NeurIPS 2026 Papers

Self-supervised Vision Transformers learn strong image-level representations, but their patch-level representations can be unreliable for dense prediction. We study this mismatch and find that the class token concentrates global semantics on a small number of representative patches, while dense prediction errors often occur on low-attention regions. We provide a local gradient analysis showing that, along the [CLS] attention readout, image-level self-distillation gradients are weighted by [CLS]-to-patch attention mass. To provide masked patches with global semantics without changing the architecture, we propose Cross Distillation (CODI), which mixes a small amount of the teacher class-token distribution into each teacher patch-token target. CODI gives every masked patch an attention-independent semantic correction while preserving the standard self-distillation architecture. Experiments on dense and image-level benchmarks show that CODI improves both dense and image-level performance under aligned pretraining settings.

Fine-tuning Large Language Models under stringent memory constraints remains a significant challenge. Recently, Zeroth-Order (ZO) optimizers have been employed to fine-tune quantized models, avoiding backpropagation and further reducing memory usage. However, quantized ZO optimizers often underperform compared to their full-precision counterparts, especially at low bit-widths. We find that existing quantized ZO methods fail to exploit the inherent low-rank structure of gradient updates, thereby limiting their effectiveness. To address this, we propose the **L**ow-rank **Q**uantized **Z**eroth-**O**rder optimizer (LQZO), which enhances quantized ZO fine-tuning by leveraging the low-rank structure inherent in gradients. Our method introduces three key innovations: (1) A low-rank parameterization of gradient perturbations using structured matrix factorization. (2) A Binary-Aggregated Gaussian perturbation strategy that employs a Rademacher distribution to constrain higher-order moments and tighten the variance bound of gradient estimates, with no additional computational overhead. (3) An Isotropy-Enhanced Estimation technique that reshapes quantization scale groups into more isotropic structures to diversify exploration directions. We provide a theoretical convergence guarantee for LQZO. Extensive experiments demonstrate that LQZO consistently outperforms quantized ZO optimizers, achieving average performance gains of 2.0%--3.3% across quantization precisions. While extreme quantization inevitably leads to information loss, LQZO significantly improves the trainability of INT2 models and narrows the performance gap compared to prior quantized ZO methods. Furthermore, we demonstrate the cross-modality versatility of LQZO by validating its superior performance on Vision Transformers.

The standard in LLM-based prediction is to use the final-layer representation as the input to a downstream predictor. However, intermediate layers may encode complementary task-relevant signals. Existing approaches therefore either search for the best layer for each task or apply expensive attention-based mechanisms to learn inter-layer aggregation. In this work, we first show that such complexity is unnecessary: a lightweight Graph Neural Network over a fully connected graph of LLM layers is more efficient and achieves significantly stronger predictive performance than existing approaches. We then introduce the Cayley-Encoder, which further improves both efficiency and predictive performance by replacing the fully connected graph with a Cayley graph over $SL(2,\mathbb{Z}_n)$. These Cayley graphs provide a mathematically grounded topology that is sparse, regular by construction, and has low diameter. This enables effective communication across layers while constraining the aggregation structure and reducing the risk of GNN overfitting. In an evaluation of Cayley-Encoder across 13 tasks and 9 LLMs, Cayley-Encoder consistently outperforms baselines, achieving improvements of up to 40 percentage points in accuracy, while introducing at most 0.1\% additional parameters relative to the LLM size. We further show that Cayley-Encoder is effective in few-shot regimes. Finally, we show that Cayley-Encoder outperforms LoRA fine-tuning while operating on the frozen LLM. We conclude with an explainability analysis showing that multiple layers contribute meaningfully to the final prediction, supporting our hypothesis.

Classifier-free guidance (CFG) is essential for improving sample quality in visual generation, but it incurs substantial sampling overhead. Existing guidance-free approaches halve this cost by approximating the CFG with a single model inference, but their performance is bounded by the CFG and inherits its limitations. In this paper, we propose self-contrastive guidance (SCG), which contrasts the model's prediction on real and self-generated samples to obtain a direct, model-based signal that pushes generation toward the true data manifold. Integrating SCG into the standard likelihood objective via reparameterization yields Self-Contrastive Alignment for Likelihood Estimation (SCALE), a guidance-free training framework that produces SCG-enhanced predictions in a single forward pass. Across class-conditional and text-to-image generation on diffusion and autoregressive models, SCALE consistently outperforms CFG-mimicking guidance-free baselines and improves further when combined with DPO, demonstrating that the self-contrastive signal yields gains complementary to both CFG and preference-based fine-tuning.

