NeurIPS 2026 Papers

Coding agents are increasingly deployed in real software development, where a single version iteration requires months of coordinated work across many files. However, most existing benchmarks focus predominantly on single-issue bug fixes from Python repositories, with coarse pass/fail evaluation outcomes, and thus fail to capture long-horizon, multi-target development at real engineering scale. To address this gap, we present ROADMAPBENCH, a benchmark of 115 long-horizon coding tasks grounded in real open-source version upgrades across 17 repositories and 5 programming languages. Each task places the agent on a source-version code snapshot and provides a multi-target roadmap instruction requiring it to implement the functionality introduced in the target version, with a median modification of 3700 lines across 51 files. We conduct a systematic evaluation on thirteen frontier models and find that even the strongest, Claude-Opus-4.7, resolves only 39.1% of tasks, while the weakest achieves merely 5.2%, in stark contrast to the high scores these models attain on existing bug-fix benchmarks, suggesting that long-horizon software development remains a largely unsolved problem.

ARC tasks can be formulated as visual transformation problems, where a model must infer a high-level transformation rule from a small set of demonstrations and apply it consistently across a grid. While dense Vision Transformers (ViTs) perform well in this setting, they entangle local visual state and global transformation logic within a shared token representation, which can hinder consistent reasoning. In this paper, we introduce RoSeViT, a role-separated Vision Transformer that explicitly decouples high-level instruction inference from visual workspace computation. RoSeViT uses controller tokens to aggregate transformation-level information and workspace tokens to represent the visual grid, with a structured attention mechanism that routes nonlocal interactions through the controller. This design is motivated by a formal analysis showing that enforcing a shared instruction representation can reduce the effective complexity of visual transformation modeling. Under the standard VARC training and evaluation protocol for ARC, RoSeViT consistently outperforms dense ViT baselines. In a controlled backbone replacement setting, it improves ARC-1 accuracy from 54.5 to 56.6, and with recurrent refinement reaches 58.8. When integrated into the full system, RoSeViT-Ensemble improves performance from 60.4 to 62.6 on ARC-1. These results demonstrate that explicitly separating instruction and workspace representations leads to more effective visual reasoning on ARC tasks.

Moving an object in a single image requires geometry-consistent spatial rearrangement, including handling occlusions, revealing previously unseen regions, and maintaining coherent shadows and reflections. Existing approaches are not well suited to this setting and often fail to preserve such scene-level consistency. We address this problem by introducing a geometry-aware object motion method that operates directly on the internal representations of diffusion transformers. Our key insight is that rotary positional embeddings (RoPE) define a structured spatial field that can be explicitly manipulated to induce controlled motion. We extend 2D RoPE into a depth-aware formulation that encodes 3D spatial structure, enabling consistent object displacement and scene-aware updates. Our model is trained using synthetic data combined with a small set of real images via parameter-efficient fine-tuning. Despite minimal real supervision, it preserves object identity under large spatial displacements, generates plausible content in newly revealed regions, and consistently updates scene-dependent effects such as shadows and illumination. Experimental results demonstrate the superiority of our method.

Learning directed causal graphs from time-series data poses significant challenges, especially in fMRI where slow sampling rate obscures fast neural interactions. This temporal mismatch leads to undersampling, which can make multiple graphs equally plausible. We address this problem by explicitly modeling undersampling effects when recovering causal graphs. Our approach employs answer set programming (ASP) to enforce domain-specific constraints and optimize soft observational constraints, thereby identifying a Markov equivalence class for the resulting graph solutions. By customizing an ASP solver to collect multiple near-optimal solutions, we obtain not only the single best-fitting graph but an equivalence class of high-scoring graphs for expert consideration. This method, called Real-world noisy RASL (RnR), can also act as a meta-solver: it refines the output of other causal discovery algorithms by accounting for undersampling biases. In synthetic data and empirical brain network data, RnR produces more accurate causal graphs than state-of-the-art methods. When applied as a meta-solver (refining outputs of existing algorithms), it improves F1 scores by an average of 46\% over baseline methods; on standalone synthetic data benchmarks, it achieves a 64\% improvement. We demonstrate that RnR is robust to varying undersampling rates, maintaining high precision and recall even as sampling becomes more sparse, whereas baseline methods degrade significantly. Finally, we test RnR on real-world settings where ground truth connectivity is unknown such as human brain fMRI data, showing that incorporating undersampling-aware constraints via ASP yields more reliable and interpretable brain connectivity estimates from fMRI time series, bridging the gap between neural dynamics and observational data.

