ICLR 2026 Papers

We present a systematic study of medical-domain interpretability in Large Language Models (LLMs). We study how the LLMs both represent and process medical knowledge through four different interpretability techniques: (1) UMAP projections of intermediate activations, (2) gradient-based saliency with respect to the model weights, (3) layer lesioning/removal and (4) activation patching. We present knowledge maps of five LLMs which show, at a coarse-resolution, where knowledge about patient's ages, medical symptoms, diseases and drugs is stored in the models. In particular for Llama3.3-70B, we find that most medical knowledge is processed in the first half of the model's layers. In addition, we find several interesting phenomena: (i) age is often encoded in a non-linear and sometimes discontinuous manner at intermediate layers in the models, (ii) the disease progression representation is non-monotonic and circular at certain layers of the model, (iii) in Llama, drugs cluster better by medical specialty rather than mechanism of action, especially for Llama and (iv) Gemma-27B and MedGemma-27B have activations that collapse at intermediate layers but recover by the final layers. These results can guide future research on fine-tuning, un-learning or de-biasing LLMs for medical tasks by suggesting at which layers in the model these techniques should be applied. We attached our source code to the paper for reproducibility.

Memory-efficient training of deep neural networks has become increasingly important as models grow larger while deployment environments impose strict resource constraints. We propose TraDy, a novel transfer learning scheme leveraging two key insights: layer importance for updates is architecture-dependent and determinable a priori, while dynamic stochastic channel selection provides superior gradient approximation compared to static approaches. We introduce a dynamic channel selection approach that stochastically resamples channels between epochs within preselected layers. Extensive experiments demonstrate TraDy achieves state-of-the-art performance across various downstream tasks and architectures while maintaining strict memory constraints, achieving up to 99\% activation sparsity, 95\% weight derivative sparsity, and 97\% reduction in FLOPs for weight derivative computation.

Bayesian optimization (BO) has been widely used to optimize expensive and black-box functions across various domains. Existing BO methods have not addressed tensor-output functions. To fill this gap, we propose a novel tensor-output BO method. Specifically, we first introduce a tensor-output Gaussian process (TOGP) with two classes of tensor-output kernels as a surrogate model of the tensor-output function, which can effectively capture the structural dependencies within the tensor. Based on it, we develop an upper confidence bound (UCB) acquisition function to select the queried points. Furthermore, we introduce a more complex and practical problem setting, named combinatorial bandit Bayesian optimization (CBBO), where only a subset of the outputs can be selected to contribute to the objective function. To tackle this, we propose a tensor-output CBBO method, which extends TOGP to handle partially observed outputs, and accordingly design a novel CMAB-UCB2 criterion to sequentially select both the queried points and the optimal output subset. Theoretical regret bounds for the two methods are established, ensuring their sublinear performance. Extensive synthetic and real-world experiments demonstrate their superiority.

Deep neural networks (DNNs) have been effectively deployed in tabular data modeling for various applications. However, these models suffer severe performance degradation when distribution shifts exist between training and test tabular data. While test-time adaptation (TTA) serves as a promising solution to distribution shifts, existing TTA methods primarily focus on visual modalities and demonstrate poor adaptation when directly applied to tabular modality. Recent efforts have proposed tabular-specific TTA approaches to mitigate distribution shifts on tabular data. Nevertheless, these methods inherently assume the accessibility of source domain or prior and fail to fundamentally address feature shift while overlooking unique characteristics of tabular data, leading to suboptimal adaptation. In this paper, we focus on the problem of \textit{prior-free tabular test-time adaptation} where no access to source data and any prior knowledge is allowed, and we propose a novel method, \underline{P}rior-\underline{F}ree \underline{T}abular \underline{T}est-\underline{T}ime \underline{A}daptation (PFT$_3$A), which has three designs to simultaneously address label shift and feature shift without source domain or prior access. Specially, PFT$_3$A contains the \textit{Class Prior Estimating} module for estimating source-target class priors to calibrate prediction, eliminating dependency on source class prior and mitigating label shift; the \textit{Robust Feature Learning} module for learning robust feature by aligning source-like and target-like features to mitigate feature shift; the \textit{Representative Subspace Exploration} module for eliminating redundant features by projecting feature into subspace to enhance feature alignment. Extensive experiments demonstrate the effectiveness and generalization of PFT$_3$A in tabular TTA tasks. The implementation is at \url{https://anonymous.4open.science/r/PFT3A/README.md}.

