ICML 2025 论文列表

Designing biological sequences with desired properties is challenging due to vast search spaces and limited evaluation budgets. Although reinforcement learning methods use proxy models for rapid reward evaluation, insufficient training data can cause proxy misspecification on out-of-distribution inputs. To address this, we propose a novel off-policy search, $\delta$-Conservative Search, that enhances robustness by restricting policy exploration to reliable regions. Starting from high-score offline sequences, we inject noise by randomly masking tokens with probability $\delta$, then denoise them using our policy. We further adapt $\delta$ based on proxy uncertainty on each data point, aligning the level of conservativeness with model confidence. Experimental results show that our conservative search consistently enhances the off-policy training, outperforming existing machine learning methods in discovering high-score sequences across diverse tasks, including DNA, RNA, protein, and peptide design.

*Warning: Contains harmful model outputs.* Despite significant advancements, the propensity of Large Language Models (LLMs) to generate harmful and unethical content poses critical challenges. Measuring value alignment of LLMs becomes crucial for their regulation and responsible deployment. Although numerous benchmarks have been constructed to assess social bias, toxicity, and ethical issues in LLMs, those static benchmarks suffer from *evaluation chronoeffect*, in which, as models rapidly evolve, existing benchmarks may leak into training data or become saturated, *overestimating* ever-developing LLMs. To tackle this problem, we propose GETA, a novel *generative evolving testing* approach based on adaptive testing methods in measurement theory. Unlike traditional adaptive testing methods that rely on a static test item pool, GETA probes the underlying moral boundaries of LLMs by dynamically generating test items tailored to model capability. GETA co-evolves with LLMs by learning a joint distribution of item difficulty and model value conformity, thus effectively addressing evaluation chronoeffect. We evaluated various popular LLMs with GETA and demonstrated that 1) GETA can dynamically create difficulty-tailored test items and 2) GETA's evaluation results are more consistent with models' performance on unseen OOD and i.i.d. items, laying the groundwork for future evaluation paradigms.

Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to infer latent associations to locate the needle within the haystack. We evaluate 13 popular LLMs that claim to support contexts of at least 128K tokens. While they perform well in short contexts (<1K), performance degrades significantly as context length increases. At 32K, for instance, 11 models drop below 50% of their strong short-length baselines. Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%. Our analysis suggests these declines stem from the increased difficulty the attention mechanism faces in longer contexts when literal matches are absent, making it harder to retrieve relevant information. Even models enhanced with reasoning capabilities or CoT prompting struggle to maintain performance in long contexts. We publicly release the dataset and evaluation code at https://github.com/adobe-research/NoLiMa.

Speculative decoding (SPD) aims to accelerate the auto-regressive token generation process of a target Large Language Model (LLM). Some approaches employ a draft model with multiple heads to predict a sequence of future tokens, where each head handles a token in the sequence. The target LLM verifies the predicted sequence and accepts aligned tokens, enabling efficient multi-token generation. However, existing methods assume that all tokens within a sequence are equally important, employing identical head structures and relying on a single-generation paradigm, either serial or parallel. To this end, we theoretically demonstrate that initial tokens in the draft sequence are more important than later ones. Building on this insight, we propose Gumiho, a hybrid model combining serial and parallel heads. Specifically, given the critical importance of early tokens, we employ a sophisticated Transformer architecture for the early draft heads in a serial configuration to improve accuracy. For later tokens, we utilize multiple lightweight MLP heads operating in parallel to enhance efficiency. By allocating more advanced model structures and longer running times to the early heads, Gumiho achieves improved overall performance. The experimental results demonstrate that our method outperforms existing approaches, fully validating its effectiveness. Our code is available at https://github.com/AMD-AIG-AIMA/Gumiho.

We revisit the problem of distribution learning within the framework of learning-augmented algorithms. In this setting, we explore the scenario where a probability distribution is provided as potentially inaccurate advice on the true, unknown distribution. Our objective is to develop learning algorithms whose sample complexity decreases as the quality of the advice improves, thereby surpassing standard learning lower bounds when the advice is sufficiently accurate. Specifically, we demonstrate that this outcome is achievable for the problem of learning a multivariate Gaussian distribution $N(\mu, \Sigma)$ in the PAC learning setting. Classically, in the advice-free setting, $\widetilde{\Theta}(d^2/\varepsilon^2)$ samples are sufficient and worst case necessary to learn $d$-dimensional Gaussians up to TV distance $\varepsilon$ with constant probability. When we are additionally given a parameter $\widetilde{\Sigma}$ as advice, we show that $\widetilde{\mathcal{O}}(d^{2-\beta}/\varepsilon^2)$ samples suffices whenever $|| \widetilde{\Sigma}^{-1/2} \Sigma \widetilde{\Sigma}^{-1/2} - I_d ||_1 \leq \varepsilon d^{1-\beta}$ (where $||\cdot||_1$ denotes the entrywise $\ell_1$ norm) for any $\beta > 0$, yielding a polynomial improvement over the advice-free setting.

