ICLR 2026 论文列表

A systematic, comparative investigation into the effects of low-quality data reveals a stark spectrum of robustness across modern probabilistic models. We find that autoregressive language models, from token prediction to sequence-to-sequence tasks, are remarkably resilient (for GPT-2, test NLL increases modestly from 2.87 to 3.59 despite 50\% token corruption). By contrast, under the same levels of data corruption, class-conditional diffusion models degrade catastrophically (image-label consistency plummets by 56.81\% relative to baseline), while classifiers show a moderate impact that diminishes with dataset scale. To explain these discrepancies, we analyze the results through a multi-perspective lens, integrating information theory, PAC learning, and gradient dynamics. \textcolor{blue}{These analyses suggest that robustness is heavily influenced by two key principles}: the \textbf{richness of conditioning information}, which constrains the learning problem, and the \textbf{absolute information content} of the training data, which allows the signal from correct information to dominate statistical noise.

Deep noise suppression (DNS) models enjoy widespread use throughout a variety of high-stakes speech applications. However, we show that four recent DNS models can each be reduced to outputting unintelligible gibberish through the addition of psychoacoustically hidden adversarial noise, even in low-background-noise and simulated over-the-air settings. For three of the models, a small transcription study with audio and multimedia experts confirms unintelligibility of the attacked audio; simultaneously, an ABX study shows that the adversarial noise is generally imperceptible, with some variance between participants and samples. While we also establish several negative results around targeted attacks and model transfer, our results nevertheless highlight the need for practical countermeasures before open-source DNS systems can be used in safety-critical applications.

Adversarial attacks have achieved widespread success in various domains, yet existing methods suffer from significant performance degradation when adversarial examples are subjected to even minor disturbances. In this paper, we propose a novel and robust attack called IGSA (**I**nverse **G**radient **S**ample-based **A**ttack), capable of generating adversarial examples that remain effective under diverse unknown disturbances. IGSA employs an iterative two-step framework: (i) inverse gradient sampling, which searches for the most disruptive direction within the neighborhood of adversarial examples, and (ii) disturbance-guided refinement, which updates adversarial examples via gradient descent along the identified disruptive disturbance. Theoretical analysis reveals that IGSA enhances robustness by increasing the likelihood of adversarial examples within the data distribution. Extensive experiments in both white-box and black-box attack scenarios demonstrate that IGSA significantly outperforms state-of-the-art attacks in terms of robustness against various unknown disturbances. Moreover, IGSA exhibits superior performance when attacking adversarially trained defense models. Code is available at https://github.com/nimingck/IGSA.

Large language models remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Defending against novel jailbreaks represents a critical challenge in AI safety. Adversarial training---designed to make models robust against worst-case perturbations---has been the dominant paradigm for adversarial robustness. However, due to optimization challenges and difficulties in defining realistic threat models, adversarial training methods often fail on newly developed jailbreaks in practice. This paper proposes a new paradigm for improving robustness against unseen jailbreaks, centered on the Adversarial Déjà Vu hypothesis: novel jailbreaks are not fundamentally new, but largely recombinations of adversarial skills from previous attacks. We study this hypothesis through a large-scale analysis of 32 attack papers published over two years. Using an automated pipeline, we extract and compress adversarial skills into a sparse dictionary of primitives, with LLMs generating human-readable descriptions. Our analysis reveals that unseen attacks can be effectively explained as sparse compositions of earlier skills, with explanatory power increasing monotonically as skill coverage grows. Guided by this insight, we introduce Adversarial Skill Compositional Training (ASCoT), which trains on diverse compositions of skill primitives rather than isolated attack instances. ASCoT substantially improves robustness to unseen attacks, including multi-turn jailbreaks, while maintaining low over-refusal rates. We also demonstrate that expanding adversarial skill coverage, not just data scale, is key to defending against novel attacks.

