ICML 2026 Papers

Large language models allocate uniform computation across all tokens, ignoring that some sequences are trivially predictable while others require deep reasoning. We introduce ConceptMoE, which dynamically merges semantically similar tokens into concepts through learnable chunking at target compression ratio $R$. The MoE architecture enables controlled evaluation: reallocating saved computation to match baseline FLOPs and parameters isolates genuine architectural benefits. ConceptMoE consistently outperforms standard MoE, achieving +0.9 points on language pretraining, +2.3 on long context, and +0.6 on multimodal tasks. Continual training conversion with layer looping gains +5.5 points. Beyond performance, at $R=2$, ConceptMoE reduces attention computation by $R^2\times$ and KV cache by $R\times$, delivering prefill speedups up to 175\% and decoding speedups up to 117\%. The minimal architectural changes enable straightforward integration, demonstrating that adaptive concept-level processing fundamentally improves LLM effectiveness and efficiency.

While large-scale text-to-image diffusion models continue to improve in visual quality, their increasing scale has widened the gap between state-of-the-art models and on-device solutions. To address this gap, we introduce NanoFLUX, a **2.4B** text-to-image flow-matching model distilled from **17B** FLUX.1-Schnell using a progressive compression pipeline designed to preserve generation quality. Our contributions include: (1) A model compression strategy driven by pruning redundant components in the diffusion transformer, reducing its size from 12B to 2B; (2) A ResNet-based token downsampling mechanism that reduces latency by allowing intermediate blocks to operate on lower-resolution tokens while preserving high-resolution processing elsewhere; (3) A novel text encoder distillation approach that leverages visual signals from early layers of the denoiser during sampling. Empirically, NanoFLUX generates $512 \times 512$ images in approximately 2.5 seconds on mobile devices, demonstrating the feasibility of high-quality on-device text-to-image generation.

Fine-tuning large pre-trained models on a target distribution often improves in-distribution (ID) accuracy, but at the cost of out-of-distribution (OOD) robustness as representations specialize to the fine-tuning data. Weight-space ensembling methods, such as Model Soups, mitigate this effect by averaging multiple checkpoints, but they are computationally prohibitive, requiring the training and storage of dozens of fine-tuned models. In this paper, we introduce MonoSoup, a simple, data-free, hyperparameter-free, post-hoc method that achieves a strong ID–OOD balance using _only a single_ checkpoint. Our method applies Singular Value Decomposition (SVD) to each layer’s update and decomposes it into high-energy directions that capture task-specific adaptation and low-energy directions that introduce noise but may still encode residual signals useful for robustness. MonoSoup then uses entropy-based effective rank to automatically re-weigh these components with layer-wise coefficients that account for the spectral and geometric structure of the model. Experiments on CLIP models fine-tuned on ImageNet and evaluated under natural distribution shifts, as well as on Qwen language models tested on mathematical reasoning and multiple-choice benchmarks, show that this plug-and-play approach is a practical and effective alternative to multi-checkpoint methods, retaining much of their benefits without their computational overhead.

Outlier detection (OD) is widely used in practice; but its effective deployment on new tasks is hindered by lack of labeled outliers, which makes algorithm and hyperparameter selection notoriously hard. Foundation models (FMs) have transformed ML, and OD is no exception: Shen et al. (2025) introduced FOMO-0D, the first FM for OD, achieving remarkable performance against numerous baselines. This work introduces OUTFORMER, which advances FOMO-0D with (1) a mixture of synthetic priors and (2) self-evolving curriculum training. OUTFORMER is pretrained solely on synthetic labeled datasets and infers test labels of a new task by using its training data as in-context input. Inference is fast and zero-shot, requiring merely forward pass and no labeled outliers. Thanks to in-context learning, it requires zero additional work—no OD model training or bespoke model selection—enabling truly plug-and-play deployment. OUTFORMER achieves state-of-the-art performance on the prominent ADBench, as well as two new large-scale OD benchmarks that we introduce, comprising over 1,500 datasets, while maintaining speedy inference.

With the rise of large-scale foundation models, efficiently adapting them to downstream tasks remains a central challenge. Linear probing, which freezes the backbone and trains a lightweight head, is computationally efficient but often restricted to last-layer representations. We show that task-relevant information is distributed across the network hierarchy rather than encoded solely in the last layers. To leverage this distribution of information, we apply an attentive probing mechanism that dynamically fuses representations from all layers of a Vision Transformer. This attentive layer fusion (ALF) learns to identify the most relevant layers for a target task and combines low-level structural cues with high-level semantic abstractions. Across 20 diverse datasets and multiple pretrained foundation models, ALF achieves consistent, substantial gains over standard linear probes. Attention heatmaps further reveal that tasks different from the pre-training domain benefit most from intermediate representations. Overall, our findings underscore the value of intermediate layers and demonstrate a principled, task-aware approach for unlocking their potential for probing-based adaptation.

