ICLR 2025 Papers

In-context learning (ICL), the remarkable ability to solve a task from only input exemplars, is often assumed to be a unique hallmark of Transformer models. By examining commonly employed synthetic ICL tasks, we demonstrate that multi-layer perceptrons (MLPs) can also learn in-context. Moreover, MLPs, and the closely related MLP-Mixer models, learn in-context comparably with Transformers under the same compute budget in this setting. We further show that MLPs outperform Transformers on a series of classical tasks from psychology designed to test relational reasoning, which are closely related to in-context classification. These results underscore a need for studying in-context learning beyond attention-based architectures, while also challenging prior arguments against MLPs' ability to solve relational tasks. Altogether, our results highlight the unexpected competence of MLPs in a synthetic setting, and support the growing interest in all-MLP alternatives to Transformer architectures. It remains unclear how MLPs perform against Transformers at scale on real-world tasks, and where a performance gap may originate. We encourage further exploration of these architectures in more complex settings to better understand the potential comparative advantage of attention-based schemes.

Diffusion Transformer, the backbone of Sora for video generation, successfully scales the capacity of diffusion models, pioneering new avenues for high-fidelity sequential data generation. Unlike static data such as images, sequential data consists of consecutive data frames indexed by time, exhibiting rich spatial and temporal dependencies. These dependencies represent the underlying dynamic model and are critical to validate the generated data. In this paper, we make the first theoretical step towards bridging diffusion transformers for capturing spatial-temporal dependencies. Specifically, we establish score approximation and distribution estimation guarantees of diffusion transformers for learning Gaussian process data with covariance functions of various decay patterns. We highlight how the spatial-temporal dependencies are captured and affect learning efficiency. Our study proposes a novel transformer approximation theory, where the transformer acts to unroll an algorithm. We support our theoretical results by numerical experiments, providing strong evidence that spatial-temporal dependencies are captured within attention layers, aligning with our approximation theory.

With the recent development and advancement of Transformer and MLP architectures, significant strides have been made in time series analysis. Conversely, the performance of Convolutional Neural Networks (CNNs) in time series analysis has fallen short of expectations, diminishing their potential for future applications. Our research aims to enhance the representational capacity of Convolutional Neural Networks (CNNs) in time series analysis by introducing novel perspectives and design innovations. To be specific, We introduce a novel time series reshaping technique that considers the inter-patch, intra-patch, and cross-variable dimensions. Consequently, we propose TVNet, a dynamic convolutional network leveraging a 3D perspective to employ time series analysis. TVNet retains the computational efficiency of CNNs and achieves state-of-the-art results in five key time series analysis tasks, offering a superior balance of efficiency and performance over the state-of-the-art Transformer-based and MLP-based models. Additionally, our findings suggest that TVNet exhibits enhanced transferability and robustness. Therefore, it provides a new perspective for applying CNN in advanced time series analysis tasks.

$k$-Nearest Neighbor search on dense vector embeddings ($k$-NN retrieval) from pre-trained embedding models is the predominant retrieval method for text and images, as well as Retrieval-Augmented Generation (RAG) pipelines. In practice, application developers often fine-tune the embeddings to improve their accuracy on the dataset and query workload in hand. Existing approaches either fine-tune the pre-trained model itself or, more efficiently, but at the cost of accuracy, train adaptor models to transform the output of the pre-trained model. We present NUDGE, a family of novel *non-parametric* embedding fine-tuning approaches that are significantly more accurate and efficient than both sets of existing approaches. NUDGE directly modifies the embeddings of data records to maximize the accuracy of $k$-NN retrieval. We present a thorough theoretical and experimental study of NUDGE's non-parametric approach. We show that even though the underlying problem is NP-Hard, constrained variations can be solved efficiently. These constraints additionally ensure that the changes to the embeddings are modest, avoiding large distortions to the semantics learned during pre-training. In experiments across five pre-trained models and nine standard text and image retrieval datasets, *NUDGE runs in minutes and often improves NDCG@10 by more than 10\% over existing fine-tuning methods. On average, NUDGE provides 3.3$\times$ and 4.3$\times$ higher increase in accuracy and runs 200$\times$ and 3$\times$ faster, respectively, over fine-tuning the pre-trained model and training adaptors.*

