ICLR 2026 Papers

Large language models (LLMs) are deployed on mobile devices to power killer applications such as intelligent assistants. LLMs pre-trained on general corpora often hallucinate when handling personalized or unseen queries, leading to incorrect or outdated responses. Knowledge editing addresses this by identifying and adjusting a small crucial portion of model weights, without compromising the general knowledge. However, prior knowledge editing methods are impractical to run on local devices due to the resource-heavy backpropagation (BP) needed for updates. We present MobiEdit, the first mobile knowledge editing framework that enables efficient LLM personalization on commercial off-the-shelf (COTS) mobile devices. MobiEdit replaces full-precision BP with quantized forward-only gradient estimation, thus compatible with the energy-efficient mobile neural processing units (NPUs). To further improve gradient estimation efficiency, we introduce two optimizations: an early stopping mechanism that adaptively terminates editing upon success and prefix activation reusing that reduce redundant computation across steps. Our approach enables real-time editing of 3B-parameter models (Qwen2.5-3B-Instruct and Llama3.2-3B-Instruct) on COTS mobile devices with 7.1$\times$ less memory, 15.8 $\times$ less energy and 3.4$\times$ less latency compared to previous knowledge editing methods.

Reward models are used throughout the post-training of language models to capture nuanced signals from preference data and provide a training target for optimization across instruction following, reasoning, safety, and more domains. The community has begun establishing best practices for evaluating reward models, from the development of benchmarks that test capabilities in specific skill areas to others that test agreement with human preferences. At the same time, progress in evaluation has not been mirrored by the effectiveness of reward models in downstream tasks -- simpler direct alignment algorithms are reported to work better in many cases. This paper introduces RewardEval, a new multi-skill reward modeling benchmark designed to bring new, challenging data for accuracy-based reward model evaluation -- models score about 20 points on average lower on RewardEval compared to RewardBench, a widely-used existing reward model evaluation-- while being highly correlated with downstream performance. Compared to most other benchmarks, RewardEval sources new human prompts instead of existing prompts from downstream evaluations, facilitating more rigorous evaluation practices. In this paper, we describe our benchmark construction process and report how existing models perform on it, while quantifying and providing new insights on how performance on the benchmark correlates with downstream use of the models in both inference-time scaling algorithms, like best-of-N sampling, and RLHF training algorithms like proximal policy optimization.

World model based planning has significantly improved decision-making in complex environments by enabling agents to simulate future states and make informed choices. This computational burden is particularly restrictive in robotics, where resources are severely constrained. To address this limitation, we propose a Sparse Imagination for Efficient Visual World Model Planning, which enhances computational efficiency by reducing the number of tokens processed during forward prediction. Our method leverages a sparsely trained vision-based world model based on transformers with randomized grouped attention strategy, allowing the model to flexibly adjust the number of tokens processed based on the computational resource. By enabling sparse imagination during latent rollout, our approach significantly accelerates planning while maintaining high control fidelity. Experimental results demonstrate that sparse imagination preserves task performance while dramatically improving inference efficiency. This general technique for visual planning is applicable from simple test-time trajectory optimization to complex real-world tasks with the latest VLAs, enabling the deployment of world models in real-time scenarios.

Diffusion models have achieved remarkable success in content generation but suffer from prohibitive computational costs due to iterative sampling. While recent feature caching methods tend to accelerate inference through temporal extrapolation, these methods still suffer from severe quality loss due to the failure in modeling the complex dynamics of feature evolution. To solve this problem, this paper presents HiCache (Hermite Polynomial-based Feature Cache), a training-free acceleration framework that fundamentally improves feature prediction by aligning mathematical tools with empirical properties. Our key insight is that feature derivative approximations in Diffusion Transformers exhibit multivariate Gaussian characteristics, motivating the use of Hermite polynomials, the potentially theoretically optimal basis for Gaussian-correlated processes. Besides, we introduce a dual-scaling mechanism that ensures numerical stability while preserving predictive accuracy, which is also effective when applied standalone to TaylorSeer. Extensive experiments demonstrate HiCache's superiority: achieving \$5.55\times\$ speedup on FLUX.1-dev while exceeding baseline quality, maintaining strong performance across text-to-image, video generation, and super-resolution tasks. Moreover, HiCache can be naturally added to the previous caching methods to enhance their performance, e.g., improving ClusCa from \$0.9480\$ to \$0.9840\$ in terms of image rewards. Our code is included in the supplementary material, and will be released on GitHub.