Deep neural networks often fail when deployed in real-world contexts due to distribution shift, a critical barrier to building safe and reliable systems. An emerging approach to address this problem relies on _disagreement discrepancy_—a measure of how the disagreement between two models changes under a shifting distribution. The process of maximizing this measure has seen applications in bounding error under shifts, testing for harmful shifts, and training more robust models. However, this optimization involves the non-differentiable zero-one loss, necessitating the use of practical surrogate losses. We prove that existing surrogates for disagreement discrepancy are not Bayes consistent, revealing a fundamental flaw: maximizing these surrogates can fail to maximize the true disagreement discrepancy. To address this, we introduce new theoretical results providing both upper and lower bounds on the optimality gap for such surrogates. Guided by this theory, we propose a novel disagreement loss that, when paired with cross-entropy, yields a provably consistent surrogate for disagreement discrepancy. Empirical evaluations across diverse benchmarks demonstrate that our method provides more accurate and robust estimates of disagreement discrepancy than existing approaches, particularly under challenging adversarial conditions.

Unit testing is a core practice in programming, enabling systematic evaluation of programs produced by human developers or large language models (LLMs). Given the challenges in writing comprehensive unit tests, LLMs have been employed to automate unit test generation, yet methods for training LLMs to produce high-quality unit tests remain underexplored. In this work, we propose UTRL, a novel reinforcement learning (RL) framework that trains an LLM to generate high-quality unit test given a programming instruction. Our key idea is to iteratively train two LLMs, the unit test generator and the code generator, in an adversarial manner via RL: (1) the unit test generator is trained to maximize a discrimination reward, encouraging it to produce tests that reveal faults in the code generator’s solutions; and (2) the code generator is trained to maximize a code reward, encouraging it to produce solutions that pass the unit tests generated by the unit test generator. In our experiment, we demonstrate that unit tests generated by Qwen3-4B trained via UTRL show higher quality compared to unit tests generated by the same model trained via supervised fine-tuning on ground-truth unit tests, yielding code evaluations that more closely align with those induced by the ground-truth tests. Moreover, Qwen3-4B trained with UTRL outperforms frontier models like GPT-4.1 and GPT-4o in generating high-quality unit tests, highlighting the effectiveness of UTRL in training LLMs for the unit test generation.

We propose $\textbf{FAME}$ (Formal Abstract Minimal Explanations), a new class of abductive explanations grounded in abstract interpretation. FAME is the first method to scale to large neural networks while reducing explanation size. Our main contribution is the design of dedicated perturbation domains that eliminate the need for traversal order. FAME progressively shrinks these domains and leverages LiRPA-based bounds to discard irrelevant features, ultimately converging to a $\textbf{formal abstract minimal explanation}$. To assess explanation quality, we introduce a procedure that measures the worst-case distance between an abstract minimal explanation and a true minimal explanation. This procedure combines adversarial attacks with an optional $VERI{\large X}+$ refinement step. We benchmark FAME against $VERI{\large X}+$ and demonstrate consistent gains in both explanation size and runtime on medium- to large-scale neural networks.

As large language models evolve, Machine Unlearning has emerged to address growing concerns around user privacy, copyright infringement, and overall safety. Yet state-of-the-art (SOTA) unlearning methods often suffer from catastrophic forgetting and metric imbalance, for example, by over-optimizing one objective (e.g., unlearning effectiveness, utility preservation, or privacy protection) at the expense of others. In addition, small perturbations in the representation or parameter space can be exploited by relearn and jailbreak attacks. To address these challenges, we propose PRISM, a unified framework that enforces dual-space smoothness in representation and parameter spaces to improve robustness and balance unlearning metrics. PRISM consists of two smoothness optimization stages: (i) a representation space stage that employs a robustly trained probe to defend against jailbreak attacks, and (ii) a parameter-space stage that decouples retain–forget gradient conflicts, reduces imbalance, and smooths the parameter space to mitigate relearning attacks. Extensive experiments on WMDP and MUSE, across conversational-dialogue and continuous-text settings, show that PRISM outperforms SOTA baselines under multiple attacks while achieving a better balance among key metrics.

