ICLR 2025 Papers

Policies developed through Reinforcement Learning (RL) and Imitation Learning (IL) have shown great potential in continuous control tasks, but real-world applications often require adapting trained policies to unforeseen requirements. While fine-tuning can address such needs, it typically requires additional data and access to the original training metrics and parameters. In contrast, an online planning algorithm, if capable of meeting the additional requirements, can eliminate the necessity for extensive training phases and customize the policy without knowledge of the original training scheme or task. In this work, we propose a generic online planning algorithm for customizing continuous-control policies at the execution time, which we call Residual-MPPI. It can customize a given prior policy on new performance metrics in few-shot and even zero-shot online settings, given access to the prior action distribution alone. Through our experiments, we demonstrate that the proposed Residual-MPPI algorithm can accomplish the few-shot/zero-shot online policy customization task effectively, including customizing the champion-level racing agent, Gran Turismo Sophy (GT Sophy) 1.0, in the challenging car racing scenario, Gran Turismo Sport (GTS) environment. Code for MuJoCo experiments is included in the supplementary and will be open-sourced upon acceptance. Demo videos are available on our website: https://sites.google.com/view/residual-mppi.

In drug discovery, highly automated high-throughput laboratories are used to screen a large number of compounds in search of effective drugs. These experiments are expensive, so one might hope to reduce their cost by only experimenting on a subset of the compounds, and predicting the outcomes of the remaining experiments. In this work, we model this scenario as a sequential subset selection problem: we aim to select the smallest set of candidates in order to achieve some desired level of accuracy for the system as a whole. Our key observation is that, if there is heterogeneity in the difficulty of the prediction problem across the input space, selectively obtaining the labels for the hardest examples in the acquisition pool will leave only the relatively easy examples to remain in the inference set, leading to better overall system performance. We call this mechanism inference set design, and propose the use of a confidence-based active learning solution to prune out these challenging examples. Our algorithm includes an explicit stopping criterion that interrupts the acquisition loop when it is sufficiently confident that the system has reached the target performance. Our empirical studies on image and molecular datasets, as well as a real-world large-scale biological assay, show that active learning for inference set design leads to significant reduction in experimental cost while retaining high system performance.

Pretraining large language models (LLMs) on vast and heterogeneous datasets is crucial for achieving state-of-the-art performance across diverse downstream tasks. However, current training paradigms treat all samples equally, overlooking the importance or relevance of individual samples throughout the training process. Existing reweighting strategies, which primarily focus on group-level data importance, fail to leverage fine-grained instance-level information and do not adapt dynamically to individual sample importance as training progresses. In this paper, we introduce novel algorithms for dynamic, instance-level data reweighting aimed at improving both the efficiency and effectiveness of LLM pretraining. Our methods adjust the weight of each training sample based on its loss value in an online fashion, allowing the model to dynamically focus on more informative or important samples at the current training stage. In particular, our framework allows us to systematically devise reweighting strategies deprioritizing redundant or uninformative data, which we find tend to work best. Furthermore, we develop a new theoretical framework for analyzing the impact of loss-based reweighting on the convergence of gradient-based optimization, providing the first formal characterization of how these strategies affect convergence bounds. We empirically validate our approach across a spectrum of tasks, from pretraining 7B and 1.4B parameter LLMs to smaller-scale language models and linear regression problems, demonstrating that our loss-based reweighting approach can lead to faster convergence and significantly improved performance.

Low-rank adaptation, also known as LoRA, has emerged as a prominent method for parameter-efficient fine-tuning of foundation models. Despite its computational efficiency, LoRA still yields inferior performance compared to full fine-tuning. In this paper, we first uncover a fundamental connection between the optimization processes of LoRA and full fine-tuning: using LoRA for optimization is mathematically equivalent to full fine-tuning using a low-rank gradient for parameter updates. And this low-rank gradient can be expressed in terms of the gradients of the two low-rank matrices in LoRA. Leveraging this insight, we introduce LoRA-Pro, a method that enhances LoRA's performance by strategically adjusting the gradients of these low-rank matrices. This adjustment allows the low-rank gradient to more accurately approximate the full fine-tuning gradient, thereby narrowing the performance gap between LoRA and full fine-tuning. Furthermore, we theoretically derive the optimal solutions for adjusting the gradients of the low-rank matrices, applying them during fine-tuning in LoRA-Pro. We conduct extensive experiments across natural language understanding, dialogue generation, mathematical reasoning, code generation, and image classification tasks, demonstrating that LoRA-Pro substantially improves LoRA's performance, effectively narrowing the gap with full fine-tuning. Our code is publicly available at https://github.com/mrflogs/LoRA-Pro.

