NeurIPS 2026 Papers

We study zero-shot coordination (ZSC), where independently developed agents must coordinate at test time. While ZSC has been well studied in the RL literature, far less is known about the performance of LLM agents despite their increasing deployment in such settings. Existing work on LLM coordination often relies on specialised scaffolding that independently developed agents are unlikely to share in practice. Moreover, evaluations on complex environments (e.g., Hanabi) make it difficult to pinpoint the sources of coordination failure. In contrast, we focus on simple, general-purpose scaffolds in minimal environments designed to isolate specific coordination challenges. Our results show that even in these controlled settings, frontier LLM agents struggle to coordinate, largely due to limited understanding of the coordination problem and weak reasoning about their partner’s beliefs. While LLMs introduce *semantic information* as an additional axis for coordination, they nonetheless fail to exploit this structure effectively. Towards this, we propose *Coordination-friendly definitions (CFDs)* as a principled approach for enabling robust coordination among LLM agents. Finally, we show that CFDs can be discovered automatically, removing the need for manual engineering.

Low-Rank Adaptation (LoRA) constrains weight updates to a low-rank subspace, making the choice of that subspace central to adaptation. We show that the value of LoRA initialization direction is regime-dependent. It helps when the representation dimension is large relative to the labeled set size and the estimated activation subspace is aligned with task structure, and is otherwise inconsequential. We formalize this view with a spiked-covariance Davis--Kahan argument and a first-step LoRA analysis showing that the initial learning signal flows through projected target-token features. From this analysis we derive an alignment coefficient $\rho$, computable from a single forward pass, that closely tracks the empirical gain from direction-aware initialization across tasks. Motivated by this view, we propose PivotLoRA, an initialization that estimates LoRA adapters from target-token PCA and applies a single output-preserving orthogonal pivot after early training drift. Across $30$ BERT-base few-shot settings and $13$ compared methods, PivotLoRA achieves the best average rank and improves over vanilla LoRA on nearly every setting, with the largest gains concentrated on high-alignment tasks. Decoder-scale classification on LLaMA-2-7B and Llama-3.1-8B reproduces the low-shot pattern, while full-data and generation controls confirm the predicted regime specificity.

Reusing a held-out benchmark adaptively should, in principle, invite overfitting. Yet benchmark-driven machine learning (ML) has produced surprisingly little overfitting in practice. An attractive hypothesis is that successful ML strategies are highly compressible. We study this in the setting of LLM-driven research agents, where the hypothesis becomes directly testable via two complementary information bottlenecks. In \emph{output compression}, an exploration agent adaptively searches for high-performance models using a validation set, and we test whether a fresh ``reproducer agent'' can reproduce its performance given only an extremely short prompt and the training data. In \emph{input compression}, the explorer receives only one-bit feedback indicating whether each submitted model improves on the running best. Across 8 datasets spanning tabular classification, vision, language modeling, diffusion modeling, and reward modeling, we find that these bottlenecks have little effect on performance: short prompts and compressible feedback are sufficient to reproduce and find high-performance models. The hypothesis is falsifiable: when we deliberately induce validation-set overfitting, the results fail to reproduce with short prompts. Taken together, our results support a description-length explanation for the lack of overfitting in benchmark-driven ML: successful strategies occupy a low-complexity region of strategy space.

Model-based reinforcement learning (MBRL) achieves strong sample efficiency by planning within learned latent dynamics, yet its performance degrades substantially under unseen visual distractions such as background variations, lighting changes, or camera shifts. Unlike model-free RL, where encoder perturbations affect only single-step predictions, MBRL suffers from a two-level vulnerability: visual distractions first push encoder outputs out of distribution, and these errors then compound through recursive latent rollouts over the planning horizon. We propose **VI**sual **G**eneralization via latent-space c**O**nsistency in model-based **R**L (**VIGOR**), a framework that enables zero-shot generalization to unseen visual distractions while retaining the sample efficiency of its MBRL backbone. VIGOR integrates three interdependent components: (i) *asymmetric weak-to-strong augmentation*, which pairs weak-only and weak-to-strong latent views within a single batch; (ii) *dynamics-level consistency*, which enforces augmentation-invariant transition predictions through direct latent regression; and (iii) *encoder-level stabilization*, which prevents encoder drift under the cross-augmentation supervision imposed by dynamics-level consistency. Evaluations on the DeepMind Control Suite (DMC) and Robosuite show that VIGOR outperforms state-of-the-art model-free and model-based baselines, surpassing the second-best baseline by $3.4\\%$ on DMC and $43.6\\%$ on Robosuite. Ablations further show that VIGOR's robustness is augmentation-agnostic: replacing the default augmentation with alternatives from distinct perturbation families preserves strong generalization, confirming that latent-space consistency, not the augmentation choice, drives robustness. The source code is available at https://anonymous.4open.science/r/vigor-B641/.