Controlling high-dimensional nonlinear systems, such as those found in biological and robotic applications, is challenging due to large state and action spaces. While deep reinforcement learning has achieved a number of successes in these domains, it is computationally intensive and time consuming, and therefore not suitable for solving large collections of tasks that require significant manual tuning. In this work, we introduce Model Predictive Control with Morphology-aware Proportional Control (MPC$^2$), a hierarchical model-based learning algorithm for zero-shot and near-real-time control of high-dimensional complex dynamical systems. MPC$^2$ uses a sampling-based model predictive controller for target posture planning, and enables robust control for high-dimensional tasks by incorporating a morphology-aware proportional controller for actuator coordination. The algorithm enables motion control of a high-dimensional human musculoskeletal model in a variety of motion tasks, such as standing, walking on different terrains, and imitating sports activities. The reward function of MPC$^2$ can be tuned via black-box optimization, drastically reducing the need for human-intensive reward engineering.

Efficiently computing approximate equilibrium strategies in large Imperfect Information Extensive-Form Games (IIEFGs) poses significant challenges due to the game tree's exponential growth. While pruning and abstraction techniques are essential for complexity reduction, existing methods face two key limitations: (i) Seamless integration of pruning with Counterfactual Regret Minimization (CFR) is nontrivial, and (ii) Pruning and abstraction approaches incur prohibitive computational costs, hindering real-world deployment. We propose Expected-Value Pruning and Abstraction (EVPA), a novel online framework that addresses these challenges through three synergistic components: (i) Expected value estimation using approximate Nash equilibrium strategies to quantify information set utilities, (ii) Minimax pruning before CFR to eliminate a large number of sub-optimal actions permanently, and (iii) Dynamic online information abstraction merging information sets based on their current and future expected values in subgames. Experiments on Heads-up No-Limit Texas Hold'em (HUNL) show EVPA outperforms DeepStack's replication and Slumbot with significant win-rate margins in multiple settings. Remarkably, EVPA requires only $1$\%-$2$\% of the solving time to reach an approximate Nash equilibrium compared to DeepStack's replication.

Sampling algorithms play an important role in controlling the quality and runtime of diffusion model inference. In recent years, a number of works (Chen et al., 2023c;b; Benton et al., 2023; Lee et al., 2022) have analyzed algorithms for diffusion sampling with provable guarantees; these works show that for essentially any data distribution, one can approximately sample in polynomial time given a sufficiently accurate estimate of its score functions at different noise levels. In this work, we propose a new scheme inspired by Shen and Lee's randomized midpoint method for log-concave sampling (Shen & Lee, 2019). We prove that this approach achieves the best known dimension dependence for sampling from arbitrary smooth distributions in total variation distance ($\widetilde O(d^{5/12})$ compared to $\widetilde O(\sqrt{d})$ from prior work). We also show that our algorithm can be parallelized to run in only $\widetilde O(\log^2 d)$ parallel rounds, constituting the first provable guarantees for parallel sampling with diffusion models. As a byproduct of our methods, for the well-studied problem of log-concave sampling in total variation distance, we give an algorithm and simple analysis achieving dimension dependence $\widetilde O(d^{5/12})$ compared to $\widetilde O(\sqrt{d})$ from prior work.

Multi-agent reinforcement learning has demonstrated significant potential in addressing complex cooperative tasks across various real-world applications. However, existing MARL approaches often rely on the restrictive assumption that the number of entities (e.g., agents, obstacles) remains constant between training and inference. This overlooks scenarios where entities are dynamically removed or $\textit{added}$ $\textit{during}$ the inference trajectory—a common occurrence in real-world environments like search and rescue missions and dynamic combat situations. In this paper, we tackle the challenge of intra-trajectory dynamic entity composition under zero-shot out-of-domain (OOD) generalization, where such dynamic changes cannot be anticipated beforehand. Our empirical studies reveal that existing MARL methods suffer $\textit{significant}$ performance degradation and increased uncertainty in these scenarios. In response, we propose FlickerFusion, a novel OOD generalization method that acts as a $\textit{universally}$ applicable augmentation technique for MARL backbone methods. FlickerFusion stochastically drops out parts of the observation space, emulating being in-domain when inferenced OOD. The results show that FlickerFusion not only achieves superior inference rewards but also $\textit{uniquely}$ reduces uncertainty vis-à-vis the backbone, compared to existing methods. Benchmarks, implementations, and model weights are organized and open-sourced at $\texttt{\href{flickerfusion305.github.io}{\textbf{flickerfusion305.github.io}}}$, accompanied by ample demo video renderings.

