ICML 2025 Papers

Decoding with autoregressive language models traditionally occurs sequentially, generating one token after another. Recent attempts to introduce parallelism require a pre-determined structure in the generated content to implement parallel generation, such as by pattern-matching on bullet points. In this work, we present a new technique to automate parallel generation by dynamically exploiting the semantic independence of generation outputs to implement asynchronous decoding. We introduce an annotation language Pasta-Lang for language models to initiate asynchronous decoding at inference time. We also develop an accompanying Pasta-Lang interpreter that performs on-the-fly asynchronous decoding, effectively implementing parallel generation and speeding up inference. We present an instruction-finetuning dataset with Pasta-Lang-annotated responses for teaching LLMs to annotate semantic independence with Pasta-Lang as well as the methodology for creating the dataset. Our evaluation shows using the interpreter with a Pasta-Lang-equipped model achieves significant speedup while maintaining the same generation quality.

Pretraining large language models (LLMs) is resource-intensive, often requiring months of training time even with high-end GPU clusters. There are two approaches of mitigating such computational demands: reusing smaller models to train larger ones (upcycling), and training computationally efficient models like mixture-of-experts (MoE). In this paper, we study the upcycling of LLMs to MoE models, of which the scaling behavior remains underexplored. Through extensive experiments, we identify empirical scaling laws that describe how performance depends on dataset size and model configuration. Particularly, we show that, while scaling these factors improves performance, there is a novel interaction term between the dense and upcycled training dataset that limits the efficiency of upcycling at large computational budgets. Based on these findings, we provide guidance to scale upcycling, and establish conditions under which upcycling outperforms from-scratch trainings within budget constraints.

The Value Iteration Network (VIN) is an end-to-end differentiable neural network architecture for planning. It exhibits strong generalization to unseen domains by incorporating a differentiable planning module that operates on a latent Markov Decision Process (MDP). However, VINs struggle to scale to long-term and large-scale planning tasks, such as navigating a $100\times 100$ maze---a task that typically requires thousands of planning steps to solve. We observe that this deficiency is due to two issues: the representation capacity of the latent MDP and the planning module's depth. We address these by augmenting the latent MDP with a dynamic transition kernel, dramatically improving its representational capacity, and, to mitigate the vanishing gradient problem, introduce an "adaptive highway loss" that constructs skip connections to improve gradient flow. We evaluate our method on 2D/3D maze navigation environments, continuous control, and the real-world Lunar rover navigation task. We find that our new method, named Dynamic Transition VIN (DT-VIN), scales to 5000 layers and solves challenging versions of the above tasks. Altogether, we believe that DT-VIN represents a concrete step forward in performing long-term large-scale planning in complex environments.

Scaling has been a major driver of recent advancements in deep learning. Numerous empirical studies have found that scaling laws often follow the power-law and proposed several variants of power-law functions to predict the scaling behavior at larger scales. However, existing methods mostly rely on point estimation and do not quantify uncertainty, which is crucial for real-world applications involving decision-making problems such as determining the expected performance improvements achievable by investing additional computational resources. In this work, we explore a Bayesian framework based on Prior-data Fitted Networks (PFNs) for neural scaling law extrapolation. Specifically, we design a prior distribution that enables the sampling of infinitely many synthetic functions resembling real-world neural scaling laws, allowing our PFN to meta-learn the extrapolation. We validate the effectiveness of our approach on real-world neural scaling laws, comparing it against both the existing point estimation methods and Bayesian approaches. Our method demonstrates superior performance, particularly in data-limited scenarios such as Bayesian active learning, underscoring its potential for reliable, uncertainty-aware extrapolation in practical applications.

When trained on large-scale object classification datasets, certain artificial neural network models begin to approximate core object recognition behaviors and neural response patterns in the primate brain. While recent machine learning advances suggest that scaling compute, model size, and dataset size improves task performance, the impact of scaling on brain alignment remains unclear. In this study, we explore scaling laws for modeling the primate visual ventral stream by systematically evaluating over 600 models trained under controlled conditions on benchmarks spanning V1, V2, V4, IT and behavior. We find that while behavioral alignment continues to scale with larger models, neural alignment saturates. This observation remains true across model architectures and training datasets, even though models with stronger inductive biases and datasets with higher-quality images are more compute-efficient. Increased scaling is especially beneficial for higher-level visual areas, where small models trained on few samples exhibit only poor alignment. Our results suggest that while scaling current architectures and datasets might suffice for alignment with human core object recognition behavior, it will not yield improved models of the brain's visual ventral stream, highlighting the need for novel strategies in building brain models.