Object-centric representations promise a key property for few-shot learning: Rather than treating a scene as a single unit, a model can decompose it into individual object-level parts that can be matched and compared across different concepts. In practice, this potential is rarely realized. Continual learners either collapse scenes into global embeddings, or train with part-level matching objectives that tie representations too closely to seen patterns, leaving them unable to generalize to truly novel concepts. In this paper, we identify this fundamental structural conflict and pioneer a new paradigm that strictly decouples representation learning from compositional inference. Leveraging the inherent patch-level semantic geometry of self-supervised Vision Transformers (ViTs), our framework employs a dual-phase strategy. During training, slot representations are optimized entirely toward holistic class identity, preserving highly generalizable, object-level geometries. At inference, preserved slots are dynamically composed to match novel scenes. We demonstrate that this paradigm offers dual structural benefits: The frozen backbone naturally prevents representation drift, while our lightweight, holistic optimization preserves the features' capacity for novel-concept transfer. Extensive experiments validate this approach, achieving state-of-the-art unseen-concept generalization and minimal forgetting across standard continual learning benchmarks.

Multinomial logistic regression (MLR) is the default last-layer classifier in fully supervised recognition, while prototype classifiers are central to few-shot recognition. Although they are usually presented as different principles, we show that both can be organized by a geometric notion of class boundaries. A Euclidean MLR logit is a signed distance to a class hyperplane, so ordinary MLR is already the hyperplane decision-boundary case of Umbilic Multinomial Logistic Regression (UMLR). Classical submanifold geometry supplies the matching spherical decision boundary: in Euclidean space, the complete connected totally umbilic hypersurfaces are hyperplanes and spheres. The spherical UMLR form keeps a prototype-like center but adds a learnable radius; its signed squared-distance form contains squared prototype scoring as the radius-zero case. Thus UMLR is not a separate replacement for MLR or prototypes, but a boundary-aware family that places hyperplane logits, spherical logits, mixed heads, and prototype scoring in one framework. Experiments on low-shot vision head swaps, prototype comparisons, and mixed hyperplane--sphere heads support this organization and show how the different decision boundaries behave across data regimes. UMLR provides a compact geometric language for studying when class evidence is better modeled by flat boundaries, spherical boundaries, or their mixture.

Diffusion-based models have demonstrated impressive accuracy and generalization in solving partial differential equations (PDEs). However, they still face significant limitations, such as high sampling costs and insufficient physical consistency, stemming from their many-step iterative sampling mechanism and lack of explicit physics constraints. To address these issues, we propose \emph{Phys-Instruct}, a novel physics-guided distillation framework which (1) not only compresses a pre-trained diffusion PDE solver into a few-step generator via matching generator and prior diffusion distributions to enable rapid sampling, (2) but also enhances the physics consistency by explicitly injecting PDE knowledge through a PDE distillation guidance. Phys-Instruct is built upon a solid theoretical foundation, leading to a practical physics-constrained training objective that admits tractable gradients. Across five PDE benchmarks, Phys-Instruct achieves orders-of-magnitude faster inference while reducing PDE Error by more than 8$\times$ compared to state-of-the-art diffusion baselines. Moreover, the resulting unconditional student model functions as a compact prior, enabling efficient and physically consistent inference for various downstream conditional tasks. Our results indicate that Phys-Instruct is a novel, effective, and efficient framework for ultra-fast PDE solving powered by deep generative models.

