ICML 2025 Papers

Designing effective reward functions remains a fundamental challenge in reinforcement learning (RL), as it often requires extensive human effort and domain expertise. While RL from human feedback has been successful in aligning agents with human intent, acquiring high-quality feedback is costly and labor-intensive, limiting its scalability. Recent advancements in foundation models present a promising alternative--leveraging AI-generated feedback to reduce reliance on human supervision in reward learning. Building on this paradigm, we introduce ERL-VLM, an enhanced rating-based RL method that effectively learns reward functions from AI feedback. Unlike prior methods that rely on pairwise comparisons, ERL-VLM queries large vision-language models (VLMs) for absolute ratings of individual trajectories, enabling more expressive feedback and improved sample efficiency. Additionally, we propose key enhancements to rating-based RL, addressing instability issues caused by data imbalance and noisy labels. Through extensive experiments across both low-level and high-level control tasks, we demonstrate that ERL-VLM significantly outperforms existing VLM-based reward generation methods. Our results demonstrate the potential of AI feedback for scaling RL with minimal human intervention, paving the way for more autonomous and efficient reward learning.

Large Language Models (LLMs) have demonstrated significant potential in code generation by following natural language instructions. Unfortunately, crucial real-world software engineering tasks, such as debugging or repository-level feature implementation, involve processing extensive contexts beyond current LLM context sizes and performing complex reasoning that is brittle using standard autoregressive decoding. Enhancing LLMs' performance in these scenarios requires careful consideration of the contextual information provided to the model, optimizing how the model leverages that, and identifying tools that enable more effective navigation of the development environment. To address these challenges, we introduce Nemotron-CORTEXA, an agentic system built on a predefined scaffold that enhances LLMs' ability to navigate and reason efficiently in complex software engineering contexts. Specifically, we develop a novel code embedding model that retrieves the most relevant files with greater precision, along with a localization agent that refines the granularity of the retrieval process. Additionally, we demonstrate that providing diverse contextual information and utilizing different prompt formats enable the model to identify and resolve issues more efficiently. We evaluate Nemotron-CORTEXA using SWE-bench, a benchmark derived from real-world GitHub issues. Compared to the widely used Agentless framework, Nemotron-CORTEXA achieves a higher issue resolution rate at a lower cost, highlighting its practical impact in addressing real-world software engineering challenges.

High-dimensional imaging of neural activity, such as widefield calcium and functional ultrasound imaging, provide a rich source of information for understanding the relationship between brain activity and behavior. Accurately modeling neural dynamics in these modalities is crucial for understanding this relationship but is hindered by the high-dimensionality, complex spatiotemporal dependencies, and prevalent behaviorally irrelevant dynamics in these modalities. Existing dynamical models often employ preprocessing steps to obtain low-dimensional representations from neural image modalities. However, this process can discard behaviorally relevant information and miss spatiotemporal structure. We propose SBIND, a novel data-driven deep learning framework to model spatiotemporal dependencies in neural images and disentangle their behaviorally relevant dynamics from other neural dynamics. We validate SBIND on widefield imaging datasets, and show its extension to functional ultrasound imaging, a recent modality whose dynamical modeling has largely remained unexplored. We find that our model effectively identifies both local and long-range spatial dependencies across the brain while also dissociating behaviorally relevant neural dynamics. Doing so, SBIND outperforms existing models in neural-behavioral prediction. Overall, SBIND provides a versatile tool for investigating the neural mechanisms underlying behavior using imaging modalities.

