ICML 2026 论文列表

In performative learning, the data distribution reacts to the deployed model—for example, because strategic users adapt their features to game it—which creates a more complex dynamic than in classical supervised learning. One should thus not only optimize the model for the current data but also take into account that the model might steer the distribution in a new direction, without knowing the exact nature of the potential shift. We explore how regularization can help cope with performative effects by studying its impact in high-dimensional ridge regression. We show that, while performative effects worsen the test risk in the population setting, when moving to the over-parameterized regime where the number of features exceeds the number of samples, the optimal regularization in the presence of performativity helps reduce the variance in the estimated parameters, thereby improving performance. We show that the optimal regularization scales with the overall strength of the performative effect, making it possible to set the regularization in anticipation of this effect. We illustrate this finding through empirical evaluations of the optimal regularization parameter on both synthetic and real-world datasets.

The integration of audio modality into Large Audio Language Models (LALMs) significantly expands their attack surface. Existing jailbreaks predominantly treat audio as a carrier for malicious payloads, relying on semantic optimization, acoustic parameter control, or additive perturbation to embed harmful content into the audio signal. In this work, we challenge this necessity and propose a new paradigm in which the role of audio shifts from content injection to safety alignment interference. We reveal that LALM safety alignment can be compromised solely by specific **Acoustic Latent Semantics (ALS)**, the underlying paralinguistic features intrinsic to the priors of audio generative models. Distinct from previous works that leverage explicit acoustic parameters to merely style malicious audio, we demonstrate that interference audio, benign in content but infused with specific ALS, can serve as a universal jailbreak trigger. Leveraging this insight, we propose **Acoustic Interference Attack (AIA)**, which decouples the attack payload from the audio. It employs a set of universal, instruction-neutral interference audio, enabling standard malicious text queries to bypass safety alignment without instance-specific optimization. Experiments on 10 LALMs across five datasets demonstrate that AIA achieves the state-of-the-art attack success rate. Furthermore, our interpretability analysis uncovers the inference path drift induced by AIA and identifies the inherent effective patterns within ALS, revealing the fundamental vulnerability of cross-modal alignment in LALMs.

While Large Language Models (LLMs) have achieved remarkable success across diverse tasks, they remain vulnerable to jailbreak attacks, which pose significant risks to their secure deployment. Driven by their inherent token-by-token autoregressive inference, LLMs exhibit semantic representations that lack robustness against spatially structured perturbations, thereby rendering current output-guardrail safety mechanisms penetrable. Exploiting the Transformer's spatial weakness, we propose SpatialJB to disrupt the model’s output generation process, allowing harmful content to bypass guardrails without detection. Comprehensive experiments on leading LLMs demonstrate that SpatialJB achieves a nearly 100\% ASR and consistently maintains a success rate exceeding 75\% even against advanced output guardrails like the OpenAI Moderation API, outperforming current jailbreak techniques by a significant margin. While **SpatialJB** advances LLM safety research by exposing guardrail weaknesses and highlighting spatial semantics, we also propose and evaluate baseline defense strategies to prevent its potential misuse. You can click [Video Link](https://1drv.ms/v/s!ApaP6YaJA87pgk-Rk7ZCvPeYcIZa?e=J8kbg0) and [Code Link](https://github.com/DongFangShenWa/SpatialJailbreak.git) to see our demo presentation and code.

Jailbreak prompts can trigger harmful completions on aligned LLMs, In accordance, safety steering has been proposed: test-time activation interventions that steer jailbreak activations to trigger refusal while preserving benign utility. However, existing steering methods are fundamentally supervised and tied to a static, limited training set, whereas real jailbreaks evolve and are often out-of-distributed from the training set, leading to failures on unseen attacks. In this paper, we tackle this failure by developping a \emph{zero-shot} defense. Base on unsupervised latent direction discovery, we directly simulate jailbroken activations without any knowledge of jailbreak strategy. To build a defense mechnism upon this, we propose a bi-level adversarial training framework. In the inner step, we simulate diverse jailbroken activations by extrapolating from refusal state harmful-request activations via unsupervised latent direction discovery. In the outer step, we train a potential-induced steering field to push these adversarial jailbroken states into refusal regions while keeping benign unchanged. Across three LLMs and six classical jailbreak families, our method achieves strong defense with attack success rates mostly below 5%, and we analyzed the increasing subspace coverage of our simulated jailbroken activations on real jailbreaks throughout training, which helps explain the increasing robustness of our defense mechnism.