Universal grasping with multi-fingered dexterous hands is a fundamental challenge in robotic manipulation. While recent approaches successfully learn closed-loop grasping policies using reinforcement learning (RL), the inherent difficulty of high-dimensional, long-horizon exploration necessitates complex reward and curriculum design, often resulting in suboptimal solutions across diverse objects. We propose DemoGrasp, a simple yet effective method for learning universal dexterous grasping. We start from a single successful demonstration trajectory of grasping a specific object and adapt to novel objects and poses by editing the robot actions in this trajectory: changing the wrist pose determines where to grasp, and changing the hand joint angles determines how to grasp. We formulate this trajectory editing as a single-step Markov Decision Process (MDP) and use RL to optimize a universal policy across hundreds of objects in parallel in simulation, with a simple reward consisting of a binary success term and a robot–table collision penalty. In simulation, DemoGrasp achieves a 95% success rate on DexGraspNet objects using the Shadow Hand, outperforming previous state-of-the-art methods. It also shows strong transferability, achieving an average success rate of 84.6% across diverse dexterous hand embodiments on six unseen object datasets, while being trained on only 175 objects. Through vision-based imitation learning, our policy successfully grasps 110 unseen real-world objects, including small, thin items. It generalizes to spatial, background, and lighting changes, supports both RGB and depth inputs, and extends to language-guided grasping in cluttered scenes.

We consider submodular function minimization using a \textit{dueling oracle}, a noisy pairwise comparison oracle that provides relative feedback on function values between two queried sets. The oracle's responses are governed by a \textit{transfer function}, which characterizes the relationship between differences in function values and the parameters of the response distribution. For a \textit{linear} transfer function, we propose an algorithm that achieves an error rate of $O(n^{\frac{3}{2}}/\sqrt{T})$, where $n$ is the size of the ground set and $T$ denotes the number of oracle calls. We establish a lower bound: Under the constraint that differences between queried sets are bounded by a constant, any algorithm incurs an error of at least $\Omega(n^{\frac{3}{2}}/\sqrt{T})$. Without such a constraint, the lower bound becomes $\Omega(n/\sqrt{T})$. These results show that our algorithm is optimal up to constant factors for constrained algorithms. For a \textit{sigmoid} transfer function, we design an algorithm with an error rate of $O(n^{\frac{7}{5}}/T^{\frac{2}{5}})$, and establish lower bounds analogous to the linear case.

We propose an accelerated exact $k$-means algorithm, Angle $k$-means. As its name suggests, the algorithm mainly leverages angular relationships between data points and cluster centers to reduce computational overhead. Although grounded in straightforward geometric principles, it delivers substantial performance improvements in empirical evaluations. In contrast to existing acceleration techniques, our model introduces no new hyperparameters, preserving full compatibility with standard $k$-means. Theoretical analysis shows that Angle $k$-means maintains linear time complexity with respect to both sample size and dimensionality, while empirical evaluations on diverse real-world datasets demonstrate significant speedup over state-of-the-art algorithms such as ball $k$-means and Exp-ns.

