ICLR 2025 Papers

Recent advancements in document understanding have been dominated by leveraging large language models (LLMs) and multimodal large models. However, enabling LLMs to comprehend complex document layouts and structural information often necessitates intricate network modifications or costly pre-training, limiting their practical applicability. In this paper, we introduce Group Position Embedding (GPE), a novel and efficient technique to enhance the layout understanding capabilities of LLMs without architectural changes or additional pre-training. GPE achieves this by strategically grouping the attention heads and feeding each group with distinct positional embeddings, effectively encoding layout information relevant to document comprehension. This simple yet powerful method allows for effective integration of layout information within the existing LLM framework. We evaluate GPE against several competitive baselines across five mainstream document tasks. We also introduce a challenging benchmark called BLADE, specifically designed to assess layout comprehension. Extensive experiments on both established and BLADE benchmarks confirm the efficacy of GPE in significantly advancing the state-of-the-art in document understanding. Our code is available at https://github.com/antgroup/GroupPositionEmbedding.git

Despite their successes, deep learning models struggle with tasks requiring complex reasoning and function composition. We present a theoretical and empirical investigation into the limitations of Structured State Space Models (SSMs) and Transformers in such tasks. We prove that one-layer SSMs cannot efficiently perform function composition over large domains without impractically large state sizes, and even with Chain-of-Thought prompting, they require a number of steps that scale unfavorably with the complexity of the function composition. Also, the language of a finite-precision SSM is within the class of regular languages. Our experiments corroborate these theoretical findings. Evaluating models on tasks including various function composition settings, multi-digit multiplication, dynamic programming, and Einstein's puzzle, we find significant performance degradation even with advanced prompting techniques. Models often resort to shortcuts, leading to compounding errors. These findings highlight fundamental barriers within current deep learning architectures rooted in their computational capacities. We underscore the need for innovative solutions to transcend these constraints and achieve reliable multi-step reasoning and compositional task-solving, which is critical for advancing toward general artificial intelligence.

Text plays a crucial role in the transmission of human civilization, and teaching machines to generate online handwritten text in various styles presents an interesting and significant challenge. However, most prior work has concentrated on generating individual Chinese fonts, leaving complete text line generation largely unexplored. In this paper, we identify that text lines can naturally be divided into two components: layout and glyphs. Based on this division, we designed a text line layout generator coupled with a diffusion-based stylized font synthesizer to address this challenge hierarchically. More concretely, the layout generator performs in-context-like learning based on the text content and the provided style references to generate positions for each glyph autoregressively. Meanwhile, the font synthesizer which consists of a character embedding dictionary, a multi-scale calligraphy style encoder and a 1D U-Net based diffusion denoiser will generate each font on its position while imitating the calligraphy style extracted from the given style references. Qualitative and quantitative experiments on the CASIA-OLHWDB demonstrate that our method is capable of generating structurally correct and indistinguishable imitation samples.

Reinforcement learning (RL) has increasingly been applied to solve real-world planning problems, with progress in handling large state spaces and time horizons. However, a key bottleneck in many domains is that RL methods cannot accommodate large, combinatorially structured action spaces. In such settings, even representing the set of feasible actions at a single step may require a complex discrete optimization formulation. We leverage recent advances in embedding trained neural networks into optimization problems to propose SEQUOIA, an RL algorithm that directly optimizes for long-term reward over the feasible action space. Our approach embeds a Q-network into a mixed-integer program to select a combinatorial action in each timestep. Here, we focus on planning over restless bandits, a class of planning problems which capture many real-world examples of sequential decision making. We introduce coRMAB, a broader class of restless bandits with combinatorial actions that cannot be decoupled across the arms of the restless bandit, requiring direct solving over the joint, exponentially large action space. We empirically validate SEQUOIA on four novel restless bandit problems with combinatorial constraints: multiple interventions, path constraints, bipartite matching, and capacity constraints. Our approach significantly outperforms existing methods—which cannot address sequential planning and combinatorial selection simultaneously—by an average of 24.8% on these difficult instances.

Multimodal Large Language Models (MLLMs) have experienced rapid progress in visual recognition tasks in recent years. Given their potential integration into many critical applications, it is important to understand the limitations of their visual perception. In this work, we study whether MLLMs can perceive small visual details as effectively as large ones when answering questions about images. We observe that their performance is very sensitive to the size of the visual subject of the question, and further show that this effect is in fact causal by conducting an intervention study. Next, we study the attention patterns of MLLMs when answering visual questions, and intriguingly find that they consistently know where to look, even when they provide the wrong answer. Based on these findings, we then propose training-free visual intervention methods that leverage the internal knowledge of any MLLM itself, in the form of attention and gradient maps, to enhance its perception of small visual details. We evaluate our proposed methods on two widely-used MLLMs and seven visual question answering benchmarks and show that they can significantly improve MLLMs' accuracy without requiring any training. Our results elucidate the risk of applying MLLMs to visual recognition tasks concerning small details and indicate that visual intervention using the model's internal state is a promising direction to mitigate this risk. Our code is available at: https://github.com/saccharomycetes/mllms_know.