Large multimodal models (LMMs) have shown strong capabilities in audio-visual understanding by jointly reasoning over visual, auditory, and linguistic inputs. However, processing long-form audio-visual content often requires a large number of tokens, leading to substantial computational and memory overhead during inference. Existing efficiency-oriented methods typically apply uniform compression or pruning strategies, which overlook modality-specific characteristics and instruction-dependent reasoning behaviors in multimodal models. In this work, we present PRIM, an inference framework for efficient multimodal reasoning that systematically compresses audio-visual representations based on attention dynamics and instruction relevance. Motivated by an attention-based analysis revealing modality imbalance and layer-wise redundancy in LMMs, PRIM introduces a cooperative compression pipeline that spans both multimodal encoders and the language model. Specifically, PRIM performs early text-conditioned audio-visual fusion to externalize cross-modal interactions, applies attention-guided and frequency-aware strategies to remove redundant audio and video tokens, and further adapts token retention inside the language model according to task demands. Extensive experiments on multiple audio-visual benchmark datasets demonstrate that PRIM consistently achieves stable and superior efficiency--accuracy trade-offs across diverse tasks and datasets. These results demonstrate that PRIM, a multimodal cooperative compression approach, provides an effective pathway toward scalable and efficient audio-visual reasoning.

Scaling laws for dense LLMs under infinite data are well explored, but how sparsity interacts with limited data is not. In this work, we study sparse training in data-constrained regimes where limited unique tokens require multi-epoch training. Our experiments span models up to 1.92B parameters in the fitting set, sparsity up to 93.75\%, unique data budgets up to 2.6B tokens, and total training tokens up to 41.6B over 16 epochs; we further validate extrapolation on held-out dense-equivalent models up to 7.68B parameters. We find that: **1. Sparse scaling in data-limited settings:** We introduce a scaling law that models loss as a function of active parameters, unique tokens, data repetition, and sparsity, accurately predicting performance across compute and data budgets. **2. Delayed data saturation:** sparse training postpones diminishing returns from repeated data, making multi-epoch training more effective. **3. Resource trade-offs:** With fixed data, loss-optimal sparsity is moderate (~ 50%), while compute-optimal sparsity is higher and grows with data scale. Overall, sparsity is not just a tool for efficiency, but a mechanism for improving scaling trade-offs under data scarcity. Our code is available at: https://github.com/boqian333/sparse-dc-scaling.

Large vision–language models (LVLMs) perform well on multimodal tasks, but their ability to reason and precisely align visual and textual information still has room for improvement. In this study, we show that external visual cues, such as symbols or grid lines, help LVLMs form more accurate connections between visual components, such as objects, and their corresponding textual descriptions, improving their grounding and reasoning abilities. We introduce the concept of Grounding IDs, which are latent identifiers that arise within the model as a result of external cues structuring both visual and textual modalities. Our analysis reveals that partition-inducing external cues lead to Grounding IDs that make better alignment between corresponding visual and text representations, helping the model focus on relevant information. We find that Grounding IDs enhance attention between related components, improving cross-modal grounding and reducing hallucinations. Overall, our results show that Grounding IDs are a key mechanism that enables external cues to improve cross-modal alignment, reduce errors, and enhance the overall performance of LVLMs across a range of multimodal tasks.

While multimodal agents can achieve strong performance through fine-tuning, their ability to generalize remains limited in complex real-world tasks such as mobile navigation, where diverse applications, frequent system changes, and customized workflows are common in practice. We argue that a fundamental bottleneck lies in whether an agent possesses sufficient task-specific procedural knowledge to accomplish a given goal. Such procedural knowledge may be provided by the general capabilities of large language models, or obtained from additional external resources such as web search when necessary. Based on this view, we propose Procedure-Aware Multimodal Agent with Meta Reasoning, a framework that explicitly represents task knowledge as natural-language procedures and trains a procedure-aware grounded agent to condition its actions on this knowledge. By learning to leverage procedural knowledge from different sources, our approach enables robust generalization across tasks, applications, interface versions, and multi-app workflows, achieving substantial improvements on challenging Android benchmarks.

