NeurIPS 2026 Papers

Distributionally robust optimization (DRO) provides a principled framework for learning under distribution shift, but its practical use is hindered by the difficulty of evaluating worst-case risks for nonconvex loss functions. We study a penalized DRO formulation in which the adversary may choose any distribution but incurs a Wasserstein penalty for deviating from the empirical distribution. We show that the adversary’s problem can be reformulated as an optimization problem over transport maps that push empirical samples to adversarial ones, and we prove that optimal maps are cyclically monotone. We also show that standard adversarial training---based on per-sample local optimization---violates cyclical monotonicity and wastes transport costs unless the adversary is severely restricted. We propose two remedies. First, we introduce multi-start particle ascent, which alternates parallel gradient ascent with reassignment to enforce cyclical monotonicity across samples. Second, we parameterize adversarial maps as gradients of input-convex neural networks, which guarantees cyclical monotonicity by construction. Experiments on robust regression, image classification, and robust control show that our methods consistently outperform standard adversarial training and state-of-the-art baselines, achieving improved robustness and better generalization under distribution shift.

Designing DNA libraries is a key challenge from drug design to protein engineering and synthetic biology. Modern generative models offer opportunities to navigate the design space and propose specific sequences predicted to be effective in-silico. Designing deterministic libraries of specific sequences is however limited by the cost of DNA synthesis -- the synthesis barrier. In contrast, high-throughput multiplexed screening can measure the function of billions of biological sequences in parallel. Harnessing this technology requires the design of randomized libraries with specific design constraints to achieve low synthesis costs. In practice, such stochastic libraries are often chosen heuristically, sacrificing control for scale. Is there a way to bridge AI-based in-silico sequence design with high-throughput experimentation? In this work, we introduce Policy Gradients for Library Design (PGLD). PGLD uses a synthesis-aware parametrization of stochastic DNA libraries and optimizes them against a specified objective function. This allows for designing massive, controlled libraries without being limited by synthesis costs. We show how PGLD enables lab-in-the-loop design of multi-round high-throughput experiments, and large-scale in-vitro DNA sampling from generative models. Finally, we use PGLD to design a library of $\sim 10^6$ unique sequences at a cost of $\sim 700$ USD to explore the mutation space of a broadly neutralizing influenza antibody.

Low-label transfer is often decided by a practical question: after a representation looks useful, how many labels are still needed to learn the downstream boundary? Existing transferability and probe-quality scores rank global alignment, likelihood, margins, or calibration, but they do not measure how much target data remains close to the fitted decision boundary, where scarce labels are most costly. We propose **effective boundary complexity**, a label-budget diagnostic for frozen representations and linear probes. It measures mass in a thin strip around the learned boundary and normalizes it by class balance, so dense ambiguous regions and rare-class difficulty are both counted. We prove that, under a smooth representation map, this boundary mass does not change because of generic volume expansion: density change and tangential surface change cancel at first order, leaving only stretch or compression in the boundary-normal direction. This yields a cached-feature estimator that fits a probe on one split, counts held-out features near its boundary, and avoids high-dimensional density estimation. Experiments support the diagnostic at three levels. Controlled transformations match the predicted law across 74 maps with 0.38% median relative error. On real frozen features, normal-direction changes move both measured boundary mass and label demand, while tangent changes do not. A differentiable surrogate reduces measured complexity by 16–36% at matched top-1 utility. Finally, CIFAR-100 and Tiny-ImageNet audits show boundary mass acting as a second-stage diagnostic for near-tied representations, with external multiclass results showing how label-grid saturation limits policy-level gains on label grids where many tasks tie.

