arXiv:2606.02512v1 Announce Type: cross
Abstract: Geometrical frustration, a central paradigm in condensed matter physics, provides a unifying language for systems in which locally preferred interactions cannot be made globally compatible. Here, we use this language to formulate a minimal theory of frustrated neural timing, mapping repulsively coupled rhythmic units onto antiferromagnetic XY models. Within this framework, the condensed-matter concepts of local constraints, degenerate ground-state manifolds, metastability, and quench dynamics become a concrete diagnostic framework for structured neural phase dynamics. We analyze a hierarchy of geometries: a triangle as the minimal frustrated motif with two chiral 120{\deg} timing states, a tetrahedron whose reduced ground-state manifold consists of intersecting continuous branches associated with antipodal pairings, and a kagome lattice on which local constraints define a constrained three-coloring manifold. The kagome lattice reveals the central dynamical result: zero-temperature relaxation suppresses global synchrony but typically selects low-energy metastable torque-balanced states rather than exact ground states. Finally, we show how the phase theory can be carried back towards biophysical neural models by treating it as an effective-interaction target, where geometrical timing frustration is realized through preferred phase lags that become incompatible around closed motifs. This perspective suggests that weak global coherence in neural systems does not necessarily signal disordered activity, but can reflect structured local timing order shaped by a frustrated dynamical landscape.
Science Journals
arXiv:1708.08723v4 Announce Type: replace
Abstract: We study straight-line drawings of graphs where the vertices are placed in convex position in the plane, i.e., \emph{convex drawings}. We consider two families of graph classes with convex drawings: \emph{outer $k$-planar} graphs, where each edge is crossed by at most $k$ other edges; and \emph{outer $k$-quasi-planar} graphs, where no $k$ edges can mutually cross.
We show that the outer $k$-planar graphs are $\lfloor3.5\sqrt{k}\rfloor$-degenerate, and consequently that every outer $k$-planar graph can be colored with $\lfloor3.5\sqrt{k}\rfloor + 1$ colors. We further show that every outer $k$-planar graph has a balanced vertex separator of size at most $2k+3$. For each fixed $k$, these small balanced separators allow us to test outer $k$-planarity in quasi-polynomial time, e.g., this implies that none of these recognition problems is NP-hard unless the Exponential Time Hypothesis fails. We also show that the class of outer 3-quasi-planar graphs and the class of planar graphs are incomparable.
Finally, we restrict outer $k$-planar and outer $k$-quasi-planar drawings to \emph{full} drawings (where no crossing appears on the boundary of the outer face) and to \emph{closed} drawings (where the vertex sequence on the boundary of the outer face is a Hamiltonian cycle in the graph). For each $k$, we express \emph{closed outer $k$-planarity} and \emph{closed outer $k$-quasi-planarity} in extended monadic second-order logic. Since every outer $k$-planar graph has treewidth $O(k)$, Courcelle's theorem implies that closed outer $k$-planarity is linear-time testable. We leverage this result to further show that full outer $k$-planarity can also be tested in linear time.
arXiv:2606.01628v1 Announce Type: cross
Abstract: Biomolecules such as proteins and small-molecule ligands play a central role in biological systems, arising from the tight interplay between sequence and three-dimensional structure. Recent generative models for biomolecular co-design aim to capture this interplay by jointly modeling coupled modalities. However, existing approaches largely adopt a parallel execution of marginal generative processes, implicitly enforcing fixed synchronous coupling. We argue that a critical but overlooked degree of freedom lies in how these marginal processes are temporally coupled during training and generation, where inappropriate coupling can introduce high-variance supervision and inconsistent intermediate states, affecting modality consistency. To address this, we introduce GeoCoupling, a systematic framework that optimizes for temporal couplings between heterogeneous modalities. Empirical results across structure-based drug design and unconditional protein design demonstrate the learned couplings consistently outperform synchronous and randomly coupled baselines, yielding biomolecules with improved physical validity and diversity.
arXiv:2606.01652v1 Announce Type: cross
Abstract: Recently, partial differential equations (PDEs) have been used to directly model the measurement process in signal processing, although their evaluation is costly. In this paper, we propose a novel alternating direction method of multipliers (ADMM)-based algorithm called physics-aware linearized ADMM (PA-LADMM) for inverse problems from PDE-based measurement processes. The key idea is the linearization of the subproblem with PDEs, leading to a cost-efficient update rule that calls only a PDE solver and its gradient evaluation per iteration. The algorithm has a theoretical convergence guarantee under certain conditions. In addition, we combine it with deep unfolding to unroll the PA-LADMM and train its internal parameters using supervised data. Two distinct experiments, compressed sensing with optical fiber communication and image restoration from noisy anisotropic diffusion, demonstrated the effectiveness of the proposed algorithms.