Recent advancements in generative modeling have significantly enhanced the reconstruction of audio waveforms from various representations. While diffusion models are adept at this task, they are hindered by latency issues due to their operation at the individual sample point level and the need for numerous sampling steps. In this study, we introduce RFWave, a cutting-edge multi-band Rectified Flow approach designed to reconstruct high-fidelity audio waveforms from Mel-spectrograms or discrete acoustic tokens. RFWave uniquely generates complex spectrograms and operates at the frame level, processing all subbands simultaneously to boost efficiency. Leveraging Rectified Flow, which targets a straight transport trajectory, RFWave achieves reconstruction with just 10 sampling steps. Our empirical evaluations show that RFWave not only provides outstanding reconstruction quality but also offers vastly superior computational efficiency, enabling audio generation at speeds up to 160 times faster than real-time on a GPU. Both an online demonstration and the source code are accessible.

Increased deployment of autonomous systems in fields like transportation and robotics have seen a corresponding increase in safety-critical failures. These failures can be difficult to model and debug due to the relative lack of data: compared to tens of thousands of examples from normal operations, we may have only seconds of data leading up to the failure. This scarcity makes it challenging to train generative models of rare failure events, as existing methods risk either overfitting to noise in the limited failure dataset or underfitting due to an overly strong prior. We address this challenge with CalNF, or calibrated normalizing flows, a self-regularized framework for posterior learning from limited data. CalNF achieves state-of-the-art performance on data-limited failure modeling and inverse problems and enables a first-of-a-kind case study into the root causes of the 2022 Southwest Airlines scheduling crisis.

Reconstructing dynamics using samples from sparsely time-resolved snapshots is an important problem in both natural sciences and machine learning. Here, we introduce a new deep learning approach for solving regularized unbalanced optimal transport (RUOT) and inferring continuous unbalanced stochastic dynamics from observed snapshots. Based on the RUOT form, our method models these dynamics without requiring prior knowledge of growth and death processes or additional information, allowing them to be learned directly from data. Theoretically, we explore the connections between the RUOT and Schrödinger bridge problem and discuss the key challenges and potential solutions. The effectiveness of our method is demonstrated with a synthetic gene regulatory network, high-dimensional Gaussian Mixture Model, and single-cell RNA-seq data from blood development. Compared with other methods, our approach accurately identifies growth and transition patterns, eliminates false transitions, and constructs the Waddington developmental landscape. Our code is available at: [https://github.com/zhenyiizhang/DeepRUOT](https://github.com/zhenyiizhang/DeepRUOT).

Document Visual Question Answering (DocVQA) has introduced a new paradigm for end-to-end document understanding, and quickly became one of the standard benchmarks for multimodal LLMs. Automating document processing workflows, driven by DocVQA models, presents significant potential for many business sectors. However, documents tend to contain highly sensitive information, raising concerns about privacy risks associated with training such DocVQA models. One significant privacy vulnerability, exploited by the membership inference attack, is the possibility for an adversary to determine if a particular record was part of the model's training data. In this paper, we introduce two novel membership inference attacks tailored specifically to DocVQA models. These attacks are designed for two different adversarial scenarios: a white-box setting, where the attacker has full access to the model architecture and parameters, and a black-box setting, where only the model's outputs are available. Notably, our attacks assume the adversary lacks access to auxiliary datasets, which is more realistic in practice but also more challenging. Our unsupervised methods outperform existing state-of-the-art membership inference attacks across a variety of DocVQA models and datasets, demonstrating their effectiveness and highlighting the privacy risks in this domain.

Training data attribution (TDA) methods aim to attribute model outputs back to specific training examples, and the application of these methods to large language model (LLM) outputs could significantly advance model transparency and data curation. However, it has been challenging to date to apply these methods to the full scale of LLM pretraining. In this paper, we refine existing gradient-based methods to work effectively at scale, allowing us to retrieve influential examples for an 8B-parameter language model from a pretraining corpus of over 160B tokens with no need for subsampling or pre-filtering. Our method combines several techniques, including optimizer state correction, a task-specific Hessian approximation, and normalized encodings, which we find to be critical for performance at scale. In quantitative evaluations on a fact tracing task, our method performs best at identifying examples that influence model predictions, but classical, model-agnostic retrieval methods such as BM25 still perform better at finding passages which explicitly contain relevant facts. These results demonstrate a misalignment between factual *attribution* and causal *influence*. With increasing model size and training tokens, we find that influence more closely aligns with factual attribution. Finally, we examine different types of examples identified as influential by our method, finding that while many directly entail a particular fact, others support the same output by reinforcing priors on relation types, common entities, and names. We release our prompt set and model outputs, along with a web-based visualization tool to explore influential examples for factual predictions, commonsense reasoning, arithmetic, and open-ended generation for an 8B-parameter LLM.