Diffusion policies have achieved strong performance in reinforcement learning (RL) due to the exceptional expressivity in capturing complex action distributions. However, existing approaches focus primarily on positive samples and lack explicit modeling of negative samples, thereby failing to fully exploit this representation capacity and ultimately limiting performance. To mitigate this problem, we propose BOBAC (BOotstrapped Bipartite Actor-Critic), a novel method for diffusion policy optimization via pairwise preference learning. Bridging the gap between distribution modeling and preference learning, we reformulate diffusion policy optimization within a preference-aware framework via the Bradley-Terry (BT) modeling, which enables the effective exploitation of negative samples through relative ranking. Since the original BT loss assigns the same weight to all pairs regardless of their varying importances, we further design a soft-gate mechanism to adaptively weight each pair according to its Q-gap. Experimental results on 6 MuJoCo tasks and 4 H1Bench tasks demonstrate that, compared with representative model-free and generative-policy RL baselines, BOBAC achieves state-of-the-art performance with consistent improvements in all continuous-control environments.

Post-training pruning of large language models (LLMs) zeroes a fraction of the weights and reconstructs the survivors without retraining, addressing the weight-memory bottleneck of LLM serving. At a fixed mask the per-row reconstruction admits a closed-form Optimal Brain Surgeon (OBS) optimum, but the cost of the exact joint solve is cubic in the layer dimension; one-pass methods such as SparseGPT and Wanda approximate it at quadratic cost without attaining the per-block joint optimum. We introduce Block-OBS-GS, which solves the per-block joint OBS reconstruction exactly within SparseGPT's complexity class via a per-row Cholesky factorisation of the in-block damped Hessian (one third the cost of an explicit inverse, reused across the algorithm) and follows it with a single cross-block Gauss--Seidel sweep. We prove three guarantees: per-layer FLOP count in SparseGPT's complexity class, with the leading $mn^{2}$ constant dropping from $1/2$ (SparseGPT) to $3\rho/2$ (Block-OBS-GS) where $\rho=1-s$, hence strictly lower for sparsity $s>2/3$; per-row reconstruction error at most SparseGPT's, strict whenever the activation Hessian has non-zero off-diagonal coupling; and a closed-form distance bound to the unregularised joint-OBS optimum that contracts geometrically per sweep. Across six Llama-3 and Qwen-3 backbones from 3B to 70B parameters, Block-OBS-GS attains the lowest WikiText-2 perplexity at every tested sparsity on the three Llama-3 backbones, with a $47\%$ perplexity reduction over SparseGPT on Llama-3.1-8B at $s=0.875$; the highest seven-task commonsense Avg7 on six of twelve populated cells; 5-shot Massive Multitask Language Understanding (MMLU) within $0.026$ of SparseGPT across populated cells; and the lowest WikiText-2 perplexity on four of five backbones under $2{:}4$ structured sparsity.

We study an endogenous nonstationary stochastic bandit problem with latent linear dynamics, where actions affect both immediate rewards and the future evolution of an unobserved latent state. Rewards are bilinear in the current action and latent state, inducing history-dependent rewards and a nontrivial long-horizon planning problem. The existing explore-then-commit approach (Choi et al., 2026) achieves $\tilde{\mathcal{O}}(T^{2/3})$ regret by uniformly exploring to estimate the latent dynamics and then committing to an optimized open-loop action sequence. We show that this rate can be improved via adaptive block-level optimism. Our key step is a cyclic approximation: under stable dynamics, the infinite-memory reward process can be truncated, and the open-loop benchmark can be approximated by optimizing a finite-memory block-level proxy. Building on this reduction, we propose a UCB-based block algorithm that maintains confidence sets for the truncated dynamics parameters and selects blocks optimistically. We prove a regret bound of order $\tilde{\mathcal{O}}(\sqrt T)$, significantly improving over the previous $\tilde{\mathcal{O}}(T^{2/3})$ guarantee for the same model. To the best of our knowledge, this is the first $\tilde{\mathcal{O}}(\sqrt T)$ regret guarantee for latent linear-dynamics bandits with bilinear reward observations and an open-loop action-sequence benchmark.