**This position paper argues that the current frontier of expert-LLM collaboration is doubly jagged: it is uneven, not only because of jagged AI capabilities but also due to jagged AI risks**. As LLM systems improve, incompetence-driven risks (e.g., misinformation) may decline, but adversarial risks can persist or even increase, leading to the likely persistence of a jagged risk–reward frontier. To show jaggedness, we introduce the first occupation-level quantification of risk-aware productivity gains for expert–LLM collaboration over a realistic task distribution, using standard O\*NET computer-occupation tasks. For each task, we leverage six frontier models (e.g., GPT-5, Claude Opus) to produce structured risk–reward ratings and use 6 human experts to verify a subset for reliability, achieving high agreement. Risk captures the possible increase in individual, organizational, or societal harm due to employing LLMs. Reward captures possible productivity gains, i.e., time/cost savings at a fixed quality target after accounting for expert oversight to detect and mitigate risks. We aggregate task-level ratings into occupation-level scores. Our analysis shows that risk varies more than reward, yielding vast differences in risk–reward tradeoff: In safety-critical roles (e.g., *Information Security Engineers*), gains are offset by risk, whereas in less safety-critical roles (e.g., *Web Developers*), gains often substantially exceed risks. At the task-level, we identify high-reward and medium-risk tasks as the "Sweet spot" for expert-LLM collaboration. The identified sweet spot is evident in real-world deployments as the majority of Claude conversations assigned to O*NET tasks (from the recent Anthropic Economic Index) are high-reward and medium-risk, whereas high-risk tasks are minimally represented, supporting both our methodology and position.

Synthetic data generation is becoming increasingly prevalent as advances in generative modeling accelerate. However, theoretical understanding of when and how synthetic data can improve downstream estimation and prediction remains underdeveloped. This question is especially relevant when synthetic samples are generated from proposal mechanisms informed by trained generators or domain knowledge, whose distributions may differ from the target population. To address this question, we propose REAL-Syn: Risk-guided Estimation-aware Acceptance for Learning with Synthetic Data, a framework for statistical learning with synthetic data selection. REAL-Syn uses selected subsets of candidate synthetic data to reduce estimator variability and often improve downstream prediction under target-proposal mismatch. The proposed procedure uses a covariance-based risk surrogate to select from the set of candidate synthetic samples; the accepted samples are then reweighted and pooled with the real observations for inference. We present theoretical results on the surrogate-optimal acceptance rule and the proposed selection procedure. We empirically show that REAL-Syn achieves competitive performance in simulations and an application to solar flare intensity forecasting.

Risk-sensitive optimal stopping arises wherever a single bad outcome dominates a sequential decision, such as option exercise, sequential medical treatment, and safety-critical control. The natural objective is the conditional value-at-risk (CVaR) of the stopped reward. We develop the first theoretical framework and algorithm for risk-sensitive deep optimal stopping. Together they resolve all four structural problems of risk-sensitive RL: state augmentation, score-function pathology, blindness to success, and bootstrapping circularity. We isolate three structural properties of optimal stopping: single-point reward, action-independent dynamics, and sum-zero gradient. They yield Markov optimality on the original state space and uniform boundedness of the per-step pathwise gradient at the deterministic-policy boundary, resolving the first two. From these, we derive DIOS (Distribution-Informed Optimal Stopping), a single-time-scale CVaR estimator on the Rockafellar-Uryasev envelope that resolves the remaining two. A soft penalty replaces the hard tail indicator, and a tail-level annealing schedule initializes training in the risk-neutral regime, where no trajectory is gated out. The inner quantile admits a closed-form update without an auxiliary critic. DIOS admits a finite-time best-iterate stationarity guarantee under standard CVaR non-degeneracy, the first explicit rate for CVaR optimal stopping. On Bermudan max-call options, DIOS achieves the highest mean CVaR and the lowest wall time against five representative CVaR baselines.

Predictive models are often deployed through existing decision policies that stakeholders are reluctant to change unless a risk constraint requires intervention. We study risk-controlled post-processing: given a deterministic baseline policy, choose a new policy that maximizes agreement with the baseline subject to a chance constraint on a user-specified loss. At the population level, we show that the optimal policy has a threshold structure: it follows the baseline except on contexts where switching to the oracle fallback policy yields a large reduction in conditional violation risk. At the finite-sample level, given a fitted fallback policy and score, we develop a post-processing algorithm that uses calibration data to select a threshold. Leveraging tools from algorithmic stability and stochastic processes, we show that under regularity conditions, in the i.i.d. setting, the expected excess risk of the post-processed policy is $O(\log n/n)$. In the special case when an exact-safe fallback policy is available, the algorithm achieves precise expected risk control under exchangeability. In this setting, we also give high-probability near-optimality guarantees on the post-processed policy. Experiments on a COVID-19 radiograph diagnosis task, an LLM routing problem, and a synthetic multiclass decision task show that targeted post-processing can meet or nearly meet risk budgets while preserving substantially more agreement with the baseline than score-blind random mixing.