Large language models (LLMs) are widely applied to data analytics over documents, yet direct reasoning over long, noisy documents remains brittle and error-prone. Hence, we study document question answering (QA) that consolidates dispersed evidence into a structured output (e.g., a table, graph, or chunks) to support reliable, verifiable QA. We propose a two-pillar framework, LiteCoST, to achieve both high accuracy and low latency with small language models (SLMs). Pillar 1: Chain-of-Structured-Thought (CoST). We introduce a CoST template—a schema-aware instruction that guides a strong LLM to produce both a step-wise CoST trace and the corresponding structured output. The process induces a minimal structure, normalizes entities/units, aligns records, serializes the output, and verifies/refines it (optionally with an LLM-as-judge), yielding auditable supervision. Pillar 2: SLM fine-tuning. We then train compact models on the LLM-generated CoST traces/structured data in two phases—Supervised Fine-Tuning for structure/format/steps, followed by Group Relative Policy Optimization with dual rewards for answer/format quality and process consistency—transferring structure-first behavior to SLMs for low-latency deployment. This approach achieves LLM-comparable quality on finance and legal long-document QA (Loong), with 3B/7B SLMs while delivering 2–4×lower latency than GPT-4o and DeepSeek-R1 (671B).

Discovering insights from a real-world data lake potentially containing unclean, semi-structured, and unstructured data requires a variety of data processing tasks, ranging from extraction and cleaning to integration, analysis, and modeling. This process often also demands domain knowledge and project-specific insight. While AI models have shown remarkable results in reasoning and code generation, their abilities to design and execute complex pipelines that solve these data-lake-to-insight challenges remain unclear. We introduce KramaBench which consists of 104 manually curated and solved challenges spanning 1700 files, 24 data sources, and 6 domains. KramaBench focuses on testing the end-to-end capabilities of AI systems to solve challenges which require automated orchestration of different data tasks. KramaBench also features a comprehensive evaluation framework assessing the pipeline design and individual data task implementation abilities of AI systems. Evaluating 8 LLMs with our single-agent reference framework DS-Guru, alongside open- and closed-source agentic systems, we find that while current single-agent systems may handle isolated data-science tasks and generate plausible draft pipelines, they struggle with producing working end-to-end pipelines. On KramaBench, the best system reaches only 50% end-to-end accuracy in the full data-lake setting. Even with perfect retrieval, the accuracy tops out at 59%. Leading LLMs can identify up to 42% of important data tasks but can only fully implement 20% of individual data tasks.

Incorporating metadata in Large Language Models (LLMs) pretraining has recently emerged as a promising approach to accelerate training. However prior work highlighted only one useful signal—URLs, leaving open the question of whether other forms of metadata could yield greater benefits. In this study, we investigate a wider range of metadata types and find other types of metadata, such as fine-grained indicators of document quality that can also accelerate pretraining when prepended. We identify a common feature among effective metadata: they encode information at a finer granularity. We further introduce metadata appending as a means of improving training efficiency, where predicting an appropriate metadata as auxiliary task can help speed up pretraining. In addition, learnable meta-tokens trained with masked loss can recover part of the speedup by inducing quality-aware latent structure. Using probing, we analyze latent representations to understand how metadata shapes learning. Together, these results yield practical guidelines for integrating metadata to improve both the efficiency and effectiveness of LLM pretraining.

Modeling non-stationary processes, where statistical properties vary across the input domain, is a critical challenge in machine learning; yet most scalable methods rely on a simplifying assumption of stationarity. This forces a difficult trade-off: use expressive but computationally demanding models like Deep Gaussian Processes, or scalable but limited methods like Random Fourier Features (RFF). We close this gap by introducing Random Wavelet Features (RWF), a framework that constructs scalable, non-stationary kernel approximations by sampling from wavelet families. By harnessing the inherent localization and multi-resolution structure of wavelets, RWF generates an explicit feature map that captures complex, input-dependent patterns. Our framework provides a principled way to generalize RFF to the non-stationary setting and comes with a comprehensive theoretical analysis, including positive definiteness, unbiasedness, and uniform convergence guarantees. We demonstrate empirically on a range of challenging synthetic and real-world datasets that RWF outperforms stationary random features and offers a compelling accuracy-efficiency trade-off against more complex models, unlocking scalable and expressive kernel methods for a broad class of real-world non-stationary problems.

