NeurIPS 2026 Papers

In adversarial multi-armed bandits, two performance measures are commonly used: static regret, which compares the learner to the best fixed arm, and dynamic regret, which compares it to the best sequence of arms. While optimal algorithms are known for each measure individually, there is no known algorithm achieving optimal bounds for both simultaneously. \cite{marinov2021pareto} first showed that such simultaneous optimality is impossible against an \emph{adaptive} adversary. Our work takes a first step to demonstrate its possibility against an \emph{oblivious} adversary when losses are \emph{deterministic}. First, we extend the impossibility result of \cite{marinov2021pareto} to the case of deterministic losses. Then, we present an algorithm achieving optimal static and dynamic regret simultaneously against an oblivious adversary. Together, they reveal a fundamental separation between adaptive and oblivious adversaries when multiple regret benchmarks are considered simultaneously. It also provides new insight into the long open problem of simultaneously achieving optimal regret against switching benchmarks of different numbers of switches. Our algorithm uses negative static regret to compensate for the exploration overhead incurred when controlling dynamic regret, and leverages Blackwell approachability to jointly control both regrets. This yields a new model selection procedure for bandits that may be of independent interest.

Designing antibodies that recognize a target antigen while avoiding non-cognate antigens remains a central challenge in computational antibody engineering. Existing antigen-conditioned generative models can jointly design antibody sequences and structures, but they are typically trained to match structural data distributions and lack an explicit alignment mechanism for antigen-specific preferences. We present AbSpecAlign, a plug-and-play specificity-reward alignment framework for antibody design. AbSpecAlign introduces a Bidirectional Multimodal Specificity Scorer (BiMSS), which integrates antibody and antigen sequence representations with SE(3)-equivariant structural embeddings to produce a contrastive reward over cognate versus non-cognate antigen pairs. We further introduce a GRPO-based post-training strategy that uses group-relative specificity rewards to align pre-trained antibody generators toward higher-scoring antigen-conditioned designs while maintaining structural regularity. To the best of our knowledge, AbSpecAlign is among the first frameworks to apply GRPO-style group-relative reward alignment to antigen-conditioned antibody sequence--structure co-design with an explicit off-target specificity reward. Experiments on RAbD CDR design benchmarks show that AbSpecAlign improves most sequence-recovery and structural-fidelity metrics when combined with diffusion-based antibody generators. Experiment results suggest that GRPO-based reward alignment with contrastive specificity signals provides an effective post-training mechanism for antigen-conditioned antibody generation, while prospective experimental validation remains necessary.

Multimodal Large Language Models (MLLMs) often exhibit significant modality preference, which is a tendency to favor one modality over another. Depending on the input, they may over-rely on linguistic priors relative to visual evidence, or conversely over-attend to visually salient cues over textual facts. Prior work has applied a uniform steering intensity to adjust the modality preference of MLLMs. However, strong steering can impair standard inference and increase error rates, whereas weak steering is often ineffective. Since steering sensitivity varies substantially across instances, a single global strength is difficult to calibrate. To address this, we introduce an instance-aware diagnostic metric Modality Contribution Ratio (MCR) that quantifies each modality’s information contribution and reveals sample-specific susceptibility to steering. Building on this signal, we propose a scaling strategy and a learnable module for instance-aware control of modality preference. Experiments show that AMPS outperforms conventional steering on $MC^2$ by improving preference control with lower generation collapse, and improves MLLM performance on general multimodal QA dataset such as MM-Vet v2.

Domain specialization of large language models (LLMs) often improves target-domain performance but degrades broad capabilities. Model merging is an attractive training-free alternative, yet common parameter-space strategies (for example, interpolation and sparsified averaging) provide limited control over task-specific interference. We introduce Adaptive Model Alignment via Ranked Inherited Subspaces (AMARIS), a covariance-driven framework that transfers specialist updates only along selected activation directions. AMARIS formulates merging as budgeted subspace inheritance: it maximizes specialist activation capture while constraining base-activation deviation, then builds a low-rank projection gate from ranked generalized eigendirections under an explicit preservation budget. This yields a single static merged checkpoint that concentrates transfer in high-utility, low-interference directions. We evaluate AMARIS on Qwen3-based merges (8B and 32B) across finance, instruction-following, and multilingual SEA specialization. In our evaluated settings, AMARIS reaches Pareto operating points that are competitive with strong training-free baselines, including 98.37\% specialist retention with 95.76\% core retention (finance), 103.49\% core retention with 90.28\% specialist retention (instruction following), and 93.69\% specialist retention with 99.77\% core retention (SEA). These results support subspace-constrained merging as a practical route to compositional capability integration without additional training.