Specialized agents collaborate by transferring memory: retrieved records, generated summaries, tool traces, and long-running state. In regulated domains this is a governance problem, not a retrieval problem: the receiver's task matters, but so does whether it was authorized to receive each item and whether the handoff exposed more than necessary. We formalize minimum-necessary context selection as a constrained submodular subset-selection problem with an additive exposure cost and a distributional task-success risk bound. Our algorithm, ScopeSelect, first hard-masks items whose layer or scope the receiver contract does not admit, then runs a cost-benefit greedy selector against a monotone submodular utility, attaining a $\\frac{1}{2}(1-1/e)$ approximation bound under a budget, and finally wraps threshold selection in Conformal Risk Control, giving a finite-sample distribution-free guarantee $\\mathbb{E}[L(S_{\\hat{\\gamma}})] \\le \\alpha$ on exchangeable calibration data. We evaluate on HealthHandoff-Bench, a synthetic benchmark with 21,600 healthcare-agent handoff tasks across four workflows, and Synthea-HH, a realism tier ingested from 340 Synthea-generated FHIR bundles. A three-source blinded label audit compares benchmark labels, model-generated labels, and clinical-informatics expert review, with high agreement for sensitivity labels and policy-boundary disagreements for necessity and authorization. Against authorized TF-IDF, embedding retrieval, a cross-encoder reranker, and a calibrated LLM-as-selector, ScopeSelect is the only method meeting $\\alpha = 0.15$ at competitive task success. The hard-mask safety property holds under 30% label noise, 100 adversarial attacks, and receiver-role shift; a live-LLM run with Claude Sonnet 4.5 exhibits a calibration-transfer gap that motivates per-workflow recalibration. We release the benchmark, Synthea-HH pipeline, CRC calibration harness, and a signed-handoff prototype.

Open-loop action repetition—hereafter also referred to as a skip—enables reinforcement learning agents to reduce decision frequency by executing selected actions for adaptive durations. A key challenge is how to evaluate the value of such a skip. Existing Skip-MDP methods typically bootstrap from the greedy value of the terminal state, implicitly assuming that the agent can immediately resume near-optimal control after the skip. This assumption is unreliable during learning, when the underlying action policy is stochastic, inaccurate, and continually changing. It can also amplify optimistic estimation errors, distorting the learned preference over skip lengths. We propose Expected Skip Evaluation, a reformulation of skip-value learning that evaluates skip terminal states through expected future control rather than idealized greedy control. This reflects imperfect control during training while preserving the optimal skip-value fixed-point as control becomes reliable. We instantiate this principle in RARe (Risk-Aware Repetition), a practical framework for discrete and continuous action spaces. Experiments across grid-world, continuous-control, and safety-critical benchmarks show improved sample efficiency and better-calibrated skip decisions.

Online POMDP planners optimize the expected cumulative cost, which can mask dangerous states when the belief places significant mass on high-cost states. Existing risk-averse methods apply static or dynamic Conditional Value at Risk (CVaR) to the value function, capturing trajectory-level risk, but share two gaps: (i) by retaining the immediate cost as an expectation of a state-dependent cost over the belief, the risk \emph{within} the belief is left unaddressed; and (ii) by modifying the value function, they require new tailored algorithms rather than reusing existing expectation-based planners. We instead apply CVaR to the immediate cost over the belief at each step, directly targeting per-step uncertainty about the current state. The standard expected cumulative return is retained as the objective, so the resulting problem has a standard MDP structure: any expectation-based POMDP planner can be made risk-sensitive by changing only the cost computation. We inherit finite-time guarantees for policy evaluation and sparse sampling---with estimation error independent of the risk level---and, as our central theoretical result, prove a finite-time bound on the gap between the particle belief MDP surrogate and the original POMDP, which together yield an end-to-end guarantee from the true POMDP value to the algorithmic estimate. In the risk-neutral limit, the formulation recovers standard expectation-based planning.