Safe global deployment of AI models requires alignment with human values that vary across cultures. Yet rater pools in safety evaluation datasets remain largely geographically homogeneous, failing to capture geo-cultural differences. Further, it remains unclear whether such differences persist after controlling for demographics such as age, gender, and ethnicity. Through a meta-analysis of safety datasets, we find that most do not report geo-cultural information, and those that do lack a unified methodology to jointly analyze geo-cultural and demographic correlates. Using the Inglehart-Welzel dimensions of cross-cultural variation, we demonstrate via multilevel modeling that cultural zone membership explains variance in safety ratings beyond standard demographics ($p<0.05$ across $6$ datasets). Moreover, our analysis indicates that roughly $10$\% of items in the datasets we examined are culturally sensitive: likely to be misclassified as safe without adequate cultural representation. We evaluate LLMs as both rater surrogates and triage tools, finding that current LLMs do not reliably stand in for raters, though they can help prioritize culturally sensitive items for human annotation. Our findings motivate more culturally pluralistic safety evaluation and offer practical takeaways to support it.

As Large Language Model (LLM) agents are increasingly deployed in complex environments, multi-turn interaction attacks have become a significant security challenge. Existing detection methods typically rely on historical context. However, this retrospective logic struggles to identify deep malicious intents that are split across turns to hide future risks. Inspired by speculative decoding, we propose the Speculative Safety Honeypot (SSH) framework. SSH uses a multi-agent simulation system composed of small LLMs to build an action-level speculate-and-verify workflow. In the speculation stage, SSH predicts future behaviors of the target agent and asynchronously builds a trajectory tree to expose potential risks in advance. In the verification stage, the system uses the target agent's real actions to calibrate and prune the trajectory tree, effectively reducing false positives. As a plug-and-playable component, SSH provides existing detectors with rich decision redundancy beyond the current interaction slice. By judging risk based on the evolution of the entire trajectory tree rather than a single point in time, the system reduces the reliance on the absolute precision of individual detection components. This improves the defense resilience and the warning lead-time of agent systems against complex temporal attacks.

Analyzing large-scale text corpora is a core challenge in machine learning, crucial for tasks like identifying undesirable model behaviors. Current methods often rely on costly LLM-based techniques (e.g. annotating dataset differences) or dense embedding models (e.g. for clustering), which lack control over the properties of interest. We propose using sparse autoencoders (SAEs) to create *SAE embeddings*: representations whose dimensions map to interpretable concepts. Through four data analysis tasks, we show that SAE embeddings are more cost-effective and reliable than LLMs and offer the controllability that dense embeddings lack. Using the large hypothesis space of SAEs, we can uncover insights such as (1) semantic differences between datasets and (2) unexpected concept correlations in documents. For instance, by comparing model responses, we find that Grok-4 clarifies ambiguities more often than nine other frontier models. Relative to LLMs, SAE embeddings uncover bigger differences at 2-8x lower cost and identify biases more reliably. Additionally, SAE embeddings are controllable: by filtering concepts, we can (3) cluster documents along axes of interest and (4) outperform dense embeddings on property-based retrieval. Using SAE embeddings, we study model behavior with two case studies: investigating how OpenAI model behavior has changed over time and finding "trigger" phrases learned by Tulu-3 (Lambert et al., 2024) from its training data. These results position SAEs as a versatile tool for unstructured data analysis and highlight the neglected importance of interpreting models through their *data*.