In this paper, we investigate the problem of recovering hidden communities in the Labeled Stochastic Block Model (LSBM) with a finite number of clusters whose sizes grow linearly with the total number of nodes. We derive the necessary and sufficient conditions under which the expected number of misclassified nodes is less than $ s $, for any number $ s = o(n) $. To achieve this, we propose IAC (Instance-Adaptive Clustering), the first algorithm whose performance matches the instance-specific lower bounds both in expectation and with high probability. IAC is a novel two-phase algorithm that consists of a one-shot spectral clustering step followed by iterative likelihood-based cluster assignment improvements. This approach is based on the instance-specific lower bound and notably does not require any knowledge of the model parameters, including the number of clusters. By performing the spectral clustering only once, IAC maintains an overall computational complexity of $ \mathcal{O}(n\, \text{polylog}(n)) $, making it scalable and practical for large-scale problems.

Language Models (LMs) are prone to ''memorizing'' training data, including substantial sensitive user information. To mitigate privacy risks and safeguard the right to be forgotten, machine unlearning has emerged as a promising approach for enabling LMs to efficiently ''forget'' specific texts. However, despite the good intentions, is textual unlearning really as effective and reliable as expected? To address the concern, we first propose Unlearning Likelihood Ratio Attack+ (U-LiRA+), a rigorous textual unlearning auditing method, and find that unlearned texts can still be detected with very high confidence after unlearning. Further, we conduct an in-depth investigation on the privacy risks of textual unlearning mechanisms in deployment and present the Textual Unlearning Leakage Attack (TULA), along with its variants in both black- and white-box scenarios. We show that textual unlearning mechanisms could instead reveal more about the unlearned texts, exposing them to significant membership inference and data reconstruction risks. Our findings highlight that existing textual unlearning actually gives a false sense of unlearning, underscoring the need for more robust and secure unlearning mechanisms.

Quantized large language models (LLMs) have gained increasing attention and significance for enabling deployment in resource-constrained environments. However, emerging studies on a few calibration dataset-free quantization methods suggest that quantization may compromise the safety capabilities of LLMs, underscoring the urgent need for systematic safety evaluations and effective mitigation strategies. In this paper, we present comprehensive safety evaluations across various mainstream quantization techniques and diverse calibration datasets, utilizing widely accepted safety benchmarks. To address the identified safety vulnerabilities, we propose a quantization-aware safety patching framework, Q-resafe, to efficiently restore the safety capabilities of quantized LLMs while minimizing any adverse impact on utility. Extensive experiment results demonstrate that Q-resafe successfully re-aligns the safety of quantized LLMs with their pre-quantization counterparts, even under challenging evaluation scenarios. Project page: https://github.com/Thecommonirin/Qresafe.

Personalized Federated Learning (PFL) has become a promising learning paradigm, enabling the training of high-quality personalized models through multiple communication rounds between clients and a central server. However, directly applying traditional PFL in real-world environments where communication is expensive, limited, or infeasible is challenging, as seen in Low Earth Orbit (LEO) satellite constellations, which face severe communication constraints due to their high mobility, limited contact windows. To address these issues, we introduce Federated Oriented Learning (FOL), a novel four-stage one-shot PFL algorithm designed to enhance local model performance by leveraging neighboring models within stringent communication constraints. FOL comprises model pretraining, model collection, model alignment (via fine-tuning, pruning, post fine-tuning, and ensemble refinement), and knowledge distillation stages. We establish two theoretical guarantees on empirical risk discrepancy between student and teacher models and the convergence of the distillation process. Extensive experiments on datasets Wildfire, Hurricane, CIFAR-10, CIFAR-100, and SVHN demonstrate that FOL consistently outperforms state-of-the-art one-shot Federated Learning (OFL) methods; for example, it achieves accuracy improvements of up to 39.24\% over the baselines on the Wildfire dataset.