Large Language Models (LLM) are increasingly trained on data generated by other LLMs, either because generated text and images become part of the pre-training corpus, or because synthetized data is used as a replacement for expensive human-annotation. This raises concerns about *model collapse*, a drop in model performance when their training sets include generated data. Considering that it is easier for both humans and machines to tell between good and bad examples than to generate high-quality samples, we investigate the use of verification on synthesized data to prevent model collapse. We provide a theoretical characterization using Gaussian mixtures, linear classifiers, and linear verifiers to derive conditions with measurable proxies to assess whether the verifier can effectively select synthesized data that leads to optimal performance. We experiment with two practical tasks -- computing matrix eigenvalues with transformers and news summarization with LLMs -- which both exhibit model collapse when trained on generated data, and show that verifiers, even imperfect ones, can indeed be harnessed to prevent model collapse and that our proposed proxy measure strongly correlates with performance.

Language-based object detection (LOD) aims to align visual objects with language expressions. A large amount of paired data is utilized to improve LOD model generalizations. During the training process, recent studies leverage vision-language models (VLMs) to automatically generate human-like expressions for visual objects, facilitating training data scaling up. In this process, we observe that VLM hallucinations bring inaccurate object descriptions (e.g., object name, color, and shape) to deteriorate VL alignment quality. To reduce VLM hallucinations, we propose an agentic workflow controlled by an LLM to re-align language to visual objects via adaptively adjusting image and text prompts. We name this workflow Real-LOD, which includes planning, tool use, and reflection steps. Given an image with detected objects and VLM raw language expressions, Real-LOD reasons its state automatically and arranges action based on our neural symbolic designs (i.e., planning). The action will adaptively adjust the image and text prompts and send them to VLMs for object re-description (i.e., tool use). Then, we use another LLM to analyze these refined expressions for feedback (i.e., reflection). These steps are conducted in a cyclic form to gradually improve language descriptions for re-aligning to visual objects. We construct a dataset that contains a tiny amount of 0.18M images with re-aligned language expression and train a prevalent LOD model to surpass existing LOD methods by around 50% on the standard benchmarks. Our Real-LOD workflow, with automatic VL refinement, reveals a potential to preserve data quality along with scaling up data quantity, which further improves LOD performance from a data-alignment perspective.

With the goal of benchmarking generative systems beyond expert software development ability, we introduce Commit0, a benchmark that challenges AI agents to write libraries from scratch. Agents are provided with a specification document outlining the library’s API as well as a suite of interactive unit tests, with the goal of producing an implementation of this API accordingly. The implementation is validated through running these unit tests. As a benchmark, Commit0 is designed to move beyond static one-shot code generation towards agents that must process long-form natural language specifications, adapt to multi-stage feedback, and generate code with complex dependencies. Commit0 also offers an interactive environment where models receive static analysis and execution feedback on the code they generate. Our experiments demonstrate that while current agents can pass some unit tests, none can yet fully reproduce full libraries. Results also show that interactive feedback is quite useful for models to generate code that passes more unit tests, validating the benchmarks that facilitate its use. We publicly release the benchmark, the interactive environment, and the leaderboard.

Messenger RNA (mRNA) plays a crucial role in protein synthesis, with its codon structure directly impacting biological properties. While Language Models (LMs) have shown promise in analyzing biological sequences, existing approaches fail to account for the hierarchical nature of mRNA's codon structure. We introduce Hierarchical Encoding for mRNA Language Modeling (HELM), a novel pre-training strategy that incorporates codon-level hierarchical structure into language model training. HELM modulates the loss function based on codon synonymity, aligning the model's learning process with the biological reality of mRNA sequences. We evaluate HELM on diverse mRNA datasets and tasks, demonstrating that HELM outperforms standard language model pre-training as well as existing foundation model baselines on six diverse downstream property prediction tasks and an antibody region annotation tasks on average by around 8%. Additionally, HELM enhances the generative capabilities of language model, producing diverse mRNA sequences that better align with the underlying true data distribution compared to non-hierarchical baselines.