Unlearning in large language models (LLMs) aims to remove specified data, but its efficacy is typically assessed with task-level metrics like accuracy and perplexity. We show that these metrics can be misleading, as models can appear to forget while their original behavior is easily restored through minimal fine-tuning. This \emph{reversibility} suggests that information is merely suppressed, not genuinely erased. To address this critical evaluation gap, we introduce a \emph{representation-level analysis framework}. Our toolkit comprises PCA similarity and shift, centered kernel alignment (CKA), and Fisher information, complemented by a summary metric, the mean PCA distance, to measure representational drift. Applying this framework across multiple unlearning methods, data domains, and LLMs, we identify four distinct forgetting regimes based on their \emph{reversibility} and \emph{catastrophicity}. We compare recovery strategies and show that relearning efficiency relies on the data source. We also find that irreversible, non-catastrophic forgetting is exceptionally challenging. By probing unlearning limits, we identify a case of seemingly irreversible, targeted forgetting, offering insights for more robust erasure algorithms. Overall, our findings expose a gap in current evaluation and establish a representation-level foundation for trustworthy unlearning.

Attribution methods reveal which input features a neural network uses for a prediction, adding transparency to their decisions. A common problem is that these attributions seem unspecific, highlighting both important and irrelevant features. We revisit the common attribution pipeline and observe that using logits as attribution target is a main cause of this phenomenon. We show that the solution is in plain sight: considering distributions of attributions over multiple classes using existing attribution methods yields specific and fine-grained attributions. On common benchmarks, including the grid-pointing game and randomization-based sanity checks, this improves the ability of 18 attribution methods across 7 architectures up to $2\times$, agnostic to model architecture.

Language models (LMs) tend to memorize portions of their training data and reproduce verbatim spans. When the underlying sources are sensitive or copyright-protected, such reproduction raises issues of consent and compensation for creators and compliance risks for developers. We propose Anchored Decoding, a plug-and-play inference-time method for suppressing verbatim reproduction: it enables decoding from any risky LM trained on mixed-license data by keeping generation in bounded proximity to a permissively trained safe LM. Anchored Decoding does so by adaptively allocating a user-chosen information budget over the generation trajectory and enforcing per-step constraints that yield a sequence-level guarantee, enabling a tunable risk–utility trade-off. To make Anchored Decoding practically useful, we introduce a new permissively trained safe model (TinyComma 1.8B), as well as Anchored-Byte Decoding, a byte-level variant of our method that enables cross-vocabulary fusion via the ByteSampler (Hayase et al., 2025) framework. Across six model pairs on long-form metrics for copying risk and utility, Anchored and Anchored-Byte Decoding define a new Pareto frontier, preserving near-original fluency and factuality while eliminating up to 75\% of the measurable copying gap between the risky baseline and a safe reference, at a modest inference overhead.

Group counterfactual explanations find a set of counterfactual instances to explain a group of input instances contrastively. However, existing methods either (i) optimize counterfactuals only for a fixed group and do not generalize to new group members, (ii) strictly rely on strong model assumptions (e.g., linearity) for tractability and/or (iii) poorly control the counterfactual group geometry distortion. We instead learn an explicit optimal transport map that sends any group instance to its counterfactual without re-optimization, minimizing the group's total transport cost. This enables generalization with fewer parameters, making it easier to interpret the common actionable recourse. For linear classifiers, we prove that functions representing group counterfactuals are derived via mathematical optimization, identifying the underlying convex optimization type (QP, QCQP, ...). Experiments show that they accurately generalize, preserve group geometry and incur only negligible additional transport cost compared to baseline methods. If model linearity cannot be exploited, our approach also significantly outperforms the baselines.

