NeurIPS 2025 Papers

We propose a testable universality hypothesis, asserting that seemingly disparate neural network solutions observed in the simple task of modular addition actually reflect a common abstract algorithm. While prior work interpreted variations in neuron-level representations as evidence for distinct algorithms, we demonstrate---through multi-level analyses spanning neurons, neuron clusters, and entire networks---that multilayer perceptrons and transformers universally implement the abstract algorithm we call the approximate Chinese Remainder Theorem. Crucially, we introduce approximate cosets and show that neurons activate exclusively on them. Furthermore, our theory works for deep neural networks (DNNs). It predicts that universally learned solutions in DNNs with trainable embeddings or more than one hidden layer require only $\mathcal{O}(\log n)$ features, a result we empirically confirm. This work thus provides the first theory‑backed interpretation of \textit{multilayer} networks solving modular addition. It advances generalizable interpretability and opens a testable universality hypothesis for group multiplication beyond modular addition.

We propose a novel gradient-based framework for learning parameterized quantum circuits (PQCs) in the presence of Pauli noise in gate operation. The key innovation in our framework is the simultaneous optimization of model parameters and learning of an inverse noise channel, specifically designed to mitigate Pauli noise. Our parametrized inverse noise model utilizes the Pauli-Lindblad equation and relies on the principle underlying the Probabilistic Error Cancellation (PEC) protocol to learn an effective and scalable mechanism for noise mitigation. In contrast to conventional approaches that apply predetermined inverse noise models during execution, our method systematically mitigates Pauli noise by dynamically updating the inverse noise parameters in conjunction with the model parameters, facilitating task-specific noise adaptation throughout the learning process. We employ proximal stochastic gradient descent (proximal SGD) to ensure that updates are bounded within a feasible range to ensure stability. This approach allows the model to converge efficiently to a stationary point, balancing the trade-off between noise mitigation and computational overhead, resulting in a highly adaptable quantum model that performs robustly in noisy quantum environments. Our framework is well-suited to near-term quantum devices in the noisy intermediate-scale quantum (NISQ) era, where noise is a significant challenge.

The conventional approach in differential privacy (DP) literature formulates the privacy-utility tradeoff with a "privacy-first" perspective: for a predetermined level of privacy, a certain utility is achievable. However, practitioners often operate under a "utility-first" paradigm, prioritizing a desired level of utility and then determining the corresponding privacy cost. Wu et al. [2019] initiated a formal study of this ``utility-first'' perspective by introducing ex-post DP. They demonstrated that by adding correlated Laplace noise and progressively reducing it on demand, a sequence of increasingly accurate estimates of a private parameter can be generated, with the privacy cost attributed only to the least noisy iterate released. This led to a Laplace mechanism variant that achieves a specified utility with minimal privacy loss. However, their work, and similar findings by Whitehouse et al. [2023], are primarily limited to simple mechanisms based on Laplace or Gaussian noise. In this paper, we significantly generalize these results. In particular, we extend the findings of Wu et al. [2019] and Liu and Talwar [2019] to support any sequence of private estimators, incurring at most a doubling of the original privacy budget. Furthermore, we demonstrate that hyperparameter tuning for these estimators, including the selection of an optimal privacy budget, can be performed without additional privacy cost. Finally, we extend our results to ex-post R\'{e}nyi DP, further broadening the applicability of utility-first privacy mechanisms.

Large language models (LLMs) are trained on a vast amount of human-written data, but data providers often remain uncredited. In response to this issue, data valuation (or data attribution), which quantifies the contribution or value of each data to the model output, has been discussed as a potential solution. Nevertheless, applying existing data valuation methods to recent LLMs and their vast training datasets has been largely limited by prohibitive compute and memory costs. In this work, we focus on influence functions, a popular gradient-based data valuation method, and significantly improve its scalability with an efficient gradient projection strategy called LoGra that leverages the gradient structure in backpropagation. We then provide a theoretical motivation of gradient projection approaches to influence functions to promote trust in the data valuation process. Lastly, we lower the barrier to implementing data valuation systems by introducing LogIX, a software package that can transform existing training code into data valuation code with minimal effort. In our data valuation experiments, LoGra achieves competitive accuracy against more expensive baselines while showing up to 6,500x improvement in throughput and 5x reduction in GPU memory usage when applied to Llama3-8B-Instruct and the 1B-token dataset.

LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs’ refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. And there exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model’s judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without suppressing the model’s internal belief of harmfulness. We also find that adversarially fine- tuning models to accept harmful instructions has minimal impact on the model’s internal belief of harmfulness. These insights lead to a practical safety application: The model’s latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs’ internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety.

Online social media provide a platform for people to discuss current events and exchange opinions with their peers. While interactions are predominantly positive, in recent years, there has been a lot of research to understand the conflicts in social networks and how they are based on different views and opinions. In this paper, we ask whether the conflicts in a network reveal a small and interpretable set of prevalent opinion ranges that explain the users' interactions. More precisely, we consider signed graphs, where the edge signs indicate positive and negative interactions of node pairs, and our goal is to infer opinion intervals that are consistent with the edge signs. We introduce an optimization problem that models this question, and we give strong hardness results and a polynomial-time approximation scheme by utilizing connections to interval graphs and the Correlation Clustering problem. We further provide scalable heuristics and show that in experiments they yield more expressive solutions than Correlation Clustering baselines. We also present a case study on a novel real-world dataset from the German parliament, showing that our algorithms can recover the political leaning of German parties based on co-voting behavior.

Data Shapley is an important tool for data valuation, which quantifies the contribution of individual data points to machine learning models. In practice, group-level data valuation is desirable when data providers contribute data in batch. However, we identify that existing group-level extensions of Data Shapley are vulnerable to \emph{shell company attacks}, where strategic group splitting can unfairly inflate valuations. We propose Faithful Group Shapley Value (FGSV) that uniquely defends against such attacks. Building on original mathematical insights, we develop a provably fast and accurate approximation algorithm for computing FGSV. Empirical experiments demonstrate that our algorithm significantly outperforms state-of-the-art methods in computational efficiency and approximation accuracy, while ensuring faithful group-level valuation.

Large Language Models (LLMs) increasingly employ alignment techniques to prevent harmful outputs. Despite these safeguards, attackers can circumvent them by crafting prompts that induce LLMs to generate harmful content. Current methods typically target exact affirmative responses, suffering from limited convergence, unnatural prompts, and high computational costs. We introduce semantic representation attacks, a novel paradigm that fundamentally reconceptualizes adversarial objectives against aligned LLMs. Rather than targeting exact textual patterns, our approach exploits the semantic representation space that can elicit diverse responses that share equivalent harmful meanings. This innovation resolves the inherent trade-off between attack effectiveness and prompt naturalness that plagues existing methods. Our Semantic Representation Heuristic Search (SRHS) algorithm efficiently generates semantically coherent adversarial prompts by maintaining interpretability during incremental search. We establish rigorous theoretical guarantees for semantic convergence and demonstrate that SRHS achieves unprecedented attack success rates (89.4% averaged across 18 LLMs, including 100% on 11 models) while significantly reducing computational requirements. Extensive experiments show that our method consistently outperforms existing approaches.

Graph Neural Networks \texttt{(GNNs)} excel at jointly modeling node features and topology, yet their \emph{black-box} nature limits their adoption in real-world applications where interpretability is desired. Inspired by the success of interpretable Neural Additive Models \texttt{(NAM)} for tabular data, Graph Neural Additive Network \texttt{(GNAN)} extends the additive modeling approach to graph data to overcome limitations of GNNs. While being interpretable, \texttt{GNAN} representation learning overlooks the importance of local aggregation and more importantly suffers from parameter complexity. To mitigate the above challenges, we introduce Graph Neural Additive Model with Random Fourier Features (\texttt{G-NAMRFF}), a lightweight, self‐interpretable graph additive architecture. \texttt{G-NAMRFF} represents each node embedding as the sum of feature‐wise contributions where contributions are modeled via a \emph{Gaussian process} \texttt{(GP)} with a graph- and feature-aware kernel. Specifically, we construct a kernel using Radial Basis Function (\texttt{RBF}) with graph structure induced by Laplacian and learnable Finite Impulse Response (\texttt{FIR}) filter. We approximate the kernel using Random Fourier Features (\texttt{RFFs}) which transforms the \texttt{GP} prior to a Bayesian formulation, which are subsequently learnt using a single layer neural network with size equal to number of \texttt{RFF} features. \texttt{G-NAMRFF} is light weight with $168\times$ fewer parameters compared to \texttt{GNAN}. Despite its compact size, \texttt{G-NAMRFF} matches or outperforms state-of-the-art \texttt{GNNs} and \texttt{GNAN} on node and graph classification tasks, delivering real-time interpretability without sacrificing accuracy.