When does an AI assistant improve human decision-making, and when does it hurt? We consider the setting in which a human worker makes a binary decision by observing her own signal and receiving a binary or probabilistic recommendation from an AI. We develop a behavioral model which considers both her belief regarding the overlap of her signal with the AI as well as the trust she has in the AI when considering its recommendation. The model yields a phase diagram in AI capability and signal overlap, partitioning the parameter space into complementarity, impairment, and automation regions. A worker who neglects signal overlap (correlation neglect) can experience performance loss when incorporating AI into her decision and is eventually dominated by AI alone as AI capability grows, with the complementarity range narrowing as overlap rises. Trust calibration shifts region boundaries: undertrust shrinks the impairment region but expands automation, while overtrust expands impairment but extends complementarity to be preferred at higher AI capability levels. Applying the framework to ChatBench (643 workers, GPT-4o and Llama-3.1-8b), AI assistance raises accuracy by 21 percentage points, but workers undertrust the AI by roughly a factor of four. The model yields design insights linking AI deployment choices, worker behavioral parameters, and outcome quality.

AI models deployed in critical domains, such as AI safety research, may subtly sabotage our efforts due to misalignment. _Sandbagging_ is a form of sabotage in which an AI intentionally underperforms on the task given to it. It is particularly pernicious on _fuzzy_ tasks, i.e. tasks which are hard to grade or require intuition. To understand sandbagging on fuzzy tasks, we introduce a novel AI control framework that considers sandbagging as an adversarial game between a blue team and a red team. The blue team uses a weak trusted model to construct a weak score against which they would train a strong, potentially sandbagging model to remove the sandbagging if it were present. The red team then tries to find model behaviors that are rated highly by the weak score, and thus might not be trained out, but actually correspond to poor performance. We test our framework on the task of writing experimental proposals for research questions from recent ML papers. We use a language model with access to the original paper as a proxy "ground-truth" scorer. Our red team discovers sandbagging behaviors using multi-objective evolutionary prompt optimization. We show that Opus 4.6 can write proposals that are worse according to the ground truth proxy than those of GPT-OSS-20B, while the weak scorer rates them as highly as the best proposals from Opus 4.6. To mitigate the threat of sandbagging, we propose an adversarial optimization algorithm for the blue team that discovers more robust prompts for the weak model. This algorithm produces a blue team prompt that our red team optimization fails to exploit.

Language model post-training is often bottlenecked by the need for human-collected preference data, which is expensive and difficult to scale. RLAIF-style approaches that leverage pseudo labels offer an abundant alternative but introduce systematic biases that degrade downstream alignment. Recent general-purpose semi-supervised methods correct for teacher bias using a small set of human-labeled examples, but suffer from high variance especially when human annotations are scarce. To this end, we propose ABC-Align, leveraging abundant pseudo label signal to minimize variance and applying a lightweight, adaptive correction grounded in the human-labeled subset. The correction strength is tuned automatically during training using plug-in estimates of the relevant bias--variance quantities. On LLM alignment with RLHF and DPO where human feedback is scarce, we empirically demonstrate that ABC-Align achieves superior performance over prior semi-supervised baselines.

Grid cells in the medial entorhinal cortex and place cells in the hippocampus together support spatial navigation. The two regions are reciprocally connected, and there is a chicken-and-egg problem for how both arise and reinforce each other during development. Current computational accounts either derive one type from the other or use network dynamics to model the emergence of one type in isolation. We introduce a unified recurrent network model that instantiates Dale's Law (every neuron is either excitatory or inhibitory), and is trained to predict the next sensory observation from masked previous sensory observations and egocentric motion. To our knowledge, this is the first single-objective model in which grid and place cells co-emerge without supervision of either type, or reliance on pre-existing spatial-cell representations. The two kinds of spatial codes coexist across 1,000 different training configurations, with their balance set by the amount of sensory noise and masking. Without retraining, the network qualitatively reproduces experimentally observed grid fragmentation in hairpin mazes, grid merging after wall removal, lattice alignment across connected rooms, locally ordered 3D fields observed in freely flying bats, as well as the developmental order in which place cells precede grid cells. We interpret these results in terms of two complementary encoding pressures within a single sensory-prediction objective: (1) correcting errors or reconstructing missing components of sensory observations, and (2) prediction of the next sensory state during navigation. Our results suggest a circuit-level account of the co-emergence of grid and place cells, and experimentally testable predictions for the two kinds of spatial codes.