Multimodal Large Language Models (MLLMs) have shown promise as interactive agents, yet precise, programmatic 3D graphics editing remains difficult to train because existing systems largely rely on closed-source model APIs and lack deterministic, state-verifiable execution environments. We present BlenderFORGE, a trainable data-environment-optimization framework for open-weight MLLM agents that edit explicit Blender scenes through Blender Python (bpy) code generation. BlenderFORGE targets four fundamental state-verifiable editing task families: object placement, shape-key editing, lighting adjustment, and material modification. It integrates three components: (1) a scalable perturbation-based data generation pipeline built from curated assets and large-scale 3D scene repositories; (2) an isolated Blender Sandbox for deterministic code execution, structured state export, and closed-loop reward computation; and (3) a two-stage training pipeline that uses offline teacher trajectories for cold-start Supervised Fine-Tuning (SFT) and sandbox-computed rewards for multi-turn tool-agentic Reinforcement Learning (AgentLoop-RL). Experiments across the four task families show that the resulting open-weight Qwen3-VL-8B-BlenderFORGE agent substantially improves over its base model and achieves competitive performance with several zero-shot proprietary MLLM baselines.

We study mathematical programs with equilibrium constraints, in which a leader knows their own cost function, but lacks a model of the followers' response. Instead, the leader can only query this response at specific points. While this setting precludes the use of gradient-based methods, existing zeroth-order approaches treat the composed objective \emph{entirely} as a black box, deploying zeroth-order tools across both the leader and follower. Such approaches are inefficient, as they discard information the leader already possesses about their own cost function. In this work we instead propose to deploy zeroth-order tools only where they are truly needed: to handle the unknown, non-smooth followers' response. Specifically, we first propose \algpzos, an algorithm that combines exact partial gradients of the leader's cost with zeroth-order Jacobian estimates of the followers' response in a chain-rule-inspired manner, and establish that it achieves a strictly lower variance bound than the black-box baseline. Second, we introduce the partial Goldstein subdifferential, a stationarity notion tailored to this composite structure, and prove convergence of our algorithm to both standard and partial Goldstein stationary points. Finally, we validate our method on two application domains -- toll optimization in routing games and defense-attack investment in security games -- demonstrating consistent improvements over black-box baselines in convergence speed, objective value, and estimator variance, with robust performance even under few queries per iteration.

A language model agent acts on a context assembled by retrieval, summarisation, prompt templates, and memory. How many decision-relevant bits must this context carry before the agent can act with low regret? We give an answer that does not depend on the model, decoder, or scaffold. Modelling the agent as a Markov chain H→C→A and treating Bayes regret (the value gap between the chosen and the optimal action) as the distortion measure, we define a regret rate-distortion (RRD) function Rℓ(ε) whose converse is a one-line consequence of the data-processing inequality (DPI). A context that carries fewer than Rℓ(ε) bits cannot drive expected regret below ε, regardless of decoder, scaffold, scale, or chain-of-thought. For finite M-ary decisions under zero-one loss the bound reduces to the Fano floor in closed form, requiring ≈3.14 bits at (M=16, ε=0.1) and ≈6.73 bits at (M=256, ε=0.1). We instantiate the theory on 441 controlled single-step tool-selection cells, a balanced 16-way classification proxy spanning three open-source large language models (Qwen3-32B, Gemma3-27B, GLM-4-32B) and three public benchmarks (API-BANK, τ-bench, ToolBench). Because the gold tool index X=f(H) is a deterministic function of the history, the action-side rate IMM(X; A) lower-bounds the context capacity by data processing, so an action-side test of the converse is strictly tighter than a context-side one. Every cell respects the predicted floor (median margin +1.07 bits). Across 629 matched-rate pairs, 97.5% share overlapping regret confidence intervals (CIs), and a chain-of-thought supplement slides cells along the frontier rather than across it. The same data exposes a tokens-versus-bits gap reaching ∼1.7×10⁶:1, with the model extracting fewer than 0.05 bits about the gold action on a 32,768-token ToolBench prompt. In the single-step regime, the framework supplies a falsifiable, model-agnostic stopping criterion at IMM=Rℓ(ε) and reframes context engineering as an information-allocation problem.