In reinforcement learning, classic on-policy evaluation methods often suffer from high variance and require massive online data to attain the desired accuracy. Previous studies attempt to reduce evaluation variance by searching for or designing proper behavior policies to collect data. However, these approaches ignore the safety of such behavior policies---the designed behavior policies have no safety guarantee and may lead to severe damage during online executions. In this paper, to address the challenge of reducing variance while ensuring safety simultaneously, we propose an optimal variance-minimizing behavior policy under safety constraints. Theoretically, while ensuring safety constraints, our evaluation method is unbiased and has lower variance than on-policy evaluation. Empirically, our method is the only existing method to achieve both substantial variance reduction and safety constraint satisfaction. Furthermore, we show our method is even superior to previous methods in both variance reduction and execution safety.

This paper considers a class of structured fractional minimization problems. The numerator consists of a differentiable function, a simple nonconvex nonsmooth function, a concave nonsmooth function, and a convex nonsmooth function composed with a linear operator. The denominator is a continuous function that is either weakly convex or has a weakly convex square root. These problems are prevalent in various important applications in machine learning and data science. Existing methods, primarily based on subgradient methods and smoothing proximal gradient methods, often suffer from slow convergence and numerical stability issues. In this paper, we introduce {\sf FADMM}, the first Alternating Direction Method of Multipliers tailored for this class of problems. {\sf FADMM} decouples the original problem into linearized proximal subproblems, featuring two variants: one using Dinkelbach's parametric method ({\sf FADMM-D}) and the other using the quadratic transform method ({\sf FADMM-Q}). By introducing a novel Lyapunov function, we establish that {\sf FADMM} converges to $\epsilon$-approximate critical points of the problem within an oracle complexity of $\mathcal{O}(1/\epsilon^{3})$. Extensive experiments on synthetic and real-world datasets, including sparse Fisher discriminant analysis, robust Sharpe ratio minimization, and robust sparse recovery, demonstrate the effectiveness of our approach.

While image-text representation learning has become very popular in recent years, existing models tend to lack spatial awareness and have limited direct applicability for dense understanding tasks. For this reason, self-supervised image-only pretraining is still the go-to method for many dense vision applications (e.g. depth estimation, semantic segmentation), despite the lack of explicit supervisory signals. In this paper, we close this gap between image-text and self-supervised learning, by proposing a novel general-purpose image-text model, which can be effectively used off the shelf for dense and global vision tasks. Our method, which we refer to as Text-Image Pretraining with Spatial awareness (TIPS), leverages two simple and effective insights. First, on textual supervision: we reveal that replacing noisy web image captions by synthetically generated textual descriptions boosts dense understanding performance significantly, due to a much richer signal for learning spatially aware representations. We propose an adapted training method that combines noisy and synthetic captions, resulting in improvements across both dense and global understanding tasks. Second, on the learning technique: we propose to combine contrastive image-text learning with self-supervised masked image modeling, to encourage spatial coherence, unlocking substantial enhancements for downstream applications. Building on these two ideas, we scale our model using the transformer architecture, trained on a curated set of public images. Our experiments are conducted on $8$ tasks involving $16$ datasets in total, demonstrating strong off-the-shelf performance on both dense and global understanding, for several image-only and image-text tasks. Code and models are released at https://github.com/google-deepmind/tips .

Geometric tempering is a popular approach to sampling from challenging multi-modal probability distributions by instead sampling from a sequence of distributions which interpolate, using the geometric mean, between an easier proposal distribution and the target distribution. In this paper, we theoretically investigate the soundness of this approach when the sampling algorithm is Langevin dynamics, proving both upper and lower bounds. Our upper bounds are the first analysis in the literature under functional inequalities. They assert the convergence of tempered Langevin in continuous and discrete-time, and their minimization leads to closed-form optimal tempering schedules for some pairs of proposal and target distributions. Our lower bounds demonstrate a simple case where the geometric tempering takes exponential time, and further reveal that the geometric tempering can suffer from poor functional inequalities and slow convergence, even when the target distribution is well-conditioned. Overall, our results indicate that the geometric tempering may not help, and can even be harmful for convergence.

