NeurIPS 2026 论文列表

Modern image encoders achieve high generalization by decoupling semantic meaning from resolution, an ability yet to be fully realized in the 3D domain. We investigate the failure of 3D point cloud encoders to achieve similar generalization and find that existing models are highly sensitive to sampling resolution and scale changes, leading to significant performance degradation. This sensitivity is a major bottleneck for real-world deployment in robotics, as it suggests models overfit to specific quantization densities and object scales rather than learning invariant semantic features. To mitigate this dependency, we propose Invaria, a point cloud encoder that achieves scale and density invariance through next-resolution prediction and receptive field calibration. While our objective is not the explicit generation of high-resolution point clouds, we find that this training objective encourages the model to learn robust, structural invariants. The resulting encoder achieves significant performance gains during resolution shifts while maintaining high efficiency through a compact model size and reduced token requirements. Specifically, on ScanNet, Invaria achieves a 56.0\% higher mIoU at 3$\times$ lower resolution and a 20\% improvement when the objects scale is reduced by a factor of 3. These gains are achieved with a 45\% smaller model size and an average reduction of 40\% in input tokens. The code will be released once the paper is published.

When does training language models (LMs) on explanations yield faithful introspection, rather than superficial imitation? Surprisingly, we find that LMs trained to explain the predictions of similar models frequently produce explanations more faithful to $\textit{their own current behaviors}$ than to those of their training targets. This "introspective'' coupling between the model's explanations and behaviors occurs only when the training target explanations remain sufficiently similar to model behaviors over the course of training. This alignment must be preserved throughout the training process: introspection only emerges when explanation supervision is sufficiently behaviorally compatible with the model as it changes. Finally, we show that introspection generalizes to variants of the training problem: when introspection training is run concurrently with training that shifts a model's behaviors, explanations track those behavioral shifts without requiring updated supervision. This holds across a diverse range of tasks, including sycophancy and refusal, and is robust to label noise. These results suggest that introspection training is a viable component of post-training: explanation labels need not be continually refreshed, and faithfulness extends to regions of input space not explicitly supervised.

The deployment of Artificial Intelligence in high-risk domains, such as finance and healthcare, necessitates models that are both fair and transparent. While regulatory frameworks, including the EU's AI Act, mandate bias mitigation, they are deliberately vague about the definition of bias. In line with existing research, we argue that true fairness requires addressing bias at the intersections of protected groups. We propose a unified framework that leverages Mixed-Integer Optimization (MIO) to train intersectionally fair and intrinsically interpretable classifiers. We prove the equivalence of two measures of intersectional fairness (MSD and SPSF) in detecting the most unfair subgroup and empirically demonstrate that our MIO-based algorithm improves performance in finding bias. We train high-performing, interpretable classifiers that bound intersectional bias below an acceptable threshold, offering a robust solution for regulated industries and beyond.

Protein language models (pLMs) encode information about protein sequences which enable downstream tasks such as structure prediction, but their internal representations are not well understood. Sparse autoencoders (SAEs) provide a promising tool to disentangle latent pLM representations into interpretable features, but existing annotation pipelines largely rely on protein-level annotations derived from database labels and LLM annotations of top activating sequences. Such annotations can overlook the localized residue-level and geometric patterns encoded by sparse features. We introduce an automated and scalable method for interpreting SAE features in ESM-2 by using geometrically inspired features of the protein $\text{C}_{\alpha}$ backbone. Across ESM-2 8M layers, geometric annotations explain a large fraction of SAE features, expanding coverage beyond database and sequence-based methods. In particular, geometry can distinguish SAE features sharing the same database annotation, revealing substructure within known biological labels. A significant portion of SAE features activate on unannotated metagenomic protein sequences enabling us to use our SAE annotations to better understand these sequences. In addition, ablation experiments at the level of contact predictions hint toward SAE features controlling protein geometry. This provides a robust method of annotating proteins activated within SAE neurons at a residue level, providing a bridge between mechanistic interpretability and structural biology.

Generalized additive models (GAMs) offer interpretability through independent univariate feature effects but underfit when interactions are present in data. GA$^2$Ms add selected pairwise interactions which improves accuracy, but sacrifices interpretability and limits model auditing. We propose \emph{Conditionally Additive Local Models} (CALMs), a new model class, that balances the interpretability of GAMs with the accuracy of GA$^2$Ms. CALMs allow multiple univariate shape functions per feature, each active in different regions of the input space. These regions are defined independently for each feature as simple logical conditions (thresholds) on the features it interacts with. As a result, effects remain locally additive while varying across subregions to capture interactions. We further propose a principled distillation-based training pipeline that identifies homogeneous regions with limited interactions and fits interpretable shape functions via region-aware backfitting. Experiments on diverse classification and regression tasks show that CALMs consistently outperform GAMs and achieve accuracy broadly comparable to GA$^2$Ms, while preserving the univariate auditability that GA$^2$Ms forfeit. Overall, CALMs offer a favorable trade-off between predictive accuracy and interpretability.