Leveraging Graph Neural Networks (GNNs) as graph encoders and aligning the resulting representations with Large Language Models (LLMs) through alignment instruction tuning has become a mainstream paradigm for constructing Graph Language Models (GLMs), combining the generalization ability of LLMs with the structural modeling capacity of GNNs. However, existing GLMs that adopt GNNs as graph encoders largely overlook the problem of aligning GNN-encoded representations across domains and tasks with the LLM token space to obtain unified graph tokens, thereby limiting their ability to generalize across diverse graph data. To bridge this gap, we aim to incorporate a multi-domain, multi-task GNN encoder into GLMs and align its representations with LLMs to enable multi-domain, multi-task graph alignment instruction tuning. This alignment problem remains underexplored and poses two key challenges: 1) learning GNN-encoded representations that are simultaneously generalizable across domains and tasks and well aligned with textual semantics is difficult, due to substantial variations in graph structures, feature distributions, and supervision signals, together with the lack of textual-semantic alignment guidance in task-specific GNN training; 2) diverse graph data and task-specific instructions can exhibit different degrees of compatibility with the LLM token space during instruction tuning, leading to varying alignment difficulty and rendering a fixed alignment strategy suboptimal. To tackle these challenges, we propose UniGraphLM, a Unified Graph Language Model that incorporates a multi-domain, multi-task GNN encoder to learn generalizable graph representations aligned with textual semantics, and then adaptively aligns these representations with the LLM. Specifically, we first develop a graph-text pair pretraining strategy with a tailored GNN encoder, trained on large-scale graph-text data spanning multiple domains and tasks to obtain generalizable representations naturally aligned with textual semantics. We further design a curriculum alignment tuning strategy that adaptively adjusts the alignment process by accounting for varying alignment difficulty across diverse graph data. Extensive experiments demonstrate that UniGraphLM consistently outperforms state-of-the-art baselines across graph datasets from different domains and tasks.

We theoretically and empirically establish a previously unexplored connection between Long-Tailed Recognition (LTR) and Continual Learning (CL). Specifically, we show that training on a long-tailed dataset drives model parameters into an $\mathcal{O}(1/\sqrt{\mathrm{IF}})$ neighborhood of the dominant-class solution, under both uniform and exponentially decaying cardinality distributions. Building on this result, we prove that the CL objective upper-bounds the balanced LTR loss, revealing that LTR can be reformulated as a sequential learning problem. Motivated by this connection, we introduce Continual Learning for Long-Tailed Recognition (CLTR), a principled framework that leverages standard off-the-shelf CL methods to sequentially learn Head and Tail classes while mitigating catastrophic forgetting. Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate our theoretical predictions and demonstrate strong performance across diverse LTR benchmarks. In the foundation-model regime, CLTR with DualPrompt on a CLIP backbone outperforms standard adaptation baselines and is competitive with specialized methods using external semantic supervision. Our work bridges LTR and CL both theoretically and empirically, providing a principled approach for addressing long-tailed learning using standard CL strategies.