Deep reinforcement learning for continuous control has recently achieved impressive progress. However, existing methods often suffer from primacy bias—a tendency to overfit early experiences stored in the replay buffer—which limits an RL agent’s sample efficiency and generalizability. A common existing approach to mitigate this issue is periodically resetting the agent during training. Yet, even after multiple resets, RL agents could still be impacted by early experiences. In contrast, humans are less susceptible to such bias, partly due to *infantile amnesia*, where the formation of new neurons disrupts early memory traces, leading to the forgetting of initial experiences. Inspired by this dual processes of forgetting and growing in neuroscience, in this paper, we propose *Forget and Grow* (**FoG**), a new deep RL algorithm with two mechanisms introduced. First, *Experience Replay Decay (ER Decay)*—"forgetting early experience''—which balances memory by gradually reducing the influence of early experiences. Second, *Network Expansion*—"growing neural capacity''—which enhances agents' capability to exploit the patterns of existing data by dynamically adding new parameters during training. Empirical results on four major continuous control benchmarks with more than 40 tasks demonstrate the superior performance of **FoG** against SoTA existing deep RL algorithms, including BRO, SimBa and TD-MPC2.

AI Scaling has traditionally been synonymous with Scaling Up, which builds larger and more powerful models. However, the growing demand for efficiency, adaptability, and collaboration across diverse applications necessitates a broader perspective. This position paper presents a holistic framework for AI scaling, encompassing Scaling Up, Scaling Down, and Scaling Out. It argues that while Scaling Up of models faces inherent bottlenecks, the future trajectory of AI scaling lies in Scaling Down and Scaling Out. These paradigms address critical technical and societal challenges, such as reducing carbon footprint, ensuring equitable access, and enhancing cross-domain collaboration. We explore transformative applications in healthcare, smart manufacturing, and content creation, demonstrating how AI Scaling can enable breakthroughs in efficiency, personalization, and global connectivity. Additionally, we highlight key challenges, including balancing model complexity with interpretability, managing resource constraints, and fostering ethical development. By synthesizing these approaches, we propose a unified roadmap that redefines the future of AI research and application, paving the way for advancements toward Artificial General Intelligence (AGI).

Inspired by the recent success of LLMs, the field of human motion understanding has increasingly shifted toward developing large motion models. Despite some progress, current efforts remain far from achieving truly generalist models, primarily due to the lack of massive high-quality data. To address this gap, we present MotionLib, the first million-level dataset for motion generation, which is at least 15$\times$ larger than existing counterparts and enriched with hierarchical text descriptions. Using MotionLib, we train a large motion model named Being-M0, demonstrating robust performance across a wide range of human activities, including unseen ones. Through systematic investigation, for the first time, we highlight the importance of scaling both data and model size for advancing motion generation, along with key insights to achieve this goal. To better integrate the motion modality, we propose Motionbook, an innovative motion encoding approach including (1) a compact yet lossless feature to represent motions; (2) a novel 2D lookup-free motion tokenizer that preserves fine-grained motion details while expanding codebook capacity, significantly enhancing the representational power of motion tokens. We believe this work lays the groundwork for developing more versatile and powerful motion generation models in the future. For further details, visit https://beingbeyond.github.io/Being-M0/.

Increasing diversity in language models is a challenging yet essential objective. A common approach is to raise the decoding temperature. In this work, we investigate this approach through a simplistic yet common case to provide insights into why decreasing temperature can improve quality (Precision), while increasing it often fails to boost coverage (Recall). Our analysis reveals that for a model to be effectively tunable through temperature adjustments, it must be trained toward coverage. To address this, we propose rethinking loss functions in language models by leveraging the Precision-Recall framework. Our results demonstrate that this approach achieves a substantially better trade-off between Precision and Recall than merely combining negative log-likelihood training with temperature scaling. These findings offer a pathway toward more versatile and robust language modeling techniques.