Building general-purpose role-playing agents that faithfully portray any character from a natural-language profile remains challenging. The dominant paradigm---supervised fine-tuning---encourages behavioral mimicry without deep, human-like internal thought processes, resulting in poor out-of-distribution generalization. Therefore, we propose Psy-CoT, a psychology-grounded chain-of-thought framework that decomposes pre-response reasoning into three role-specific steps---Interaction Perception, Psychological Empathy, and Logical Construction---so that the model thinks dynamically from the profile rather than merely mimicking surface patterns. While structured reasoning provides a foundation, it alone is insufficient; reinforcement learning is essential to further align the model with character fidelity. However, we observe that under LLM-based reward models, both generic phrases that hack the reward model and genuinely role-specific phrases receive identical gradient signals---this hacking accumulates over training, misleading the model into treating both as equally optimal choices. To address this, we propose Role-Aware Policy Optimization (RAPO), which uses profile--token mutual information to weight gradients asymmetrically---amplifying role-specific tokens under positive advantage while attenuating them under negative advantage. Experiments on CoSER, CharacterBench, and CharacterEval demonstrate that Psy-CoT outperforms existing role-playing CoT methods, and RAPO consistently surpasses GRPO across multiple model scales.

Generative models for function-valued data, such as time series and solutions to partial differential equations, must learn distributions over infinite-dimensional spaces rather than over finite-dimensional vectors. Functional Flow Matching (FFM) is a recent extension of Flow Matching to this setting, learning a velocity field whose flow transports a Gaussian prior to the data distribution. However, FFM inherits a structural limitation from standard Flow Matching: in each training batch, prior and data samples are paired independently, so the conditional bridge between them must simultaneously traverse the shared global structure of the dataset and instance-specific residuals. This limitation is more consequential in function space than in finite dimensions: directly formulating optimal transport on function spaces is technically delicate, and a flat Euclidean surrogate ignores the function-space geometry that distinguishes function-valued data. We propose kernel Functional Flow Matching (kFFM), which replaces the independent pairing by entropic optimal transport in a kernel-induced Hilbert space via the Hilbert Sinkhorn Divergence, leaving the FFM neural-operator architecture unchanged. Theoretically, we prove the HSD objective is well-posed on Banach ambient spaces with bounded kernels, derive compact-metric approximation bounds against quadratic-cost OT with an explicit kernel-cost mismatch term, and quantify the gap between the function-space objective and the truncated representation computed on a grid, at a rate governed by data regularity. Empirically, kFFM consistently improves MMD-RBF distributional matching over FFM, diffusion, adversarial, and finite-dimensional OT baselines on time-series and PDE benchmarks, with function-space-aware kernels emerging as the strongest choices on every PDE and path-valued benchmarks.

Diffusion-based posterior sampling (PS) is a leading framework for imaging inverse problems, combining learned priors with measurement constraints. Yet, its standard formulations rely on instantaneous data-consistent estimates, which induce temporal variability in the reverse dynamics. We reinterpret PS from a dynamical perspective, showing that the standard PS update corresponds to a first-order discretization of the diffusion dynamics plus a residual correction capturing the mismatch between the denoised prediction and the data-consistent estimate. A second-order discretization, however, naturally introduces a temporal correction based on the variation of consecutive estimates. Building on this, we propose LAMP, combining the second-order update with the residual correction characterizing a PS technique. LAMP thus inherits a lagged temporal correction, and it can be implemented as a modular plug-in over the PS backbone. We show that LAMP preserves the structure of a posterior sampler, and we perform a one-step risk analysis to characterize when LAMP improves the reverse transition via a bias-variance trade-off. Experiments across multiple imaging tasks demonstrate consistent improvements over strong baselines such as DiffPIR and DDRM, without increasing the number of denoising evaluations.