Optical Music Recognition (OMR), the task of transcribing sheet music into a structured textual representation, is currently bottlenecked by a lack of large-scale, annotated datasets of real scans. This forces models to rely on either few-shot transfer or synthetic training pipelines that remain overly simplistic. A secondary challenge is encoding non-uniqueness: in the popular Humdrum **kern format for transcribing music, multiple different text encodings can render into the same visual sheet music. This one-to-many mapping creates a harder learning task and introduces high uncertainty during decoding. We propose Transcoda, an OMR system built on (i) an advanced synthetic data generation pipeline, (ii) a normalization of the **kern encoding to enforce a unique normal form and (iii) grammar-based decoding to ensure the syntactic correctness of the output. This approach allows us to train a compact 59M-parameter model in just 6 hours on a single GPU that outperforms billion-parameter baselines. Transcoda achieves the best score among state of the art baselines on a newly curated benchmark of synthetically rendered scores at 18.46% OMR-NED (compared to 43.91% for the next-best system, Legato) and reduces the error rate on historical Polish scans to 63.97% OMR-NED (down from 80.16% for SMT++). We will make our model, data and generation pipeline publicly available upon publication of the paper.

Discrete flow matching generates text by iteratively transforming noise tokens into coherent language, but may require hundreds of forward passes. Distillation uses the multi-step trajectory to train a student to reproduce the process in a few steps. When the student underperforms, the usual explanation is insufficient capacity. We argue the opposite: the trajectory is the bottleneck, not the student. Each training trajectory is built through a chain of blind stochastic jumps with no evaluation of sequence quality; a single bad decision at an early midpoint propagates through subsequent steps, yet the student must imitate the result. Trajectory-Shaped Discrete Flow Matching (TS-DFM) replaces these blind jumps with guided navigation: a lightweight energy compass evaluates candidate continuations at each midpoint, selecting the most coherent. All shaping is training-only; inference cost is unchanged. On 170M-parameter language modeling, the shaped student at 8 steps achieves 32% lower perplexity than the 1 024-step teacher while being 128×faster, with gains consistent across source distributions and three evaluators of increasing scale. TS-DFM achieves the best perplexity of any discrete-generation baseline we compare against, including methods trained on 6×more data or using 5×larger models.

Across scientific disciplines, Laplacian eigenvectors serve as a fundamental basis for simplifying complex systems, from signal processing to quantum mechanics. In reinforcement learning (RL), they similarly form a basis over the state space, enabling reward functions to be approximated by projection onto a small set of eigenvectors. This projection makes zero-shot control possible, but it also imposes a fundamental limitation: the induced policies are only as expressive as the linear span of the chosen eigenvectors. We introduce the Laplacian Keyboard (LK), a hierarchical framework that goes beyond this linear span. LK constructs a task-agnostic library of behaviors from these eigenvectors, forming a behavior basis guaranteed to contain the optimal policy for any reward within the linear span. A meta-policy learns to stitch these behaviors dynamically, enabling efficient learning of policies outside the original linear constraints. We establish theoretical bounds on zero-shot approximation error and demonstrate empirically that LK improves over the zero-shot solution while achieving better sample efficiency compared to standard RL methods.

Offline meta-reinforcement learning seeks to learn a policy that generalizes to new related tasks online. Context-based methods infer a task representation from transition histories, yet learning an effective task representation without supervision remains challenging. Existing methods relying on contrastive learning learn discriminative task representations, but fail to identify task-specific dynamics, while relying on reconstruction can be insufficient to model long-horizon dependencies, limiting generalization to new tasks. We investigate the impact of temporal consistency in latent space on task representation learning, showing that enforcing multi-step predictions in latent space encourages task representations that are able to capture task-dependent dynamics while preventing representation collapse. We provide theoretical analysis characterizing sources of error in value estimation and show through extensive experiments on MuJoCo, Contextual DeepMind Control, and MetaWorld benchmarks that temporal consistency significantly improves both zero-shot and few-shot generalization.

Few-shot object detection (FSOD) requires detectors to adapt to novel categories from only a few labeled instances, where prototype-based transfer methods have become a strong and efficient paradigm. However, we observe that such methods remain unstable in extreme low-shot regimes. We attribute this instability to two coupled factors: semantically unanchored query initialization, which yields high-variance cold-start optimization, and deterministic prototype matching, which over-trusts noisy or occluded support regions. To address these limitations, we propose Semantic Fine-Grained Prototype Distillation (SFPD), a language-conditioned probabilistic adaptation framework for prototype-based FSOD. SFPD introduces language-conditioned feature queries to provide semantic anchors for novel-class adaptation, uncertainty-aware prototype distillation to down-weight unreliable support evidence through heteroscedastic Gaussian modeling, and component-guided part-aware prototypes to refine fine-grained semantic-visual alignment. These modules act during adaptation and preserve the original detector path at inference, introducing no additional inference FLOPs. Experiments on PASCAL VOC and MS-COCO show that SFPD consistently improves a strong FPD baseline, with especially clear gains in the most challenging low-shot settings, e.g., a 3.9-point nAP50 improvement on VOC Split 2 under 1-shot. Further ablations, multi-seed evaluation, and convergence analysis indicate that SFPD improves both accuracy and adaptation stability.