Large language models (LLMs) have demonstrated remarkable in-context learning (ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits two limitations: \textbf{(a) Limited \textit{i.i.d.} Setting.} Most studies focus on supervised function learning tasks where prompts are constructed with \textit{i.i.d.} input-label pairs. This \textit{i.i.d.} assumption diverges significantly from real language learning scenarios where prompt tokens are interdependent. \textbf{(b) Lack of Emergence Explanation.} Most literature answers \textbf{\textit{what}} ICL does from an implicit optimization perspective but falls short in elucidating \textbf{\textit{how}} ICL emerges and the impact of pre-training phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm, \textbf{\textit{auto-regressive next-token prediction (AR-NTP)}}, which closely aligns with the actual training of language models. Specifically, within AR-NTP, we emphasize prompt token-dependency, which involves predicting each subsequent token based on the preceding sequence. To address (b), we formalize a systematic pre-training and ICL framework, highlighting the layer-wise structure of sequences and topics, alongside a two-level expectation. In conclusion, we present data-dependent, topic-dependent and optimization-dependent PAC-Bayesian generalization bounds for pre-trained LLMs, investigating that \textbf{\textit{ICL emerges from the generalization of sequences and topics}}. Our theory is supported by experiments on numerical linear dynamic systems, synthetic GINC and real-world language datasets.

Reinforcement learning algorithms typically necessitate extensive exploration of the state space to find optimal policies. However, in safety-critical applications, the risks associated with such exploration can lead to catastrophic consequences. Existing safe exploration methods attempt to mitigate this by imposing constraints, which often result in overly conservative behaviours and inefficient learning. Heavy penalties for early constraint violations can trap agents in local optima, deterring exploration of risky yet high-reward regions of the state space. To address this, we introduce a method that explicitly learns state-conditioned safety representations. By augmenting the state features with these safety representations, our approach naturally encourages safer exploration without being excessively cautious, resulting in more efficient and safer policy learning in safety-critical scenarios. Empirical evaluations across diverse environments show that our method significantly improves task performance while reducing constraint violations during training, underscoring its effectiveness in balancing exploration with safety.

A mechanistic understanding of how MLPs do computation in deep neural net- works remains elusive. Current interpretability work can extract features from hidden activations over an input dataset but generally cannot explain how MLP weights construct features. One challenge is that element-wise nonlinearities introduce higher-order interactions and make it difficult to trace computations through the MLP layer. In this paper, we analyze bilinear MLPs, a type of Gated Linear Unit (GLU) without any element-wise nonlinearity that neverthe- less achieves competitive performance. Bilinear MLPs can be fully expressed in terms of linear operations using a third-order tensor, allowing flexible analysis of the weights. Analyzing the spectra of bilinear MLP weights using eigendecom- position reveals interpretable low-rank structure across toy tasks, image classifi- cation, and language modeling. We use this understanding to craft adversarial examples, uncover overfitting, and identify small language model circuits directly from the weights alone. Our results demonstrate that bilinear layers serve as an interpretable drop-in replacement for current activation functions and that weight- based interpretability is viable for understanding deep-learning models.

Diffusion models have emerged as a successful approach for molecular docking, but they often cannot model protein flexibility or generate nonphysical poses. We argue that both these challenges can be tackled by framing the problem as a transport between distributions. Still, existing paradigms lack the flexibility to define effective maps between such complex distributions. To address this limitation, we propose Unbalanced Flow Matching, a generalization of Flow Matching (FM) that allows trading off sample efficiency with approximation accuracy and enables more accurate transport. Empirically, we apply Unbalanced FM on flexible docking and structure relaxation, demonstrating our ability to model protein flexibility and generate energetically favorable poses. On the PDBBind docking benchmark, our method FlexDock improves the docking performance while increasing the proportion of energetically favorable poses from 30% to 73%.

Self-consistency-based approaches, which involve repeatedly sampling multiple outputs and selecting the most consistent one as the final response, prove to be remarkably effective in improving the factual accuracy of large language models. Nonetheless, existing methods usually have strict constraints on the task format, largely limiting their applicability. In this paper, we present Integrative Decoding (ID), to unlock the potential of self-consistency in open-ended generation tasks. ID operates by constructing a set of inputs, each prepended with a previously sampled response, and then processes them concurrently, with the next token being selected by aggregating of all their corresponding predictions at each decoding step. In essence, this simple approach implicitly incorporates self-consistency in the decoding objective. Extensive evaluation shows that ID consistently enhances factuality over a wide range of language models, with substantial improvements on the TruthfulQA (+11.2%), Biographies (+15.4%) and LongFact (+8.5%) benchmarks. The performance gains amplify progressively as the number of sampled responses increases, indicating the potential of ID to scale up with repeated sampling.

