NeurIPS 2026 Papers

Generative agents have proven to be powerful assistants in a wide variety of contexts. Given this success, users are now deploying agents with minimal restrictions in open ended, multi-agent environments. Current methods for measuring the dynamics of open-ended multi-agent systems are limited to qualitative inspection. In this paper, we extend the process-theoretic notion of adaptive control charts to multi-agent systems to enable automated monitoring. Using simulation, we demonstrate that adaptive control charts are necessary for monitoring multi-agent systems that can learn from their environment. We further demonstrate, both empirically and theoretically, that adaptive control charts are susceptible to adversarial agents that defect sufficiently slowly. These results illustrate a fundamental tradeoff in multi-agent system control: either agents in a system cannot learn \textit{or} the system is susceptible to adversaries.

Model merging integrates task-specific fine-tuned models into a single multi-task model, but often suffers from parameter interference caused by conflicting task-vector updates. Existing methods typically mitigate conflicts by pruning task vectors based on weight magnitude or random heuristics, treating Transformers as unstructured ``bags of parameters'' and overlooking their inherent modularity. In this paper, we propose \textbf{C}ontribution-\textbf{A}ware \textbf{S}tructured \textbf{S}parsity (CASS), a unified framework that reduces parameter interference by identifying and preserving task-specific components. At the core of CASS is a contribution-aware structured mask that identifies task-relevant attention heads and FFN neurons. We instantiate this mask in two settings: CASS-Merging, the primary post-hoc setting where masks serve as a plug-and-play denoising filter for existing merging operators, and CASS-Tuning, an extension for scenarios with fine-tuning access where masks constrain gradients to reduce structural overlap between task vectors. Our analysis shows that task-relevant components are sparse and partially disjoint, supporting structured component-level filtering as an effective way to reduce merging interference. Extensive experiments across vision (ViT, 20 tasks) and language (RoBERTa, 8 tasks; Qwen2.5, 4 tasks) benchmarks demonstrate that CASS improves a range of representative merging baselines.

Large reasoning models separate inference into an explicit thinking phase followed by a final answer phase, yet how the answer phase uses the reasoning trace remains unclear. We present a systematic study of this trace-to-answer transition across multiple LRMs and find that (i) final answers diverge from correct reasoning traces at rates between 6\% and 48\%, even when traces terminate cleanly at full budget, (ii) naively copying the trace answer underperforms standard decoding in some settings, revealing that the answer phase has independent corrective capacity, and (iii) the same phenomenon extends beyond structured tasks to settings where reasoning traces and final answers can diverge under social-pressure framing (sycophancy). These findings motivate Contrastive Thinking Decoding (CTD), a training-free, single-model decoding method that contrasts answer-phase logits under the primary reasoning trace against those from a deliberately degraded noisy trace. CTD selectively amplifies trace-consistent tokens while preserving the answer phase's corrective capacity, concentrating CTD-internal joint mass on the both-correct cell (T$\checkmark$A$\checkmark$) without inducing trace-correct-to-answer-wrong drift relative to the base distribution. Across math reasoning, science, code, social-pressure resistance (sycophancy), and a broad knowledge benchmark (MMLU), \ctd reduces trace-answer disagreement and improves the accuracy--compute trade-off without parameter updates or auxiliary models at decoding time.

The strong zero-shot and long-context capabilities of recent Large Language Models (LLMs) have enabled highly effective list-wise reranking systems. Attention-based rerankers leverage transformer attention patterns as retrieval signals, but not all attention heads contribute equally: many introduce noise and redundancy, limiting both reranking quality and efficiency. In this work, we introduce CoRe heads, a small subset of retrieval-specialized heads identified through a contrastive scoring objective that rewards attention to relevant documents while penalizing attention to distractors. This contrastive criterion isolates highly discriminative retrieval heads and yields a strong training-free list-wise reranker. Our analysis reveals that retrieval-relevant computation in transformer rerankers is highly sparse and structurally localized: fewer than 1% of attention heads account for most reranking performance across models and datasets, and these heads consistently concentrate in middle transformer layers. This localization enables aggressive layer pruning for efficient reranking: removing the final 50% of transformer layers preserves nearly identical reranking accuracy while substantially reducing inference latency and GPU memory usage.