Vision-Language Models (VLMs), such as CLIP, have shown strong zero-shot generalization but remain highly vulnerable to adversarial perturbations, posing serious risks in real-world applications. Test-time defenses for VLMs have recently emerged as a promising and efficient approach to defend against adversarial attacks without requiring costly large-scale retraining. In this work, we uncover a surprising phenomenon: under diverse input transformations, adversarial images in CLIP’s feature space consistently shift along a dominant direction, in contrast to the dispersed patterns of clean images. We hypothesize that this dominant shift, termed the Defense Direction, opposes the adversarial shift, pointing features back toward their correct class centers. Building on this insight, we propose Directional Bias-guided Defense (DBD), a test-time framework that estimates the Defense Direction and employs a DB-score–based two-stream reconstruction strategy to recover robust representations. Experiments on 15 datasets demonstrate that DBD not only achieves SOTA adversarial robustness while preserving clean accuracy, but also reveals the counterintuitive result that adversarial accuracy can even surpass clean accuracy. This demonstrates that adversarial perturbations inherently encode directional priors about the true decision boundary.

In this paper, we uniquely study the adversarial robustness of deep neural networks (NN) for classification tasks against that of optimal classifiers. We look at the smallest magnitude of possible additive perturbations that can change a classifier's output. We provide a matrix-theoretic explanation of the adversarial fragility of deep neural networks for classification. In particular, our theoretical results show that a neural network's adversarial robustness can degrade as the input dimension $d$ increases. Analytically, we show that neural networks' adversarial robustness can be only $1/\sqrt{d}$ of the best possible adversarial robustness of optimal classifiers. Our theories match remarkably well with numerical experiments of practically trained NN, including NN for ImageNet images. The matrix-theoretic explanation is consistent with an earlier information-theoretic feature-compression-based explanation for the adversarial fragility of neural networks.

Jailbreak transferability is the surprising phenomenon when an adversarial attack compromising one model also elicits harmful responses from other models. Despite widespread demonstrations, there is little consensus on why transfer is possible: is it a quirk of safety training, an artifact of model families, or a more fundamental property of representation learning? We present evidence that transferability emerges from shared representations rather than incidental flaws. Across 20 open-weight models and 33 jailbreak attacks, we find two factors that systematically shape transfer: (1) representational similarity under benign prompts, and (2) the strength of the jailbreak on the source model. To move beyond correlation, we show that deliberately increasing similarity through benign-only distillation causally increases transfer. Qualitative analysis reveal systematic patterns; for example, persona-style jailbreaks transfer far more often than cipher-based prompts, consistent with the idea that natural-language attacks exploit models’ shared representation space, whereas cipher-based attacks rely on idiosyncratic quirks that do not generalize. Together, these results reframe jailbreak transfer as a consequence of representation alignment rather than a fragile byproduct of safety training.

Influence estimation methods promise to explain and debug machine learning by estimating the impact of individual samples on the final model. Yet, existing methods collapse under training randomness: the same example may appear critical in one run and irrelevant in the next. Such instability undermines their use in data curation or cleanup since it is unclear if we indeed deleted/kept the correct datapoints. To overcome this, we introduce *f-influence* -- a new influence estimation framework grounded in hypothesis testing that explicitly accounts for training randomness, and establish desirable properties that make it suitable for reliable influence estimation. We also design a highly efficient algorithm *f*-*IN*fluence *E*stimation (**f-INE**) that computes f-influence in a **in a single training run**. Finally, we scale up f-INE to estimate influence of instruction tuning data on Llama 3.1 8B and show it can reliably detect poisoned samples that steer model opinions, demonstrating its utility for data cleanup and attributing model behavior.