We study \textbf{code-to-metric regression}: predicting numeric outcomes of code executions, a challenging task due to the open-ended nature of programming languages. While prior methods have resorted to heavy and domain-specific feature engineering, we show that a single unified Regression Language Model (RLM) using a frozen LLM encoder can simultaneously predict directly from text, (i) the memory footprint of code across multiple high-level languages such as Python and C++, (ii) the latency of Triton GPU kernels, and (iii) the accuracy and speed of trained neural networks represented in ONNX. In particular, a relatively small 300M parameter RLM based on T5Gemma, obtains $>$0.9 Spearman-rank on competitive programming submissions from APPS, and a single unified model achieves $>$0.5 average Spearman-rank across 24 different programming languages from CodeNet. Furthermore, the RLM can obtain the highest average Kendall-Tau of 0.46 on five classic NAS design spaces previously dominated by graph neural networks, and simultaneously predict architecture latencies on numerous hardware platforms.

Any-to-any modeling aims to flexibly relate arbitrary modalities within a single system, a requirement that arises across multimodal learning and scientific domains such as ecology and astronomy. However, existing any-to-any approaches are typically trained from scratch using encoder–decoder or diffusion architectures, limiting empirical performance and the use of pretrained models. We investigate decoder-only any-to-any multimodal modeling, which treats all modalities symmetrically and supports arbitrary modalities as inputs and outputs without modality-specific heads, losses, or task pipelines. As a consequence of this unified design, the resulting model MODUS naturally enables chained generation through intermediate modalities, cross-modal consistency verification, and analysis of visual representations by combining semantic and reconstruction features. Across a range of benchmarks, MODUS demonstrates strong out-of-the-box performance and flexible multimodal composition within a single model.

Existing voice deepfake detection and localization models rely heavily on representations extracted from speech foundation models (SFMs). However, downstream finetuning has now reached a state of diminishing returns. In this paper, we shift the focus to pretraining and propose a novel recipe that combines *bottleneck masked embedding prediction with flow-matching based spectrogram reconstruction*. The outcome, *Alethia*, is the first foundational audio encoder for various voice deepfake detection and localization tasks. We evaluate on 5 different tasks with 56 benchmark datasets, and note *Alethia* significantly outperforms state-of-the-art SFMs with superior robustness to real-world perturbations and zero-shot generalization to unseen domains (e.g., singing deepfakes). We also demonstrate the limitation of discrete targets in masked token prediction, and show the importance of *continuous embedding* prediction and *generative pretraining* for capturing deepfake artifacts.

In this work, we aim to develop effective data synthesis techniques that autonomously synthesize multimodal training data for enhancing MLLMs in solving complex real-world tasks. To this end, we propose Collective Adversarial Data Synthesis (CADS), a novel and general approach to synthesize high-quality, diverse and challenging multimodal data for MLLMs. The core idea of CADS is to leverage collective intelligence to ensure high-quality and diverse generation, while exploring adversarial learning to synthesize challenging samples for effectively driving model improvement. Specifically, CADS operates with two cyclic phases, i.e., Collective Adversarial Data Generation (CAD-Generate) and Collective Adversarial Data Judgment (CAD-Judge). CAD-Generate leverages collective knowledge to jointly generate new and diverse multimodal data, while CAD-Judge collaboratively assesses the quality of synthesized data. In addition, CADS introduces an Adversarial Context Optimization mechanism to optimize the generation context to encourage challenging and high-value data generation. With CADS, we construct MMSynthetic-20K and train our model R1-SyntheticVL, which demonstrates superior performance on various benchmarks.

Frontier models are transitioning from _multimodal large language models_ (MLLMs) that merely ingest visual information to _unified multimodal models_ (UMMs) capable of native interleaved generation. This shift has sparked interest in using intermediate visualizations as a reasoning aid, akin to human _mental imagery_. Central to this idea is the ability to form, maintain, and manipulate visual representations in a goal-oriented manner. To evaluate and probe this capability, we develop MentisOculi, a procedural, stratified suite of multi-step reasoning problems amenable to visual solution, tuned to challenge frontier models. Evaluating visual strategies ranging from latent tokens to explicit generated imagery, we find they generally fail to improve performance. Analysis of UMMs specifically exposes a critical limitation: While they possess the textual reasoning capacity to solve a task and can sometimes generate correct visuals, they suffer from compounding generation errors and fail to leverage even ground-truth visualizations. Our findings suggest that despite their inherent appeal, _visual thoughts do not yet benefit model reasoning_. MentisOculi establishes the necessary foundation to analyze and close this gap across diverse model families.