Reinforcement learning (RL) research requires diverse, challenging environments that are both tractable and scalable. While modern video games may offer rich dynamics, they are computationally expensive and poorly suited for large-scale experimentation due to their CPU-bound execution. We introduce Octax, a high-performance suite of classic arcade game environments implemented in JAX, based on CHIP-8 emulation, a predecessor to Atari, which is widely adopted as a benchmark in RL research. Octax provides the JAX community with a long-awaited end-to-end GPU alternative to Atari games, offering image-based environments, spanning puzzle, action, and strategy genres, all executable at massive scale on modern GPUs. Our JAX-based implementation achieves orders-of-magnitude speedups over traditional CPU emulators. We demonstrate Octax's capabilities by training RL agents across multiple games, showing significant improvements in training speed and scalability compared to existing solutions. The environment's modular design enables researchers to easily extend the suite with new games or generate novel environments using large language models, making it an ideal platform for large-scale RL experimentation.

As large language models (LLMs) are increasingly deployed in the real world, the ability to ``unlearn'', or remove specific pieces of knowledge post hoc, has become essential for a variety of reasons ranging from privacy regulations to correcting outdated or harmful content. Prior work has proposed unlearning benchmarks and algorithms, and has typically assumed that the training process and the target model are fixed. In this work, we empirically investigate how learning-time encoding in knowledge encoding impact the effectiveness of unlearning factual knowledge. We conduct two studies: (i) examining how paraphrased descriptions influence unlearning performance, and (ii) analyzing unlearning when multiple facts are embedded within the same training text chunk. Our empirical study reveals two important implications: a new perspective for interpreting unlearning performance and practical strategies for improving LLM unlearning.

Recent work increasingly adopts AI for Science (AI4S) models to replace expensive PDE solvers as simulation environments for reinforcement learning (RL), enabling faster training in complex physical control tasks. However, using approximate simulators introduces modeling errors that affect the learned policy. In this paper, we introduce a unified $\varepsilon$-$N$ framework that quantifies the minimal computational cost $N^*(\varepsilon)$ required for an AI4S model to ensure that tabular RL can estimate the value function with unbiasedness, with probability at least $1 - \delta$. This characterization allows us to connect surrogate accuracy, grid resolution, and RL policy quality under a shared probabilistic language. We analyze how the discretization level $K$ of AI4S and RL space governs both PDE surrogate error and RL lattice approximation error, and we employ spectral theory and Sobolev estimates to derive optimal grid strategies that minimize total cost while preserving learning fidelity. Our theory reveals that different systems, such as ODE- and PDE-governed environments, require different allocations of effort between physical simulation and RL optimization. Overall, our framework offers a principled foundation for designing efficient, scalable, and cost-aware AI4S-RL systems with provable learning guarantees.

Distribution shift poses a significant challenge in machine learning, particularly in biomedical applications using data collected across different subjects, institutions, and recording devices, such as sleep data. While existing normalization layers, BatchNorm, LayerNorm and InstanceNorm, help mitigate distribution shifts, when applied over the time dimension they ignore the dependencies and auto-correlation inherent to the vector coefficients they normalize. In this paper, we propose PSDNorm that leverages Monge mapping and temporal context to normalize feature maps in deep learning models for signals. Evaluations with architectures based on U-Net or transformer backbones trained on 10K subjects across 10 datasets, show that PSDNorm achieves state-of-the-art performance on unseen left-out datasets while being 4-times more data-efficient than BatchNorm.

We characterize the effectiveness of Sharpness-aware minimization (SAM) under machine unlearning scheme, where unlearning forget signals interferes with learning retain signals. While previous work prove that SAM improves generalization with noise memorization prevention, we show that SAM abandons such denoising property when fitting the forget set, leading to altered generalization depending on signal strength. We further characterize the signal surplus of SAM in the order of signal strength, which enables learning from less retain signals to maintain model performance and putting more weight on unlearning the forget set. Empirical studies show that SAM outperforms SGD with relaxed requirement for retain signals and can enhance various unlearning methods either as pretrain or unlearn algorithm. Motivated by our refined characterization of SAM unlearning and observing that overfitting can benefit more stringent sample-specific unlearning, we propose Sharp MinMax, which splits the model into two to learn retain signals with SAM and unlearn forget signals with sharpness maximization, achieving best performance. Extensive experiments show that SAM enhances unlearning across varying difficulties measured by memorization, yielding decreased feature entanglement between retain and forget sets, stronger resistance to membership inference attacks, and a flatter loss landscape. Our observations generalize to more noised data, different optimizers, and different architectures.