Predicting future clinical events from longitudinal electronic health records (EHRs) requires selecting plausible outcomes from a large and structured event space under sparse observations. While clinical coding systems provide hierarchical organization of events, cross-modal and temporal relationships are not explicitly specified and must instead be inferred from data, making prediction difficult for weakly observed longitudinal transitions. We introduce Risk Horizons, a geometry-aware framework for constructing patient-specific candidate spaces for multi-modal next-visit prediction. Risk Horizons combines deterministic coding hierarchies with data-driven lagged cross-modal associations, embeds the resulting clinical graph in hyperbolic space, and retrieves candidate futures using directional risk cones. This reframes longitudinal prediction as ranking within a compact, clinically coherent hypothesis space rather than scoring an unconstrained vocabulary. Experiments on MIMIC-IV and eICU demonstrate competitive next-visit prediction performance, with consistently improved hierarchy consistency across diagnoses, procedures, and medications. Further analysis suggests that hyperbolic structured candidate retrieval is the primary driver of performance, while LLMs are effective as constrained inference-time rerankers operating over clinically grounded candidate sets.

Rip currents are a serious and often under-addressed threat to beach safety, and the leading cause of coastal drownings worldwide. They are difficult to detect due to their amorphous structure, similarity to the background, and large variability in viewpoints and environments. Despite their dangers and growing interest in automated detection, existing research remains fragmented across datasets, metrics, models and approaches, leading to an incomplete understanding of model performance. We introduce RipBench, a unified benchmark that enables controlled evaluation of rip current detection spanning multiple tasks, including classification, axis-aligned and oriented object detection, and instance, semantic, and panoptic segmentation, on the same data with standardized splits. This allows direct comparison of model performance on all levels of visual abstraction. Across $304$ videos ($303{,}491$ frames) collected from diverse coastlines, our results expose a clear performance gap between coarse recognition and precise spatial understanding. While models achieve near-saturated classification performance, accurate localization proves to be a substantially more challenging task, with performance varying across tasks. All tasks are supported with carefully curated annotations and evaluated using both standard and safety-critical metrics, with a focus on the $F_2$ score to emphasize recall in this safety-critical setting. RipBench, along with multiple baseline models per task, is publicly available at \url{https://RipBench.ai} to support progress in real-world, multi-task vision for beach safety.

Reconstructing multi-camera videos is a fundamental task in autonomous driving. However, recent learned methods either rely on costly attention-based networks for global geometry modeling or require complex post-hoc alignment for chunked predictions, limiting their scalability to long multi-view rig sequences. We propose RigRecon, a generalized rig-aware reconstruction framework that enables efficient cross-view interaction in a low-resolution space while asynchronously decoding high-resolution depth maps in a frame-wise manner. Input images are first down-sampled and processed by a transformer backbone to capture coarse global structure, which is then provided to a decoder to reconstruct frame-wise high-resolution depth maps. By fully leveraging calibration and low-resolution feature interaction, RigRecon achieves consistent-scale geometry and accurate pose estimation with significantly reduced memory cost. It achieves state-of-the-art performance on 3D reconstruction, depth estimation, and camera pose estimation, and can efficiently process videos with over 1000 images in a single feedforward pass.

Static neural point reconstructions capture a subject at high fidelity from posed images. Given such a reconstruction and a monocular fixed-viewpoint driving video of the subject, whether captured or produced by image-to-video (I2V) generation, we recover a rigged, re-posable 3D asset. Existing methods deform Gaussian splats through direct linear blend skinning (LBS) or mesh proxies, both of which are prone to joint-boundary artifacts under articulation, even with per-primitive corrections. We trace the artifact to the representation: each splat carries an individual shape calibrated in the canonical pose to tile with its neighbours. Under rigid LBS, each splat moves with its bone but cannot bend, so the canonical tiling breaks at joint boundaries into gaps and spikes. Proximity attention point rendering (PAPR) instead carries no per-primitive shape; each pixel is recomposed at render time from the deformed primitives' positions, so the surface re-forms naturally with the articulation. We present RigPAPR, which auto-rigs a static PAPR cloud and drives it under direct LBS from a single fixed-viewpoint video, without mesh proxy, pose-dependent correction, or category template. On synthetic subjects, RigPAPR matches the strongest baseline at the supervised view and exceeds mesh-based and Gaussian-splatting baselines at novel views by 3+ dB PSNR, with cleaner joint-boundary renderings of both synthetic and real subjects.