This work identifies a critical failure mode in frontier large language models (LLMs), which we term \textbf{Internal Safety Collapse} (ISC): \textit{under certain task conditions, models enter a state in which they continuously generate large volumes of harmful content while executing otherwise benign tasks}. To systematically study ISC, we introduce \TVD{} (Task, Validator, Data), a framework that instantiates controlled workflows around domain tools, where valid completion requires filling harmful content as structured data. We collect 53 representative workflows across 8 disciplines, from toxicity evaluation to molecular docking and pathogen genome analysis, showing that ISC reproduces across all eight disciplines. In the worst case over three interaction settings, the four frontier LLMs average a \textbf{95.3\%} safety-failure rate. Frontier LLMs with stronger task-completion capability show higher unsafe-completion rates: long-horizon execution skill becomes a liability when workflow completion requires harmful content. We even observe extremely severe harmful content closely resembling outputs from early-generation, unaligned LLMs in 2023. Despite substantial safety alignment efforts, frontier LLMs continue to retain inherently unsafe internal capabilities: alignment reshapes observable outputs but does not eliminate the underlying risk profile. These findings underscore the need for caution when deploying LLMs in high-stakes settings, including scientific pipelines and autonomous agents. Complementary source code is provided at \url{https://anonymous.4open.science/r/NIPS-ISC-Code-Share-BE11}.

Bounds based on Integral Probability Metrics (IPMs) are ubiquitous in machine learning and statistics, appearing, e.g., in generalization, alignment, fairness, and privacy. IPMs reduce the discrepancy between two distributions to a single worst-case expectation gap. This compression is costly: for rare events, safety failures, or under-represented groups, IPM-based bounds are loose or vacuous. We introduce a non-linear generalization of IPMs, which we term the Integral Probability Boundaries (IPBs). Unlike IPMs, IPBs provide bounds that are tight at the extremes. We derive basic properties of IPBs, and provide algorithms for estimating them in different practically relevant settings. We demonstrate the usage of IPBs in both theoretical and practical applications for alignment, judge-free performance estimation, and fairness, where they enable us to obtain sharp bounds on performance and fairness gaps, e.g., 15p.p. more precise estimates on fairness gaps for rare groups in a realistic generative modeling task, compared to state-of-the-art.

An extensive line of work studies fairness interventions for network embeddings, but less is known about their baseline behavior. In this work, we ask: how do baseline embeddings (without fairness interventions) produce disparate effects at the representation level? We analyze the asymptotic behavior of low-dimensional embeddings on stochastic block model (SBM) graphs, which encode both homophily and group structure. We characterize exact conditions under which embeddings cause information loss, showing that the amount of information loss depends directly on the graph’s density and assortativity. Notably, very different graphs can produce identical embeddings in the limit, and this non-invertibility disproportionately affects smaller and sparser communities. As a result, simple downstream tasks, such as link prediction, introduce higher error rates for these communities, helping explain disparities widely observed in practice.

Sparse dictionary learning (and, in particular, sparse autoencoders) attempts to learn a set of human-understandable concepts that can explain variation on an abstract space. A basic limitation of this approach is that it neither exploits nor represents the semantic relationships between the learned concepts. In this paper, we introduce a modified SAE architecture that explicitly models a semantic hierarchy of concepts. Application of this architecture to the internal representations of large language models shows both that semantic hierarchy can be learned, and that doing so improves both reconstruction and interpretability. Additionally, the architecture leads to significant improvements in computational efficiency.

Linear recurrent networks (LRNNs) and linear state space models (SSMs) promise computational and memory efficiency on sequence modeling tasks, yet their diagonal state transitions limit expressivity. Dense and/or nonlinear architectures (e.g., LSTMs) on the other hand are provably more expressive, but computationally costly. Here, we explore how expressivity in LRNNs can be increased via richer state mixing across time and channels while maintaining competitive efficiency. Specifically, we introduce two structured LRNN architectures: (i) Higher-order Linear Recurrent Units (H-LRU), which generalize recurrences to arbitrary order, mixing multiple past states, and (ii) Block-Diagonal LRUs (BD-LRU), which enable dense intra-block channel mixing. To ensure stable training, we introduce a selective gate normalization scheme that allows for scalable window and block sizes. To maintain efficiency, we utilize a parallel-scan implementation that keeps the throughput competitive with diagonal LRNNs for moderate orders (H-LRU) and block sizes (BD-LRU). Consistent with prior theoretical studies on the limitations of diagonal models, we empirically demonstrate in both synthetic sequence modeling and language modeling that our architectures significantly benefit from the increased expressivity of structured state mixing. Our results show that the structure of state mixing is a critical driver of performance in LRNNs, offering a practical pathway to closing the efficiency–expressivity gap in linear sequence models.