Recent advances in gigapixel-level imaging have brought High-Resolution Wide shots to the forefront of research. However, these images present significant challenges: extreme sparsity of foreground, gigapixel-level resolutions and interleaving of foreground and background. This causes traditional detectors and attention mechanisms to be hindered by background, resulting in inefficiency and inaccuracy. To tackle this problem, we propose BiHashFormer, a dual sparsified branch transformer built on hashing. RoI Selector will predict the foreground proportion for top-k selecting target-containing windows. The compression branch processes windows selected and refines attention into HashAttention by discarding half of the key–value pairs, reducing cost and improving efficiency. The compensation branch uses HashMiner to perform low-cost hash searches in the remaining regions, recovering features of missed objects. Experiments confirm that the selection relationship in HashAttention is one-way and reveal the preference of different queries for keys. In experiments on the gigapixel benchmark PANDA, BiHashFormer reduces 37.5\% backbone FLOPs while improving $\text{AP}_{50}$ to 81.0\%. Moreover, this parallel dual-branch structure allows the model to avoid inference delays caused by inter-branch dependencies, while achieving much greater stability than single-branch models, specially with low rates of RoI selection.

Deep learning models achieve high accuracy in signal processing but remain difficult to interpret. Classical matching pursuit and sparse coding offer better transparency but suffer from spectral leakage and grid mismatch when signal frequencies do not align with discrete dictionary atoms. We introduce Continuous Dictionary Pursuit (CDP), a framework for interpretable signal decomposition that learns the parameters of analytical functions directly from data. Unlike traditional methods, CDP optimizes over a continuous manifold of differentiable atoms to resolve the limitations of fixed grids. The algorithm employs a greedy iterative strategy using data-driven priors and gradient-based optimization to isolate individual signal components. We provide a library of parametric atoms that covers periodic, trend, and transient structures, including discontinuous functions modeled through spectral annealing. To ensure a sparse atom set for reconstruction, the framework utilizes a stopping criterion based on a predefined atom budget and residual error convergence. Evaluations on synthetic and real-world datasets show that CDP resolves spectral leakage for the parameterized atom families considered and recovers exact mathematical components. Beyond decomposition, we demonstrate that the extracted atoms can serve as structured curriculum targets for training sequence models, leading to consistent improvements in forecasting accuracy compared to standard end-to-end training. The method offers a glass-box alternative for signal decomposition by providing high reconstruction fidelity alongside direct physical interpretability.

Activation steering has emerged as a promising alternative for controlling language-model behavior at inference time by modifying intermediate representations while keeping model parameters frozen. However, large-scale evaluations such as AxBench show that existing steering methods are often outperformed by simple in-context prompting and generalize poorly to unseen concepts. We hypothesize that these limitations arise from unvalidated simplifying assumptions shared across prior methods, which typically restrict steering interventions to fixed, single-step, position-invariant transforms. We propose FLAS (Flow-based Activation Steering), which learns a general, concept-conditioned velocity field $v_t(h,t,c)$ that transports unsteered activations to steered ones without relying on these assumptions. On AxBench, FLAS is the first learned method to consistently outperform prompting, reaching held-out harmonic means of $1.015$ on Gemma-2-2B-IT and $1.113$ on Gemma-2-9B-IT without per-concept tuning. Analysis of the learned flow shows curved, multi-step, token-varying trajectories, which suggests that previous hypotheses on activation space geometry might be incomplete. Our code is available at https://anonymous.4open.science/r/FLAS.