Large language model (LLM) unlearning aims to surgically remove the influence of undesired data or knowledge from an existing model while preserving its utility on unrelated tasks. This paradigm has shown promise in addressing privacy and safety concerns. However, recent findings reveal that unlearning effects are often *fragile*: post-unlearning manipulations such as weight quantization or fine-tuning can quickly neutralize the intended forgetting. Prior efforts to improve robustness primarily reformulate unlearning objectives by explicitly assuming the role of vulnerability sources. In this work, we take a different perspective by investigating the role of the *optimizer*, independent of unlearning objectives and formulations, in shaping unlearning robustness. We show that the "*grade*" of the optimizer, defined by the level of information it exploits, ranging from zeroth-order (gradient-free) to first-order (gradient-based) to second-order (Hessian-based), is tightly linked to the resilience of unlearning. Surprisingly, we find that downgrading the optimizer, such as using zeroth-order methods or compressed-gradient variants (*e.g.,* gradient sign-based optimizers), often leads to stronger robustness. While these optimizers produce noisier and less precise updates, they encourage convergence to harder-to-disturb basins in the loss landscape, thereby resisting post-training perturbations. By connecting zeroth-order methods with randomized smoothing, we further highlight their natural advantage for robust unlearning. Motivated by these insights, we propose a *hybrid optimizer* that combines first-order and zeroth-order updates, preserving unlearning efficacy while enhancing robustness. Extensive experiments on the MUSE and WMDP benchmarks, across multiple LLM unlearning algorithms, validate that our approach achieves more resilient forgetting without sacrificing unlearning quality.

We propose a concept-based adversarial attack framework that extends beyond single-image perturbations by adopting a probabilistic perspective. Rather than modifying a single image, our method operates on an entire concept - represented by a distribution - to generate diverse adversarial examples. Preserving the concept is essential, as it ensures that the resulting adversarial images remain identifiable as instances of the original underlying category or identity. By sampling from this concept-based adversarial distribution, we generate images that maintain the original concept but vary in pose, viewpoint, or background, thereby misleading the classifier. Mathematically, this framework remains consistent with traditional adversarial attacks in a principled manner. Our theoretical and empirical results demonstrate that concept-based adversarial attacks yield more diverse adversarial examples and effectively preserve the underlying concept, while achieving higher attack efficiency.

Quantifying uncertainty is crucial in safety-critical learning systems where decisions are made based on uncertainty. Conformal prediction is one promising way to quantify uncertainty that comes with a theoretical guarantee. However, the theoretical guarantee comes with a price of the assumption on data generation process, including exchangeability or full feedback. In this paper, we propose a novel conformal prediction method with less data generation assumption, i.e., a learning method for online conformal prediction with partial feedback from an adaptive adversary. In particular, we leverage matured literatures in sequential prediction and adversarial bandits to design our online conformal prediction algorithm. The great benefit of the reliance on adversarial bandits is that we can exploit theoretical regret bounds for conformal prediction guarantees. Here, we explicitly connect the regret and a desired miscoverage guarantee in conformal prediction such that our algorithm via adversarial bandits can naturally provide a miscoverage guarantee from the regret bounds. Furthermore, we extend an existing adversarial bandit method to leverage the properties of conformal prediction, resulting in a bandit method with a tighter regret bound. We empirically demonstrate the efficacy of our conformal prediction method over various learning setups, including a stochastic setup and covariate shift setup, showing a controllability on a micoverage rate while achieving a reasonable conformal set size.

While diffusion models achieve state-of-the-art generation quality, they still suffer from computationally expensive sampling. Recent works address this issue with gradient-based optimization methods that distill a few-step ODE diffusion solver from the full sampling process, reducing the number of function evaluations from dozens to just a few. However, these approaches often rely on intricate training techniques and do not explicitly focus on preserving fine-grained details. In this paper, we introduce the Generalized Solver: a simple parameterization of the ODE sampler that does not require additional training tricks and improves quality over existing approaches. We further combine the original distillation loss with adversarial training, which mitigates artifacts and enhances detail fidelity. We call the resulting method the Generalized Adversarial Solver and demonstrate its superior performance compared to existing solver training methods under similar resource constraints.