Error correcting codes (ECCs) are indispensable for reliable transmission in communication systems. Recent advancements in deep learning have catalyzed the exploration of ECC decoders based on neural networks. Among these, transformer-based neural decoders have achieved state-of-the-art decoding performance. In this paper, we propose a novel Cross-Attention Message-Passing Transformer (CrossMPT), which shares key operational principles with conventional message-passing decoders. While conventional transformer-based decoders employ a self-attention mechanism without distinguishing between magnitude and syndrome embeddings, CrossMPT updates these two types of embeddings separately and iteratively via two masked cross-attention blocks. The mask matrices are determined by the code's parity-check matrix, which explicitly captures and removes irrelevant relationships between the magnitude and syndrome embeddings. Our experimental results show that CrossMPT significantly outperforms existing neural network-based decoders for various code classes. Notably, CrossMPT achieves this decoding performance improvement while significantly reducing memory usage, computational complexity, inference time, and training time.

With the increasing interest in Quantum Machine Learning, Quantum Neural Networks (QNNs) have emerged and gained significant attention. These models have, however, been shown to be notoriously difficult to train, which we hypothesize is partially due to the architectures, called ansatzes, that are hardly studied at this point. Therefore, in this paper, we take a step back and analyze ansatzes. We initially consider their expressivity, i.e., the space of operations they are able to express, and show that the closeness to being a 2-design, the primarily used measure, fails at capturing this property. Hence, we look for alternative ways to characterize ansatzes, unrelated to expressivity, by considering the local neighborhood of the model space, in particular, analyzing model distinguishability upon small perturbation of parameters. We derive an upper bound on their distinguishability, showcasing that QNNs using the Hardware Efficient Ansatz with few parameters are hardly discriminable upon update. Our numerical experiments support our bounds and further indicate that there is a significant degree of variability, which stresses the need for warm-starting or clever initialization. Altogether, our work provides an ansatz-centric perspective on training dynamics and difficulties in QNNs, ultimately suggesting that iterative training of small quantum models may not be effective, which contrasts their initial motivation.

To build safe and reliable graph machine learning systems, unsupervised graph-level anomaly detection (GLAD) and unsupervised graph-level out-of-distribution (OOD) detection (GLOD) have received significant attention in recent years. Though these two lines of research share the same objective, they have been studied independently in the community due to distinct evaluation setups, creating a gap that hinders the application and evaluation of methods from one to the other. To bridge the gap, in this work, we present a Unified Benchmark for unsupervised Graph-level OOD and anomaly Detection (UB-GOLD), a comprehensive evaluation framework that unifies GLAD and GLOD under the concept of generalized graph-level OOD detection. Our benchmark encompasses 35 datasets spanning four practical anomaly and OOD detection scenarios, facilitating the comparison of 18 representative GLAD/GLOD methods. We conduct multi-dimensional analyses to explore the effectiveness, generalizability, robustness, and efficiency of existing methods, shedding light on their strengths and limitations. Furthermore, we provide an open-source codebase of UB-GOLD to foster reproducible research and outline potential directions for future investigations based on our insights.

Maximizing the spectral gap through graph rewiring has been proposed to enhance the performance of message-passing graph neural networks (GNNs) by addressing over-squashing. However, as we show, minimizing the spectral gap can also improve generalization. To explain this, we analyze how rewiring can benefit GNNs within the context of stochastic block models. Since spectral gap optimization primarily influences community strength, it improves performance when the community structure aligns with node labels. Building on this insight, we propose three distinct rewiring strategies that explicitly target community structure, node labels, and their alignment: (a) community structure-based rewiring (ComMa), a more computationally efficient alternative to spectral gap optimization that achieves similar goals; (b) feature similarity-based rewiring (FeaSt), which focuses on maximizing global homophily; and (c) a hybrid approach (ComFy), which enhances local feature similarity while preserving community structure to optimize label-community alignment. Extensive experiments confirm the effectiveness of these strategies and support our theoretical insights.

Local learning, which trains a network through layer-wise local targets and losses, has been studied as an alternative to backpropagation (BP) in neural computation. However, its algorithms often become more complex or require additional hyperparameters due to the locality, making it challenging to identify desirable settings where the algorithm progresses in a stable manner. To provide theoretical and quantitative insights, we introduce maximal update parameterization ($\mu$P) in the infinite-width limit for two representative designs of local targets: predictive coding (PC) and target propagation (TP). We verify that $\mu$P enables hyperparameter transfer across models of different widths. Furthermore, our analysis reveals unique and intriguing properties of $\mu$P that are not present in conventional BP. By analyzing deep linear networks, we find that PC's gradients interpolate between first-order and Gauss-Newton-like gradients, depending on the parameterization. We demonstrate that, in specific standard settings, PC in the infinite-width limit behaves more similarly to the first-order gradient. For TP, even with the standard scaling of the last layer differing from classical $\mu$P, its local loss optimization favors the feature learning regime over the kernel regime.