Procedural Pretraining: Warming Up Language Models with Abstract Data

Pretraining directly on web-scale corpora is the de facto paradigm for building language models. We study an alternative setting where the model is initially exposed to abstract structured data, as a means to ease the subsequent acquisition of rich semantic knowledge, much like humans learn simple logic and mathematics before higher reasoning. We specifically focus on procedural data, generated by formal languages and other simple algorithms, as such abstract data. We first diagnose the algorithmic skills that different forms of procedural data can improve, often significantly. For example, on context recall (Needle-in-a-haystack), the accuracy jumps from 10 to 98% when pretraining on Dyck sequences (balanced brackets). Second, we study how these gains are reflected in pretraining larger models (up to 1.3B). We find that front-loading as little as 0.1% procedural data significantly outperforms standard pretraining on natural language, code, and informal mathematics (C4, CodeParrot, and DeepMind-Math datasets). Notably, this procedural pretraining enables the models to reach the same loss value with only 55, 67, 86% of the original data. Third, we explore the mechanisms behind and find that procedural pretraining instils non-trivial structure in both attention and MLP layers. The former is particularly important for structured domains (e.g. code), and the latter for language. Finally, we lay a path for combining multiple forms of procedural data. Our results show that procedural pretraining is a simple, lightweight means to improving performance and accelerating language model pretraining, ultimately suggesting the promise of disentangling knowledge acquisition from reasoning in LLMs.

Meta-RL Induces Exploration in Language Agents

Reinforcement learning (RL) has enabled the training of large language model (LLM) agents to interact with the environment and to solve multi-turn long-horizon tasks. However, the RL-trained agents often struggle in tasks that require active exploration and fail to efficiently adapt from trial-and-error experiences. In this paper, we present LaMer, a general Meta-RL framework that enables LLM agents to actively explore and learn from the environment feedback at test time. LaMer consists of two key components: (i) a cross-episode training framework to encourage exploration and long-term rewards optimization; and (ii) in-context policy adaptation via reflection, allowing the agent to adapt their policy from task feedback signal without gradient update. Experiments across diverse environments show that LaMer significantly improves performance over RL baselines, with 11%, 14%, and 19% performance gains on Sokoban, MineSweeper and Webshop, respectively. Moreover, LaMer also demonstrates better generalization to more challenging or previously unseen tasks compared to the RL-trained agents. Overall, our results demonstrate that Meta-RL provides a principled approach to induce exploration in language agents, enabling more robust adaptation to novel environments through learned exploration strategies.

Can You Learn to See Without Images? Procedural Warm-Up for Vision Transformers

Transformers show remarkable versatility across domains, suggesting the existence of inductive biases beneficial across modalities. In this work, we explore a new way to instil such generic biases in vision transformers (ViTs) by pretraining on procedurally-generated data devoid of visual or semantic content. We generate this data with simple algorithms such as formal grammars, so the results bear no relationship to either natural or synthetic images. We use this procedurally-generated data to pretrain ViTs in a warm-up phase that bypasses their visual patch embedding mechanisms, thus encouraging the models to internalise abstract computational priors. When followed by standard image-based training, this warm-up significantly improves data efficiency, convergence speed, and downstream performance. On ImageNet-1k for example, allocating just 1% of the training budget to procedural data improves final accuracy by over 1.7%. In terms of its effect on performance, 1% procedurally generated data is thus equivalent to 28% of the ImageNet-1k data. These findings suggest a promising path toward new data-efficient and domain-agnostic pretraining strategies.

How does the optimizer implicitly bias the model merging loss landscape?

Model merging methods combine models with different capabilities into a single one while maintaining the same inference cost. Two popular approaches are linear interpolation, which linearly interpolates between model weights, and task arithmetic, which combines task vectors obtained by the difference between finetuned and base models. While useful in practice, what properties make merging effective are poorly understood. This paper explores how the optimization process affects the loss landscape geometry and its impact on merging success. We show that a single quantity -- the effective noise scale -- unifies the impact of optimizer and data choices on model merging. Across architectures and datasets, the effectiveness of merging success is a non-monotonic function of effective noise, with a distinct optimum. Decomposing this quantity, we find that larger learning rates, stronger weight decay, smaller batch sizes, and data augmentation all independently modulate the effective noise scale, exhibiting the same qualitative trend. Unlike prior work that connects optimizer noise to the flatness or generalization of individual minima, we show that it also affects the global loss landscape, predicting when independently trained solutions can be merged. Our findings broaden the understanding of how optimization shapes the loss landscape geometry and its downstream consequences for model merging, suggesting the possibility of further manipulating the training dynamics to improve merging effectiveness.

Transformers Pretrained on Procedural Data Contain Modular Structures for Algorithmic Reasoning

Pretraining on large, semantically rich datasets is key for developing language models. Surprisingly, recent studies have shown that even synthetic data, generated procedurally through simple semantic-free algorithms, can yield some of the same benefits as natural language pretraining. It is unclear what specific capabilities such simple synthetic data instils in a model, where these capabilities reside in the architecture, and how they manifest within its weights. In this short paper, we identify several beneficial forms of procedural data, together with specific algorithmic reasoning skills that improve in small transformers. Our core finding is that different procedural rules instil distinct but complementary inductive structures in the model. With extensive ablations and partial-transfer experiments, we discover that these structures reside in different parts of the model. Attention layers often carry the most transferable information, but some pretraining rules impart useful structure to MLP blocks instead. Most interestingly, the structures induced by multiple rules can be composed to jointly reinforce multiple capabilities. These results suggest an exciting possibility of disentangling the acquisition of knowledge from reasoning in language models, with the goal of improving their robustness and data efficiency.