Transformer LLMs have been shown to exhibit strong reasoning ability that scales with inference-time compute, most prominently through token-space “thinking” (i.e., chains of thought). A growing line of work pushes this extra computation into the model’s latent space (adjacent to standard decoding) which we term Auxiliary Latent-Space Computation (ALSC). Existing ALSC methods largely fall into three buckets: (i) token-mediated latent or special-token rollouts, (ii) residual/activation steering, and (iii) memory compression via cache pruning, merging, or summarization. An underexplored alternative is memory consolidation and reconsolidation, two processes in the brain that are responsible for stabilising newly formed memory traces, and, upon recall, transiently rendering established traces plastic such they can integrate new contextual information before restabilising. In a Transformer LLM, this can be seen as analogous to performing in-place global rewrites of incoming KV segments, and rewrites of past segments conditioned on newly observed tokens. In this work, we give a theoretical justification as to why memory (re)consolidation via KV cache rewrites is beneficial for improved reasoning. We do this through the lens of Information Bottleneck (IB) theory, which posits that model generalisation emerges from an optimal balance between input information compression and retention of predictive information in latent representations. We prove using IB theory that Vanilla decoder-only Transformers are inherently constrained in their ability to form task-optimal sequence representations. We then introduce the Bottlenecked Transformer, which augments a decoder-only backbone LLM with a lightweight Cache Processor, an auxiliary Transformer that performs periodic, non-causal, in-place KV rewrites at newline-delimited reasoning step boundaries. The processor consolidates recently written KV entries and reconsolidates a small, top-$k$ attention-selected set of prior entries, conditioned on recent context. We evaluate our Bottlenecked Transformer architecture on seven mathematical reasoning benchmarks, with four backbone LLMs. Our model sees consistent performance gains over vanilla Transformers and pause-token augmented Transformer baselines, with gains of up to +6.6pp for selected tasks and backbones.

Spatial reasoning is a fundamental aspect of human cognition, yet it remains a major challenge for contemporary vision–language models (VLMs). Prior work largely relied on synthetic or LLM-generated environments with limited task designs and puzzle-like setups, failing to capture the real-world complexity, visual noise, and diverse spatial relationships that VLMs encounter. To address this, we introduce **_SpatiaLab_**, a comprehensive benchmark for evaluating VLMs’ spatial reasoning in realistic, unconstrained contexts. **_SpatiaLab_** comprises 1,400 visual question–answer pairs across six major categories: *Relative Positioning, Depth & Occlusion, Orientation, Size & Scale, Spatial Navigation,* and *3D Geometry*, each with five subcategories, yielding 30 distinct task types. Each subcategory contains at least 25 questions, and each main category includes at least 200 questions, supporting both multiple-choice and open-ended evaluation. Experiments across diverse state-of-the-art VLMs, including open- and closed-source models, reasoning-focused, and specialized spatial reasoning models, reveal a substantial gap in spatial reasoning capabilities compared with humans. In the multiple-choice setup, InternVL3.5-72B achieves 54.93% accuracy versus 87.57% for humans. In the open-ended setting, all models show a performance drop of around 10–25%, with GPT-5-mini scoring highest at 40.93% versus 64.93% for humans. These results highlight key limitations in handling complex spatial relationships, depth perception, navigation, and 3D geometry. By providing a diverse, real-world evaluation framework, **_SpatiaLab_** exposes critical challenges and opportunities for advancing VLMs’ spatial reasoning, offering a benchmark to guide future research toward robust, human-aligned spatial understanding. We will release **_SpatiaLab_**.

State-of-the-art machine learning often follows a two-stage process: $(i)$ pre-training on large, general-purpose datasets; $(ii)$ fine-tuning on task-specific data. In fine-tuning, selecting training examples that closely reflect the target distribution is crucial. However, it is often the case that only a few samples are available from the target distribution. Existing data selection methods treat these target samples as a validation set and estimate the effect of adding or removing a single sample from the training pool by performing inference on the validation set. We propose a simpler and faster alternative that inverts the usual role of train and validation: we perform inference on the training pool before and after fine-tuning on the validation set. We then select samples whose predictions change the most. Our key insight is that the training samples most affected by fine-tuning on a small validation set tend to be the most beneficial for reducing test loss on the target distribution. Experiments on instruction tuning and named entity recognition tasks show that, in most cases, our method achieves lower test log-loss than state-of-the-art approaches. We support our findings with theoretical analysis.

