ICML 2025 Papers

While the activations of neurons in deep neural networks usually do not have a simple human-understandable interpretation, sparse autoencoders (SAEs) can be used to transform these activations into a higher-dimensional latent space which can be more easily interpretable. However, SAEs can have millions of distinct latents, making it infeasible for humans to manually interpret each one. In this work, we build an open-source automated pipeline to generate and evaluate natural language interpretations for SAE latents using LLMs. We test our framework on SAEs of varying sizes, activation functions, and losses, trained on two different open-weight LLMs. We introduce five new techniques to score the quality of interpretations that are cheaper to run than the previous state of the art. One of these techniques, intervention scoring, evaluates the interpretability of the effects of intervening on a latent, which we find explains latents that are not recalled by existing methods. We propose guidelines for generating better interpretations that remain valid for a broader set of activating contexts, and discuss pitfalls with existing scoring techniques. Our code is available at https://github.com/EleutherAI/delphi.

Submodular functions, defined on continuous or discrete domains, arise in numerous applications. We study the minimization of the difference of submodular (DS) functions, over both domains, extending prior work restricted to set functions. We show that all functions on discrete domains and all smooth functions on continuous domains are DS. For discrete domains, we observe that DS minimization is equivalent to minimizing the difference of two convex (DC) functions, as in the set function case. We propose a novel variant of the DC Algorithm (DCA) and apply it to the resulting DC Program, obtaining comparable theoretical guarantees as in the set function case. The algorithm can be applied to continuous domains via discretization. Experiments demonstrate that our method outperforms baselines in integer compressive sensing and integer least squares.

Autoregressive models (ARMs) have become the workhorse for sequence generation tasks, since many problems can be modeled as next-token prediction. While there appears to be a natural ordering for text (i.e., left-to-right), for many data types, such as graphs, the canonical ordering is less obvious. To address this problem, we introduce a variant of ARM that generates high-dimensional data using a probabilistic ordering that is sequentially inferred from data. This model incorporates a trainable probability distribution, referred to as an order-policy, that dynamically decides the autoregressive order in a state-dependent manner. To train the model, we introduce a variational lower bound on the exact log-likelihood, which we optimize with stochastic gradient estimation. We demonstrate experimentally that our method can learn meaningful autoregressive orderings in image and graph generation. On the challenging domain of molecular graph generation, we achieve state-of-the-art results on the QM9 and ZINC250k benchmarks, evaluated using the Fréchet ChemNet Distance (FCD), Synthetic Accessibility Score (SAS), Quantitative Estimate of Drug-likeness (QED).

Optimizing high-dimensional and complex black-box functions is crucial in numerous scientific applications. While Bayesian optimization (BO) is a powerful method for sample-efficient optimization, it struggles with the curse of dimensionality and scaling to thousands of evaluations. Recently, leveraging generative models to solve black-box optimization problems has emerged as a promising framework. However, those methods often underperform compared to BO methods due to limited expressivity and difficulty of uncertainty estimation in high-dimensional spaces. To overcome these issues, we introduce \textbf{DiBO}, a novel framework for solving high-dimensional black-box optimization problems. Our method iterates two stages. First, we train a diffusion model to capture the data distribution and deep ensembles to predict function values with uncertainty quantification. Second, we cast the candidate selection as a posterior inference problem to balance exploration and exploitation in high-dimensional spaces. Concretely, we fine-tune diffusion models to amortize posterior inference. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines across synthetic and real-world tasks. Our code is publicly available \href{https://github.com/umkiyoung/DiBO}{here}.

Sparse tensor programs are essential in deep learning and graph analytics, driving the need for optimized processing. To meet this demand, specialized hardware accelerators are being developed. Optimizing these programs for accelerators is challenging for two reasons: program performance is highly sensitive to variations in sparse inputs, and early-stage accelerators rely on expensive simulators. Therefore, ML-based cost models used for optimizing such programs on general-purpose hardware are often ineffective for early-stage accelerators, as they require large datasets for proper training. To this end, we introduce COGNATE, a novel framework that leverages inexpensive data samples from general-purpose hardware (e.g., CPUs) to train cost models, followed by few-shot fine-tuning on emerging hardware. COGNATE exploits the homogeneity of input features across hardware platforms while effectively mitigating heterogeneity, enabling cost model training with just 5% of the data samples needed by accelerator-specific models to achieve comparable performance. We conduct extensive experiments to demonstrate that COGNATE outperforms existing techniques, achieving average speedups of 1.47× (up to 5.46×) for SpMM and 1.39× (up to 4.22×) for SDDMM.