Recent years have seen a surge in interest in digital content watermarking techniques, driven by the proliferation of generative models and increased legal pressure. With an ever-growing percentage of AI-generated content available online, watermarking plays an increasingly important role in ensuring content authenticity and attribution at scale. There have been many works assessing the robustness of watermarking to removal attacks, yet, watermark forging, the scenario when a watermark is stolen from genuine content and applied to malicious content, remains underexplored. In this work, we investigate watermark forging in the context of widely used post-hoc image watermarking. Our contributions are as follows. First, we introduce a preference model to assess whether an image is watermarked. The model is trained using a ranking loss on purely procedurally generated images without any need for real watermarks. Second, we demonstrate the model's capability to remove and forge watermarks by optimizing the input image through backpropagation. This technique requires only a single watermarked image and works without knowledge of the watermarking model, making our attack much simpler and more practical than attacks introduced in related work. Third, we evaluate our proposed method on a variety of post-hoc image watermarking models, demonstrating that our approach can effectively forge watermarks, questioning the security of current watermarking approaches. Our code and further resources are publicly available.

Image watermarking embeds imperceptible signals into AI-generated images for deepfake detection and provenance verification. Although recent semantic-level watermarking methods demonstrate strong resistance against conventional pixel-level removal attacks, their robustness against more advanced removal strategies remains underexplored, raising concerns about their reliability in practical scenarios. Existing removal attacks primarily operate in the pixel domain without altering image semantics, which limits their effectiveness against semantic-level watermarks. In this paper, we propose Next Frame Prediction Attack (NFPA), the first semantic-level removal attack. Unlike pixel-level attacks, NFPA formulates watermark removal as a video generation task: it treats the watermarked image as the initial frame and aims to subtly manipulate the image semantics to generate the next-frame image, i.e., the unwatermarked image. We conduct a comprehensive evaluation on eight state-of-the-art image watermarking schemes, demonstrating that NFPA consistently outperforms thirteen removal attack baselines in terms of the trade-off between watermark removal and image quality. Our results reveal the vulnerabilities of current image watermarking methods and highlight the urgent need for more robust watermarks.

Model diffing is the study of how fine-tuning changes a model's representations and internal algorithms. Many behaviors of interest are introduced during fine-tuning, and model diffing offers a promising lens to interpret such behaviors. Crosscoders are a recent model diffing method that learns a shared dictionary of interpretable concepts represented as latent directions in both the base and fine-tuned models, allowing us to track how concepts shift or emerge during fine-tuning. Notably, prior work has observed concepts with no direction in the base model, and it was hypothesized that these model-specific latents were concepts introduced during fine-tuning. However, we identify two issues which stem from the crosscoders L1 training loss that can misattribute concepts as unique to the fine-tuned model, when they really exist in both models. We develop Latent Scaling to flag these issues by more accurately measuring each latent's presence across models. In experiments comparing Gemma 2 2B base and chat models, we observe that the standard crosscoder suffers heavily from these issues. Building on these insights, we train a crosscoder with BatchTopK loss and show that it substantially mitigates these issues, finding more genuinely chat-specific and highly interpretable concepts. We recommend practitioners adopt similar techniques. Using the BatchTopK crosscoder, we successfully identify a set of chat-specific latents that are both interpretable and causally effective, representing concepts such as false information and personal question, along with multiple refusal-related latents that show nuanced preferences for different refusal triggers. Overall, our work advances best practices for the crosscoder-based methodology for model diffing and demonstrates that it can provide concrete insights into how chat-tuning modifies model behavior.