We propose X-NeMo, a novel zero-shot diffusion-based portrait animation pipeline that animates a static portrait using facial movements from a driving video of a different individual. Our work first identifies the root causes of the limitations in prior approaches, such as identity leakage and difficulty in capturing subtle and extreme expressions. To address these challenges, we introduce a fully end-to-end training framework that distills a 1D identity-agnostic latent motion descriptor from driving image, effectively controlling motion through cross-attention during image generation. Our implicit motion descriptor captures expressive facial motion in fine detail, learned end-to-end from a diverse video dataset without reliance on any pre-trained motion detectors. We further disentangle motion latents from identity cues with enhanced expressiveness by supervising their learning with a dual GAN decoder, alongside spatial and color augmentations. By embedding the driving motion into a 1D latent vector and controlling motion via cross-attention instead of additive spatial guidance, our design effectively eliminates the transmission of spatial-aligned structural clues from the driving condition to the diffusion backbone, substantially mitigating identity leakage. Extensive experiments demonstrate that X-NeMo surpasses state-of-the-art baselines, producing highly expressive animations with superior identity resemblance. Our code and models will be available for research.

Diffusion models, the most popular generative paradigm so far, can inject conditional information into the generation path to guide the latent towards desired directions. However, existing text-to-image diffusion models often fail to maintain high image quality and high prompt-image alignment for those challenging prompts. To mitigate this issue and enhance existing pretrained diffusion models, we mainly made three contributions in this paper. First, we propose **diffusion self-reflection** that alternately performs denoising and inversion and demonstrate that such diffusion self-reflection can leverage the guidance gap between denoising and inversion to capture prompt-related semantic information with theoretical and empirical evidence. Second, motivated by theoretical analysis, we derive Zigzag Diffusion Sampling (Z-Sampling), a novel self-reflection-based diffusion sampling method that leverages the guidance gap between denosing and inversion to accumulate semantic information step by step along the sampling path, leading to improved sampling results. Moreover, as a plug-and-play method, Z-Sampling can be generally applied to various diffusion models (e.g., accelerated ones and Transformer-based ones) with very limited coding and computational costs. Third, our extensive experiments demonstrate that Z-Sampling can generally and significantly enhance generation quality across various benchmark datasets, diffusion models, and performance evaluation metrics. For example, DreamShaper with Z-Sampling can self-improve with the HPSv2 winning rate up to **94%** over the original results. Moreover, Z-Sampling can further enhance existing diffusion models combined with other orthogonal methods, including Diffusion-DPO. The code is publicly available at [github.com/xie-lab-ml/Zigzag-Diffusion-Sampling](https://github.com/xie-lab-ml/Zigzag-Diffusion-Sampling).

Recent advancements in operator-type neural networks have shown promising results in approximating the solutions of spatiotemporal Partial Differential Equations (PDEs). However, these neural networks often entail considerable training expenses, and may not always achieve the desired accuracy required in many scientific and engineering disciplines. In this paper, we propose a new learning framework to address these issues. A new spatiotemporal adaptation is proposed to generalize any Fourier Neural Operator (FNO) variant to learn maps between Bochner spaces, which can perform an arbitrary-length temporal super-resolution for the first time. To better exploit this capacity, a new paradigm is proposed to refine the commonly adopted end-to-end neural operator training and evaluations with the help from the wisdom from traditional numerical PDE theory and techniques. Specifically, in the learning problems for the turbulent flow modeled by the Navier-Stokes Equations (NSE), the proposed paradigm trains an FNO only for a few epochs. Then, only the newly proposed spatiotemporal spectral convolution layer is fine-tuned without the frequency truncation. The spectral fine-tuning loss function uses a negative Sobolev norm for the first time in operator learning, defined through a reliable functional-type a posteriori error estimator whose evaluation is exact thanks to the Parseval identity. Moreover, unlike the difficult nonconvex optimization problems in the end-to-end training, this fine-tuning loss is convex. Numerical experiments on commonly used NSE benchmarks demonstrate significant improvements in both computational efficiency and accuracy, compared to end-to-end evaluation and traditional numerical PDE solvers under certain conditions. The source code is publicly available at https://github.com/scaomath/torch-cfd.