The LLM-as-a-judge paradigm is widely used in both evaluating free-text model responses and reward modeling for model alignment and finetuning. Recently, finetuning judges with judge-specific data has emerged as an often preferred choice over directly prompting frontier models as judges, as the former achieves better performance with smaller model sizes while being more robust to common biases. However, the standard evaluation ignores several practical concerns of finetuned judges regarding their real world deployment. In this paper, we identify and formalize three aspects that affect the *shelf life* of these judges: *future-proofing* and *backward-compatibility* – how well judges finetuned on responses by today's generator models perform on responses by future models or past models, as well as *question generalization* – how well judges generalize to unseen questions at test time. We study these three aspects under a unified framework with varying train and test distributions in two reasoning datasets, three SFT- and DPO-based fine-tuning algorithms, and three different backbone models. Experiments suggest that future-proofing is challenging for most models, while backward-compatibility is relatively easy, with DPO-trained models consistently *improving* performance. We further find that continual learning provides a more balanced adaptation to shifts between older and newer response distributions than training solely on stronger or weaker responses. Moreover, all models observe certain degrees of performance degradation when moving from questions seen during training to unseen ones, showing that current judges do not fully generalize to unseen questions. These findings provide insights into practical considerations for developing and deploying judge models in the face of ever-changing generators.

Prototypical self-supervised learning methods consistently suffer from partial prototype collapse, where multiple prototypes converge to nearly identical representations. This undermines their central purpose—providing diverse and informative targets to guide encoders toward rich representations—and has led practitioners to over-parameterize prototype sets or add ad-hoc regularizers, which mitigate symptoms rather than address the root cause. We empirically trace the collapse to the joint optimization of encoders and prototypes, which encourages a type of shortcut learning: early in training prototypes drift toward redundant representations that minimize loss without necessarily enhancing representation diversity. To break the joint optimization, we introduce a fully decoupled training strategy that learns prototypes and encoders under separate objectives. Concretely, we model prototypes as a Gaussian mixture updated with an online EM-style procedure, independent of the encoder's loss. This simple yet principled decoupling eliminates prototype collapse without explicit regularization and yields consistently diverse prototypes, which in several settings translate to improved downstream performance.

In this paper, we propose a 3D asset-referenced diffusion model for image generation, exploring how to integrate 3D assets into image diffusion models. Existing reference-based image generation methods leverage large-scale pretrained diffusion models and demonstrate strong capability in generating diverse images conditioned on a single reference image. However, these methods are limited to single-image references and cannot leverage 3D assets, constraining their practical versatility. To address this gap, we present a cross-domain diffusion model with dual-branch perception that leverages multi-view RGB images and point maps of 3D assets to jointly model their colors and canonical-space coordinates, achieving precise consistency between generated images and the 3D references. Our spatially aligned dual-branch generation architecture and domain-decoupled generation mechanism ensure the simultaneous generation of two spatially aligned but content-disentangled outputs, RGB images and point maps, linking 2D image attributes with 3D asset attributes. Experiments show that our approach effectively uses 3D assets as references to produce images consistent with the given assets, opening new possibilities for combining diffusion models with 3D content creation.

Model merging has emerged as a promising approach for unifying independently fine-tuned models into an integrated framework, significantly enhancing computational efficiency in multi-task learning. Recently, several SVD-based techniques have been introduced to exploit low-rank structures for enhanced merging, but their reliance on heuristically designed rank selection often leads to inter-task interference and suboptimal performance. In this paper, we propose AdaRank, a model merging framework that replaces this heuristic selection by adaptively selecting the beneficial singular components of task vectors to merge multiple models. We first show empirically that (i) selecting only the top singular components of task vectors can cause critical interference with other tasks, and (ii) assigning fixed ranks does not align with the varying complexity of tasks and layers. AdaRank addresses both issues by adapting per-component masks, indicating the selection of the component, to the unlabeled test data with entropy minimization. Our experimental findings show that AdaRank consistently improves existing merging methods across diverse backbones from different modalities, largely narrowing the performance gap against individually fine-tuned models.