Most LLM unlearning methods aim to approximate retrain-from-scratch behaviors with minimal distribution shift, often via alignment-style objectives defined in the prediction space. While effective at reducing forgotten content generation, such approaches may act as suppression: forgotten concepts can persist in representations and remain entangled with retained knowledge. We introduce CLReg, a contrastive representation regularizer that identifies forget features while pushing them away from retain features, reducing forget--retain interference while empirically preserving the scale and shape of retain features. As light motivation for the mechanism, we provide a one-step analysis showing that CLReg decreases a simple entanglement proxy in the embedding space. Across unlearning benchmarks and LLMs of different sizes, CLReg decreases forget-retain representation entanglement to enhance mainstream unlearning methods without extra privacy risks, inspiring future unlearning work to remove forget concepts via representation shaping.

This paper concerns the problem of scaling agents to large action spaces demanding pretrained world knowledge. What type of representation pretraining would enable policies such as language models to generalize rewards to appropriate actions? Contrastive learning aligns similar inputs and otherwise enforces distancing of representations, while non-contrastive alternatives omit distancing terms and so learn anisotropic representations that collapse distinctive clusters. Though contrastive pretraining has been proven to promote supervised learning efficiency, non-contrastive pretraining nevertheless remains canonical for language agents. Here we examine the extent to which these pretraining types support on-policy finetuning, in which efficiency is coextensive with online performance. Broadly, we argue that contrastive pretraining optimizes efficient exploration of large spaces. We prove that anisotropic representations collapse distinct actions and necessitate complex policies, whereas contrastive pretraining retains distinctions and enables reward transfer through spectral regularization. We validate our conclusions empirically by providing spectral analyses of language data and simulations involving multimodal personalization policies as well as self-improving reasoning agents finetuned online using a policy gradient. Our latter simulations suggest contrastive pretraining for language agents exploring diverse reasoning paths.

In many critical domains, features are not freely available at inference time: each measurement may come with a cost of time, money, and risk. Longitudinal prediction further complicates this setting because both features and labels evolve over time, and missing measurements at earlier timepoints may become permanently unavailable. We propose CAPE (Contrastive Acquisition-Plan Embeddings), a partitioning acquisition policy for longitudinal active feature acquisition. CAPE first defines NOCT, an oracle-based objective that scores a set of future feature-time acquisitions by its expected predictive loss together with its acquisition cost. Based on NOCT, CAPE learns a contrastive acquisition-utility manifold of masked partial observations, where proximity reflects the expected utility of future acquisitions. Offline, training states are embedded and partitioned at each rollout timepoint; candidate future plans are scored within each partition using plug-in NOCT, and the best partition-level acquisition plans are cached. During inference, CAPE retrieves the cached plan via nearest-centroid lookup, enabling low-latency adaptive acquisitions. Experiments on synthetic and real-world healthcare datasets demonstrate that CAPE outperforms nonparametric, greedy, and RL-based alternatives, achieving higher accuracy at lower acquisition costs.

In the classical *identification in the limit* model of Gold [Inf. Control, 1967], a stream of positive examples is presented round by round, and the learner must eventually recover the target hypothesis. Recently, Kleinberg and Mullainathan [NeurIPS 2024] introduced *generation in the limit*, where the learner instead must eventually output novel elements of the target's support. Both lines of work focus on positive-only or fully labeled data. Yet many natural supervision signals are inherently relational rather than singleton: comparative experiments, A/B tests, side-by-side judgments, and similarity–dissimilarity annotations produce observations that encode relationships between examples rather than labels of individual ones. This motivates us to initiate the learning-theoretic study of *contrastive identification and generation in the limit*, where the learner observes a *contrastive presentation* of data: a stream of unordered pairs $\\{x,y\\}$ satisfying $h(x)\ne h(y)$ for an unknown target binary hypothesis $h$, but which element is positive is hidden from the learner. We first present three results in the noiseless setting: an exact characterization of contrastive identifiable classes (a one-line geometric refinement of Angluin's tell-tale condition [Angluin, Inf. Control 1980]), a combinatorial dimension called *contrastive closure dimension* extending the closure dimension of Raman et al. [COLT 2025] and exactly characterizing uniform contrastive generation with tight sample complexity, and a strict hierarchy in which contrastive generation and text identification are mutually incomparable. We then prove a sharp *reversal* under finite adversarial corruption: there exist classes identifiable from contrastive pairs under any finite corruption budget by a single budget-independent algorithm, yet not identifiable from positive examples under even one corrupted observation. The unifying technical object is the *common crossing graph*, which encodes pairwise ambiguity, family-level generation obstructions, and corruption defects in a single coverage-and-incidence language.