LongRoPE2 is a novel approach that extends the effective context window of pre-trained large language models (LLMs) to the target length, while preserving the performance on the original shorter context window. This is achieved by three contributions: (1) a hypothesis that insufficient training in higher RoPE dimensions contributes to the persistent out-of-distribution (OOD) issues observed in existing methods; (2) an effective RoPE rescaling algorithm that adopts evolutionary search guided by "needle-driven" perplexity to address the insufficient training problem; (3) a mixed context window training approach that fine-tunes model weights to adopt rescaled RoPE for long-context sequences while preserving the short-context performance with the original RoPE. Extensive experiments on LLaMA3-8B and Phi3-mini-3.8B across various benchmarks validate the hypothesis and demonstrate the effectiveness of LongRoPE2. Remarkably, LongRoPE2 extends LLaMA3-8B to achieve a 128K effective context length while retaining over 98.5% of short-context performance, using only 10B tokens -- 80x fewer than Meta's approach, which fails to reach the target effective context length.

Domain Generalized Semantic Segmentation (DGSS) trains a model on a labeled source domain to generalize to unseen target domains with consistent contextual distribution and varying visual appearance. Most existing methods rely on domain randomization or data generation but struggle to capture the underlying scene distribution, resulting in the loss of useful semantic information. Inspired by the diffusion model's capability to generate diverse variations within a given scene context, we consider harnessing its rich prior knowledge of scene distribution to tackle the challenging DGSS task. In this paper, we propose a novel agent \textbf{Query}-driven learning framework based on \textbf{Diff}usion model guidance for DGSS, named QueryDiff. Our recipe comprises three key ingredients: (1) generating agent queries from segmentation features to aggregate semantic information about instances within the scene; (2) learning the inherent semantic distribution of the scene through agent queries guided by diffusion features; (3) refining segmentation features using optimized agent queries for robust mask predictions. Extensive experiments across various settings demonstrate that our method significantly outperforms previous state-of-the-art methods. Notably, it enhances the model's ability to generalize effectively to extreme domains, such as cubist art styles. Code is available at https://github.com/FanLiHub/QueryDiff.

Large Language Models (LLMs) have been observed to process non-human-readable text sequences, such as jailbreak prompts, often viewed as a bug for aligned LLMs. In this work, we present a systematic investigation challenging this perception, demonstrating that unnatural languages - strings that appear incomprehensible to humans but maintain semantic meanings for LLMs - contain latent features usable by models. Notably, unnatural languages possess latent features that can be generalized across different models and tasks during inference. Furthermore, models fine-tuned on unnatural versions of instruction datasets perform on-par with those trained on natural language, achieving \(49.71\) win rates in Length-controlled AlpacaEval 2.0 in average across various base models. In addition, through comprehensive analysis, we demonstrate that LLMs process unnatural languages by filtering noise and inferring contextual meaning from filtered words. Our code is publicly available at https://github.com/John-AI-Lab/Unnatural_Language.

With the rapid scaling of large language models (LLMs), structured pruning has become a widely used technique to learn efficient, smaller models from larger ones, delivering superior performance compared to training similarly sized models from scratch. In this paper, we move beyond the traditional static pruning approach of determining a fixed pruning mask for a model, and propose a dynamic approach to structured pruning. In our method, the pruning mask is input-dependent and adapts dynamically based on the information described in a user instruction. Our approach, termed "instruction-following pruning'', introduces a sparse mask predictor that takes the user instruction as input and dynamically selects the most relevant model parameters for the given task. To identify and activate effective parameters, we jointly optimize the sparse mask predictor and the LLM, leveraging both instruction-following data and the pre-training corpus. Experimental results demonstrate the effectiveness of our approach on a wide range of evaluation benchmarks. For example, our 3B activated model improves over the 3B dense model by 5-8 points of absolute margin on domains such as math and coding, and rivals the performance of a 9B model.

