Tokenizers provide the fundamental basis through which text is represented and processed by language models (LMs). Despite the importance of tokenization, its role in LM performance and behavior is poorly understood due to the challenge of measuring the impact of tokenization in isolation. To address this need, we present TokSuite, a collection of models and a benchmark that supports research into tokenization's influence on LMs. Specifically, we release fourteen pre-trained models that use different tokenizers but are otherwise identical, using the same architecture, dataset, training budget, and initialization. We also release a multilingual robustness benchmark that measures model performance under real-world perturbations in English, Chinese, Farsi, Italian, and Turkish, curated by native annotators. Together, TokSuite allows robust decoupling of the influence of a model's tokenizer, supporting a series of novel findings that elucidate the respective benefits and shortcomings of a wide range of popular tokenizers.
论文检索
输入标题、作者或关键词，从 13,033 篇学术成果中精准定位
Applications · Computer Vision
Diffusion Transformers (DiTs) have emerged as a powerful backbone for image generation, offering superior scalability over U-Nets. However, their practical deployment is hindered by significant computational costs. While Quantization-Aware Training (QAT) shows promise, its application to DiTs is challenged by the high sensitivity and complex distributions of activations. Identifying activation quantization as the primary bottleneck for low-bit settings, we propose **RobuQ**, a systematic QAT framework. We first establish a strong ternary weight (W1.58A4) baseline. Building on this, we introduce **RobustQuantizer**, which utilizes the Hadamard transform to convert unknown per-token distributions into normal distributions. Furthermore, we propose **AMPN**, the first **A**ctivation-only **M**ixed-**P**recision **N**etwork pipeline, applying ternary weights globally while allocating layer-specific activation precisions to eliminate information bottlenecks. Extensive experiments demonstrate that **RobuQ** achieves state-of-the-art performance on **ImageNet-1K**, representing the first stable image generation with activations quantized to an average of 2 bits.
Semi-LAR: Semi-supervised Contrastive Learning with Linear Attention for Removal of Nighttime Flares
Applications · Computer Vision
Lens flare removal is challenging due to the large spatial extent of flare artifacts and their entangle-ment with scene structures, while existing meth-ods heavily rely on large-scale paired data. We propose a semi-supervised flare removal frame-work that enables stable learning from unlabeled images by jointly addressing pseudo-label relia-bility and representation discrimination. We pro-pose an adaptive pseudo-label repository that pro-gressively refines pseudo supervision through no-reference quality assessment, momentum-based updates, and invalid label filtering, effectively mit-igating error accumulation. Moreover, we pro-pose a flare-aware contrastive loss that explic-itly treats flare-contaminated inputs as negatives and performs patch-level contrastive learning, en-couraging representations that are discriminative against flare patterns while remaining consistent with reliable pseudo targets. Extensive experi-ments on multiple flare benchmarks demonstrate that the proposed framework is model-agnostic and consistently improves performance and ro-bustness.
While multi-agent systems (MAS) promise elevated intelligence through coordination of agents, current approaches to automatic MAS design under-deliver. Such shortcomings stem from two key factors: (1) methodological complexity – agent orchestration is performed using sequential, code-level execution that limits global system-level holistic reasoning and scales poorly with agent complexity – and (2) efficacy uncertainty – MAS are deployed without understanding if there are tangible benefits compared to single-agent systems (SAS). We propose MASOrchestra, a training-time framework that formulates MAS orchestration as a function-calling reinforcement learning problem with holistic orchestration, generating an entire MAS at once. In MAS-Orchestra, complex, goal-oriented subagents are abstracted as callable functions, enabling global reasoning over system structure while hiding internal execution details. To rigorously study when and why MAS are beneficial, we introduce MASBENCH, a controlled benchmark that characterizes tasks along five axes: Depth, Horizon, Breadth, Parallel, and Robustness. Our analysis reveals that MAS gains depend critically on task structure, verification protocols, and the capabilities of both orchestrator and subagents, rather than holding universally. Guided by these insights, MAS-Orchestra achieves consistent improvements on public benchmarks including mathematical reasoning, multi-hop QA, and search-based QA, while achieving more than 10× efficiency over strong baselines. Together, MAS-Orchestra and MASBENCH enable better training and understanding of MAS in the pursuit of multi-agent intelligence.