Large language models (LLMs) exhibit social biases that reinforce harmful stereotypes, limiting their safe deployment. Most existing debiasing methods adopt a suppressive paradigm by modifying parameters, prompts, or neurons associated with biased behavior; however, such approaches are often brittle, weakly generalizable, data-inefficient, and prone to degrading general capability. We propose \textbf{KnowBias}, a lightweight and conceptually distinct framework that mitigates bias by strengthening, rather than suppressing, neurons encoding bias-knowledge. KnowBias identifies neurons encoding bias knowledge using a small set of bias-knowledge questions via attribution-based analysis, and selectively enhances them at inference time. This design enables strong debiasing while preserving general capabilities, generalizes across bias types and demographics, and is highly data efficient, requiring only a handful of simple yes/no questions and no retraining. Experiments across multiple benchmarks and LLMs demonstrate consistent state-of-the-art debiasing performance with minimal utility degradation. Data and code are available at \url{https://github.com/JP-25/KnowBias}.

Token-Pruning accelerates Vision-Language Models by removing redundant visual tokens, yet its safety implications remain underexplored. In this work, we present the first comprehensive safety evaluation of Token-Pruning mechanisms and find that: most pruning strategies significantly degrade safety as pruning ratios increase, whereas Query-based Compression shows the opposite, with extreme pruning (up to 99.8%), unexpectedly improves model safety. This sharp contrast prompts a key question: How do different Token-Pruning strategies reshape model safety behavior, and is it possible to enhance safety without sacrificing acceleration? To answer this, we identify an unrecognized mechanism, termed Pruning-Induced Malicious Amplification, where removal of background tokens triggers a side effect: forcing the model's attention to collapse onto a few retained malicious anchors within the foreground, inadvertently amplifying their toxic semantics under jailbreak. To address that, we propose an inference-time and plug-and-play Safety-Aware Pruning (SAP) mechanism that counteracts such dominance via three steps: (1) identifying malicious anchors, (2) restoring pruned benign tokens, and (3) reallocating excessive attention from malicious anchors to benign tokens. Extensive experiments across three safety and four utility benchmarks demonstrate that SAP mitigates pruning-induced vulnerabilities, i.e., reducing ASR by up to 62%, without compromising efficiency or utility.

Reliable uncertainty quantification (UQ) is crucial for deploying graph neural networks (GNNs) in safety-critical settings, yet dominant solutions either rely on costly multi-pass sampling or require retraining—often using *black-box auxiliary* models—to obtain evidential semantics. We propose **X-EviProbe**, a simple and parameter-free *post-hoc* framework that turns a *frozen* GNN into an evidential predictor with a decomposable view of epistemic vs. aleatoric uncertainty. X-EviProbe constructs class-wise Dirichlet evidence by probing the frozen latent space and the model’s native outputs, and incorporates graph structure via lightweight evidence-strength propagation. This yields a transparent evidential representation without retraining or additional neural components. Extensive experiments on seven benchmarks show that X-EviProbe consistently ranks among the top methods for both OOD detection and misclassification detection, improving AUROC by up to **33.4%** and **8.7%** over the strongest baselines.

Local post-hoc explanations are widely used to understand black-box models on tabular data, with Local Interpretable Model-agnostic Explanations (LIME) being a popular approach. LIME approximates a black-box model using a sparse linear surrogate in a local neighborhood, implicitly assuming feature-wise linear homogeneity. However, this assumption often fails when local feature effects exhibit heterogeneous or nonlinear behaviors, resulting in unfaithful and unstable explanations. Moreover, LIME relies on a decoupled feature selection procedure that is not aligned with the surrogate modeling objective, further exacerbating instability under local sampling. To address these limitations, we propose PL-LIME, a two-stage sparse local explanation framework that ensures objective consistency across stages. PL-LIME models feature-wise local effects using instance-anchored piecewise linear functions, providing a minimal yet principled extension beyond linear surrogates under a fixed explanation budget. Sparsity is enforced through a decoupled nonnegative shrinkage procedure that directly scales the estimated local effects, improving stability while preserving interpretability. Experiments on synthetic and real-world datasets demonstrate that PL-LIME achieves higher local fidelity and stability, and provides more reliable local explanations that capture finer-grained local effect structures than LIME.