LLM agents are increasingly used for open-ended, data-driven scientific discovery, yet most systems remain pointwise: they evaluate one candidate hypothesis at a time for surprise, novelty, or evidential support. This is information-inefficient: rejecting a hypothesis reveals that one explanation is unlikely, but not which alternatives are better supported. We argue that the proper epistemic unit of discovery is not a single hypothesis but a scientific question together with the mutually exclusive and collectively exhaustive (MECE) explanations that could answer it. We introduce Contrastive discovery, which scores candidate discoveries by how evidence redistributes belief across competing explanations. This allows a single experiment to both rule out hypotheses and identify the most plausible alternative. Across six scientific domains, Contrastive discovery yields substantially more resolved scientific questions than pointwise baselines under the same evaluation budget. High-surprisal pointwise candidates are overwhelmingly rejection-heavy, yet many become resolved alternatives when evaluated contrastively, showing that an explicit representation of alternatives can turn negative evidence into resolved discoveries.

Graph Neural Networks (GNNs) are effective for modeling relational data but are vulnerable to label poisoning, where a small number of corrupted training labels can propagate errors across the graph via message-passing. Despite this risk, defenses against label poisoning remain underexplored: existing methods are primarily designed for label noise and often rely on robust losses or heuristic data cleaning that fail to distinguish adversarial poisoning from naturally hard examples. In this paper, we propose CoLAT (Contrastive Label-flipping for Adversarial Training), a novel contrastive adversarial training framework tailored for robust node classification on graphs. The core of our approach is a structure-aware detection model that uses contrastive learning to identify label–structure inconsistencies. Unlike prior contrastive methods that focus on representation learning, CoLAT leverages contrastive embeddings to select high-risk nodes that guide adversarial label-flipping during training. This alternating optimization not only performs structure-aware adversarial training but also implicitly sanitizes corrupted labels, improving robustness against diverse attacks from the literature. Extensive experiments on multiple benchmarks show that CoLAT consistently outperforms existing defenses under various poisoning intensities while scaling efficiently to large graphs.

Model distillation provides an alternative when computing resources are limited. How to transfer the reasoning ability from teacher model to student model influences the performance of model distillation. We develop a mathematical framework for reasoning distillation grounded in dynamical systems theory. By modeling the teacher model's residual stream as a discrete dynamical system, we define a precise notion of *orbital stability*: perturbations transversal to the teacher's reasoning trajectory should contract, while the tangential reasoning signal propagates without active suppression. We derive the **Contractive Restoring Flow (CRF)** loss from first principles and prove that, at the loss minimum, it achieves strict transversal contraction at a constant rate $1-\alpha$, leaves the tangential dynamics unconstrained (tangential agnosticism), and produces a global basin of attraction with bounded steady-state error within the linearisation regime. Empirically, reasoning-trajectory distillation with the CRF loss outperforms other distillation methods, and shows greater robustness on long-chain reasoning tasks, supporting the predicted effect of contracting transversal error. The code is avaiable at: https://anonymous.4open.science/r/CRF-5945/.

Graph neural networks exhibit a fragile relationship with depth, as adding layers leads to rapid homogenization of node representations (over-smoothing). Recent manifold-constrained approaches using doubly stochastic matrices address this issue, but the fundamental principle underlying their success remains unclear. This study investigates whether doubly stochasticity is essential or exemplifies a broader geometric principle for stable residual propagation in deep GNNs.We formalize three axioms to define residual mixing matrices as contractive monoids. We instantiate this framework through three manifolds (Orthogonal Group, Diagonal Contraction Monoid, Spectral Norm Ball) and evaluate them across nine node-classification benchmarks using four GNN backbones. Results demonstrate that contractive-monoid constraints maintain stable accuracy and low representational collapse at extreme depths, while standard GNNs degrade sharply beyond moderate depth. This work establishes that stable residual propagation arises from contractive, compositional structure anchored at identity, revealing a significantly larger design space for deep GNN architectures while maintaining theoretical guarantees.