General Machine Learning · Evaluation
Large language models (LLMs) commonly boost reasoning via sample-evaluate-ensemble decoders (e.g., majority voting), achieving label free gains without ground truth. However, prevailing strategies score candidates using only external outputs such as token probabilities, entropies, or self evaluations, and these signals can be poorly calibrated after post training. We instead analyze internal behavior based on neuron activations and uncover three findings: (1) external signals are low dimensional projections of richer internal dynamics; (2) correct responses activate substantially fewer unique neurons than incorrect ones throughout generation; and (3) activations from correct responses exhibit stronger cross sample agreement, whereas incorrect ones diverge. Motivated by these observations, we propose Neuron Agreement Decoding (NAD), an unsupervised best of N method that selects candidates using activation sparsity and cross sample neuron agreement, operating solely on internal signals and without requiring comparable textual outputs. NAD enables early correctness prediction within the first 32 generated tokens and supports aggressive early stopping. Across math and science benchmarks with verifiable answers, NAD matches majority voting; on open ended coding benchmarks where majority voting is inapplicable, NAD consistently outperforms Avg@64. By pruning unpromising trajectories early, NAD reduces token usage by 99% with minimal loss in generation quality, showing that internal signals provide reliable, scalable, and efficient guidance for label free ensemble decoding.
Applications · Everything Else
Despite the success of large transformers at modeling variable-length protein sequences, most statistical phylogenetics analyses use relatively simple continuous-time finite-state Markov models of point substitution to describe molecular evolution, keeping sequence length fixed and ignoring insertions and deletions (indels) entirely. The simplistic assumptions of these models limit the realism of such analyses. We extend the TKF92 model - the canonical hierarchical model combining an outer birth-death process for indels with an inner finite-state Markov chain for substitutions - by introducing additional nesting and latent states. We compare these TKF92 extensions (which are exactly solvable, and in which evolutionary time naturally appears as a matrix exponential coefficient) to two classes of neural seq2seq models that take evolutionary time as an input feature: the first class of model being constrained to enforce a TKF92-like structure, and the second lacking any such constraint. We evaluate the per-character perplexities of all models on splits of the PFam database of aligned protein domains. A nested TKF-based model with only 32,000 parameters is highly competitive with neural networks containing tens of millions of parameters, outperforming all but two of the neural architectures tested. Our results indicate that approaches grounded in molecular evolutionary theory may be more parameter-efficient and provide a better fit to real alignments than unconstrained alternatives, supporting the incorporation of classical model structure within future neural phylogenetic approaches.
Deep Learning · Attention Mechanisms
Scaling language models to long contexts is often bottlenecked by the size of the key-value (KV) cache. In deployed settings, long contexts are typically managed through *compaction* in token space via summarization. However, summarization can be highly lossy, substantially harming downstream performance. Recent work on Cartridges (Eyuboglu et al., 2025) has shown that it is possible to *train* highly compact KV caches in latent space that closely match full-context performance, but at the cost of slow and expensive end-to-end optimization. This work describes an approach for *fast* context compaction in latent space through **Attention Matching**, which constructs compact keys and values to reproduce attention outputs and preserve attention mass at a per-KV-head level. We show that this formulation naturally decomposes into simple subproblems, some of which admit efficient closed-form solutions. Within this framework, we develop a family of methods that significantly push the Pareto frontier of compaction time versus quality, achieving up to $50\times$ compaction in seconds on some datasets with little quality loss.
Deep Learning · Foundation Models
Unified Multimodal Modeling aims to integrate visual understanding and generation within a single system. However, existing approaches typically rely on two disparate visual tokenizers, which splits the representation space and hinder truly unified modeling. We propose UniAR, a unified autoregressive framework where a single discrete visual tokenizer serves as the key bridge between understanding and generation, enabling a shared context in which the model can directly interpret its own generated visual tokens without additional re-encoding. UniAR adapts a pretrained vision encoder with multi-level feature fusion and a lookup-free bitwise quantization scheme, preserving both high-level semantics and low-level details while scaling the effective visual vocabulary at minimal cost. Building on this, the unified autoregressive model adopts parallel-bitwise-prediction to jointly predict spatially grouped, multi-level visual codes, substantially reducing visual sequence length and accelerating generation. Finally, a diffusion-based visual decoder operates on discrete visual tokens to reconstruct high-fidelity images. Through large-scale pre-training on 1T multimodal tokens, followed by supervised fine-tuning and reinforcement learning, UniAR achieves state-of-the-art performance on text-to-image generation and image editing while remaining competitive on multimodal understanding benchmarks.