Referring MLLMs extend conventional multimodal large language models by allowing them to receive referring visual prompts and generate responses tailored to the indicated regions. However, these models often suffer from suboptimal performance due to incorrect responses tailored to misleading areas adjacent to or similar to the target region. This work introduces CPCF, a novel framework to address this issue and achieve superior results. CPCF contrasts outputs generated from the indicated visual prompt with those from contrastive prompts sampled from misleading regions, effectively suppressing the influence of erroneous information outside the target region on response generation. To further enhance the effectiveness and efficiency of our framework, several novel designs are proposed, including a prompt extraction network to automatically identify suitable contrastive prompts, a self-training method that leverages unlabeled data to improve training quality, and a distillation approach to reduce the additional computational overhead associated with contrastive decoding. Incorporating these novel designs, CPCF achieves state-of-the-art performance, as demonstrated by extensive experiments across multiple benchmarks. Project page: https://lanyunzhu.site/CPCF/

Inspired by the principle of deliberate practice in human learning, we propose Deliberate Practice for Synthetic Data Generation (DP), a novel framework that improves sample efficiency through dynamic synthetic data generation. Prior work has shown that scaling synthetic data is inherently challenging, as naively adding new data leads to diminishing returns. To address this, pruning has been identified as a key mechanism for improving scaling, enabling models to focus on the most informative synthetic samples. Rather than generating a large dataset and pruning it afterward, DP efficiently approximates the direct generation of informative samples. We theoretically show how training on challenging, informative examples improves scaling laws and empirically validate that DP achieves better scaling performance with significantly fewer training samples and iterations. On ImageNet-100, DP generates 3.4x fewer samples and requires six times fewer iterations, while on ImageNet-1k, it generates 8x fewer samples with a 30% reduction in iterations, all while achieving superior performance compared to prior work.

One explanation for the strong generalization ability of neural networks is implicit bias. Yet, the definition and mechanism of implicit bias in non-linear contexts remains little understood. In this work, we propose to characterize implicit bias by the count of connected regions in the input space with the same predicted label. Compared with parameter-dependent metrics (e.g., norm or normalized margin), region count can be better adapted to nonlinear, overparameterized models, because it is determined by the function mapping and is invariant to reparametrization. Empirically, we found that small region counts align with geometrically simple decision boundaries and correlate well with good generalization performance. We also observe that good hyper-parameter choices such as larger learning rates and smaller batch sizes can induce small region counts. We further establish the theoretical connections and explain how larger learning rate can induce small region counts in neural networks.

There is a growing interest in the ability of neural networks to execute algorithmic tasks (e.g., arithmetic, summary statistics, and sorting). The goal of this work is to better understand the role of attention in Transformers for algorithmic execution. Its importance for algorithmic execution has been studied theoretically and empirically using parallel computational models. Notably, many parallel algorithms communicate between processors solely using positional information. Inspired by this observation, we investigate how Transformers can execute algorithms using positional attention, where attention weights depend exclusively on positional encodings. We prove that Transformers with positional attention (positional Transformers) maintain the same expressivity of parallel computational models, incurring a logarithmic depth cost relative to the input length. We analyze their in-distribution learnability and explore how parameter norms in positional attention affect sample complexity. Our results show that positional Transformers introduce a learning trade-off: while they exhibit better theoretical dependence on parameter norms, certain tasks may require more layers, which can, in turn, increase sample complexity. Finally, we empirically explore the out-of-distribution performance of positional Transformers and find that they perform well in tasks where their underlying algorithmic solution relies on positional information.

The Latent Stochastic Differential Equation (SDE) is a powerful tool for time series and sequence modeling. However, training Latent SDEs typically relies on adjoint sensitivity methods, which depend on simulation and backpropagation through approximate SDE solutions, which limit scalability. In this work, we propose SDE Matching, a new simulation-free method for training Latent SDEs. Inspired by modern Score- and Flow Matching algorithms for learning generative dynamics, we extend these ideas to the domain of stochastic dynamics for time series modeling, eliminating the need for costly numerical simulations. Our results demonstrate that SDE Matching achieves performance comparable to adjoint sensitivity methods while drastically reducing computational complexity.