As LLM-based agentic systems grow more complex, they increasingly rely on meta-agents: higher-order agents that act on other agents, much like managers supervise employees. Yet existing agentic runtimes expose execution only as static environmental states, limiting the kinds of live and post-hoc interventions a meta-agent requires. To unlock these capabilities, we introduce Shepherd, a functional programming model that formalizes meta-agent operations on target agents as functions, with core operations mechanized in Lean. Shepherd records every agent–environment interaction as a typed event in a principled Git-like execution trace where any past state can be cheaply forked and replayed. Even at scale, Shepherd forks the agent process and its filesystem 5x faster than Docker, with >95% prompt-cache reuse on replay. We exemplify Shepherd's versatility in three use cases: (1) runtime intervention, where a live supervisor improves pair coding pass rate from 28.8% to 54.7% on CooperBench; (2) counterfactual meta-optimization, where a meta-agent branches to explore alternative paths, beating MetaHarness and GEPA across four benchmarks by up to 11 points with up to 58% lower wall-clock; and (3) Tree-RL training, where a meta-agent forks rollouts at chosen turns, lifting TerminalBench-2 from 34.2% to 39.4% on Qwen3.5-35B-A3B. These use cases show Shepherd is a performant and efficient substrate for programming meta-agents; we open-source it to advance future research.

Federated intrusion detection must adapt to rare attacks from only a few labeled flows per client, while client distributions are highly non-IID and raw traffic cannot be centralized. Existing few-shot intrusion detectors typically construct class priors in centralized settings, and standard federated training does not provide an explicit mechanism for reusing cross-client class statistics during local low-shot adaptation. We propose a semantic-statistical prior bank for federated low-shot intrusion detection. Clients first participate in federated representation learning and submit protected class-wise moment statistics through secure aggregation, from which the server constructs a global prior bank without accessing raw traffic. During local episodic adaptation, label-phrase embeddings retrieve relevant priors, which are combined with scarce support statistics through a maximum a posteriori (MAP) shrinkage estimator in the feature space. The adapted class distributions are then used to train lightweight episode-specific classifiers on each client. Experiments on CICIDS2017 and CICDDoS2019 under 5-way 1-shot and 5-shot non-IID settings show consistent improvements over FedAvg and backbone-matched baselines, with the largest gains appearing in 1-shot and highly skewed client distributions. With ResNet1D, the proposed adaptation improves 1-shot F1 from 92.97 to 95.55 on CICIDS2017 and from 82.13 to 88.25 on CICDDoS2019.

What does it mean to create a new concept, rather than retrieve a familiar one? Diffusion models sampled many times at the same prompt vary only within the narrow stylistic envelope set by a fixed prompt embedding. We propose an operational definition of creativity as the generation of concepts that are initially unfamiliar yet rapidly learnable to an adaptive observer, and formalize it as a bilevel optimization between a Creator that generates and an Appraiser that adapts: the Appraiser's improvement under a brief inner-loop adaptation provides the reward signal that the Creator maximizes. We call this framework CAMEL (Creator-Appraiser Meta-Learning), and instantiate it on MNIST with an autoencoder Appraiser and on natural images with a CLIP Appraiser based on a low-rank adapter over the text projection. CAMEL produces outputs that lie outside the basin reachable by vanilla diffusion sampling at any sample count, while remaining recognizable as the prompt's class to an off-the-shelf observer.