We introduce a unified framework that reinterprets first order optimization on a smooth parameter space as a closed loop controlled dynamical system on a Riemannian manifold. Within this framework, optimizers are specified by a quadruple $(M, g, \Phi, \eta)$ consisting of a Riemannian metric, an internal state transition map, a direction field generator, and a lifting parameter. Classical and modern methods including SGD, Adam, AdamW, Lion and Shampoo are recovered as specific instantiations of this quadruple. The central structural object of the framework is a time indexed family of target submanifolds in the extended state space of parameters, velocities, and internal memory, which we call the Normally Attracting Invariant Manifold (NAIM) family. These submanifolds organize training dynamics into two timescales: a fast contraction of the velocity state onto the current target submanifold, followed by a slow descent along it. Using Riemannian backstepping, we construct a strict Lyapunov function on the extended state space and prove Uniform Ultimate Boundedness (UUB) of the optimization trajectory under a descent alignment assumption and bounded stochastic disturbances. Under the additional Polyak-Lojasiewicz (PL) condition, the Lyapunov function converges linearly to a bounded neighborhood of the optimum, defined as the drift of the target direction field between consecutive iterates after parallel transport. To validate the framework's generativity, we derive three concrete instantiations of the quadruple, each exercising a distinct design axis, and show on geometric diagnostics, image classification benchmarks, and language modeling that their empirical behavior matches the theoretical predictions. The framework casts optimizer design as controller synthesis, complementing heuristic perspectives by providing a constructive template with provable guarantees.

This paper studies bilevel optimization on Riemannian manifolds where the upper-level objective is nonconvex and the lower-level problem satisfies a Riemannian Polyak - Łojasiewicz condition rather than geodesic strong convexity. The classical hypergradient formula then breaks down, since the lower-level Hessian may be singular or indefinite away from the minimizer. We address this with an intrinsic regularized tangent-space formulation based on spectral clipping, and develop a Riemannian bilevel algorithm that avoids Hessian-inverse solves and inner-loop differentiation. We establish an $O(1/T)$ rate on the averaged squared Riemannian gradient mapping and $O(\varepsilon^{-1})$ iteration complexity, and extend the method to the stochastic finite-sum setting. Experiments on Stiefel, Grassmannian, and Poincar\'e-ball problems with rank-deficient Hessians (BCI~IV-2a, UCI Superconductivity) and Stiefel-constrained meta-learning on MiniImageNet show that the proposed method is the only one whose linear-system error stays at machine precision and whose Riemannian gradient norm decreases monotonically, while Hessian-inversion and Neumann-series baselines fail to descend; the stochastic complexity exhibits the predicted $\Theta(1/B)$ decay.

In learnable logic gate networks, each neuron implements a $k$-input Boolean function. The dominant setting $k{=}2$ was chosen for tractability. We ask: when does wider fan-in help? At fixed parameter budget, each gate costs $2^k$ parameters, so increasing $k$ exponentially reduces width. Combined with the depth--fan-in bound $L \geq \lceil \log_k d \rceil$, this partitions tasks into two regimes: width-dominated (image classification), where more gates beat richer gates, and degree-dominated (parity, arithmetic), where $k{>}2$ is necessary. Our main finding is that representability does not imply learnability: at the regime boundary, the selection mechanism determines training success at identical architecture, and the optimal mechanism inverts between regimes. We formalize this via a Jacobian analysis: the softmax mechanism's positive-semidefinite (PSD) structure gives basin attraction absent from sigmoid. We train multilinear coefficients with exact gradients and Möbius-transform snapping (zero deployment error), and validate on 28 tasks and 5 real-world datasets across multiple selection mechanisms.

Moving beyond evaluations that collapse performance across heterogeneous prompts toward fine-grained evaluation at the prompt level, or within relatively homogeneous subsets, is necessary to diagnose generative models' strengths and weaknesses. Such fine-grained evaluations, however, suffer from a data bottleneck: human gold-standard labels are too costly at this scale, while automated ratings are often misaligned with human judgment. To resolve this challenge, we propose a novel statistical model based on tensor factorization that merges cheap autorater data with a limited set of human gold-standard labels. Specifically, our approach uses autorater scores to pretrain latent representations of prompts and generative models, and then aligns those pretrained representations to human preferences using a small calibration set. This sample-efficient methodology is robust to autorater quality, more accurately predicts human preferences on a per-prompt basis than standard baselines, and provides tight confidence intervals for key statistical parameters of interest. We also showcase the practical utility of our method by constructing granular leaderboards based on prompt qualities and by estimating model performance solely from autorater scores, entirely eliminating the need for additional human annotations.