Tool-augmented LLM agents call APIs whose intermediate outputs, such as presigned URLs, session tokens, and OAuth state parameters, are observation contracts: artifacts whose later use is constrained by the external system that produced them. We show that observation contract compliance (preserving the temporal validity and byte-level integrity) is an emergent, regression-prone capability: it is neither guaranteed by general tool-use ability nor consistently improved by larger or newer models. To measure this, we introduce ContractBench, a benchmark of 33 dual-axis tasks that probe two orthogonal failure modes no existing benchmark evaluates: validity failures (using an artifact after expiry) and integrity failures (corrupting an artifact's bytes through the observation-to-action pipeline). Our evaluation is deterministic and programmatic, with a virtual clock controls time and SHA-256 hashes verify byte integrity. We assign the outcomes failure labels drawn from real-world API specifications. We evaluate 38 models and report four findings: (i) no evaluated model clears 80%, with Claude-Opus-4.6 leading at 77.8%, revealing that current frontier models still fail to comply with observation contracts; (ii) a sharp within-family capability cliff in Qwen 3.5 between 4B (0%) and 9B (56.6%), smoothing to 70.7% at 397B-A17B: what emerges across the cliff is mid-trajectory restraint, not tool-call competence; (iii) non-monotonic scaling across GPT-5 family: agentic post-training can erode compliance through sycophancy-driven regression; (iv) our failure taxonomy works as an actionable in-context reward signal, yielding +7.1 pp on 42 paired GPT-5.1 failures.

Efficient sampling from complex probability distributions is an important task across Bayesian inference and certain classes of deep generative models. Markov chain Monte Carlo (MCMC) remains one of the most widely used tools for this. However, even state-of-the-art samplers, such as the NUTS implementations in Stan and Turing.jl, can converge slowly on ill-conditioned, heavy-tailed, or near-degenerate targets. We identify that sampling error can be decomposed into two terms: mismatch in the potential energy (PE) marginal, and average conditional mismatch across PE level sets. We propose *Contour Monte Carlo (CMC)*, which corrects each of these in turn: first by drawing energies $u_i$ from the PE marginal, and then by sampling states on the corresponding level sets. For a broad family of radial targets, including the Gaussian and Student's $t$ distributions, both stages admit closed-form draws, so CMC is exact and requires no Markov chain. Beyond this, for general targets, we estimate the PE marginal using umbrella sampling with MBAR, then run constrained MCMC on the sampled level sets, initialised from energy-matched umbrella samples (which tend to lie in regions of high conditional mass). We show that this procedure essentially reduces the global sampling problem to approximating the one-dimensional PE marginal, after which posterior draws can be generated readily on demand. Furthermore, the level-set chains are embarrassingly parallel, and therefore well-suited to modern accelerator hardware. On a suite of challenging benchmarks -- composed of `posteriordb` models, synthetic targets, and real-world inference problems -- CMC converges substantially faster than HMC and NUTS baselines at a fixed computational budget.

Missing data is pervasive in real-world datasets, arising from sensor failures, image occlusions, and incomplete reporting. Imputation is a conditional generation problem that must preserve observed entries exactly while producing plausible values for missing ones. Yet most diffusion-based imputers adopt Gaussian noising whose global perturbations are structurally misaligned with this local hard constraint, so consistency is typically enforced through conditional training, clamping, projection, or repainting-style corrections. We address this mismatch with Continuously-Augmented Hybrid Masked Diffusion (CAHMD), a masked diffusion framework whose conditional sampler leaves observed coordinates unchanged by construction. To extend masked diffusion beyond purely discrete data, CAHMD attaches coordinate-wise auxiliary continuous latents to masked entries, providing a denoising channel before values are revealed in continuous and hybrid data. For incomplete training data, we introduce observed-mask gating, an observed-data masked reconstruction principle that uses available entries as pseudo-missing targets and avoids repeated EM-style imputation steps. Across tabular, image, and image--caption benchmarks, CAHMD gives a strong reconstruction--perception--efficiency trade-off while reducing the training overhead of EM-style self-imputation and the inference overhead of repainting-style inner loops.