Omnidirectional images (ODIs) provide full 360$^{\circ} \times$ 180$^{\circ}$ view which are widely adopted in VR, AR and embodied intelligence applications. While multi-modal large language models (MLLMs) have demonstrated remarkable performance on conventional 2D image and video understanding benchmarks, their ability to comprehend the immersive environments captured by ODIs remains largely unexplored. To address this gap, we first present ODI-Bench, a novel comprehensive benchmark specifically designed for omnidirectional image understanding. ODI-Bench contains 2,000 high-quality omnidirectional images and over 4,000 manually annotated question-answering (QA) pairs across 10 fine-grained tasks, covering both general-level and spatial-level ODI understanding. Extensive experiments are conducted to benchmark 20 representative MLLMs, including proprietary and open-source models, under both close-ended and open-ended settings. Experimental results reveal that current MLLMs still struggle to capture the immersive context provided by ODIs. To this end, we further introduce Omni-CoT, a training-free method which significantly enhances MLLMs’ comprehension ability in the omnidirectional environment through chain-of-thought reasoning across both textual information and visual cues. Both the benchmark and the code will be released upon the publication.

Modern conversational agents such as ChatGPT and Alexa+ have become indispensable in everyday life. To handle diverse business requirements and enable agentic capabilities, these LLM-based systems often rely on predefined policies, which specify instructions such as model metadata, response styles, and tool-using rules. These policies, typically implemented as in-context prompts, are becoming increasingly complex and lengthy, posing challenges for models in faithfully following them. Moreover, they impose a large fixed computational cost regardless of the input query. As multimodal conversational agents emerge, complex policies that govern multimodal tasks and even involve visual instructions are becoming increasingly necessary, yet they have been rarely studied in previous work. In particular, prior work on prompt compression has focused solely on reducing the length of task templates and demonstrations, which require limited reasoning compared to policies. Meanwhile, related work on policy alignment has been limited to internalizing text-only safety instructions. To bridge this gap, we introduce Multimodal Policy Internalization (MPI), a new task that aims to internalize reasoning-intensive multimodal policies into the parameters of a large multimodal model, enabling stronger policy-following behavior without requiring the policy to be included in-context during inference. MPI presents unique challenges from both data and algorithmic perspectives. We construct two new datasets that cover complex decision-making and tool-using tasks across both synthetic and real-world visual inputs. We investigate diverse internalization strategies and propose a novel three-stage training framework, TriMPI, which enables stronger guidance from the original policy during internalization. Specifically, we first introduce a continual pretraining stage before supervised finetuning, which directly injects policy knowledge into the model. We then propose PolicyRollout, a simple yet effective extension to GRPO-style RL algorithms, which enables more grounded exploration by augmenting the rollout space with policy-aware responses. We show significant improvements of TriMPI over strong baselines in end-to-end performance, generalization capability, and robustness to catastrophic forgetting. As the first work on multimodal policy internalization, we aim to build a strong foundation for future research by providing datasets, training recipes, and comprehensive evaluations.

Large language models (LLMs) are increasingly being used to generate comprehensive, knowledge-intensive reports. However, while these models are trained on diverse academic papers and reports, they are not exposed to the reasoning processes and intents that guide authors in crafting these documents. We hypothesize that enhancing a model's intent awareness can significantly improve the quality of generated long-form reports. We develop and employ structured, tag-based schemes to better elicit underlying implicit intents to write or cite. We demonstrate that these extracted intents enhance both zero-shot generation capabilities in LLMs and enable the creation of high-quality synthetic data for fine-tuning smaller models. Our experiments reveal improved performance across various challenging scientific report generation tasks, with an average improvement of +2.9 and +12.3 absolute points for large and small models over baselines, respectively. Furthermore, our analysis illuminates how intent awareness enhances model citation usage and substantially improves report readability.

Generalist visual captioning goes beyond a simple appearance description task, but requires integrating a series of visual cues into a caption and handling various visual domains. In this task, current open-source models present a large performance gap with commercial ones, which limits various applications such as data synthesis. To bridge the gap, this paper proposes CapFlow, a novel multi-agent collaboration workflow. CapFlow demonstrates for the first time that, by capitalizing on open-source models, it is possible to achieve caption quality on par with GPT-4.1 in various domains with an 89.5\% reduction in costs. By leveraging CapFlow as the data synthesizer, we produce high-quality visual captions from image and video domains at scale, and obtain a generalist visual captioner via fine-tuning, namely MetaCaptioner. Through extensive experiments, we show that MetaCaptioner not only achieves comparable captioning capabilities with commercial models but also reaches top-tier multimodal performance in the open-source community. We hope CapFlow and MetaCaptioner can benefit future multimodal research by providing a strong and cost-effective visual captioning solution. Our source code and models will be publicly released.