Solving mathematics problems has been an intriguing capability of large language models, and many efforts have been made to improve reasoning by extending reasoning length, such as through self-correction and extensive long chain-of-thoughts. While promising in problem-solving, advanced long reasoning chain models exhibit an undesired single-modal behavior, where trivial questions require unnecessarily tedious long chains of thought. In this work, we propose a way to allow models to be aware of inference budgets by formulating it as utility maximization with respect to an inference budget constraint, hence naming our algorithm Inference Budget-Constrained Policy Optimization (IBPO). In a nutshell, models fine-tuned through IBPO learn to ``understand'' the difficulty of queries and allocate inference budgets to harder ones. With different inference budgets, our best models are able to have a $4.14$\% and $5.74$\% absolute improvement ($8.08$\% and $11.2$\% relative improvement) on MATH500 using $2.16$x and $4.32$x inference budgets respectively, relative to LLaMA3.1 8B Instruct. These improvements are approximately $2$x those of self-consistency under the same budgets.

Implicit feedback recommendation is challenged by the missing negative feedback essential for effective model training. Existing methods often resort to negative sampling, a technique that assumes unlabeled interactions as negative samples. This assumption risks misclassifying potential positive samples within the unlabeled data, thereby undermining model performance. To address this issue, we introduce PURL, a model-agnostic framework that reframes implicit feedback recommendation as a weakly supervised learning task, eliminating the need for negative samples. However, its unbiasedness hinges on the accurate estimation of the class prior. To address this challenge, we propose Progressive Proximal Transport (PPT), which estimates the class prior by minimizing the proximal transport cost between positive and unlabeled samples. Experiments on three real-world datasets validate the efficacy of PURL in terms of improved recommendation quality. Code is available at https://github.com/HowardZJU/weakrec.

Inspired by the well-known coreset in clustering algorithms, we introduce the definition of the core kernel for multiple kernel clustering (MKC) algorithms. The core kernel refers to running MKC algorithms on smaller-scale base kernel matrices to obtain kernel weights similar to those obtained from the original full-scale kernel matrices. Specifically, the core kernel refers to a set of kernel matrices of size $\widetilde{\mathcal{O}}(1/\varepsilon^2)$ that perform MKC algorithms on them can achieve a $(1+\varepsilon)$-approximation for the kernel weights. Subsequently, we can leverage approximated kernel weights to obtain a theoretically guaranteed large-scale extension of MKC algorithms. In this paper, we propose a core kernel construction method based on singular value decomposition and prove that it satisfies the definition of the core kernel for three mainstream MKC algorithms. Finally, we conduct experiments on several benchmark datasets to verify the correctness of theoretical results and the efficiency of the proposed method.

Unsupervised domain adaptation (UDA) for time series data remains a critical challenge in deep learning, with traditional pseudo-labeling strategies failing to capture temporal patterns and channel-wise shifts between domains, producing sub-optimal pseudo labels. As such, we introduce TransPL, a novel approach that addresses these limitations by modeling the joint distribution $P(X,y)$ of the source domain through code transition matrices, where the codes are derived from vector quantization (VQ) of time series patches. Our method constructs class- and channel-wise code transition matrices from the source domain and employs Bayes' rule for target domain adaptation, generating pseudo-labels based on channel-wise weighted class-conditional likelihoods. TransPL offers three key advantages: explicit modeling of temporal transitions and channel-wise shifts between different domains, versatility towards different UDA scenarios (e.g., weakly-supervised UDA), and explainable pseudo-label generation. We validate TransPL's effectiveness through extensive analysis on four time series UDA benchmarks and confirm that it consistently outperforms state-of-the-art pseudo-labeling methods by a strong margin (6.1\% accuracy improvement, 4.9\% F1 improvement), while providing interpretable insights into the domain adaptation process through its learned code transition matrices.