Effective Continual Video Instance Segmentation (CVIS) requires balancing catastrophic forgetting with the plasticity to learn new tasks. Existing methods typically adapt pre-trained VIS models using trainable visual prompts, but suffer from two key limitations: they overlook the distinct degradation of frame-level spatial knowledge (e.g., misclassification) and video-level temporal knowledge (e.g., trajectory disruption), and rely on rigid prompt retrieval that struggles under task ambiguity and distribution shifts. To address these issues, we propose STEAM, a CVIS framework with Spatial-Temporal Balanced Mixture-of-Experts Adapters. Instead of explicit retrieval, we introduce a Distribution-Aware spatial-temporal Scaling and sHifting (DASH) mechanism to align the pre-trained feature space with sequential tasks and reduce the domain gap. Furthermore, our method adaptively learns Task-Adaptive Spatial-Temporal mixture-of-Experts (TASTE) at both the frame and video levels, while updating them with Subspace-Constrained Orthogonal Adaptation (SCOA) to mitigate interference. Extensive experiments across diverse CVIS settings show that STEAM preserves temporal-spatial coherence and achieves state-of-the-art performance.

Image classification models often encounter objects in diverse and unpredictable contexts at test time, which may differ substantially from those seen during training. This phenomenon, termed *context shift*, exposes a key fragility in neural networks: when models rely on non-causal contextual cues for prediction, shifts in context can lead to significant performance degradation. In this paper, we propose a novel training framework to improve the robustness of convolutional neural networks under context shift. Our approach leverages spatial feature attribution to guide models toward making predictions that rely less on contextual regions and more on object-relevant features. As a first step, we identify a theoretical limitation of existing feature attribution methods and introduce a new variant, *ContrastiveCAMs*, which produces more faithful attribution maps of model predictions. Building on ContrastiveCAMs, we further propose *Context-Regularized Cross-Entropy* (CR-CE), a modification of the standard cross-entropy loss that regularizes the model’s attention by suppressing the influence of contextual regions, thereby improving robustness to context shift. We evaluate the effectiveness of our approach on several medium-to-large scale datasets (Waterbirds, Spawrious, ImageNet/ImageNet-BG, Hard-ImageNet), and report consistent improvements in context-shift robustness.

Explaining model predictions in high-stakes settings requires knowing *why* a prediction was made, a central challenge in modern AI. Yet existing methods lack a principled causal framework for explanation, leading to a disagreement problem: explanation methods produce conflicting attributions with no way to determine which is correct. Based on the literature in cognitive science, law, and philosophy, and using the language of causality, we formalize four desiderata for explanation methods that assign numerical credit to variables. We show that no prior explanation method including SHAP and LIME satisfies all desiderata, and prove the Explanatory Impossibility Theorem, which shows that even in principle, no method can satisfy all desiderata without causal assumptions. We introduce *counterfactual Shapley values*, an aggregation of a novel counterfactual quantity we call the *natural total effect*, and prove they satisfy all desiderata. We develop an algorithm for computing tight bounds on counterfactual Shapley values. Experiments corroborate our theory.

A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety audit can support. We present two techniques that make audits in Petri, where an auditor model red-teams a target over multiple turns to elicit concerning behavior, harder to distinguish from real deployments. Our first technique, \emph{critique refinement}, spends additional inference-time compute on each auditor action: an instance of the target model scores candidate actions for realism, provides feedback, and the auditor iterates before the highest-scoring candidate is selected. On Sonnet-4.6, realism win rate (the fraction of pairings in which the audit transcript is judged more realistic than a real deployment transcript) rises monotonically from 12\% to 39\% as refinement depth increases, and verbalized evaluation awareness drops to near zero; Haiku-4.5 and Opus-4.7 show a similar pattern. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in a Claude Code agent harness, reducing the gap between auditor-simulated and real deployment environments; in coding settings DISH raises realism win rate from 7\% to 18\% on Sonnet-4.6. The gains are additive on Sonnet-4.6 (12pp over either alone) but not on Opus-4.7.