Attention-based models, such as Transformer, excel across various tasks but lack a comprehensive theoretical understanding, especially regarding token-wise sparsity and internal linear representations. To address this gap, we introduce the single-location regression task, where only one token in a sequence determines the output, and its position is a latent random variable, retrievable via a linear projection of the input. To solve this task, we propose a dedicated predictor, which turns out to be a simplified version of a non-linear self-attention layer. We study its theoretical properties, by showing its asymptotic Bayes optimality and analyzing its training dynamics. In particular, despite the non-convex nature of the problem, the predictor effectively learns the underlying structure. This work highlights the capacity of attention mechanisms to handle sparse token information and internal linear structures.

We study continual learning on multiple linear classification tasks by sequentially running gradient descent (GD) for a fixed budget of iterations per each given task. When all tasks are jointly linearly separable and are presented in a cyclic/random order, we show the directional convergence of the trained linear classifier to the joint (offline) max-margin solution. This is surprising because GD training on a single task is implicitly biased towards the individual max-margin solution for the task, and the direction of the joint max-margin solution can be largely different from these individual solutions. Additionally, when tasks are given in a cyclic order, we present a non-asymptotic analysis on cycle-averaged forgetting, revealing that (1) alignment between tasks is indeed closely tied to catastrophic forgetting and backward knowledge transfer and (2) the amount of forgetting vanishes to zero as the cycle repeats. Lastly, we analyze the case where the tasks are no longer jointly separable and show that the model trained in a cyclic order converges to the unique minimum of the joint loss function.

The generative modeling of data on manifolds is an important task, for which diffusion models in flat spaces typically need nontrivial adaptations. This article demonstrates how a technique called `trivialization' can transfer the effectiveness of diffusion models in Euclidean spaces to Lie groups. In particular, an auxiliary momentum variable was algorithmically introduced to help transport the position variable between data distribution and a fixed, easy-to-sample distribution. Normally, this would incur further difficulty for manifold data because momentum lives in a space that changes with the position. However, our trivialization technique creates a new momentum variable that stays in a simple fixed vector space. This design, together with a manifold preserving integrator, simplifies implementation and avoids inaccuracies created by approximations such as projections to tangent space and manifold, which were typically used in prior work, hence facilitating generation with high-fidelity and efficiency. The resulting method achieves state-of-the-art performance on protein and RNA torsion angle generation and sophisticated torus datasets. We also, arguably for the first time, tackle the generation of data on high-dimensional Special Orthogonal and Unitary groups, the latter essential for quantum problems. Code is available at https://github.com/yuchen-zhu-zyc/TDM.

To create useful reinforcement learning (RL) agents, step zero is to design a suitable reward function that captures the nuances of the task. However, reward engineering can be a difficult and time-consuming process. Instead, human-in-the-loop RL methods hold the promise of learning reward functions from human feedback. Despite recent successes, many of the human-in-the-loop RL methods still require numerous human interactions to learn successful reward functions. To improve the feedback efficiency of human-in-the-loop RL methods (i.e., require less human interaction), this paper introduces Sub-optimal Data Pre-training, SDP, an approach that leverages reward-free, sub-optimal data to improve scalar- and preference-based RL algorithms. In SDP, we start by pseudo-labeling all low-quality data with the minimum environment reward. Through this process, we obtain reward labels to pre-train our reward model without requiring human labeling or preferences. This pre-training phase provides the reward model a head start in learning, enabling it to recognize that low-quality transitions should be assigned low rewards. Through extensive experiments with both simulated and human teachers, we find that SDP can at least meet, but often significantly improve, state of the art human-in-the-loop RL performance across a variety of simulated robotic tasks.

Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) have demonstrated impressive language/vision reasoning abilities, igniting the recent trend of building agents for targeted applications such as shopping assistants or AI software engineers. Recently, many data science benchmarks have been proposed to investigate their performance in the data science domain. However, existing data science benchmarks still fall short when compared to real-world data science applications due to their simplified settings. To bridge this gap, we introduce DSBench, a comprehensive benchmark designed to evaluate data science agents with realistic tasks. This benchmark includes 466 data analysis tasks and 74 data modeling tasks, sourced from Eloquence and Kaggle competitions. DSBench offers a realistic setting by encompassing long contexts, multimodal task backgrounds, reasoning with large data files and multi-table structures, and performing end-to-end data modeling tasks. Our evaluation of state-of-the-art LLMs, LVLMs, and agents shows that they struggle with most tasks, with the best agent solving only 34.12% of data analysis tasks and achieving a 34.74% Relative Performance Gap (RPG). These findings underscore the need for further advancements in developing more practical, intelligent, and autonomous data science agents.