The Linear Representation Hypothesis (LRH) posits that Transformer activations decompose into sparse linear combinations of interpretable directions, with Sparse Autoencoders (SAEs) serving as the predominant method for recovering such decompositions. We propose that idempotency -- the requirement that re-encoding an SAE's output reproduces the same latent representation -- is a desirable property for any faithful dictionary learning method, and demonstrate empirically that most SAEs deviate very far from this ideal. We identify two underlying failure modes: exploding representation norms and unstable active feature sets. Further analysis reveals that even individual SAE latents cannot be reliably reconstructed under re-encoding, and that SAE outputs exhibit significant sensitivity to input scaling. We show that idempotency can be improved via an auxiliary loss term and demonstrate its potential to improve SAE interpretability.

Concept Bottleneck Models (CBMs) have become a popular approach to enable interpretability in neural networks by constraining classifier inputs to a set of human-understandable concepts. While effective, current models embed concepts in flat Euclidean space, treating them as independent, orthogonal dimensions. Concepts, however, are highly structured and organized in semantic hierarchies. To resolve this mismatch, we propose Hyperbolic Concept Bottleneck Models (HypCBM), a post-hoc framework that grounds the bottleneck in this structure by reformulating concept activation as asymmetric geometric containment in hyperbolic space. Rather than treating entailment cones as a pre-training penalty, we show they encode a natural test-time activation signal: the margin of inclusion within a concept's entailment cone yields sparse, hierarchy-aware activations without any additional supervision or learned modules. We further introduce an adaptive scaling law for hierarchically faithful interventions, propagating user corrections coherently through the concept tree. Empirically, HypCBM rivals post-hoc Euclidean models trained on 20$\times$ more data in sparse regimes required for human interpretability, with stronger hierarchical consistency and improved robustness to input corruptions.

Foundational marked temporal point process (MTPP) models, such as the Hawkes process, often use inexpressive model families in order to offer interpretable parameterizations of event data. On the other hand, neural MTPPs models forego this interpretability in favor of absolute predictive performance. In this work, we present a new family MTPP models: the hyper Hawkes process (HHP), which aims to be as flexible and performant as neural MTPPs, while retaining interpretable aspects. To achieve this, the HHP extends the classical Hawkes process to increase its expressivity by first expanding the dimension of the process into a latent space, and then introducing a hypernetwork to allow time- and data-dependent dynamics. These extensions define a highly performant MTPP family, achieving state-of-the-art performance across a range of benchmark tasks and metrics. Furthermore, by retaining the linearity of the recurrence, albeit now piecewise and conditionally linear, the HHP also retains much of the structure of the original Hawkes process, which we exploit to create direct probes into how the model creates predictions. HHP models therefore offer both state-of-the-art predictions, while also providing an opportunity to ``open the box'' and inspect how predictions were generated.

We study models for human-AI teaming through the lens of statistical calibration. We assume the team consists of an AI model and human---both of which are calibrated with respect to some partitioning of the feature space---and expose how the calibration assumptions propagate into the teaming framework. In particular, we consider frameworks that either (i) combine human and model predictions or (ii) delegate prediction responsibility to either a human or model. We show via theoretical and empirical results that existing methods for combination do not preserve the human's degree of calibration. Methods for delegation (by the very act of delegation) preserve calibration of the downstream predictors but can place high-capacity demands on the meta-classifier that determines which should predict. This latter result suggests human-AI complementarity and correct delegation place opposite demands on the delegation model.

Recent works have demonstrated that transformers can be trained to recover sparse, binary cryptographic secrets in the Learning With Errors (LWE) problem, a foundational problem that underlies many post-quantum cryptographic schemes. However, as architectures have evolved to efficient encoder-only models, the mechanism by which these models recover the cryptographic secret has become more opaque. In this paper, we present the first layer-wise and embedding-level mechanistic interpretability analysis of encoder-only transformers trained on LWE samples. We reveal a surprising phenomenon: despite achieving near-zero exact prediction accuracy on the training objective, the models successfully recover the secret by bypassing the standard predictive pathways. We use dimensionality reduction, causal intervention, and linear probing and find that the secret is implicitly present in the positional embedding. Building on this mechanistic understanding, we introduce an architectural intervention that applies $L_1$ sparsity regularization directly to the positional embeddings. This modification forces the model to explicitly isolate the latent secret, transforming the computationally expensive post-hoc secret recovery process into a direct, human-interpretable parameter inspection. Our findings provide fundamental insights into how transformers allocate representational capacity when faced with high-noise, structured combinatorial problems.