We present a principled, per-instance approach to quantifying the difficulty of unlearning via fine-tuning. We begin by sharpening an analysis of noisy gradient descent for unlearning (Chien et al., 2024), obtaining a better utility–unlearning trade-off by replacing worst-case privacy loss bounds with per-instance privacy losses (Thudi et al., 2024), each of which bounds the (R ´enyi) divergence to retraining without an individual datapoint. To demonstrate the practical applicability of our theory, we present empirical results showing that our theoretical predictions are born out both for Stochastic Gradient Langevin Dynamics (SGLD) as well as for standard fine-tuning without explicit noise. We further demonstrate that per-instance privacy losses correlate well with several existing data difficulty metrics, while also identifying harder groups of data points, and introduce novel evaluation methods based on loss barriers. All together, our findings provide a foundation for more efficient and adaptive unlearning strategies tailored to the unique properties of individual data points.

In generative modelling and stochastic optimal control, a central computational task is to modify a reference diffusion process to maximise a given terminal-time reward. Most existing methods require this reward to be differentiable, using gradients to steer the diffusion towards favourable outcomes. However, in many practical settings, like diffusion bridges, the reward is singular, taking an infinite value if the target is hit and zero otherwise. We introduce a novel framework, based on Malliavin calculus and centred around a generalisation of the Tweedie score formula to nonlinear stochastic differential equations, that enables the development of methods robust to such singularities. This allows our approach to handle a broad range of applications, like diffusion bridges, or adding conditional controls to an already trained diffusion model. We demonstrate that our approach offers stable and reliable training, outperforming existing techniques. As a byproduct, we also introduce a novel score matching objective. Our loss functions are formulated such that they could readily be extended to manifold-valued and infinite dimensional diffusions.

This paper focuses on implanting multiple heterogeneous backdoor triggers in bridge-based diffusion models designed for complex and arbitrary input distributions. Existing backdoor formulations mainly address single-attack scenarios and are limited to Gaussian noise input models. To fill this gap, we propose MixBridge, a novel diffusion Schrödinger bridge (DSB) framework to cater to arbitrary input distributions (taking I2I tasks as special cases). Beyond this trait, we demonstrate that backdoor triggers can be injected into MixBridge by directly training with poisoned image pairs. This eliminates the need for the cumbersome modifications to stochastic differential equations required in previous studies, providing a flexible tool to study backdoor behavior for bridge models. However, a key question arises: can a single DSB model train multiple backdoor triggers? Unfortunately, our theory shows that when attempting this, the model ends up following the geometric mean of benign and backdoored distributions, leading to performance conflict across backdoor tasks. To overcome this, we propose a Divide-and-Merge strategy to mix different bridges, where models are independently pre-trained for each specific objective (Divide) and then integrated into a unified model (Merge). In addition, a Weight Reallocation Scheme (WRS) is also designed to enhance the stealthiness of MixBridge. Empirical studies across diverse generation tasks speak to the efficacy of MixBridge. The code is available at: https://github.com/qsx830/MixBridge.

Advancements in generative models have promoted text- and image-based multi-context image generation. Brain signals, offering a direct representation of user intent, present new opportunities for image customization. However, it faces challenges in brain interpretation, cross-modal context fusion and retention. In this paper, we present MindCustomer to explore the blending of visual brain signals in multi-context image generation. We first design shared neural data augmentation for stable cross-subject brain embedding by introducing the Image-Brain Translator (IBT) to generate brain responses from visual images. Then, we propose an effective cross-modal information fusion pipeline that mask-freely adapts distinct semantics from image and brain contexts within a diffusion model. It resolves semantic conflicts for context preservation and enables harmonious context integration. During the fusion pipeline, we further utilize the IBT to transfer image context to the brain representation to mitigate the cross-modal disparity. MindCustomer enables cross-subject generation, delivering unified, high-quality, and natural image outputs. Moreover, it exhibits strong generalization for new subjects via few-shot learning, indicating the potential for practical application. As the first work for multi-context blending with brain signal, MindCustomer lays a foundational exploration and inspiration for future brain-controlled generative technologies.

This paper presents CPL-IQA, a novel semi-supervised blind image quality assessment (BIQA) framework for authentic distortion scenarios. To address the challenge of limited labeled data in IQA area, our approach leverages confidence-quantifiable pseudo-label learning to effectively utilize unlabeled authentically distorted images. The framework operates through a preprocessing stage and two training phases: first converting MOS labels to vector labels via entropy minimization, followed by an iterative process that alternates between model training and label optimization. The key innovations of CPL-IQA include a manifold assumption-based label optimization strategy and a confidence learning method for pseudo-labels, which enhance reliability and mitigate outlier effects. Experimental results demonstrate the framework's superior performance on real-world distorted image datasets, offering a more standardized semi-supervised learning paradigm without requiring additional supervision or network complexity.