Applications · Computer Vision
Generating high-fidelity 3D indoor scenes remains a significant challenge due to data scarcity and the complexity of modeling intricate spatial relations. Current methods often struggle to scale beyond training distribution to dense scenes or rely on Large Language Models (LLMs) that lack the ability for precise spatial reasoning. Building on top of the observation that object placement relies mainly on local dependencies instead of information-redundant global distributions, in this paper, we propose **Pair2Scene**, a novel procedural generation framework, for scene generation based on a set of *learned procedural rules*. These rules mainly capture two types of inter-object relations, namely *support relations* that follow physical hierarchies, and *functional relations* that reflect semantic links. We model these rules through a network, which estimates spatial position distributions of dependent objects conditioned on position and geometry of the anchor ones. Accordingly, we curate a dataset **3D-Pairs** from existing scene data to train the model. During inference, our framework can generate scenes by recursively applying our model. Extensive experiments demonstrate that our framework outperforms existing methods in generating complex environments that go beyond training data while maintaining physical and semantic plausibility.
Applications · Computer Vision
Distribution matching distillation (DMD) aligns a multi-step generator with its few-step counterpart to enable high-quality generation under low inference cost. However, DMD tends to suffer from mode collapse, as its reverse-KL formulation inherently encourages mode-seeking behavior, for which existing remedies typically rely on perceptual or adversarial regularization, thereby incurring substantial computational overhead and training instability. In this work, we propose a role-separated distillation framework that explicitly disentangles the roles of distilled steps: the first step is dedicated to preserving sample diversity via a target-prediction (e.g., v-prediction) objective, while subsequent steps focus on quality refinement under the standard DMD loss, with gradients from the DMD objective blocked at the first step. We term this approach Diversity-Preserved DMD (**DP-DMD**), which, despite its simplicity—no perceptual backbone, no discriminator, no auxiliary networks, and no additional ground-truth images—preserves sample diversity while maintaining visual quality on par with state-of-the-art methods in extensive text-to-image experiments. Code and models will be released publicly.
General Machine Learning · Representation Learning
Connectionist models and symbolic models have long embodied two divergent paradigms: the former excel at differentiable representation learning yet struggle with transparency, while the latter deliver explicit rule-based reasoning but resist gradient-based optimization. We introduce Arboreal Neural Networks (ArbNN), a neural–symbolic framework that unifies these paradigms both computationally and conceptually. At the design level, ArbNN departs fundamentally from prior neuralized-tree models through a depth-aware routing mechanism and a topology-informed softmax aggregation, which together enable one-shot full-path gradient propagation and consequently achieving rapid and well-conditioned optimization dynamics and high parallel inference efficiency. At the conceptual level, ArbNN reveals that decision-tree branching and self-attention routing are two realizations of the same conditional computation primitive. We prove a structural isomorphism between a decision tree and a single-query attention head, enabling a differentiable architecture that faithfully preserves symbolic decision logic. The defining property of ArbNN is Bidirectional Fidelity, ensuring that the neural module can be compiled from—and losslessly decompiled back into—a symbolic tree, yielding both ordering consistency in ranking behavior and explicit, auditable interpretability via reconstructed if–else rules. ArbNN further supports GBDT-based initialization, allowing it to inherit strong inductive biases and integrate seamlessly with existing production workflows. Empirically, ArbNN achieves state-of-the-art performance on various public tabular benchmarks and delivers consistent gains under temporal distribution shift in large-scale industrial credit-risk systems. To support realistic evaluation, we additionally contribute TabCredit, a feature-rich, temporally partitioned dataset built from millions of real-world loan applications. Together, these results demonstrate that ArbNN forms a unified, reversible, and practically deployable bridge between symbolic reasoning and neural computation for high-stakes tabular domains.
