traces, and assign comparable weights to multiple such traces to explore when it is uncertain about which solution is correct, which results in superposition. Our experimental results tracking the growth of logits further validate our theory.

Recent advances in diffusion-based image generation models (IGMs), such as Stable Diffusion (SD), have substantially improved the quality and diversity of AI-generated content. However, these models also pose ethical, legal, and societal risks, including the generation of harmful, misleading, or copyright-infringing material. Machine unlearning (MU) has emerged as a promising mitigation by selectively removing undesirable concepts from pretrained models, yet the robustness of existing methods, particularly under multi-modal adversarial inputs, remains insufficiently explored. To address this gap, we propose RECALL, a multi-modal adversarial framework for systematically evaluating and compromising the robustness of unlearned IGMs. Unlike prior approaches that primarily optimize adversarial text prompts, RECALL exploits the native multi-modal conditioning of diffusion models by efficiently optimizing adversarial image prompts guided by a single semantically relevant reference image. Extensive experiments across ten state-of-the-art unlearning methods and diverse representative tasks show that RECALL consistently surpasses existing baselines in adversarial effectiveness, computational efficiency, and semantic fidelity to the original prompt. These results reveal critical vulnerabilities in current unlearning pipelines and underscore the need for more robust, verifiable unlearning mechanisms. More than just an attack, RECALL also serves as an auditing tool for model owners and unlearning practitioners, enabling systematic robustness evaluation. Code and data are available at https://anonymous.4open.science/r/RECALL.

While existing multivariate time series forecasting models have advanced significantly in modeling periodicity, they largely neglect the periodic heterogeneity common in real-world data, where variables exhibit distinct and dynamically changing periods. To effectively capture this periodic heterogeneity, we propose PHAT (Period Heterogeneity-Aware Transformer). Specifically, PHAT arranges multivariate inputs into a three-dimensional "periodic bucket" tensor, where the dimensions correspond to variable group characteristics with similar periodicity, time steps aligned by phase, and offsets within the period. By restricting interactions within buckets and masking cross-bucket connections, PHAT effectively avoids interference from inconsistent periods. We also propose a positive-negative attention mechanism, which captures periodic dependencies from two perspectives: periodic alignment and periodic deviation. Additionally, the periodic alignment attention scores are decomposed into positive and negative components, with a modulation term encoding periodic priors. This modulation constrains the attention mechanism to more faithfully reflect the underlying periodic trends. A mathematical explanation is provided to support this property. We evaluate PHAT comprehensively on 14 real-world datasets against 18 baselines, and the results show that it significantly outperforms existing methods, achieving highly competitive forecasting performance. Our sources is available at GitHub.

There has been significant recent interest on understanding the capacity of Transformers for in-context learning (ICL), yet most theory focuses on supervised settings with explicitly labeled pairs. In practice, Transformers often perform well even when labels are sparse or absent, suggesting crucial structure within unlabeled contextual demonstrations. We introduce and study in-context semi-supervised learning (IC-SSL), where a small set of labeled examples is accompanied by many unlabeled points, and show that Transformers can leverage the unlabeled context to learn a robust, context-dependent representation. This representation enables accurate predictions and markedly improves performance in low-label regimes, offering foundational insights into how Transformers exploit unlabeled context for representation learning within the ICL framework. Our code is available at https://github.com/Jason-fan20/ICL_Semi.

Large Language Models (LLMs) achieve state-of-the-art performance but are costly to deploy in resource-constrained environments. Pruning with $n:m$ semi-structured sparsity reduces computation and enables hardware acceleration, yet existing methods face a trade-off: one-shot approaches are efficient but heuristic, while optimization-based methods are accurate but expensive. We introduce \textbf{HyperPrune}, a resource-efficient framework that directly optimizes $n:m$ sparsity. A lightweight hypernetwork, shared across layers and conditioned on learnable embeddings, generates structured masks in a one-shot, layer-wise manner. \textit{Continual pruning} preserves cross-layer knowledge, and \textit{feature outlier regularization} retains critical activations, unifying the strengths of heuristic and optimization-based methods. Experiments on LLaMA-7B to 70B show state-of-the-art accuracy–sparsity trade-offs on a single A100 GPU, achieving higher efficiency, accuracy, and scalability than prior approaches. HyperPrune offers a practical, scalable, and hardware-friendly solution for structured LLM pruning.

Large Language Models (LLMs) excel across diverse applications but remain impractical for edge deployment due to severe memory bottlenecks at the edge devices. We propose Equilibrium Language Models (ELMs), a novel compression framework that replaces groups of Transformer layers with a lightweight fixed-point network, reinterpreting deep computation as solving for an equilibrium state. To achieve ELMs, We introduce *Group Pruning Policy Optimization*, which automatically learns optimal pruning intervals. Moreover, we propose *One-Step KV-Cache*, which drastically reduces memory overhead by storing only the final iteration cache without compromising the accuracy, to enable effective deployment at the edge devices. Across different tasks such as common sense reasoning, mathematical problem solving, and code generation, ELMs prune 28\% of parameters while retaining 99\% of the accuracy of dense fine-tuned LLMs, establishing a new direction for memory-efficient edge deployment of large models.