Federated learning (FL), aimed at leveraging vast distributed datasets, confronts a crucial challenge: the heterogeneity of data across different silos. While previous studies have explored discrete representations to enhance model generalization across minor distributional shifts, these approaches often struggle to adapt to new data silos with significantly divergent distributions. In response, we have identified that models derived from FL exhibit markedly increased uncertainty when applied to data silos with unfamiliar distributions. Consequently, we propose an innovative yet straightforward iterative framework, termed \emph{Uncertainty-Based Extensible-Codebook Federated Learning (UEFL)}. This framework dynamically maps latent features to trainable discrete vectors, assesses the uncertainty, and specifically extends the discretization dictionary or codebook for silos exhibiting high uncertainty. Our approach aims to simultaneously enhance accuracy and reduce uncertainty by explicitly addressing the diversity of data distributions, all while maintaining minimal computational overhead in environments characterized by heterogeneous data silos. Extensive experiments across multiple datasets demonstrate that UEFL outperforms state-of-the-art methods, achieving significant improvements in accuracy (by 3\%--22.1\%) and uncertainty reduction (by 38.83\%--96.24\%). The source code is available at https://github.com/destiny301/uefl.

Aligning large language models (LLMs) with human preferences requires balancing policy optimization with computational stability. While recent offline methods like DPO and SimPO bypass reinforcement learning’s complexity, they face critical limitations: DPO relies on static reference models that degrade with policy updates, and SimPO assumes a uniform target reward margin that ignores instance-wise preference strength. We propose AlphaDPO, an adaptive preference optimization framework that dynamically reparameterizes the reference distribution to address these issues. Our key innovation lies in an implicit reference model \(\hat{\pi}_{\text{ref}} \propto U(y|x)(\pi_\theta/\pi_{\text{ref}})^\alpha\), which interpolates between policy-driven specialization and uniform exploration while enabling instance-adaptive reward margins. Theoretically, we prove AlphaDPO implicitly controls sequential KL divergence between iterative policy updates, ensuring stability even with poorly calibrated reference models. Empirically, AlphaDPO achieves state-of-the-art performance on AlpacaEval 2 (58.7\% LC win rate) and Arena-Hard (35.7\% win rate) across Mistral2-7B, Llama3-8B, and Gemma2-9B, demonstrating robust alignment without multi-stage training. Our work establishes adaptive reference reparameterization as a principled mechanism for preference optimization.

Benchmark contamination has become a significant concern in the LLM evaluation community. Previous Agents-as-an-Evaluator address this issue by involving agents in the generation of questions. Despite their success, the biases in Agents-as-an-Evaluator methods remain largely unexplored. In this paper, we present a theoretical formulation of evaluation bias, providing valuable insights into designing unbiased evaluation protocols. Furthermore, we identify two type of bias in Agents-as-an-Evaluator through carefully designed probing tasks on a minimal Agents-as-an-Evaluator setup. To address these issues, we propose the Unbiased Evaluator, an evaluation protocol that delivers a more comprehensive, unbiased, and interpretable assessment of LLMs. Extensive experiments reveal significant room for improvement in current LLMs. Additionally, we demonstrate that the Unbiased Evaluator not only offers strong evidence of benchmark contamination but also provides interpretable evaluation results.

Self-Supervised Learning (SSL) powers many current AI systems. As research interest and investment grow, the SSL design space continues to expand. The Platonic view of SSL, following the Platonic Representation Hypothesis (PRH), suggests that despite different methods and engineering approaches, all representations converge to the same Platonic ideal. However, this phenomenon lacks precise theoretical explanation. By synthesizing evidence from Identifiability Theory (IT), we show that the PRH can emerge in SSL. There is a gap between SSL theory and practice: Current IT cannot explain SSL's empirical success, though it has practically relevant insights. Our work formulates a blueprint for SSL research to bridge this gap: we propose expanding IT into what we term Singular Identifiability Theory (SITh), a broader theoretical framework encompassing the entire SSL pipeline. SITh would allow deeper insights into the implicit data assumptions in SSL and advance the field towards learning more interpretable and generalizable representations. We highlight three critical directions for future research: 1) training dynamics and convergence properties of SSL; 2) the impact of finite samples, batch size, and data diversity; and 3) the role of inductive biases in architecture, augmentations, initialization schemes, and optimizers.

Cyclic peptides offer inherent advantages in pharmaceuticals. For example, cyclic peptides are more resistant to enzymatic hydrolysis compared to linear peptides and usually exhibit excellent stability and affinity. Although deep generative models have achieved great success in linear peptide design, several challenges prevent the development of computational methods for designing diverse types of cyclic peptides. These challenges include the scarcity of 3D structural data on target proteins and associated cyclic peptide ligands, the geometric constraints that cyclization imposes, and the involvement of non-canonical amino acids in cyclization. To address the above challenges, we introduce CpSDE, which consists of two key components: AtomSDE, a generative structure prediction model based on harmonic SDE, and ResRouter, a residue type predictor. Utilizing a routed sampling algorithm that alternates between these two models to iteratively update sequences and structures, CpSDE facilitates the generation of cyclic peptides. By employing explicit all-atom and bond modeling, CpSDE overcomes existing data limitations and is proficient in designing a wide variety of cyclic peptides. Our experimental results demonstrate that the cyclic peptides designed by our method exhibit reliable stability and affinity.