Social Aspects · Security
The integration of external tools has substantially expanded the capabilities of large language model (LLM) agents, but also introduced new attack surfaces beyond prompt injection. In particular, cross-tool description poisoning can manipulate planner-visible tool metadata to steer an agent’s trajectory, even if the poisoned tool itself is never chosen. To understand the effectiveness of existing attacks against this emerging threat, we evaluate several existing agent defenses against prompt-injection and find they transfer poorly to cross-tool description poisoning. Building on this insight, we propose Tool-Guard, a novel defense based on a new concept called isolated planning, in which tool invocations that are detected as misaligned or suspicious cause the corresponding tool to be placed in a quarantined list (the influenced list), breaking further influence from poisoned descriptions. With this influence isolated, the tool can continue to be used to support the task, enabling a robust defense that preserves legitimate tool utility. Experiments on the AgentDojo and ASB benchmarks show that \sysname substantially reduces attack success while maintaining high task utility.
Applications · Computer Vision
3D Scene Graph Generation (3DSGG) aims to create a structured representation of 3D environment by identifying objects as nodes and their relations as edges. Existing 3DSGG methods based on RGB-D sequences typically put much focus on the adaption of neural networks to robust node and edge feature extraction in complex 3D scenes, yet ignoring the inherent intra-class diversity within each class and inter-class similarity between different categories associated with nodes and edges. In this work, we develop GMPSSG, a novel Gaussian Mixture-distributed Prototype mining framework for 3DSGG. Specifically, we model different categories with independent Gaussian Mixture-distributed Prototype to effectively mitigate inter-class similarity, while employing multiple Gaussian components within each prototype to capture intra-class diversity. Moreover, Prototype-anchored Representation Learning is introduced to construct a well-structured and mutually independent category space; Topology-aware Prototype Interaction is devised to capture implicit co-occurrence priors within the scene, and leverage them to calibrate prototype distributions, thereby ensuring the plausibility of node-edge matching. Experiments on 3DSSG dataset demonstrate GMPSSG outperforms various top-leading methods. Source code will be released.
Theory · Game Theory
When organizations delegate text generation tasks to AI providers via pay-for-performance contracts, expected payments rise when evaluation is noisy. As evaluation methods become more elaborate, the economic benefits of decreased noise are often overshadowed by increased evaluation costs. In this work, we introduce adaptive contracts for AI delegation, which allow detailed evaluation to be performed selectively after observing an initial coarse signal in order to conserve resources. We make three sets of contributions: First, we provide efficient algorithms for computing optimal adaptive contracts under natural assumptions or when core problem dimensions are small, and prove hardness of approximation in the general unstructured case. We then formulate alternative models of randomized adaptive contracts and discuss their benefits and limitations. Finally, we empirically demonstrate the benefits of adaptivity over non-adaptive baselines using question-answering and code-generation datasets.
Transformers have demonstrated remarkable in-context learning (ICL) capabilities. The strong ICL performance of transformers is commonly believed to arise from their ability to implicitly execute certain algorithms on the context, thereby enhancing prediction and generation. In this work, we investigate how transformers with softmax attention perform in-context learning on linear classification data. We first construct a class of multi-layer transformers that can perform in-context logistic regression, with each layer exactly performing one step of normalized gradient descent on an in-context loss. Then, we show that our constructed transformer can be obtained through (i) training a single self-attention layer supervised by one-step gradient descent, and (ii) recurrently applying the trained layer to obtain a looped model. Training convergence guarantees of the self-attention layer and out-of-distribution generalization guarantees of the looped model are provided. Our results advance the theoretical understanding of ICL mechanism by showcasing how softmax transformers can effectively act as in-context learners.
Reinforcement Learning · Deep RL
On-policy deep reinforcement learning remains a dominant paradigm for continuous control, yet standard implementations rely on Gaussian actors and relatively shallow MLP policies, often leading to brittle optimization when gradients are noisy and policy updates must be conservative. In this paper, we revisit policy representation as a first-class design choice for on-policy optimization. We study discretized categorical actors that represent each action dimension with a distribution over bins, yielding a policy objective that resembles a cross-entropy loss. Building on architectural advances from supervised learning, we further propose regularized actor networks, while keeping critic design fixed. Our results show that simply replacing the standard actor network with our discretized regularized actor yields consistent gains and achieve the state-of-the-art performance across diverse continuous-control benchmarks.