Inference speed and tracking performance are two critical evaluation metrics in the field of visual tracking. However, high-performance trackers often suffer from slow processing speeds, making them impractical for deployment on resource-constrained devices. To alleviate this issue, we propose $\textbf{FARTrack}$, a $\textbf{F}$ast $\textbf{A}$uto-$\textbf{R}$egressive $\textbf{T}$racking framework. Since autoregression emphasizes the temporal nature of the trajectory sequence, it can maintain high performance while achieving efficient execution across various devices. FARTrack introduces $\textbf{Task-Specific Self-Distillation}$ and $\textbf{Inter-frame Autoregressive Sparsification}$, designed from the perspectives of $\textbf{shallow-yet-accurate distillation}$ and $\textbf{redundant-to-essential token optimization}$, respectively. Task-Specific Self-Distillation achieves model compression by distilling task-specific tokens layer by layer, enhancing the model's inference speed while avoiding suboptimal manual teacher-student layer pairs assignments. Meanwhile, Inter-frame Autoregressive Sparsification sequentially condenses multiple templates, avoiding additional runtime overhead while learning a temporally-global optimal sparsification strategy. FARTrack demonstrates outstanding speed and competitive performance. It delivers an AO of 70.6\% on GOT-10k in real-time. Beyond, our fastest model achieves a speed of 343 FPS on the GPU and 121 FPS on the CPU. The code will be released.

Axis-aligned decision trees are fast and stable but struggle on datasets with rotated or interaction-dependent decision boundaries, where informative splits require linear combinations of features rather than single-feature thresholds. Oblique forests address this with per-node hyperplane splits, but at added computational cost. We propose a simple alternative: JARF, Jacobian-Aligned Random Forests. Concretely, we fit a random forest to estimate class probabilities or regression outputs, compute finite-difference gradients with respect to each feature, form an expected Jacobian outer product/expected gradient outer product, and use it as a single global linear preconditioner for all inputs. This preserves the simplicity of axisaligned trees while applying a single global rotation to capture oblique boundaries and feature interactions that would otherwise require many axis-aligned splits to approximate. On tabular benchmarks, our preconditioned forest matches or surpasses oblique baselines while training faster. Our results suggest that supervised preconditioning can deliver the accuracy of oblique forests while keeping the simplicity of axis-aligned trees.

Speculative sampling (SpS) has been successful in accelerating the decoding throughput of auto-regressive large language models by leveraging smaller draft models. SpS strictly enforces the generated distribution to match that of the verifier LLM. This is unnecessarily restrictive as slight variation of the verifier's distribution, such as sampling with top-$k$ or temperature, would also be acceptable. Typical acceptance sampling (TAS) alleviates this issue by accepting more tokens using entropy-based heuristics. However, this approach distorts the verifier distribution, potentially degrading output quality when the verifier encodes critical information. In this work, we formalize the speculative sampling algorithm through the lens of constrained optimization. Based on this formulation, we propose **Cactus** (**c**onstrained **ac**cep**t**ance spec**u**lative **s**ampling), a method that guarantees controlled divergence from the verifier distribution and increasing acceptance rates. Empirical results across a wide range of benchmarks confirm the effectiveness of our approach.

Inspired by the human learning and memory system, particularly the interplay between the hippocampus and cerebral cortex, this study proposes a dual-learner framework comprising a fast learner and a meta learner to address continual Reinforcement Learning~(RL) problems. These two learners are coupled to perform distinct yet complementary roles: the fast learner focuses on knowledge transfer, while the meta learner ensures knowledge integration. In contrast to traditional multi-task RL approaches that share knowledge through average return maximization, our meta learner incrementally integrates new experiences by explicitly minimizing catastrophic forgetting, thereby supporting efficient cumulative knowledge transfer for the fast learner. To facilitate rapid adaptation in new environments, we introduce an adaptive meta warm-up mechanism that selectively harnesses past knowledge. We conduct experiments in various pixel-based and continuous control benchmarks, revealing the superior performance of continual learning for our proposed dual-learner approach relative to baseline methods.

In the era of AI-driven science and engineering, we often want to design discrete objects (e.g., circuits, proteins, materials) in silico according to user-specified properties (e.g., that a protein binds its target). Given a property predictive model, in silico design typically involves training a generative model over the design space (e.g., over the set of all length-L proteins) to concentrate on designs with the desired properties. Distributional optimization, formalized as an estimation of distribution algorithm or as reinforcement learning policy optimization, maximizes an objective function in expectation over samples. Optimizing a distribution over discrete-valued designs is in general challenging due to the combinatorial nature of the design space. However, many property predictors in scientific applications are decomposable in the sense that they can be factorized over design variables in a way that will prove useful. For example, the active site amino acids in a catalytic protein may need to only loosely interact with the rest of the protein for maximal catalytic activity. Current distributional optimization algorithms are unable to make use of such structure, which could dramatically improve the optimization. Herein, we propose and demonstrate use of a new distributional optimization algorithm, Decomposition-Aware Distributional Optimization (DADO), that can leverage any decomposability defined by a junction tree on the design variables. At its core, DADO employs a factorized “search distribution”—a learned generative model—for efficient navigation of the search space, and invokes graph message passing to coordinate optimization across all variables.