Vision encoders for retrieval are typically trained with class-label supervision: each training pair reduces to a scalar that uniformly pushes the embedding apart or pulls it together, as if every visual attribute either differed or matched. A multimodal large language model (MLLM), shown the same pair, can articulate those attributes and use them to predict whether the images share a class. We propose \textbf{SAGA}, a framework that turns this language-grounded, attribute-aware perception into a training signal for the encoder itself. Specifically, we use Group Relative Policy Optimization (GRPO) to reward the MLLM for correct predictions on the vision encoder's tokens. Since correct predictions require those tokens to expose the specific attributes that differ or match between the pair, the gradient pushes the encoder to encode them, replacing the uniform pair-level scalar with attribute-resolved supervision. An auxiliary attention-distillation loss anchors the encoder's embedding to tokens the MLLM attended to, and a standard metric-learning loss shapes the embedding geometry for nearest-neighbour retrieval. The MLLM is frozen throughout and discarded at inference, matching the deployment cost of a metric-learning baseline. SAGA improves Recall@1 by 3 to 6 points over state-of-the-art baselines on CUB-200-2011, Cars-196, FGVC-Aircraft, and iNaturalist Aves on zero-shot image retrieval.

Modern end-to-end autonomous driving systems often suffer from fragile out-of-distribution (OOD) generalization and limited decision transparency, largely due to their reliance on raw sensory observations and susceptibility to spurious visual correlations. Although memory-augmented approaches have been explored to improve generalization, existing methods typically store raw observations, symbolic rules, or entangled latent features, making retrieval sensitive to surface-level similarity rather than decision-relevant driving structure. To address this limitation, we propose DriveMem, a plug-and-play framework that moves beyond raw observations by distilling and storing invariant driving memories for generalizable driving systems. DriveMem integrates two core components: Invariant Feature Abstraction that first distills perturbation-stable, decision-relevant representations from visual tokens, and Prototype Memory Bank that then stores and retrieves prototypical driving experiences grounded in invariant representations. Extensive open-loop and closed-loop experiments demonstrate that DriveMem improves OOD generalization and safety-related planning metrics across diverse unseen scenarios. Notably, under severe sensor disruptions, DriveMem achieves up to a 39.5\% relative reduction in Collision Rate compared with a strong domain-generalization baseline. These results suggest that building prototype memory in an invariant representation space is a promising approach for improving robustness, while offering prototype-grounded traceability for model behavior.

Visual foundation models (VFMs) provide strong region-level representations for semantic segmentation, substantially improving high-level semantic understanding. However, accurate boundary recovery remains challenging, especially for fine structures and semantic transitions. A key reason is that conventional pixel-level supervision largely treats pixels independently and lacks explicit modeling of local structure and neighboring-pixel relationships. To address this issue, we propose a Local Structure (LS) regularization framework that constrains the second-order local organization of semantic transition regions. Unlike existing boundary losses that mainly supervise contour location, distance, overlap, or topology, LS models boundary neighborhoods through a structure-tensor representation of semantic probability fields. It decomposes local boundary structure into two complementary cues: Local Structure Magnitude (LSM), which measures the strength of anisotropic class transitions, and Local Structure Orientation (LSO), which describes their dominant geometric orientation. LS is model-agnostic, compatible with different VFM-backbone-decoder combinations, and introduces no extra inference cost. Extensive experiments on four diverse datasets show that LS consistently improves boundary-sensitive performance while also enhancing overall IoU, demonstrating its effectiveness for boundary-structure learning in semantic segmentation.

Model merging has emerged as a promising paradigm for composing large language models directly in weight space, enabling training-free integration of specialized models. However, existing methods rely on parameter-space averaging that systematically induces generation dysregulation---a spectrum of failure modes ranging from structural repetition and termination failure to circular reasoning, and at the extreme, full semantic collapse into incoherent symbols. Even when source models exhibit near-zero dysregulation, existing methods introduce it at $>$97\% (14B/32B scales), collapsing reasoning benchmarks by up to 70 points. We propose Sparse Complementary Fusion with Reverse KL (\textbf{SCF-RKL}), a data-free merging framework that selects complementary parameters via reverse KL divergence on parameter-group proxy distributions. This structure-preserving, sparsity-inducing design maintains proxy-space geometry---theoretically motivated via entropy and subspace bounds---and empirically suppresses generation dysregulation while integrating new capabilities. Extensive experiments on 24 benchmarks across 7B--32B models spanning reasoning, instruction following, safety, and vision demonstrate that SCF-RKL achieves the best overall performance across scales while maintaining near-zero dysregulation ($<$1\%) and strong generalization.