Institutions with limited data and computing resources often outsource model training to third-party providers in a semi-honest setting, assuming adherence to prescribed training protocols with pre-defined learning paradigm (e.g., supervised or semi-supervised learning). However, this practice can introduce severe security risks, as adversaries may poison the training data to embed backdoors into the resulting model. Existing detection approaches predominantly rely on statistical analyses, which often fail to maintain universally accurate detection accuracy across different learning paradigms. To address this challenge, we propose a unified backdoor detection framework in the semi-honest setting that exploits cross-examination of model inconsistencies between two independent service providers. Specifically, we integrate central kernel alignment to enable robust feature similarity measurements across different model architectures and learning paradigms, thereby facilitating precise recovery and identification of backdoor triggers. We further introduce backdoor fine-tuning sensitivity analysis to distinguish backdoor triggers from adversarial perturbations, substantially reducing false positives. Extensive experiments demonstrate that our method achieves superior detection performance, improving accuracy by 4.4%, 1.7%, and 10.6% over SoTA baselines across supervised, semi-supervised, and autoregressive learning tasks, respectively. Notably, it is the first to effectively detect backdoors in multimodal large language models, further highlighting its broad applicability and advancing secure deep learning.

In the approximate quantiles problem, the goal is to output $m$ quantile estimates, the ranks of which are as close as possible to $m$ given quantiles $0 \leq q_1 \leq\dots \leq q_m \leq 1$. We present a mechanism for approximate quantiles that satisfies $\varepsilon$-differential privacy for a dataset of $n$ real numbers where the ratio between the distance between the closest pair of points and the size of the domain is bounded by $\psi$. As long as the minimum gap between quantiles is sufficiently large, $|q_i-q_{i-1}|\geq \Omega\left(\frac{m\log(m)\log(\psi)}{n\varepsilon}\right)$ for all $i$, the maximum rank error of our mechanism is $O\left(\frac{\log(\psi) + \log^2(m)}{\varepsilon}\right)$ with high probability. Previously, the best known algorithm under pure DP was due to Kaplan, Schnapp, and Stemmer (ICML '22), who achieved a bound of $O\left(\frac{\log(\psi)\log^2(m) + \log^3(m)}{\varepsilon}\right)$. Our improvement stems from the use of continual counting techniques which allows the quantiles to be randomized in a correlated manner. We also present an $(\varepsilon,\delta)$-differentially private mechanism that relaxes the gap assumption without affecting the error bound, improving on existing methods when $\delta$ is sufficiently close to zero. We provide experimental evaluation which confirms that our mechanism performs favorably compared to prior work in practice, in particular when the number of quantiles $m$ is large.

In this paper, we address the challenges associated with merging low-rank adaptations of large neural networks. With the rise of parameter-efficient adaptation techniques, such as Low-Rank Adaptation (LoRA), model fine-tuning has become more accessible. While fine-tuning models with LoRA is highly efficient, existing merging methods often sacrifice this efficiency by merging fully-sized weight matrices. We propose the Core Space merging framework, which enables the merging of LoRA-adapted models within a common alignment basis, thereby preserving the efficiency of low-rank adaptation while substantially improving accuracy across tasks. We further provide a formal proof that projection into Core Space ensures no loss of information and provide a complexity analysis showing the efficiency gains. Extensive empirical results demonstrate that Core Space significantly improves existing merging techniques and achieves state-of-the-art results on both vision and language tasks while utilizing a fraction of the computational resources. Codebase is available at https://github.com/apanariello4/core-space-merging.