It has recently been argued that AI models' representations are becoming aligned as their scale and performance increase. Empirical analyses have been designed to support this idea and conjecture the possible alignment of different representations toward a shared statistical model of reality. In this paper, we propose a learning-theoretic perspective to representation alignment. First, we review and connect different notions of alignment based on metric, probabilistic, and spectral ideas. Then, we focus on stitching, a particular approach to understanding the interplay between different representations in the context of a task. Our main contribution here is to relate the properties of stitching to the kernel alignment of the underlying representation. Our results can be seen as a first step toward casting representation alignment as a learning-theoretic problem.

Recent advances in GPU-based parallel simulation have enabled practitioners to collect large amounts of data and train complex control policies using deep reinforcement learning (RL), on commodity GPUs. However, such successes for RL in robotics have been limited to tasks sufficiently simulated by fast rigid-body dynamics. Simulation techniques for soft bodies are comparatively several orders of magnitude slower, thereby limiting the use of RL due to sample complexity requirements. To address this challenge, this paper presents both a novel RL algorithm and a simulation platform to enable scaling RL on tasks involving rigid bodies and deformables. We introduce Soft Analytic Policy Optimization (SAPO), a maximum entropy first-order model-based actor-critic RL algorithm, which uses first-order analytic gradients from differentiable simulation to train a stochastic actor to maximize expected return and entropy. Alongside our approach, we develop Rewarped, a parallel differentiable multiphysics simulation platform that supports simulating various materials beyond rigid bodies. We re-implement challenging manipulation and locomotion tasks in Rewarped, and show that SAPO outperforms baselines over a range of tasks that involve interaction between rigid bodies, articulations, and deformables. Additional details at https://rewarped.github.io/.

Parametric differential equations of the form $\frac{\partial u}{\partial t} = f(u, x, t, p)$ are fundamental in science and engineering. While deep learning frameworks like the Fourier Neural Operator (FNO) efficiently approximate differential equation solutions, they struggle with inverse problems, sensitivity calculations $\frac{\partial u}{\partial p}$, and concept drift. We address these challenges by introducing a novel sensitivity loss regularizer, demonstrated through Sensitivity-Constrained Fourier Neural Operators (SC-FNO). Our approach maintains high accuracy for solution paths and outperforms both standard FNO and FNO with Physics-Informed Neural Network regularization. SC-FNO exhibits superior performance in parameter inversion tasks, accommodates more complex parameter spaces (tested with up to 82 parameters), reduces training data requirements, and decreases training time while maintaining accuracy. These improvements apply across various differential equations and neural operators, enhancing their reliability without significant computational overhead (30%–130% extra training time per epoch). Models and selected experiment code are available at: [https://github.com/AMBehroozi/SC_Neural_Operators](https://github.com/AMBehroozi/SC_Neural_Operators).

We devise a novel method for nowcasting implied volatility based on neural operators. Better known as implied volatility smoothing in the financial industry, nowcasting of implied volatility means constructing a smooth surface that is consistent with the prices presently observed on a given option market. Option price data arises highly dynamically in ever-changing spatial configurations, which poses a major limitation to foundational machine learning approaches using classical neural networks. While large models in language and image processing deliver breakthrough results on vast corpora of raw data, in financial engineering the generalization from big historical datasets has been hindered by the need for considerable data pre-processing. In particular, implied volatility smoothing has remained an instance-by-instance, hands-on process both for neural network-based and traditional parametric strategies. Our general *operator deep smoothing* approach, instead, directly maps observed data to smoothed surfaces. We adapt the graph neural operator architecture to do so with high accuracy on ten years of raw intraday S&P 500 options data, using a single model instance. The trained operator adheres to critical no-arbitrage constraints and is robust with respect to subsampling of inputs (occurring in practice in the context of outlier removal). We provide extensive historical benchmarks and showcase the generalization capability of our approach in a comparison with classical neural networks and SVI, an industry standard parametrization for implied volatility. The operator deep smoothing approach thus opens up the use of neural networks on large historical datasets in financial engineering.

In this study, we investigate the problem of dynamic multi-product selection and pricing by introducing a novel framework based on a *censored multinomial logit* (C-MNL) choice model. In this model, sellers present a set of products with prices, and buyers filter out products priced above their valuation, purchasing at most one product from the remaining options based on their preferences. The goal is to maximize seller revenue by dynamically adjusting product offerings and prices, while learning both product valuations and buyer preferences through purchase feedback. To achieve this, we propose a Lower Confidence Bound (LCB) pricing strategy. By combining this pricing strategy with either an Upper Confidence Bound (UCB) or Thompson Sampling (TS) product selection approach, our algorithms achieve regret bounds of $\tilde{O}(d^{\frac{3}{2}}\sqrt{T/\kappa})$ and $\tilde{O}(d^{2}\sqrt{T/\kappa})$, respectively. Finally, we demonstrate the performance of our methods through simulations.