Applications · Computer Vision
Dataset condensation aims to construct compact datasets from real data via synthesis or selection. However, existing approaches are ill-suited for diffusion model training: synthetic data generation often yields low-fidelity samples unsuitable for authentic modeling, while real subset selection typically fails to preserve the distributional geometry required by diffusion likelihood objectives. To address this, we propose to reformulate real subset selection as a geometry-aware distribution alignment problem. By incorporating one-sided partial optimal transport, our method selectively aligns a compact subset with the full data distribution while allowing unmatched mass in low-density regions, ensuring the preserved geometric structure necessary for effective diffusion model training. To further ensure distributional fidelity, we complement geometric alignment with lightweight feature-statistics and semantic consistency regularization. An efficient two-stage discrete optimization strategy is proposed to achieve this alignment objective. Extensive experiments across diffusion variants, subset sizes, image resolutions, and training rounds show that our approach achieves superior fidelity and distributional coverage in diffusion model training. Codes are available at https://anonymous.4open.science/r/ICML2026_4092.
Deep Learning · Large Language Models
Low-Rank Adaptation (LoRA) is a standard tool for parameter-efficient finetuning of large models. While it induces a small memory footprint, its training dynamics can be surprisingly complex as they depend on several hyperparameters such as initialization, adapter rank, and learning rate. In particular, it is unclear *how the optimal learning rate scales with adapter rank*, which forces practitioners to re-tune the learning rate whenever the rank is changed. In this paper, we introduce *Maximal-Update Adaptation* ($\mu$A), a theoretical framework that characterizes how the "optimal" learning rate should scale with model width and adapter rank to produce stable, non-vanishing feature updates under standard configurations. Our analysis leverages techniques from hyperparameter transfer and reveals that the optimal learning rate exhibits different scaling patterns depending on initialization and LoRA scaling factor. Specifically, we identify two regimes: one where the optimal learning rate remains roughly invariant across ranks, and another where it scales inversely with rank. We further identify a configuration that allows learning rate transfer from LoRA to full finetuning, drastically reducing the cost of learning rate tuning for full finetuning. Experiments across language, vision, vision–language, image generation, and reinforcement-learning tasks validate our scaling rules and show that learning rates tuned on LoRA transfer reliably to full finetuning.
Reinforcement Learning · Deep RL
Accelerating exploration in visual reinforcement learning under sparse rewards remains challenging due to the substantial task-irrelevant variations. Despite advances in intrinsic exploration, many methods either assume access to low-dimensional states or lack task-aware exploration strategies, thereby rendering them fragile in visual domains. To bridge this gap, we present TEB, a Task-aware Exploration approach that tightly couples task-relevant representations with exploration through a predictive Bisimulation metric. Specifically, TEB leverages the metric not only to learn behaviorally grounded task representations but also to measure behaviorally intrinsic novelty over the learned latent space. To realize this, we first theoretically mitigate the representation collapse of degenerate bisimulation metrics under sparse rewards by internally introducing a simple but effective predicted reward differential. Building on this robust metric, we design potential-based exploration bonuses, which measure the relative novelty of adjacent observations over the latent space. Extensive experiments on MetaWorld and Maze2D show that TEB achieves superior exploration ability and outperforms recent baselines.
Deep Learning · Generative Models and Autoencoders
Recently, sparse autoencoders (SAEs) have emerged as an attractive tool for interpreting and interacting with representations in practical neural networks. While it is common empirical folklore, we also show theoretically that SAEs are highly unstable: different training runs are likely to produce different concept dictionaries and sparse codes. We characterize the model characteristics that get in the way of the stability of real-world SAEs, and address each of these problems through minimal changes to the architecture and training procedure. Together, these changes yield iSAE, a variant of the standard TopK SAE with lower reconstruction error and improved stability. We explain this improvement theoretically by connecting SAEs with traditional dictionary learning approaches, and show that the dictionaries learned in practice satisfy an approximate restricted isometry condition, rendering the corresponding sparse codes in those models near-identifiable.