Backdoor attacks implant a trigger-target association into a model, causing malicious behavior at test time while largely preserving clean performance. Despite extensive empirical study, a unified explanation for why standard training dynamics learn backdoors so effectively remains largely missing. We argue that backdoor learning can be understood as a consequence of simplicity-biased optimization dynamics: trigger features are dynamically simpler than semantic features under stochastic gradient descent (SGD) because they induce more coherent gradient alignment, more stable gating behavior, and stronger early-stage amplification. To formalize this perspective, we introduce a directional data model that separates semantic and trigger features and study a one-hidden-layer ReLU network trained by SGD. Our analysis identifies a frozen-gate signal that governs group-wise early-stage loss decrease up to controlled gate-drift and gate-flip errors, yielding a quantitative explanation for faster poisoned-sample fitting and its dependence on the poisoning ratio. Beyond optimization dynamics, we show that this early-stage bias can induce trigger-dominant neurons with selective activation on poisoned inputs, explaining why vanilla clean fine-tuning can attenuate backdoor behavior without necessarily erasing the underlying trigger-related representation. Empirically, we validate the predicted signatures of early-stage optimization bias, representation-level trigger dominance, and post-training residual behavior using ResNet-18 on standard image classification benchmarks.

We present a systematic benchmark of out-of-distribution (OOD) detection CSFs through a representation-centric lens. Our study spans CNN and ViT backbones, multiple training paradigms, four image-classification source datasets (CIFAR-10, CIFAR-100, SuperCIFAR-100, and TinyImageNet), and OOD datasets grouped into near, mid, and far regimes using CLIP-derived semantic distances. To compare CSFs across these settings, we employ a multiple-comparison-controlled rank pipeline that identifies top cliques of statistically indistinguishable winners under threshold-free ranking metrics (AURC and AUGRC). The main empirical finding is that the competitive detector family depends more on the learned representation than on score design alone. For both CNNs and ViTs, simple probabilistic scores dominate misclassification detection. On CNNs, margin-based scores are strongest in near-OOD regimes, while geometry-aware scores such as NNGuide, fDBD, and CTM become more competitive as shift severity increases. On fine-tuned ViTs, the top cliques are led mainly by reconstruction- and residual-based scores. To interpret these ranking shifts, we analyze the last-layer representation using Neural Collapse (NC) metrics. The resulting picture is consistent across architectures: prototype- and boundary-aware scores become stronger when the representation is more collapsed and better aligned with classifier weights, whereas weaker-collapse regimes favor gradient- and manifold-based scores. Building on these insights, we propose two contributions: a simple PCA-based projection-filtering procedure that improves detector performance, and an approach that uses NC measurements computed from a trained classifier to predict its competitive out-of-distribution detector shortlist, without requiring any additional OOD data.

Safety alignment in language models operates through two mechanistically distinct systems: refusal neurons that gate whether harmful knowledge is expressed, and concept neurons that encode the harmful knowledge itself. By targeting a single neuron in each system, we demonstrate both directions of failure --- bypassing safety on explicit harmful requests via suppression, and inducing harmful content from innocent prompts via amplification --- across seven models spanning two families and 1.7B to 70B parameters, without any training or prompt engineering. Our findings suggest that safety alignment is not robustly distributed across model weights but is mediated by individual neurons that are each causally sufficient to gate refusal behavior --- suppressing any refusal neuron bypasses safety alignment across diverse harmful requests.

Gromov--Wasserstein optimal transport (GWOT) aligns metric measure spaces by matching their within-domain relational structures, but large-scale GWOT remains challenging because its objective is nonconvex and projection onto the transport polytope is often solved only approximately in practice. This leads to a gap between practical projected-gradient implementations and convergence theory, which typically assumes exact projections. For squared-loss GWOT, we propose an inexact projected-gradient framework with a verifiable feasibility-residual-based inexact condition for the projection subproblem. This condition is directly computable and avoids unknown quantities such as the exact projection point. Under this implementable condition, we prove subsequential convergence to stationary points and, with a mild tolerance-decay condition, convergence of the whole sequence. The resulting method retains the simplicity and sparsity of projected-gradient schemes while providing rigorous convergence guarantees, turning projected-gradient methods into a principled and scalable approach for GWOT with provable reliability.

While the maximal coding rate reduction (MCR$^2$) objective has become a useful principle for learning compact and discriminative representations, a theoretical understanding of its optimization geometry remains limited, especially on why first-order methods often exhibit fast convergence near its local maximizers. In this work, we show that the MCR$^2$ objective satisfies an error-bound property in a neighborhood of its local maximizers. This property guarantees that the distance to the local maximizer set can be bounded above by the gradient norm at any point in the neighborhood, directly implying regularity conditions such as the Polyak-Lojasiewicz inequality and quadratic growth. Leveraging this property, we prove that first-order methods converge linearly to a local maximizer of the MCR$^2$ objective under mild conditions. Numerical experiments validate the predicted local convergence behavior and provide empirical evidence that optimizing the MCR$^2$ objective improves representation geometry while maintaining competitive downstream accuracy.