Despite the remarkable success of diffusion models (DMs) in data generation, they exhibit specific failure cases with unsatisfactory outputs. We focus on one such limitation: the ability of DMs to learn hidden rules between image features. Specifically, for image data with dependent features ($\mathbf{x}$) and ($\mathbf{y}$) (e.g., the height of the sun ($\mathbf{x}$) and the length of the shadow ($\mathbf{y}$)), we investigate whether DMs can accurately capture the inter-feature rule ($p(\mathbf{y}|\mathbf{x})$). Empirical evaluations on mainstream DMs (e.g., Stable Diffusion 3.5) reveal consistent failures, such as inconsistent lighting-shadow relationships and mismatched object-mirror reflections. Inspired by these findings, we design four synthetic tasks with strongly correlated features to assess DMs' rule-learning abilities. Extensive experiments show that while DMs can identify coarse-grained rules, they struggle with fine-grained ones. Our theoretical analysis demonstrates that DMs trained via denoising score matching (DSM) exhibit constant errors in learning hidden rules, as the DSM objective is not compatible with rule conformity. To mitigate this, we introduce a common technique - incorporating additional classifier guidance during sampling, which achieves (limited) improvements. Our analysis reveals that the subtle signals of fine-grained rules are challenging for the classifier to capture, providing insights for future exploration.

In this work, we explore a cost-effective framework for multilingual image generation. We find that, unlike models tuned on high-quality images with multilingual annotations, leveraging text encoders pre-trained on widely available, noisy Internet image-text pairs significantly enhances data efficiency in text-to-image (T2I) generation across multiple languages. Based on this insight, we introduce MuLan, Multi-Language adapter, a lightweight language adapter with fewer than 20M parameters, trained alongside a frozen text encoder and image diffusion model. Compared to previous multilingual T2I models, this framework offers: (1) Cost efficiency. Using readily accessible English data and off-the-shelf multilingual text encoders minimizes the training cost; (2) High performance. Achieving comparable generation capabilities in over 110 languages with CLIP similarity scores nearly matching those in English (39.57 for English vs. 39.61 for other languages); and (3) Broad applicability. Seamlessly integrating with compatible community tools like LoRA, LCM, ControlNet, and IP-Adapter, expanding its potential use cases.

Harmful fine-tuning (HFT), performed directly on open-source LLMs or through Fine-tuning-as-a-Service, breaks safety alignment and poses significant threats. Existing methods aim to mitigate HFT risks by learning robust representation on alignment data or making harmful data unlearnable, but they treat each data sample equally, leaving data vulnerability patterns understudied. In this work, we reveal that certain subsets of alignment data are consistently more prone to forgetting during HFT across different fine-tuning tasks and exhibit lower robustness compared to other subsets. Inspired by these findings, we propose Vulnerability-Aware Alignment (VAA), which calculates data vulnerability, partitions data into "vulnerable" and "invulnerable" groups, and encourages balanced learning using a group distributionally robust optimization (Group DRO) framework. Specifically, VAA learns an adversarial sampler that samples examples from the currently underperforming group and then applies group-dependent adversarial perturbations to the data during training, aiming to encourage a balanced learning process across groups. Experiments across four fine-tuning tasks demonstrate that VAA significantly reduces harmful scores while preserving downstream task performance, outperforming state-of-the-art baselines.

Adversarial Training (AT) impacts different architectures in distinct ways: vision models gain robustness but face reduced generalization, encoder-based models exhibit limited robustness improvements with minimal generalization loss, and recent work in latent-space adversarial training demonstrates that decoder-based models achieve improved robustness by applying AT across multiple layers. We provide the first explanation for these trends by leveraging the manifold conjecture: off-manifold adversarial examples (AEs) enhance robustness, while on-manifold AEs improve generalization. We show that vision and decoder-based models exhibit low intrinsic dimensionality in earlier layers (favoring off-manifold AEs), whereas encoder-based models do so in later layers (favoring on-manifold AEs). Exploiting this property, we introduce SMAAT, which improves the scalability of AT for encoder-based models by perturbing the layer with the lowest intrinsic dimensionality. This reduces the projected gradient descent (PGD) chain length required for AE generation, cutting GPU time by 25–33% while significantly boosting robustness. We validate SMAAT across multiple tasks, including text generation, sentiment classification, safety filtering, and retrieval augmented generation setups, demonstrating superior robustness with comparable generalization to standard training.