We study the problem of procurement auctions, in which an auctioneer seeks to acquire services from a group of strategic sellers with private costs. The quality of the services is measured through some submodular function that is known to the auctioneer. Our goal is to design computationally efficient procurement auctions that (approximately) maximize the difference between the quality of the acquired services and the total cost of the sellers, in a way that is incentive compatible (IC) and individual rational (IR) for the sellers, and generates non-negative surplus (NAS) for the auctioneer. {Our contribution is twofold: \textbf{i)} we provide an improved analysis of existing algorithms for non-positive submodular function maximization and \textbf{ii)} we design computationally efficient frameworks that transform submodular function optimization algorithms to mechanisms that are IC and IR for the sellers, NAS for the auctioneer, and approximation-preserving.} Our frameworks are general and work both in the offline setting where the auctioneer can observe the bids and the services of all the sellers simultaneously, and in the online setting where the sellers arrive in an adversarial order and the auctioneer has to make an irrevocable decision whether to purchase their service or not. We further investigate whether it is possible to convert state-of-art submodular optimization algorithms into descending auctions. We focus on the adversarial setting, meaning that the schedule of the descending prices is determined by an adversary. We show that a submodular optimization algorithm satisfying bi-criteria $(1/2,1)$-approximation in welfare can be effectively converted to a descending auction in this setting. We further establish a connection between descending auctions and online submodular optimization. Finally, we demonstrate the practical applications of our frameworks by instantiating them with different state-of-the-art submodular optimization algorithms and comparing their welfare performance through empirical experiments on publicly available datasets that consist of thousands of sellers.

While the capabilities of generative foundational models have advanced rapidly in recent years, methods to prevent harmful and unsafe behaviors remain underdeveloped. Among the pressing challenges in AI safety, machine unlearning (MU) has become increasingly critical to meet upcoming safety regulations. Most existing MU approaches focus on altering the most significant parameters of the model. However, these methods often require fine-tuning substantial portions of the model, resulting in high computational costs and training instabilities, which are typically mitigated by access to the original training dataset. In this work, we address these limitations by leveraging Singular Value Decomposition (SVD) to create a compact, low-dimensional projection that enables the selective forgetting of specific data points. We propose Singular Value Decomposition for Efficient Machine Unlearning (SEMU), a novel approach designed to optimize MU in two key aspects. First, SEMU minimizes the number of model parameters that need to be modified, effectively removing unwanted knowledge while making only minimal changes to the model's weights. Second, SEMU eliminates the dependency on the original training dataset, preserving the model's previously acquired knowledge without additional data requirements. Extensive experiments demonstrate that SEMU achieves competitive performance while significantly improving efficiency in terms of both data usage and the number of modified parameters.

Self-supervised learning (SSL) has proven to be a powerful approach for extracting biologically meaningful representations from single-cell data. To advance our understanding of SSL methods applied to single-cell data, we present scSSL-Bench, a comprehensive benchmark that evaluates nineteen SSL methods. Our evaluation spans nine datasets and focuses on three common downstream tasks: batch correction, cell type annotation, and missing modality prediction. Furthermore, we systematically assess various data augmentation strategies. Our analysis reveals task-specific trade-offs: the specialized single-cell frameworks, scVI, CLAIRE, and the finetuned scGPT excel at uni-modal batch correction, while generic SSL methods, such as VICReg and SimCLR, demonstrate superior performance in cell typing and multi-modal data integration. Random masking emerges as the most effective augmentation technique across all tasks, surpassing domain-specific augmentations. Notably, our results indicate the need for a specialized single-cell multi-modal data integration framework. scSSL-Bench provides a standardized evaluation platform and concrete recommendations for applying SSL to single-cell analysis, advancing the convergence of deep learning and single-cell genomics.

Conformal prediction is a popular framework of uncertainty quantification that constructs prediction sets with coverage guarantees. To uphold the exchangeability assumption, many conformal prediction methods necessitate an additional hold-out set for parameter tuning. Yet, the impact of violating this principle on coverage remains underexplored, making it ambiguous in practical applications. In this work, we empirically find that the tuning bias - the coverage gap introduced by leveraging the same dataset for tuning and calibration, is negligible for simple parameter tuning in many conformal prediction methods. In particular, we observe the scaling law of the tuning bias: this bias increases with parameter space complexity and decreases with calibration set size. Formally, we establish a theoretical framework to quantify the tuning bias and provide rigorous proof for the scaling law of the tuning bias by deriving its upper bound. In the end, we discuss how to reduce the tuning bias, guided by the theories we developed.