Treating patients with combinations of drugs reduces the risk of resistance to any individual drug. Finding effective combinations is difficult because the large search space makes combinatorial screens prohibitively expensive, time consuming, and often technically infeasible. Predictive models can fill this gap, yet existing methods typically require molecular profiling of each sample and per-cohort training, limiting their applicability when time and tissue are scarce. To address this challenge, we introduce ScreenShot, a hierarchical transformer pretrained on 40 drug screening datasets covering 3,700 drugs and 6,000 biological samples, whose architecture mirrors the nested structure of screening data. Given a few-shot context of observations from a new patient, ScreenShot predicts the response of the sample to combination therapies through in-context learning, operating directly on functional measurements with no fine-tuning and no molecular profiling. On four held-out datasets, ScreenShot outperforms all baselines in both prediction accuracy and identification of selectively effective treatments. ScreenShot's internal representations are directly useful for experimental design: we use them to drive a weighted $k$-means++ active learning strategy that selects which experiments to run, achieving the same hit detection as uniform screening with a third of the budget. Source code and interactive dashboard: https://github.com/AnonymousAccount3/screenshot

Few-shot adaptation of vision--language models must balance two competing goals: learning fine-grained visual distinctions from very limited supervision, and preserving the efficiency that makes frozen CLIP-like models attractive for deployment. Existing lightweight adaptation methods, including prompt tuning, LoRA, and cache-based adapters, typically supervise global image embeddings and therefore underuse token-level structure. Conversely, structure-aware methods can exploit local evidence but often introduce additional inference-time computation. We propose SD-LoRA, a structural distillation framework that uses token-level graph reasoning only during training and absorbs its effect into a lightweight LoRA-adapted CLIP model. During training, a heterogeneous graph teacher models patch-patch, global-local, and visual-text relations over high-resolution ViT tokens. Rather than keeping this graph at test time, SD-LoRA transfers its structural bias into the shared CLIP-LoRA backbone through implicit distillation and Prototype Predictive Alignment, which aligns graph-conditioned representations with support-derived class prototypes. At inference, the entire graph teacher is discarded; prediction uses only a single CLIP-LoRA forward pass and a cache-based classifier. Across 11 few-shot benchmarks, SD-LoRA improves over a strong CLIP-LoRA baseline by 1.3/1.2/2.4/1.9/2.0% in the 1/2/4/8/16-shot settings, while adding no graph computation at test time. Further results on base-to-novel generalization, cross-dataset transfer, OOD robustness, and calibration show that training-time structural supervision improves not only accuracy but also transfer reliability.

Accurate monocular metric depth estimation, i.e., predicting absolute depth values, is critical for applications such as robotics and augmented reality. However, existing methods often suffer from poor transferability to in-the-wild scenarios due to the lack of knowledge of camera intrinsics. In this paper, we propose a novel framework, Roll2Depth, to recover metric scale in unseen environments with unknown camera parameters. Our key idea is to leverage scene-independent physical cues induced by the rolling shutter effect of standard CMOS sensors as metric priors. When a scene is illuminated by a flickering light, the rolling shutter produces distinctive stripe-like spatial patterns, revealing the scene's shading with and without the flickering light source. This contrast encodes absolute depth cues that can be exploited for metric estimation. Roll2Depth isolates these patterns as metric priors to guide affine-invariant models on metric depth estimation in a zero-shot setting. Our evaluations on standard benchmark datasets, as well as a real-world dataset collected using commercial smartphones and our prototype system, demonstrate that Roll2Depth, as a low-cost active solution, outperforms state-of-the-art passive baselines in zero-shot settings. These results highlight its potential impact in applications such as indoor robotics and embodied perception, where affordable and effective active sensing would be valuable.

Probing studies what information is encoded in a frozen LLM's layer representations by training a lightweight predictor on top of them. Beyond analysis, probes are often used operationally in probe-then-steer pipelines: a learned concept vector is extracted from a probe and then injected via additive activation steering by adding it to a layer representation during the forward pass. The effectiveness of this pipeline hinges on estimating concept vectors that are accurate, directionally stable under ablation, and inexpensive to obtain. Motivated by these desiderata, we propose RAPTOR (Ridge-Adaptive Logistic Probe), a simple $\ell_2$-regularized logistic probe whose validation-tuned ridge strength yields concept vectors from normalized weights. Across extensive experiments on instruction-tuned LLMs and human-written concept datasets, RAPTOR matches or exceeds strong baselines in accuracy while achieving competitive directional stability and substantially lower training cost; these quantitative results are supported by qualitative downstream steering demonstrations. Finally, using the Convex Gaussian Min-max Theorem (CGMT), we provide a mechanistic characterization of ridge logistic regression in an idealized Gaussian teacher-student model in the high-dimensional few-shot regime, explaining how penalty strength $\lambda$ mediates probe accuracy and concept-vector stability, yielding structural predictions that qualitatively align with trends observed on real LLM embeddings.