We address differentially private stochastic bandit problems by leveraging Thompson Sampling with Gaussian priors and Gaussian differential privacy (GDP). We propose DP-TS-UCB, a novel parametrized private algorithm that enables trading off privacy and regret. DP-TS-UCB satisfies $ \tilde{O} \left(T^{0.25(1-\alpha)}\right)$-GDP and achieves $O \left(K\ln^{\alpha+1}(T)/\Delta \right)$ regret bounds, where $K$ is the number of arms, $ \Delta$ is the sub-optimality gap, $T$ is the learning horizon, and $\alpha \in [0,1]$ controls the trade-off between privacy and regret. Theoretically, DP-TS-UCB relies on anti-concentration bounds for the Gaussian distributions, linking the exploration mechanisms of Thompson Sampling and Upper Confidence Bound, which may be of independent research interest.

The success of large language models (LLMs) in text generation has inspired their application to image generation. However, existing methods either rely on specialized designs with inductive biases or adopt LLMs without fully exploring their potential in vision tasks. In this work, we systematically investigate the design space of LLMs for image generation and demonstrate that LLMs can achieve near state-of-the-art performance without domain-specific designs, simply by making proper choices in tokenization methods, modeling approaches, scan patterns, vocabulary design, and sampling strategies. We further analyze autoregressive models' learning and scaling behavior, revealing how larger models effectively capture more useful information than the smaller ones. Additionally, we explore the inherent differences between text and image modalities, highlighting the potential of LLMs across domains. The exploration provides valuable insights to inspire more effective designs when applying LLMs to other domains. With extensive experiments, our proposed model, **ELM** achieves an FID of 1.54 on 256$\times$256 ImageNet and an FID of 3.29 on 512$\times$512 ImageNet, demonstrating the powerful generative potential of LLMs in vision tasks.

While Multi-modal Large Language Models (MLLMs) demonstrate impressive abilities over high-level perception and reasoning, their robustness in the wild remains limited, often falling short on tasks that are intuitive and effortless for humans. We examine the hypothesis that these deficiencies stem from the absence of core knowledge—rudimentary cognitive abilities innate to humans from early childhood. To explore the core knowledge representation in MLLMs, we introduce CoreCognition, a large-scale benchmark encompassing 12 core knowledge concepts grounded in developmental cognitive science. We evaluate 230 models with 11 different prompts, leading to a total of 2,530 data points for analysis. Our experiments uncover four key findings, collectively demonstrating core knowledge deficits in MLLMs: they consistently underperform and show reduced, or even absent, scalability on low-level abilities relative to high-level ones. Finally, we propose Concept Hacking, a novel controlled evaluation method that reveals MLLMs fail to progress toward genuine core knowledge understanding, but instead rely on shortcut learning as they scale.

Uncertainty quantification (UQ) provides a resource-efficient solution for on-device monitoring of tinyML models deployed remotely without access to true labels. However, existing UQ methods impose significant memory and compute demands, making them impractical for ultra-low-power, KB-sized tinyML devices. Prior work has attempted to reduce overhead by using early-exit ensembles to quantify uncertainty in a single forward pass, but these approaches still carry prohibitive costs. To address this, we propose QUTE, a novel resource-efficient early-exit-assisted ensemble architecture optimized for tinyML models. QUTE introduces additional output blocks at the final exit of the base network, distilling early-exit knowledge into these blocks to form a diverse yet lightweight ensemble. We show that QUTE delivers superior uncertainty quality on tiny models, achieving comparable performance on larger models with 59% smaller model sizes than the closest prior work. When deployed on a microcontroller, QUTE demonstrates a 31% reduction in latency on average. In addition, we show that QUTE excels at detecting accuracy-drop events, outperforming all prior works.

As large language models (LLMs) become increasingly versatile, numerous large scale benchmarks have been developed to thoroughly assess their capabilities. These benchmarks typically consist of diverse datasets and prompts to evaluate different aspects of LLM performance. However, comprehensive evaluations on hundreds or thousands of prompts incur tremendous costs in terms of computation, money, and time. In this work, we investigate strategies to improve evaluation efficiency by selecting a subset of examples from each benchmark using a learned policy. Our approach models the dependencies across test examples, allowing accurate prediction of the evaluation outcomes for the remaining examples based on the outcomes of the selected ones. Consequently, we only need to acquire the actual evaluation outcomes for the selected subset. We rigorously explore various subset selection policies and introduce a novel RL-based policy that leverages the captured dependencies. Empirical results demonstrate that our approach significantly reduces the number of evaluation prompts required while maintaining accurate performance estimates compared to previous methods.