In multimodal classification, late-fusion approaches classify concatenated modality-specific features extracted by unimodal neural networks. When modality imbalance is pronounced, various regularization techniques have been proposed to balance the learning process and overcome the inferior performance of late-fusion networks. In contrast, this work demonstrates that multimodal data can be effectively classified without any explicit modality fusion, using deep ensembles of unimodal networks. We systematically compare deep ensembles to late-fusion networks at equal parameter count and show that ensembles consistently outperform state-of-the-art late-fusion methods designed to address modality imbalance. This advantage also holds over intermediate-fusion techniques we evaluated and over hybrid methods that combine unimodal and multimodal predictions. We propose and empirically validate a method for selecting the number of models per modality in an ensemble, avoiding computationally expensive exhaustive search. Under extreme modality imbalance and small ensemble sizes, the heuristic indicates that ensembles of unimodal models trained solely on the stronger modality are preferable; as the ensemble scales up, incorporating models from the weaker modality becomes beneficial. Both predictions align with our empirical findings. To systematically explore the challenges of optimizing multimodal models, we propose a synthetic multimodal framework that allows control over both the number of modalities and their predictive strength; our findings are consistent across synthetic and real-world datasets. Finally, by fitting scaling laws to bimodal datasets, we estimate the asymptotic performance of ensembles.

Generalized linear models (GLMs) are fundamental tools for statistical modeling, with maximum likelihood estimation (MLE) serving as the classical approach for parameter inference. While MLE performs well for canonical GLMs, it can become computationally challenging in more general settings with non-canonical, non-smooth, or nonlinear link functions, where the resulting optimization landscape may be ill-conditioned, non-convex, or non-differentiable. In this paper, we study an alternative estimation framework based on variational inequalities (VIs), which formulates GLM estimation through an operator-based equilibrium condition rather than likelihood minimization. We analyze the VI estimator from a statistical perspective and establish finite-sample error bounds and asymptotic normality under mild regularity conditions, together with convergence guarantees for fixed-point and stochastic approximation algorithms. The framework accommodates a broad class of link functions, including non-canonical and non-monotone cases satisfying a strong Minty-type condition, and extends naturally to generalized additive models via basis expansion. Numerical experiments demonstrate that the VI approach achieves competitive finite-sample accuracy and improved numerical stability relative to MLE, particularly in GLMs and GAMs with non-canonical or non-smooth link functions.

Deep neural networks continue to grow in parameter count, driving up training and inference cost on GPUs. Sparse neural networks and Dynamic Sparse Training (DST) promise to reduce these costs, but most implementations rely on binary masks over dense tensors and recover little of the theoretical compute, memory, or energy savings. We propose SNACK, a truly sparse GPU layer that stores and computes only non-zero connections. SNACK exposes a simple PyTorch API for restructuring connections and backpropagating gradients entirely in the sparse paradigm, and uses a custom COO-format SpMM CUDA kernel with a batch-to-Streaming-Multiprocessor mapping tuned for the small-batch, high-sparsity regime typical of large-model training and single-stream inference. At the kernel level, SNACK is up to $7\times$ faster than the masked dense baseline (Dense+Mask) and competitive with cuSPARSE, Sputnik, and FlashSparse at 95\% sparsity. At 90\% sparsity, a single SNACK layer accelerates training by $8\times$ and $3.7\times$, and inference by $4\times$ and $2\times$, over Dense+Mask and fully dense layers, respectively, while using 25\% less memory than dense and substantially less energy. End-to-end, SNACK reduces GPT-2 peak training memory by up to 40\% and graph-style inference latency by $4.8\times$ over Dense+Mask at 99\% sparsity.