\emph{Scaling} has been the central driver of progress in modern AI, yet data curation remains an exception: it still relies on manual heuristics rather than systematic, scalable methods. In practice, data recipes (e.g., filter thresholds, deduplication strength, domain mixing ratios) are determined through many small-scale ablation training runs. This work demonstrates that the performance of the best-discovered recipe improves \emph{predictably}--as a \emph{power law}--in the number of such ablation runs. Theoretically, the power-law form arises naturally from regret bounds in zero-order optimization theory. Empirically, we validate it across diverse training stages, including pretraining, supervised fine-tuning, preference optimization, and reinforcement learning. Building on this finding, we derive a principled approach to \emph{data quality--quantity tradeoff} under a fixed compute budget, yielding a Pareto frontier that dominates standard heuristic pipelines (e.g., DCLM). Together, this work provides a general framework for scaling data-curation compute, establishing it as a new axis alongside the well-studied pretraining and test-time scaling.

Online data selection is typically driven by per-sample scores such as loss, uncertainty, or gradient norm, which measure how strongly the model reacts to a candidate but not what parameter update the candidate would induce. What an online learner should select is therefore not the most striking example, but the one whose update is most worth taking. To make this concrete, we score a candidate by the movement induced by its update, using a Fisher-whitened information gain that evaluates the update in the loss geometry rather than by raw gradient size. Movement, however, is not yet progress, so we pass this score through a \emph{Descent Alignment Gate}, a soft factor that retains it only when the induced update agrees with the descent direction. This yields a sample-level score favoring updates that are both informative and descent-aligned. To make the criterion compatible with training, we extend it from samples to batches, where selected updates should not only score well individually but also complement one another. We capture this with the \emph{Gated Fisher Volume}, a log-determinant batch objective that reduces to the per-sample score on a singleton, grows with the volume spanned by the chosen updates in the Fisher geometry, and admits a greedy algorithm with a constant-factor approximation guarantee. Across supervised fine-tuning benchmarks, our selector consistently outperforms online selection baselines under matched training budgets. More broadly, the framework reframes data selection itself: rather than a spotlight on conspicuous examples, it acts as a compass toward updates worth taking.

Projection-based optimization offers a gradient-free alternative to conventional gradient-based methods, but existing frameworks suffer from high computational costs and poor scaling with depth. We introduce $\mathcal{P}$Torch, a projection-based framework that builds directly on top of PyTorch's autograd engine, eliminating the overhead of prior approaches. We further improve performance through nonlinear relaxation and memory-efficient projections. Together, these contributions make the algorithm competitive with gradient-based methods for shallow MLPs on MNIST and CIFAR-10. Finally, we prove a vanishing target theorem showing that the layer-wise learning signal decays with increased depth, serving as an analogue of the vanishing gradients problem for projection-based methods.

Optimizing non-convex functions is a fundamental challenge across machine learning and combinatorial optimization. We introduce and study $\gamma$-weakly $\theta$-up-concavity, a novel first-order condition that characterizes a broad class of such functions. This condition provides a powerful unifying framework, strictly generalizing both DR-submodular and One-Sided Smooth (OSS) functions while capturing broader forms of scale-dependent curvature, including accumulating-then-diminishing returns and flat-start behavior. Our central theoretical contribution demonstrates that $\gamma$-weakly $\theta$-up-concave functions are upper-linearizable: for any feasible point, we can construct a linear surrogate whose gains provably approximate the original non-linear objective. A key technical contribution is a nonuniform upper-linearization argument yielding approximation coefficients that depend explicitly on the curvature parameters and the geometry of the feasible region. This linearizability yields immediate and unified approximation guarantees for a wide range of problems. Specifically, we obtain unified approximation guarantees for offline optimization as well as static and dynamic regret bounds in online settings via standard reductions to linear optimization. Moreover, our framework recovers the optimal approximation coefficient for DR-submodular maximization and improves existing approximation coefficients for OSS optimization, particularly over matroid constraints.