Text-to-image diffusion models have shown an impressive ability to generate high-quality images from input textual descriptions/prompts. However, concerns have been raised about the potential for these models to create content that infringes on copyrights or depicts disturbing subject matter. Removing specific concepts from these models is a promising solution to this issue. However, existing methods for concept removal do not work well in practical but challenging scenarios where concepts need to be continuously removed. Specifically, these methods lead to poor alignment between the text prompts and the generated image after the continuous removal process. To address this issue, we propose a novel concept removal approach called CCRT that includes a designed knowledge distillation paradigm. CCRT constrains the text-image alignment behavior during the continuous concept removal process by using a set of text prompts. These prompts are generated through our genetic algorithm, which employs a designed fuzzing strategy. To evaluate the effectiveness of CCRT, we conduct extensive experiments involving the removal of various concepts, algorithmic metrics, and human studies. The results demonstrate that CCRT can effectively remove the targeted concepts from the model in a continuous manner while maintaining the high image generation quality (e.g., text-image alignment). The code of CCRT is available at https://github.com/wssun/CCRT.

Neural networks are valuable intellectual property due to the significant computational cost, expert labor, and proprietary data involved in their development. Consequently, protecting their parameters is critical not only for maintaining a competitive advantage but also for enhancing the model's security and privacy. Prior works have demonstrated the growing capability of cryptanalytic attacks to scale to deeper models. In this paper, we present the first defense mechanism against cryptanalytic parameter extraction attacks. Our key insight is to eliminate the neuron uniqueness necessary for these attacks to succeed. We achieve this by a novel, extraction-aware training method. Specifically, we augment the standard loss function with an additional regularization term that minimizes the distance between neuron weights within a layer. Therefore, the proposed defense has zero area-delay overhead during inference. We evaluate the effectiveness of our approach in mitigating extraction attacks while analyzing the model accuracy across different architectures and datasets. When re-trained with the same model architecture, the results show that our defense incurs a marginal accuracy change of less than 1\% with the modified loss function. Moreover, we present a theoretical framework to quantify the success probability of the attack. When tested comprehensively with prior attack settings, our defense demonstrated empirical success for sustained periods of extraction, whereas unprotected networks are extracted between 14 minutes to 4 hours.

As generative AI systems become competent and democratized in science, business, and government, deeper insight into their failure modes now poses an acute need. The occasional volatility in their behavior, such as the propensity of transformer models to hallucinate, impedes trust and adoption of emerging AI solutions in high-stakes areas. In the present work, we establish how and when hallucinations arise in pre-trained transformer models through concept representations captured by sparse autoencoders, under scenarios with experimentally controlled uncertainty in the input space. Our systematic experiments reveal that the number of semantic concepts used by the transformer model grows as the input information becomes increasingly unstructured. In the face of growing uncertainty in the input space, the transformer model becomes prone to activate coherent yet input-insensitive semantic features, leading to hallucinated output. At its extreme, for pure-noise inputs, we identify a wide variety of robustly triggered and meaningful concepts in the intermediate activations of pre-trained transformer models, whose functional integrity we confirm through targeted steering. We also show that hallucinations in the output of a transformer model can be reliably predicted from the concept patterns embedded in transformer layer activations. This collection of insights on transformer internal processing mechanics has immediate consequences for aligning AI models with human values, AI safety, opening the attack surface for potential adversarial attacks, and providing a basis for automatic quantification of a model’s hallucination risk.

Recent proposals for regulating frontier AI models have sparked concerns about the cost of safety regulation, and most such regulations have been shelved due to the safety-innovation tradeoff. This paper argues for an alternative regulatory approach that ensures AI safety while actively \textit{promoting} innovation: mandating that large AI laboratories release small, openly accessible "analog models"—scaled-down versions trained similarly to and distilled from their largest proprietary models. Analog models serve as public proxies, allowing broad participation in safety verification, interpretability research, and algorithmic transparency without forcing labs to disclose their full-scale models. Recent research demonstrates that safety and interpretability methods developed using these smaller models generalize effectively to frontier-scale systems. By enabling the wider research community to directly investigate and innovate upon accessible analogs, our policy substantially reduces the regulatory burden and accelerates safety advancements. This mandate promises minimal additional costs, leveraging reusable resources like data and infrastructure, while significantly contributing to the public good. Our hope is not only that this policy be adopted, but that it illustrates a broader principle supporting fundamental research in machine learning: deeper understanding of models relaxes the safety-innovation tradeoff and lets us have more of both.