Confronted by the existing prototype-based methods that often fail to capture complex intra-class variations and inter-class semantic relationships, weakly supervised semantic segmentation (WSSS) faces persistent challenges due to the inherent incompleteness and boundary inaccuracies of Class Activation Maps (CAMs). To tackle these critical limitations, we propose a relationally optimized prototype Memory bank (RO-PMB) framework in this paper, where the integration of Graph Neural Networks (GNNs) with a dynamically optimized global prototype memory bank is pioneered to explicitly model and learn the intricate semantic relationships among all prototypes. Specifically, our Prototype Relationship Diagram Construction (PRDC) module leverages GNNs for contextual message passing over a learned prototype graph, enriching global prototypes with crucial associative knowledge. As a result, a subsequent context-aware refinement process is sustained to inject structured semantic information into local image-specific prototypes, thereby generating CAMs with unparalleled completeness and boundary precisions. Extensive experiments validate that, compared with the existing representative state of the arts in multi-stage WSSS baselines, our proposed RO-PMB achieves competitive and superior performances on a range of challenging benchmarks, such as PASCAL VOC and MS COCO etc. Code will be released.

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.

We study the problem of learning in multi-player general-sum Markov games. In the simulator setting, where the learner can sample the next state conditioned on an arbitrary state action pair, the previous best-known upper bound of the sample complexity for learning an $\varepsilon$-approximate coarse correlated equilibrium (CCE) is $\widetilde{O}(H^4S\sum_{i=1}^m A_i/\varepsilon^2)$ (Li et al., 2022), where $H$ is the horizon, $S$ is the number of states, and $A_i$ denotes the number of actions for the $i$-th player. This work improves the upper bound to $\widetilde{O}(H^4S\max_{i\in[m]} A_i/\varepsilon^2)$, matching the lower bound of $\Omega(H^4S\max_{i\in[m]} A_i/\varepsilon^2)$ (Li et al., 2022). In the online setting, in which the learner can only sample a trajectory, the previous best-known sample complexity upper bound for learning CCE in multi-player general-sum Markov games is $\widetilde{O}(H^6S\max_{i\in[m]} A_i/\varepsilon^2)$ (Song et al., 2021; Jin et al., 2024; Mao et al., 2022). In this work, we improve this to $\widetilde{O}(H^5S\max_{i\in[m]} A_i/\varepsilon^2)$. To our knowledge, this is the tightest upper bound that breaks the curse of multi-agency. The core of our algorithmic design and analysis is the new gap decomposition and, in particular, a bandit algorithm with a high-probability empirical variance regret bound, which might be of independent interest.

We study regret minimization for time-homogeneous episodic tabular reinforcement learning with $S$ states and $A$ actions under non-interactive $\epsilon$-local differential privacy (LDP). In each episode $k \in [K]$, the decision-maker sends a policy to a user, the user executes the policy for $H$ steps, and sends an $\epsilon$-locally differentially private message back to the decision-maker. When $H=1$, the problem reduces to an LDP contextual bandit with $S$ contexts, $A$ arms and horizon $K$. We prove that the minimax regret is $\min\bigg(K, \frac{ \sqrt{\max( S, \exp(\epsilon) + 1) S A K (\exp(\epsilon) + 1)}}{\exp(\epsilon) - 1}\bigg)$, up to constants and a $\sqrt{\log(A)}$ factor. The upper bound is achieved by an adaptation of EXP3 and a unary encoding mechanism. When $H > 1$, and for small $\epsilon$, we show a lower bound in $\Omega(HS\frac{\sqrt{A K}}{\epsilon})$ and an upper bound in $\tilde{O}(H^2 S^2 \frac{\sqrt{ AK}}{\epsilon})$. The upper bound is achieved by an algorithm based on UCBVI and a unary encoding mechanism. For small $\epsilon$, these bounds improve the previously known lower bound by a factor $\sqrt{S}$ and the previously known upper bound by a factor of $H \sqrt{A}$.