Consider episodic Markov decision processes (MDPs) with adversarially chosen transition kernels, where the transition kernel is adversarially chosen at each episode. Prior works have established regret upper bounds of $\widetilde{\mathcal{O}}(\sqrt{T} + C^P)$, where $T$ is the number of episodes and $C^P$ quantifies the degree of adversarial change in the transition dynamics. This regret bound may scale as large as $\mathcal{O}(T)$, leading to a linear regret. This raises a fundamental question: *Can sublinear regret be achieved under fully adversarial transition kernels?* We answer this question affirmatively. First, we show that the optimal policy for MDPs with adversarial transition kernels must be history-dependent. We then design an algorithm of Adversarial Dynamics Follow-the-Regularized-Leader (AD-FTRL), and prove that it achieves a sublinear regret of $\mathcal{O}(\sqrt{(|\mathcal{S}||\mathcal{A}|)^K T})$, where $K$ is the horizon length, $|\mathcal{S}|$ is the number of states, and $|\mathcal{A}|$ is the number of actions. Such a regret cannot be achieved by simply solving this problem as a contextual bandit. We further construct a hard MDP instance and prove a matching lower bound on the regret, which thereby demonstrates the **minimax optimality** of our algorithm.

Models are susceptible to adversarially out-of-distribution (OOD) data despite large training-compute investments into their robustification. Zaremba et al. (2025) make progress on this problem at test time, showing LLM reasoning improves satisfaction of model specifications designed to thwart attacks, resulting in a correlation between reasoning effort and robustness to jailbreaks. However, this benefit of test compute fades when attackers are given access to gradients or multimodal inputs. We address this gap, clarifying that inference-compute offers benefits even in such cases. Our approach argues that compositional generalization, through which OOD data is understandable via its in-distribution (ID) components, enables adherence to defensive specifications on adversarially OOD inputs. Namely, we posit the Robustness from Inference Compute Hypothesis (RICH): inference-compute defenses profit as the model's training data better reflects the attacked data’s components. We empirically support this hypothesis across vision language model and attack types, finding robustness gains from test-time compute if specification following on OOD data is unlocked by compositional generalization. For example, InternVL 3.5 gpt-oss 20B gains little robustness when its test compute is scaled, but such scaling adds significant robustness if we first robustify its vision encoder. This correlation of inference-compute's robustness benefit with base model robustness is the rich-get-richer dynamic of the RICH: attacked data components are more ID for robustified models, aiding compositional generalization to OOD data. Thus, we advise layering train-time and test-time defenses to obtain their synergistic benefit.

We study online adversarial imitation learning (AIL), where an agent learns from offline expert demonstrations and interacts with the environment online without access to rewards. Despite strong empirical results, the benefits of online interaction and the impact of stochasticity remain poorly understood. We address these gaps by introducing a model-based AIL algorithm (MB-AIL) and establish its horizon-free, second-order sample-complexity guarantees under general function approximations for both expert data and reward-free interactions. These second-order bounds provide an instance-dependent result that can scale with the variance of returns under the relevant policies and therefore tighten as the system approaches determinism. Together with second-order, information-theoretic lower bounds on a newly constructed hard-instance family, we show that MB-AIL attains minimax-optimal sample complexity for online interaction (up to logarithmic factors) with limited expert demonstrations and matches the lower bound for expert demonstrations in terms of the dependence on horizon $H$, precision $\epsilon$ and the policy variance $\sigma^2$. Experiments further validate our theoretical findings and demonstrate that a practical implementation of MB-AIL matches or surpasses the sample efficiency of existing methods.