Vision-Language Models (VLMs) inherit the auto-regressive generation paradigm and cache the keys and values (KV) of all previous tokens to accelerate inference, resulting in memory consumption that scales linearly with context length. This issue is particularly pronounced in VLMs due to substantial redundancy in the visual modality. Although KV cache eviction approaches can effectively reduce inference memory, they often incur significant performance degradation in VLMs, as most are designed for language models and overlook the inherent gap between text and vision. By systematically analyzing the modality gap in VLMs in this work, we argue that the importance of visual information should be grounded in textual guidance and accordingly propose a **T**ext-**G**rounded KV Eviction method for **V**LMs (**TGV-KV**). TGV-KV comprises three submodules: *(1) Text-Vision Budgeting (TVB)* assigns budget to each layer based on the mutual information interaction. *(2) Text-Weighted Ranking (TWR)* assesses the priority of text and ranks vision importance based on weighted text-image attention. *(3) Text-Prioritised Retention (TPR)* policy strategically preserves text KV to avoid acute information loss. We evaluate TGV-KV across five models with different sizes and architectures, showing that TGV-KV preserves 99.2% full-KV accuracy on the VizWiz-VQA task with LLaVA-NeXT and boosts end-to-end throughput by 52.6% with an extreme retention budget of 5%. [Code Link](https://github.com/Danielement321/TGV-KV).

Large diffusion vision–language models (LDVLMs) have recently emerged as a promising alternative to autoregressive models, enabling parallel decoding for efficient inference and leveraging bidirectional attention for global context. Despite these advances, their behavior under long-form generation remains underexplored. In this work, we show that existing LDVLMs suffer from repetitive generation and degraded visual grounding, and identify two underlying causes. First, repetitive generation originates from a mask token prior: since generation tokens are initialized as mask tokens, their hidden representations progressively drift toward a shared prior direction over generation steps. Second, a fundamental misalignment between the positional attention bias and the iterative unmasking process suppresses attention toward informative visual tokens, degrading visual grounding. Based on these insights, we propose a training-free approach, introducing Mask Prior Suppression and Monotonic RoPE Scaling to mitigate mask prior drift and positional attention collapse during decoding. Experiments on general multimodal benchmarks and visual grounding tasks demonstrate improvements over baseline LDVLMs, with robust gains on long-form description benchmarks. Our results show that these failures can be effectively addressed with a lightweight, plug-and-play strategy that requires no additional training and generalizes across diverse LDVLM architectures.

Relational Deep Learning aims to learn directly on multi-table databases, yet current methods face a fundamental tension: Transformers' quadratic complexity prohibits the large contexts relational data demands, while GNNs sacrifice global context for efficiency. We introduce Ramba, the first selective state-space model for relational databases. Our approach features two innovations: (1) Topology-Aware Linearization, which processes cells via global columnar serialization in O(L) complexity while recovering relational structure through sparse entity and foreign-key attention masks; and (2) Schema Dynamic Gating, which modulates SSM state transitions based on semantic alignment between the currently scanned attribute and the prediction target, enabling cross-table relevance filtering without relying on value distributions. Together, these enable Ramba to ingest vast relational contexts while selectively retaining semantically relevant information, a combination existing architectures cannot achieve. Experiments demonstrate state-of-the-art performance with linear scalability across diverse relational benchmarks.

While scaling laws guide compute allocation for LLM pre-training, analogous prescriptions for reinforcement learning (RL) post-training of LLMs remain poorly understood. We study the compute-optimal allocation of sampling compute for on-policy RL methods in LLMs, framing scaling as a compute-constrained optimization over three resources: parallel rollouts per problem, number of problems per batch, and number of update steps. We find that the compute-optimal number of parallel rollouts per problem increases predictably with compute budget and then saturates. This trend holds across both easy and hard problems, though driven by different mechanisms: solution sharpening on easy problems and coverage expansion on hard problems. We further show that increasing the number of parallel rollouts mitigates interference across problems, while the number of problems per batch primarily affects training stability and can be chosen within a broad range. Validated across base models and data distributions, our results recast RL scaling laws as prescriptive allocation rules and provide practical guidance for compute-efficient LLM RL post-training.

Real-world multivariate time series are inherently in flux: different variables evolve asynchronously and interact in complex, time-varying ways, yet accurate forecasting requires these dispersed signals to converge into a single unified prediction. This structural mismatch between dynamic, heterogeneous inputs and a unified forecasting objective poses a fundamental challenge for building general-purpose multivariate forecasting models, especially in zero-shot and large-scale settings. To this end, inspired by the idea that ``\emph{all rivers run into the sea}'', we propose \textbf{ConFlux}, a \emph{general-purpose foundation model for multivariate time-series forecasting} by learning to adaptively integrate cross-channel information under a unified forecasting objective. Specifically, ConFlux first reorders variables to reduce cross-variable entanglement, then aggregates adjacent variables into compact patches that can be processed by a Vision Transformer-style architecture. This design shortens the effective context, reduces attention complexity, and provides a unified token representation for pre-training and downstream tasks. Experiments on 25 public datasets show that ConFlux achieves state-of-the-art performance in zero-shot, fine-tuning, and from-scratch settings, while offering faster inference and lower memory usage.