Graph neural networks (GNNs) have demonstrated remarkable success in graph representation learning and various sampling approaches have been proposed to scale GNNs to applications with large-scale graphs. A class of promising GNN training algorithms take advantage of historical embeddings to reduce the computation and memory cost while maintaining the model expressiveness of GNNs. However, they incur significant computation bias due to the stale feature history. In this paper, we provide a comprehensive analysis of their staleness and inferior performance on large-scale problems. Motivated by our discoveries, we propose a simple yet highly effective training algorithm (REST) to effectively reduce feature staleness, which leads to significantly improved performance and convergence across varying batch sizes, especially when staleness is predominant. The proposed algorithm seamlessly integrates with existing solutions, boasting easy implementation, while comprehensive experiments underscore its superior performance and efficiency on large-scale benchmarks. Specifically, our improvements to state-of-the-art historical embedding methods result in a 2.7\% and 3.6\% performance enhancement on the ogbn-papers100M and ogbn-products dataset respectively, accompanied by notably accelerated convergence. The code can be found at https://github.com/RXPHD/REST.

The residual stream acts as a memory bus where transformer layers both store and access features (Elhage et al., 2021). We consider changing the mechanism for retrieving and storing information in the residual stream, and replace the residual stream of the transformer with an outer product memory matrix (Kohonen, 1972, Anderson, 1972). We call this model the Residual Matrix Transformer (RMT). We find that the RMT enjoys a number of attractive properties: 1) the size of the residual stream can be scaled independently of compute and model size, improving performance, 2) the RMT can achieve the same loss as the transformer with 58% fewer FLOPS, 25% fewer parameters, and 41% fewer training tokens tokens, and 3) the RMT outperforms the transformer on downstream evaluations. We theoretically analyze the transformer and the RMT, and show that the RMT allows for more efficient scaling of the residual stream, as well as improved variance propagation properties.

Recent research across mathematical problem solving, proof assistant programming and multimodal jailbreaking documents a striking finding: when (multimodal) language model tackle a suite of tasks with multiple attempts per task -- succeeding if any attempt is correct -- then the negative log of the average success rate scales a power law in the number of attempts. In this work, we identify an apparent puzzle: a simple mathematical calculation predicts that on each problem, the failure rate should fall exponentially with the number of attempts. We confirm this prediction empirically, raising a question: from where does aggregate polynomial scaling emerge? We then answer this question by demonstrating per-problem exponential scaling can be made consistent with aggregate polynomial scaling if the distribution of single-attempt success probabilities is heavy tailed such that a small fraction of tasks with extremely low success probabilities collectively warp the aggregate success trend into a power law - even as each problem scales exponentially on its own. We further demonstrate that this distributional perspective explains previously observed deviations from power law scaling, and provides a simple method for forecasting the power law exponent with an order of magnitude lower relative error, or equivalently, ${\sim}2-4$ orders of magnitude less inference compute. Overall, our work contributes to a better understanding of how neural language model performance improves with scaling inference compute and the development of scaling-predictable evaluations of (multimodal) language models.

This paper formalizes an emerging learning paradigm that uses a trained model as a reference to guide and enhance the training of a target model through strategic data selection or weighting, named **model steering**. While ad-hoc methods have been used in various contexts, including the training of large foundation models, its underlying principles remain insufficiently understood, leading to sub-optimal performance. In this work, we propose a theory-driven framework for model steering called **DRRho risk minimization**, which is rooted in Distributionally Robust Optimization (DRO). Through a generalization analysis, we provide theoretical insights into why this approach improves generalization and data efficiency compared to training without a reference model. To the best of our knowledge, this is the first time such theoretical insights are provided for the new learning paradigm, which significantly enhance our understanding and practice of model steering. Building on these insights and the connection between contrastive learning and DRO, we introduce a novel method for Contrastive Language-Image Pretraining (CLIP) with a reference model, termed DRRho-CLIP. Extensive experiments validate the theoretical insights, reveal a superior scaling law compared to CLIP without a reference model, and demonstrate its strength over existing heuristic approaches. Code is released at [github.com/Optimization-AI/DRRho-CLIP](https://github.com/Optimization-AI/DRRho-CLIP)

The leakage of benchmark data into the training data has emerged as a significant challenge for evaluating the capabilities of large language models (LLMs). In this work, we challenge the common assumption that small-scale contamination renders benchmark evaluations invalid. First, we experimentally quantify the magnitude of benchmark overfitting based on scaling along three dimensions: The number of model parameters (up to 1.6B), the number of times an example is seen (up to 144), and the number of training tokens (up to 40B). If model and data follow the Chinchilla scaling laws, minor contamination indeed leads to overfitting. At the same time, even 144 times of contamination can be forgotten if the training data is scaled beyond five times Chinchilla, a regime characteristic of many modern LLMs. Continual pre-training of OLMo-7B corroborates these results. Next, we study the impact of the weight decay parameter on example forgetting, showing that empirical forgetting occurs faster than the cumulative weight decay. This allows us to gauge the degree of example forgetting in large-scale training runs, indicating that many LLMs, including Llama 3 405B, have forgotten the data seen at the beginning of training.