As a core component of Transformer-like deep architectures, a feed-forward network (FFN) for channel mixing is responsible for learning features of each token. Recent works show channel mixing can be enhanced by increasing computational burden or can be slimmed at the sacrifice of performance. Although some efforts have been made, existing works are still struggling to solve the paradox of performance and complexity trade-offs. In this paper, we propose an Asymmetric Factorized Bilinear Operation (AFBO) to replace FFN of vision transformer (ViT), which attempts to efficiently explore rich statistics of token features for achieving better performance and complexity trade-off. Specifically, our AFBO computes second-order statistics via a spatial-channel factorized bilinear operation for feature learning, which replaces a simple linear projection in FFN and enhances the feature learning ability of ViT by modeling second-order correlation among token features. Furthermore, our AFBO presents two structured-sparsity channel mapping strategies, namely Grouped Cross Channel Mapping (GCCM) and Overlapped Cycle Channel Mapping (OCCM). They decompose bilinear operation into grouped channel features by considering information interaction between groups, significantly reducing computational complexity while guaranteeing model performance. Finally, our AFBO is built with GCCM and OCCM in an asymmetric way, aiming to achieve a better trade-off. Note that our AFBO is model-agnostic, which can be flexibly integrated with existing ViTs. Experiments are conducted with twenty ViTs on various tasks, and the results show our AFBO is superior to its counterparts while improving existing ViTs in terms of generalization and robustness.

Discrete diffusion has achieved state-of-the-art performance, outperforming or approaching autoregressive models on standard benchmarks. In this work, we introduce *Discrete Diffusion with Planned Denoising* (DDPD), a novel framework that separates the generation process into two models: a planner and a denoiser. At inference time, the planner selects which positions to denoise next by identifying the most corrupted positions in need of denoising, including both initially corrupted and those requiring additional refinement. This plan-and-denoise approach enables more efficient reconstruction during generation by iteratively identifying and denoising corruptions in the optimal order. DDPD outperforms traditional denoiser-only mask diffusion methods, achieving superior results on language modeling benchmarks such as *text8*, *OpenWebText*, and token-based generation on *ImageNet 256 × 256*. Notably, in language modeling, DDPD significantly reduces the performance gap between diffusion-based and autoregressive methods in terms of generative perplexity. Code is available at [github.com/liusulin/DDPD](https://github.com/liusulin/DDPD).

Why do larger language models generalize better? To explore this question, we develop generalization bounds on the pretraining objective of large language models (LLMs) in the compute-optimal regime, as described by the Chinchilla scaling laws. We introduce a novel, fully empirical Freedman-type martingale concentration inequality that tightens existing bounds by accounting for the variance of the loss function. The generalization bound can be broken into three contributions: the number of parameters per token, the loss variance, and the quantization error at a fixed bitrate. As language models are scaled up, the number of parameters per data point stays constant; however, both the loss variance and the quantization error decrease, implying that larger models should have \emph{smaller} generalization gaps. We examine why larger models tend to be more quantizable from an information theoretic perspective, showing that the rate at which they can integrate new information grows slower than their capacity on the compute optimal frontier. From these findings we produce a scaling law for the generalization gap, showing that our bounds decrease in a predictable way.

Understanding how features evolve across layers in deep neural networks is a fundamental challenge in mechanistic interpretability, particularly due to polysemanticity and feature superposition. While Sparse Autoencoders (SAEs) have been used to extract interpretable features from individual layers, aligning these features across layers has remained an open problem. In this paper, we introduce SAE Match, a novel, data-free method for aligning SAE features across different layers of a neural network. Our approach involves matching features by minimizing the mean squared error between the folded parameters of SAEs, a technique that incorporates activation thresholds into the encoder and decoder weights to account for differences in feature scales. Through extensive experiments on the Gemma 2 language model, we demonstrate that our method effectively captures feature evolution across layers, improving feature matching quality. We also show that features persist over several layers and that our approach can approximate hidden states across layers. Our work advances the understanding of feature dynamics in neural networks and provides a new tool for mechanistic interpretability studies.