Spatio-temporal point process (STPP) data appear in many domains. A natural way to model them is to describe how the instantaneous event rate varies over space and time given the observed history which enables interpretation, interaction detection, and forecasting. Traditional parametric kernel-based models, while historically dominant, struggle to capture complex nonlinear patterns. In contrast, deep learning methods leverage the representational power of neural networks to aggregate historical events and integrate spatio-temporal point processes. However, existing deep learning methods often process space and time independently, overlooking the spatio-temporal dependencies. To address this limitation, we propose a novel method called Topology-ENhanced Diffusion Model (TEN-DM), including two key components namely spatio-temporal graph construction and multimodal topological feature representation learning. Further, we use temporal query technique to effectively capture periodic temporal patterns for learning effective temporal representations. Extensive experiments show the effectiveness of TEN-DM on multiple STPP datasets compared to state-of-the-art methods.

Body-surface health conditions, spanning diverse clinical departments, represent some of the most frequent diagnostic scenarios and a primary target for medical multimodal large language models (MLLMs). Yet existing medical benchmarks are either built from publicly available sources with limited expert curation or focus narrowly on disease classification, failing to reflect the stepwise recognition and reasoning processes physicians follow in real practice. To address this gap, we introduce MedLesionVQA, the first benchmark explicitly designed to evaluate MLLMs on the visual diagnostic workflow for body-surface conditions in large scale. All questions are derived from authentic clinical visual diagnosis scenarios and verified by medical experts with over 20 years of experience, while the data are drawn from 10k+ real patient visits, ensuring authenticity, clinical reality and diversity. MedLesionVQA consists of 12K in-house volunteer images (never publicly leaked) and 19K expert-verified question–answer pairs, with fine-grained annotations of 94 lesion types, 110 body regions, and 96 diseases. We evaluate 20+ state-of-the-art MLLMs against human physicians: the best model reaches 56.2% accuracy, far below primary physicians (61.4%) and senior specialists (73.2%). These results expose the persistent gap between MLLMs and clinical expertise, underscoring the need for the multimodal benchmarks to drive trustworthy medical AI.

Dynamical systems modeling is a core pillar of scientific inquiry across natural and life sciences. Increasingly, dynamical system models are learned from data, rendering identifiability a paramount concept. For systems that are not identifiable from data, no guarantees can be given about their behavior under new conditions and inputs, or about possible control mechanisms to steer the system. It is known in the community that "linear ordinary differential equations (ODE) are almost surely identifiable from a single trajectory." However, this only holds for dense matrices. The sparse regime remains underexplored, despite its practical relevance with sparsity arising naturally in many biological, social, and physical systems. In this work, we address this gap by characterizing the identifiability of sparse linear ODEs. Contrary to the dense case, we show that sparse systems are unidentifiable with a positive probability in practically relevant sparsity regimes and provide lower bounds for this probability. We further study empirically how this theoretical unidentifiability manifests in state-of-the-art methods to estimate linear ODEs from data. Our results corroborate that sparse systems are also practically unidentifiable. Theoretical limitations are not resolved through inductive biases or optimization dynamics. Our findings call for rethinking what can be expected from data-driven dynamical system modeling and allows for quantitative assessments of how much to trust a learned linear ODE.