We study online multiclass classification with surrogate losses. We identify and exploit a structural property of margin-based classifiers: on many rounds predictions are stable in the sense that after an update of the parameters the algorithms do not change the predicted label on the current example. By exploiting this stability we improve upon the state of the art in several ways. We provide an improved surrogate regret bound for the perceptron, develop a parameter-free version of $\texttt{GAPTRON}$, develop improved results for the delayed feedback setting, and extend our results to the batch setting. These results are complimented by empirical evaluations.

Value iteration (VI) is the unit-step explicit-Euler discretization of a continuous-time *Bellman residual flow* $\dot V=\operatorname{BR}(V)$, and its standard linear dependence on the effective horizon $1/(1-\gamma)$ can be viewed as reflecting the step-size restriction that explicit integration imposes on stiff dynamics. We propose *Implicit-Euler Value Iteration* (IE-VI): a planning routine equivalent to the implicit-Euler discretization of the same flow, in which each step is unconditionally stable and contracts at rate $1/(1+h(1-\gamma))$ for every step size $h>0$. The implicit step requires solving an equation in the true Bellman operator, which we realize at finite cost by *iterative defect correction*: a short sequence of cheap planning rounds in an approximate model $\hat{\mathcal{P}} \approx \mathcal{P}$ (a simulator or a learned dynamics model), with the true model queried once per round to refresh a residual correction term. For Policy Evaluation and Control, IE-VI converges to the value function of $\mathcal{P}$ for any $\hat{\mathcal{P}}$, however inaccurate---in contrast to model-based planners, which generally converge to a biased fixed point. When the model error shrinks at a sublinear-in-horizon rate, IE-VI attains polylogarithmic iteration complexity in $1/(1-\gamma)$, an exponential improvement over VI; outside this regime its complexity stays within a logarithmic factor of VI. We also give an adaptive variant A-IE-VI, which removes the need to know the model-error scale, and a Dyna-style sample-based variant IE-Dyna for tabular RL.

Recognizing human value preferences is crucial for building reliable AI agents in user-centric fields like psychological counseling and strategic negotiation. Existing benchmarks rely on explicit questionnaires, assuming (1) users are always cooperative and willing to self-report; (2) value preferences remain static. However, humans may conceal true preferences if they feel probed, and may adopt distinct value preferences depending on the context. Consequently, explicit queries are insufficient for predicting value preferences reliably. To address this, we introduce Implicit Value Probing (IVP), a benchmark simulating realistic scenarios where agents implicitly infer value preferences through strategic, multi-turn conversations. IVP grounds these interactions using a diverse set of Persona Agents, covering a wide spectrum of personalities and values. We evaluate mainstream LLMs on IVP, including Qwen3, DeepSeek, Doubao, and GPT-5, observing substantial performance gaps. We further develop Value-Prober-8B, a specialized model internalizing structured cognitive reasoning via distillation. It employs strategic questioning to elicit concealed value preferences. Experimental results demonstrate that Value-Prober-8B can effectively balance probing accuracy with social acceptability, achieving capabilities comparable to frontier models.

We propose a neural exploratory framework for *variational problems over graphons*, i.e., symmetric measurable functions $[0,1]^2 \to [0,1]$, that arise as limits of dense graph sequences. Such problems are central to two parts of the mathematical literature: extremal graph theory, which studies graph parameter optimisation under given restrictions, often for those graphs with a large number of vertices, and the large deviation theory of dense random graphs, which characterises the structure of rare events through constrained graphon optimisation. We represent graphons by implicit neural networks and optimise graphon objectives by gradient descent. Our design combines three ingredients: a multi-scale sinusoidal residual architecture biased toward sharp, step-like graphons; an embedded solver that enforces a single empirical density constraint and is differentiated by implicit differentiation; and symmetry-aware Monte Carlo estimators. On generalised Turán problems that previously required substantial human effort, our framework rediscovers known optimal graphons without human intervention. Applied to open instances, it produces candidate optima, including a previously unreported family of extremal structures. Applied to the variational problems from the large deviation theory of Erdős–Rényi random graphs, the framework produces new candidate optimal graphons across both upper- and lower-tail regimes for multiple pattern graphs; for the case of the triangle graph and the upper-tail regime, these candidates improve upon the best-known reference construction of Lubetzky and Zhao.