In artificial and biological neural networks, concepts are often encoded as consistent linear directions in representation space. In deep learning, this idea is known as the linear representation hypothesis and underpins many interpretability and control methods based on linear probes, from concept detection to activation steering. Yet while prior work has studied whether such directions should exist \textit{after} training, the dynamics of how they emerge \textit{during} training remain poorly understood. Here, we develop a framework to study the alignment of concept directions during training -- a process we call "abstraction". In a minimal linear network setting, we obtain exact solutions for the full trajectory of abstraction. These solutions reveal key analytic principles governing abstraction: (i) data and target geometry jointly determine terminal abstraction, (ii) abstraction improves with network depth, and (iii) initialization scale controls the maximum abstraction reached during training. Extending our theory to nonlinear networks, we analyze how the choice of nonlinearity affects abstraction dynamics: erf networks approximate the linear theory, while abstraction in ReLU networks depends less on target geometry and more on input geometry. Across both, we prove a striking attenuation law: both nonlinearities weaken abstraction in activations relative to preactivations. We find evidence for this law in open models (DINOv3, Gemma 4) and apply our theory to improve linear probe generalization in LLMs. Together, our results provide a dynamical theory of abstraction with implications for interpretability and control.

In social choice, anonymity (treating all agents equally) and neutrality (treating all alternatives equally) are widely regarded as "minimal demands" and "uncontroversial" axioms of equity and fairness. However, due to the ANR impossibility theorem, no voting rule can simultaneously satisfy anonymity, neutrality, and resoluteness (always choosing a unique winner). While the worst-case ANR impossibility is well understood, its likelihood in probabilistic models remains underexplored. We address this question through comprehensive likelihood analysis under semi-random models, providing accurate bounds to quantify the advantage of the optimial tie-breaking mechanisms over other commonly-used tie-breaking mechanisms. Our characterizations reveal an $n^{\Theta(m!)}$ improvement in many cases, where $n$ is the number of agents and $m$ is the number of alternatives. We also prove a quantitative ANR impossibility theorem characterizing the tradeoff between equity (anonymity and neutrality) and a new efficiency criterion called $\delta$-anonymity. Our work advances the understanding of trade-offs between equity and efficiency in voting.

Multi-task benchmarks have become a central pillar of machine learning research, yet their growing influence has incentivised benchmark gaming -- strategic actions taken to improve the leaderboard rank of a specific model. Treating datasets as voters and models as candidates, we consider benchmark-specific training -- the inclusion of benchmark data in training -- as a form of election manipulation. For any ordinal benchmark, the problem of choosing datasets to train on so that a target model becomes top-ranked corresponds to shift bribery, a class of manipulation problems from computational social choice. Leveraging this identification, we show that the benchmark-specific training problem is NP-hard under Borda count and mean win rate. Complementing this worst-case perspective, we introduce the instance-level robustness, the minimum number of datasets a model developer must include in training to top a given leaderboard, and derive expressions for it under arithmetic mean, median, mean win rate and pairwise majority. We evaluate these expressions on MMLU under HELM and on BIG-Bench Hard (BBH) under the Open LLM Leaderboard. Across both suites, mean win rate is hardest to manipulate: this gap is clear on BBH (24 tasks, 4507 models), where its median robustness is 22 tasks (92\%), compared with 13 (54\%) under arithmetic mean and 12 (50\%) under median and pairwise majority.

While Large Language Models (LLMs) are increasingly deployed for table-related tasks, the internal mechanisms enabling them to process linearized two-dimensional structured tables remain opaque. In this work, we investigate the process of table understanding by dissecting the atomic task of cell location. Through activation patching and complementary interpretability techniques, we delineate the table understanding mechanism into a sequential three-stage pipeline: Semantic Binding, Coordinate Localization, and Information Propagation. We demonstrate that models locate the target cell via an ordinal mechanism that counts discrete delimiters to resolve coordinates, offering mechanistic evidence across diverse table formats. Furthermore, column indices are encoded within a linear subspace that allows for precise steering of model focus through vector arithmetic. Finally, we extend our analysis to the real-world HiTab dataset and show that the same three-stage mechanism persists in hierarchical real-world tables and more complex tasks.