Pre-trained transformers have demonstrated remarkable generalization abilities, at times extending beyond the scope of their training data. Yet, real-world deployments often face unexpected or adversarial data that diverges from training data distributions. Without explicit mechanisms for handling such shifts, model reliability and safety degrade, urging more disciplined study of out-of-distribution (OOD) settings for transformers. By systematic experiments, we present a mechanistic framework for delineating the precise contours of transformer model robustness. We find that OOD inputs, including subtle typos and jailbreak prompts, drive language models to operate on an increased number of fallacious concepts in their internals. We leverage this device to quantify and understand the degree of distributional shift in prompts, enabling a mechanistically grounded fine-tuning strategy to robustify LLMs. Expanding the very notion of OOD from input data to a model’s private computational processes—a new transformer diagnostic at inference time—is a critical step toward making AI systems safe for deployment across science, business, and government.

Conformal Prediction (CP) has emerged as a powerful statistical framework for reliable classification, which generates a prediction set, guaranteed to include the true label with a pre-specified probability. The performance of CP methods is typically assessed by their average prediction set size. In setups where the classes can be partitioned into semantic groups, e.g., based on shared downstream actions or more interpretable coarse labels, users can benefit from prediction sets that are not only small but also contain a limited number of groups. This paper begins by addressing this problem and ultimately offers a widely applicable tool for boosting any CP method on any dataset. First, given a class partition, we propose augmenting the CP score function with a term that penalizes predictions with "out-of-group" errors. We theoretically analyze this strategy and prove its advantages for group-related metrics. Surprisingly, we show mathematically that, for common class partitions, it can also reduce the average set size of *any* CP score function. Our analysis reveals the class-similarity factors behind this improvement and motivates a variant that can further reduce prediction set size by leveraging the model's embeddings, *without requiring any human semantic partition*. Finally, we present an extensive empirical study, encompassing prominent CP methods, multiple models, and several datasets, which demonstrates that our class-similarity-based approach consistently enhances CP methods.

Recent work shows that LLMs can sometimes detect when steering vectors are injected into their residual stream and identify the injected concept, a phenomenon cited as evidence of "introspective awareness." But what mechanisms underlie this capability, and do they reflect genuine introspective circuitry or more shallow heuristics? We investigate these questions in open-source models and establish three main findings. First, introspection is behaviorally robust: detection achieves moderate true positive rates with 0% false positives across diverse prompts. We also find this capability emerges specifically from post-training rather than pretraining. Second, introspection is not reducible to a single linear confound: anomaly detection relies on distributed MLP computation across multiple directions, implemented by interpretable gate and evidence-carrier features. Third, models possess greater introspective capability than is elicited by default: ablating refusal directions improves detection by ~50% and a trained steering vector improves detection by ~75%. Overall, our results suggest that introspective awareness is behaviorally robust, grounded in nontrivial internal anomaly detection, and likely could be substantially improved in future models.

A key tool in developing safe AI models is *data auditing*, i.e., using statistical tools to determine whether harmful content may have been used in the training data of a black-box model. Unfortunately, most *membership inference attacks* (MIAs) used to perform this type of auditing themselves assume *access* to examples of harmful content from the same distribution as the query data. In real-world auditing scenarios, auditors often face legal and ethical restrictions preventing them from accessing a representative set of samples of harmful content to train MIA models effectively. We abstract and formalize this setting into a new data access model, the “unseen class” setting, and show that the state-of-the-art MIAs fail due to the lack of access to the full target distribution. We show that in this setting, *quantile regression attacks* outperform approaches typically considered to be state of the art. We demonstrate this both empirically and theoretically, showing that quantile regression attacks achieve up to **11× the TPR** of shadow model-based approaches in practice, and providing a theoretical model that outlines the generalization properties required for this approach to succeed. Our work identifies an important failure mode in existing MIAs and provides a cautionary tale for practitioners who aim to directly use existing tools for real-world applications of AI safety.