Proteolysis-targeting chimeras (PROTACs) are a groundbreaking technology for targeted protein degradation, but designing effective linkers that connect two molecular fragments to form a drug-candidate PROTAC molecule remains a key challenge. While diffusion models show promise in molecular generation, current diffusion models for PROTAC linker design are typically trained on small molecule datasets, introducing distribution mismatches in the chemical space between small molecules and target PROTACs. Direct fine-tuning on limited PROTAC datasets often results in overfitting and poor generalization. In this work, we propose DAD-PROTAC, a domain-adapted diffusion model for PROTAC linker design, which addresses this distribution mismatch in chemical space through density ratio estimation to bridge the gap between small-molecule and PROTAC domains. By decomposing the target score estimator into a pre-trained score function and a lightweight score correction term, DAD-PROTAC achieves efficient fine-tuning without full retraining. Experimental results demonstrate its superior ability to generate high-quality PROTAC linkers.

With the growing interest in embodied and spatial intelligence, accurately predicting trajectories in 3D environments has become increasingly critical. However, no datasets have been explicitly designed to study 3D trajectory prediction. To this end, we contribute a 3D motion trajectory (3DMoTraj) dataset collected from unmanned underwater vehicles (UUVs) operating in oceanic environments. Mathematically, trajectory prediction becomes significantly more complex when transitioning from 2D to 3D. To tackle this challenge, we analyze the prediction complexity of 3D trajectories and propose a new method consisting of two key components: decoupled trajectory prediction and correlated trajectory refinement. The former decouples inter-axis correlations, thereby reducing prediction complexity and generating coarse predictions. The latter refines the coarse predictions by modeling their inter-axis correlations. Extensive experiments show that our method significantly improves 3D trajectory prediction accuracy and outperforms state-of-the-art methods. Both the 3DMoTraj dataset and the method are available at https://github.com/zhouhao94/3DMoTraj.

In this work, we propose a new architecture-agnostic method for training idempotent neural networks. An idempotent operator satisfies $f(x) = f(f(x))$, meaning it can be applied iteratively with no effect beyond the first application. Some neural networks used in data transformation tasks, such as image generation and augmentation, can represent non-linear idempotent projections. Using methods from perturbation theory we derive the recurrence relation ${\mathbf{K}' \leftarrow 3\mathbf{K}^2 - 2\mathbf{K}^3}$ for iteratively projecting a real-valued matrix $\mathbf{K}$ onto the manifold of idempotent matrices. Our analysis shows that for linear, single-layer MLP networks this projection 1) has idempotent fixed points, and 2) is attracting only around idempotent points. We give an extension to non-linear networks by considering our approach as a substitution of the gradient for the canonical loss function, achieving an architecture-agnostic training scheme. We provide experimental results for MLP- and CNN-based architectures with significant improvement in idempotent error over the canonical gradient-based approach. Finally, we demonstrate practical applications of the method as we train a generative network successfully using only a simple reconstruction loss paired with our method.

Digital twins are models of real-world systems that can simulate their dynamics in response to potential actions. In complex settings, the state and action variables, and available data and knowledge relevant to a system can constantly change, requiring digital twins to continuously update with these changes to remain relevant. Current approaches struggle in this regard, as they require fixed, well-defined modelling environments, and they cannot adapt to novel variables without re-designs, or incorporate new information without re-training. To address this, we frame digital twinning as an in-context learning problem using large language models, enabling seamless updates to the twin at inference time. We develop CALM-DT, a Context-Adaptive Language Model-based Digital Twin that can accurately simulate across diverse state-action spaces using in-context learning alone by utilising fine-tuned encoders for sample retrieval. We empirically demonstrate CALM-DT's competitive performance with existing digital twin approaches, and its unique ability to adapt to changes in its modelling environment without parameter updates.