Sparse autoencoders (SAEs) are a popular tool for interpreting large language model activations, but their utility in addressing open questions in interpretability remains unclear. In this work, we demonstrate their effectiveness by using SAEs to deepen our understanding of the mechanism behind in-context learning (ICL). We identify abstract SAE features that (i) encode the model’s knowledge of which task to execute and (ii) whose latent vectors causally induce the task zero-shot. This aligns with prior work showing that ICL is mediated by task vectors. We further demonstrate that these task vectors are well approximated by a sparse sum of SAE latents, including these task-execution features. To explore the ICL mechanism, we scale the sparse feature circuits methodology of Marks et al. (2024) to the Gemma 1 2B model for the more complex task of ICL. Through circuit finding, we discover task-detecting features with corresponding SAE latents that activate earlier in the prompt, that detect when tasks have been performed. They are causally linked with task-execution features through the attention and MLP sublayers.

Compositional understanding is crucial for human intelligence, yet it remains unclear whether contemporary vision models exhibit it. The dominant machine learning paradigm is built on the premise that scaling data and model sizes will improve out-of-distribution performance, including compositional generalization. We test this premise through controlled experiments that systematically vary data scale, concept diversity, and combination coverage. We find that compositional generalization is driven by data diversity, not mere data scale. Increased combinatorial coverage forces models to discover a linearly factored representational structure, where concepts decompose into additive components. We prove this structure is key to efficiency, enabling perfect generalization from few observed combinations. Evaluating pretrained models (DINO, CLIP), we find above-random yet imperfect performance, suggesting partial presence of this structure. Our work motivates stronger emphasis on constructing diverse datasets for compositional generalization, and considering the importance of representational structure that enables efficient compositional learning.

Recent theories suggest that Neural Scaling Laws arise whenever the task is linearly decomposed into units that are power-law distributed. Alternatively, scaling laws also emerge when data exhibit a hierarchically compositional structure, as is thought to occur in language and images. To unify these views, we consider classification and next-token prediction tasks based on probabilistic context-free grammars—probabilistic models that generate data via a hierarchy of production rules. For classification, we show that having power-law distributed production rules results in a power-law learning curve with an exponent depending on the rules’ distribution and a large multiplicative constant that depends on the hierarchical structure. By contrast, for next-token prediction, the distribution of production rules controls the fine details of the learning curve, but not the exponent describing the large-scale behaviour.

We study universal properties in real-world complex and synthetically generated datasets. Our approach is to analogize data to a physical system and employ tools from statistical physics and Random Matrix Theory (RMT) to reveal their underlying structure. Examining the local and global eigenvalue statistics of feature-feature covariance matrices, we find: (i) bulk eigenvalue power-law scaling vastly differs between uncorrelated Gaussian and real-world data, (ii) this power law behavior is reproducible using Gaussian data with long-range correlations, (iii) all dataset types exhibit chaotic RMT universality, (iv) RMT statistics emerge at smaller dataset sizes than typical training sets, correlating with power-law convergence, (v) Shannon entropy correlates with RMT structure and requires fewer samples in strongly correlated datasets. These results suggest natural image Gram matrices can be approximated by Wishart random matrices with simple covariance structure, enabling rigorous analysis of neural network behavior.

Decomposing hard problems into subproblems often makes them easier and more efficient to solve. With the high cost of running LLMs at scale, there is an increasing effort to decompose systems into sets of LLM-based agents, each of whom can be delegated sub-tasks. However, this decomposition (even when automated) is often intuitive, e.g., based on how a human might assign roles to members of a human team. How close are these role decompositions to optimal? This position paper argues that asymptotic analysis with LLM primitives is needed to reason about the efficiency of such problem decompositions, and that insights from such analysis will unlock opportunities for scaling such systems. By treating the LLM forward pass as the atomic unit of computational cost, one can separate out the (often opaque) inner workings of a particular LLM from the inherent efficiency of how a set of LLMs are orchestrated to solve hard problems. In other words, if we want to scale the deployment of LLMs to the limit, instead of anthropomorphizing LLMs, asymptotic analysis with LLM primitives should be used to reason about and develop more powerful decompositions of large problems into LLM agents.