Machine learning (ML) has shown promise for tackling combinatorial optimization (CO), but much of the reported progress relies on small-scale, synthetic benchmarks that fail to capture real-world structure and scale. A core limitation is that ML methods are typically trained and evaluated on synthetic instance generators, leaving open how they perform on irregular, competition-grade, or industrial datasets. We present FrontierCO, a benchmark for evaluating ML-based CO solvers under real-world structure and extreme scale. FrontierCO spans eight CO problems, including routing, scheduling, facility location, and graph problems, with instances drawn from competitions and public repositories (e.g., DIMACS, TSPLib). Each task provides both easy sets (historically challenging but now solvable) and hard sets (open or computationally intensive), alongside standardized training/validation resources. Using FrontierCO, we evaluate 16 representative ML solvers---graph neural approaches, hybrid neural–symbolic methods, and LLM-based agents---against state-of-the-art classical solvers. We find a persistent performance gap that widens under structurally challenging and large instance sizes (e.g., TSP up to 10M nodes; MIS up to 8M), while also identifying cases where ML methods outperform classical solvers. By centering evaluation on real-world structure and orders-of-magnitude larger instances, FrontierCO provides a rigorous basis for advancing ML for CO.

Knowledge editing (KE) is crucial for making precise modifications to factual knowledge within large language models (LLMs). Existing KE methods, however, are primarily designed for dense architectures, limiting their applicability to the increasingly popular sparse Mixture-of-Experts (MoE) models that power modern scalable LLMs. While MoEs offer remarkable efficiency and capacity scaling, their unique structure introduces new challenges for KE. Naively adapting dense-model editors to MoEs is not only computationally expensive but also induces routing distribution shifts that degrade model stability and consistency. To address these challenges, we introduce MoEEdit, the first systematic framework for routing-stable knowledge editing in MoE LLMs. Our approach reparameterizes expert updates through per-expert null-space projections, ensuring router inputs remain invariant to suppress these shifts, and solves the resulting block-structured optimization with an efficient block coordinate descent (BCD) solver. Experiments demonstrate that MoEEdit achieves state-of-the-art efficacy and generalization, while maintaining high specificity, routing stability, and superior computational and memory efficiency. Our work establishes a robust foundation for scalable and precise knowledge editing in modern sparse LLMs by highlighting the necessity of routing-stable interventions.

Reinforcement learning has emerged as a popular method for post-training large language models (LLMs). While improving the model's performance on downstream tasks, it often reduces the model's output diversity, leading to narrow, canonical responses. Existing methods to enhance diversity are limited, either by operating at inference time or by focusing on lexical differences. We propose a novel training method named DQO (Diversity Quality Optimization) based on determinantal point processes (DPPs) to jointly optimize LLMs for quality and semantic diversity. Our approach samples and embeds a group of responses for each prompt, then uses the determinant of a kernel-based similarity matrix to measure diversity as the volume spanned by the embeddings of these responses. Experiments across instruction-following, summarization, story generation, and reasoning tasks demonstrate that our method substantially improves semantic diversity without sacrificing model quality.

Featurizing odorants to enable robust prediction of their properties is difficult due to the complex activation patterns that odorants evoke in the olfactory system. Structurally similar odorants can elicit distinct activation patterns in both the sensory periphery (i.e., at the receptor level) and downstream brain circuits (i.e., at a perceptual level). Despite efforts to design or discover features for odorants to better predict how they activate the olfactory system, we lack a universally accepted way to featurize odorants. In this work, we demonstrate that feature-based approaches that rely on pre-trained foundation models $\textit{do not}$ significantly outperform classical hand-designed features, but that targeted foundation model fine-turning can increase model performance beyond these limits. To show this, we introduce a new model that creates olfaction-specific representations: $\textbf{L}$oRA-based $\textbf{O}$dorant-$\textbf{R}$eceptor $\textbf{A}$ffinity prediction with $\textbf{CROSS}$-attention ($\textbf{LORAX}$). We compare existing chemical foundation model representations to hand-designed physicochemical descriptors using feature-based methods and identify large information overlap between these representations, highlighting the necessity of fine-tuning to generate novel and superior odorant representations. We show that LORAX produces a feature space more closely aligned with olfactory neural representation, enabling it to outperform existing models on predictive tasks.