Leaner Transformers: More Heads, Less Depth

Transformers have reshaped machine learning by utilizing attention mechanisms to capture complex patterns in large datasets, leading to significant improvements in performance. This success has contributed to the belief that "bigger means better", leading to ever-increasing model sizes. This paper challenge this ideology by showing that many existing transformers might be unnecessarily oversized. We discover a theoretical principle that redefines the role of multi-head attention. An important benefit of the multiple heads is in improving the conditioning of the attention block. We exploit this theoretical insight and redesign popular architectures with an increased number of heads. The improvement in the conditioning proves so significant in practice that model depth can be decreased, reducing the parameter count by up to 30-50% while maintaining accuracy. We obtain consistent benefits across a variety of transformer-based architectures of various scales, on tasks in computer vision (ImageNet-1k) as well as language and sequence modeling (GLUE benchmark, TinyStories, and the Long-Range Arena benchmark).

AugGen: Synthetic Augmentation Can Improve Discriminative Models

The increasing dependence on large-scale datasets in machine learning introduces significant privacy and ethical challenges. Synthetic data generation offers a promising solution; however, most current methods rely on external datasets or pre-trained models, which add complexity and escalate resource demands. In this work, we introduce a novel self-contained synthetic augmentation technique that strategically samples from a conditional generative model trained exclusively on the target dataset. This approach eliminates the need for auxiliary data sources. Applied to face recognition datasets, our method achieves 1--12\% performance improvements on the IJB-C and IJB-B benchmarks. It outperforms models trained solely on real data and exceeds the performance of state-of-the-art synthetic data generation baselines. Notably, these enhancements often surpass those achieved through architectural improvements, underscoring the significant impact of synthetic augmentation in data-scarce environments. These findings demonstrate that carefully integrated synthetic data not only addresses privacy and resource constraints but also substantially boosts model performance. Project page https://parsa-ra.github.io/auggen

Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild

Neural architectures tend to fit their data with relatively simple functions. This "simplicity bias" is widely regarded as key to their success. This paper explores the limits of this principle. Building on recent findings that the simplicity bias stems from ReLU activations [96], we introduce a method to meta-learn new activation functions and inductive biases better suited to specific tasks. Findings: We identify multiple tasks where the simplicity bias is inadequate and ReLUs suboptimal. In these cases, we learn new activation functions that perform better by inducing a prior of higher complexity. Interestingly, these cases correspond to domains where neural networks have historically struggled: tabular data, regression tasks, cases of shortcut learning, and algorithmic grokking tasks. In comparison, the simplicity bias induced by ReLUs proves adequate on image tasks where the best learned activations are nearly identical to ReLUs and GeLUs. Implications: Contrary to popular belief, the simplicity bias of ReLU networks is not universally useful. It is near-optimal for image classification, but other inductive biases are sometimes preferable. We showed that activation functions can control these inductive biases, but future tailored architectures might provide further benefits. Advances are still needed to characterize a model's inductive biases beyond "complexity", and their adequacy with the data.

OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable?

Out-of-distribution (OOD) generalization is challenging because distribution shifts come in many forms. A multitude of learning algorithms exist and each can improve performance in specific OOD situations. We posit that much of the challenge of OOD generalization lies in choosing the right algorithm for the right dataset. However, such algorithm selection is often elusive under complex real-world shifts. In this work, we formalize the task of algorithm selection for OOD generalization and investigate whether it could be approached by learning. We propose a solution, dubbed OOD-Chameleon that treats the task as a supervised classification over candidate algorithms. We construct a dataset of datasets to learn from, which represents diverse types, magnitudes and combinations of shifts (covariate shift, label shift, spurious correlations). We train the model to predict the relative performance of algorithms given a dataset's characteristics. This enables a priori selection of the best learning strategy, i.e. without training various models as needed with traditional model selection. Our experiments show that the adaptive selection outperforms any individual algorithm and simple selection heuristics, on unseen datasets of controllable and realistic image data. Inspecting the model shows that it learns non-trivial data/algorithms interactions, and reveals the conditions for any one algorithm to surpass another. This opens new avenues for (1) enhancing OOD generalization with existing algorithms instead of designing new ones, and (2) gaining insights into the applicability of existing algorithms with respect to datasets' properties.

Scalable Ensemble Diversification for OOD Generalization and Detection

Training a diverse ensemble of models has several practical applications such as providing candidates for model selection with better out-of-distribution (OOD) generalization, and enabling the detection of OOD samples via Bayesian principles. An existing approach to diverse ensemble training encourages the models to disagree on provided OOD samples. However, the approach is computationally expensive and it requires well-separated ID and OOD examples, such that it has only been demonstrated in small-scale settings. $\textbf{Method.}$ This work presents a method for Scalable Ensemble Diversification (SED) applicable to large-scale settings (e.g. ImageNet) that does not require OOD samples. Instead, SED identifies hard training samples on the fly and encourages the ensemble members to disagree on these. To improve scaling, we show how to avoid the expensive computations in existing methods of exhaustive pairwise disagreements across models. $\textbf{Results.}$ We evaluate the benefits of diversification with experiments on ImageNet. First, for OOD generalization, we observe large benefits from the diversification in multiple settings including output-space (classical) ensembles and weight-space ensembles (model soups). Second, for OOD detection, we turn the diversity of ensemble hypotheses into a novel uncertainty score estimator that surpasses a large number of OOD detection baselines. Code is available here: https://github.com/AlexanderRubinstein/diverse-universe-public.