Deep Gaussian processes (DGPs) are appealing Bayesian regression models, but their standard discrete-depth parameterization often becomes harder to optimize as depth grows. We argue that depth is better treated as a latent continuous evolution than as a long discrete stack. We therefore propose a continuous-depth formulation in which the latent representation follows a stochastic differential equation with a depth-indexed drift family equipped with a GP-based prior. This reframes the core design choice as the prior placed on that family, rather than adopting a purely neural continuous-depth parameterization as in neural ODE or neural SDE models. We study two instantiations of this idea: CDGP, a direct GP prior over the state-depth drift, and FlowDGP, a flow-evolved prior that captures richer depth dependence at lower training cost. To motivate the redesign, we show that the single-sample DGP objective can become increasingly ill-conditioned under repeated layer composition, while residual DGP is a useful first repair that still remains a finite discrete construction. We also give pathwise sensitivity analysis for the continuous-depth formulation. Experiments on synthetic, benchmark, and large-scale regression tasks show that residual DGP improves over plain DGP, while CDGP and FlowDGP provide a stronger overall story across the regimes we study.

Step distillation has become a leading technique for accelerating diffusion models, among which Distribution Matching Distillation (DMD) and Consistency Distillation are two representative paradigms. While consistency methods enforce self-consistency along the full PF-ODE trajectory to steer it toward the clean data manifold, vanilla DMD relies on sparse supervision at a few predefined discrete timesteps. This restricted discrete-time formulation and mode-seeking nature of the reverse KL divergence tends to exhibit visual artifacts and over-smoothed outputs, often necessitating complex auxiliary modules---such as GANs or reward models---to restore visual fidelity. _In this work, we introduce **C**ontinuous-Time **D**istribution **M**atching (**CDM**), migrating the DMD framework from discrete anchoring to continuous optimization for the first time._ CDM achieves this through two continuous-time designs. First, we replace the fixed discrete schedule with a dynamic continuous schedule of random length, so that distribution matching is enforced at arbitrary points along sampling trajectories rather than only at a few fixed anchors. Second, we propose a continuous-time alignment objective that performs active off-trajectory matching on latents extrapolated via the student's velocity field, improving generalization and preserving fine visual details. Extensive experiments on different architectures, including SD3-Medium and Longcat-Image, demonstrate that CDM provides highly competitive visual fidelity for few-step image generation without relying on complex auxiliary objectives.

We present the first method for intrinsic, continuous gradient-based optimization for models with $p$-adic parameters. To overcome the vanishing gradients caused by the discrete valuation of the $p$-adics, we lift parameters to the Berkovich affine line, a canonical analytification of the $p$-adic numbers that provides a path-connected space. This enables gradient descent that respects the non-Archimedean metric with a local transition cost that is linear. By extending losses to the Berkovich line, we get piecewise linear functions, corresponding to the cell decomposition of a tropical complex; gradients are efficiently computed via max-plus operations. We prove local convergence within tropical cells for linear models on the $\mathbb{Q}_p$ subtree, and demonstrate the success of our method on synthetic tasks and on the $p$-adic Quillian semantic benchmark [Martins, 2025], where we achieve performance equal or better than $\mathbb{R}$-valued counterparts while training in comparable time.

Existing conditional diffusion models typically inject condition information into the reverse-time denoising process, either as model inputs or through guidance. Although effective for conditional fidelity, this design burdens the denoiser when compact continuous signals, such as user preference ratios or mixture degrees, must support both image reconstruction and condition-wise separation, often degrading sample quality. In this work, we reinterpret this limitation not as a problem of condition representation, but as a problem of condition placement. We propose the Continuous Personalized Diffusion Model (CPDM), which shifts the locus of conditioning from reverse-time signal injection to forward geometry formation by coupling bounded spinor-component coordinates with normalized image-space drift directions. The resulting condition-dependent forward drift forms condition-specific terminal geometry, from which continuous changes along the spinor-component coordinate are reflected as continuous visual transitions rather than endpoint interpolation. Experiments show that, even with a single scalar condition, CPDM achieves competitive performance compared with standard and higher-capacity reverse-conditioning baselines and supports continuous generation under compact scalar control. These results suggest forward geometry design as a practical alternative to conventional reverse-time conditioning for compact continuous personalization.