arXiv:2606.01655v1 Announce Type: cross
Abstract: The Bayesian paradigm offers principled tools for sequential decision-making under uncertainty, but its reliance on a probabilistic model for all parameters can hinder the incorporation of complex structural constraints. We introduce a minimalist Bayesian framework that places a prior only on the location of the optimum, while eliminating nuisance parameters through profile likelihood. This yields a generalized posterior that naturally accommodates structural constraints. As a direct instantiation, we develop MINimalist Thompson Sampling (MINTS). For multi-armed bandits with mean constraints, we establish near-optimal non-asymptotic regret guarantees and sharp almost-sure asymptotic regret characterizations. In particular, MINTS attains the classical Lai--Robbins constant in the unstructured setting and automatically adapts to unimodal structure, achieving the sharp constant determined only by the immediate neighbors of the optimal arm.
arXiv:2606.01709v1 Announce Type: cross
Abstract: The ephemeral regions (ERs), which are short-lived bipolar magnetic regions that emerge across the solar cycle but do not appear as sunspots, play a crucial role in the Sun's magnetic flux budget. However, their properties, particularly the tilt distribution, are poorly constrained by observations. In this study, we isolate ERs from the Automatic Tracking Algorithm for Bipolar Magnetic Regions (AutoTAB) catalog during Solar Cycles 24 and 25 by applying flux and footpoint-separation thresholds. Although AutoTAB was designed to track high-flux regions, it also records ephemeral regions with fluxes of 10^19 to 10^20 Mx, placing them at the upper end of the ER spectrum. The isolated ERs have an average lifetime of 1.2 days. Footpoint separation begins at supergranular scales (about 20 Mm), grows during the first half of the lifetime, and then saturates.
ERs occur most frequently near solar minima, consistent with earlier studies and likely reflecting AutoTAB's greater sensitivity to weaker regions when strong BMRs are scarce. Tilt properties reveal a more complex picture. For lifetimes shorter than two days, ERs show a broad, noisy distribution with no systematic latitude dependence. Including longer-lived ERs produces a weak, though statistically insignificant, increasing trend with latitude, suggesting that short-lived ERs are shaped by turbulent convection, while stronger, longer-lived ERs may retain Coriolis-imparted tilts. Overall, these results support the view that ERs occupy the low-flux end of the BMR spectrum and contribute meaningfully to the solar dynamo.
arXiv:2606.01724v1 Announce Type: cross
Abstract: LEO satellite constellations, led by deployments such as Starlink, are playing an increasingly pivotal role in enabling global broadband connectivity. However, the reliability and performance of these space-based networks are highly sensitive to environmental dynamics, particularly localized weather phenomena that exhibit strong spatio-temporal variability. In this study, we present a continental-scale geospatial analysis of weather-induced performance degradation in the Starlink LEO network, with a focus on the contiguous United States. Leveraging a unique dataset comprising more than 870,000 terminal hours of minute-level telemetry from 1,292 Starlink terminals, we integrate high-resolution localized weather observations to quantify the impact of various meteorological conditions. We evaluated key performance indicators (KPIs)-including ping latency, ping drop rate, and signal quality-using spatial join techniques and time-aligned correlation with classified weather events. Our analysis reveals that severe weather events, such as thunderstorms with heavy rain or snow, have a pronounced effect on network performance. In particular, more than 55% affected terminals experienced substantial degradation. Temporal continuity analysis at the minute level shows that such degradation can lead to sustained impairments or full service outages lasting from several minutes to multiple hours.This work contributes to the first large-scale empirical study linking LEO satellite Internet performance with fine-grained weather data in both space and time. Our findings offer actionable insights for geospatial predictive modeling, weather-aware network provisioning, and resilient satellite communication system design. We also propose a framework for incorporating weather-inferred performance variability into future geospatial planning and service-level forecasting tools for LEO-based Internet systems.
arXiv:2509.07126v2 Announce Type: replace
Abstract: Gaze prediction is essential for addressing motion-to-photon latency and ensuring seamless foveated rendering in Virtual Reality. The reliability of gaze forecasting is highly sensitive to individual differences and the eye movements being predicted. We evaluate recurrent, transformer-based, and classification-guided architectures to assess their generalization capabilities across oculomotor events. Using the GazeBase VR and Meta Quest Pro datasets, we analyzed the relationship between the median (P50) and high-percentile (P95) error profiles across subjects. The analysis reveals significant performance variability, showing that subjects with low P50 errors do not always exhibit the lowest extreme-case errors. Consequently, low median errors do not guarantee the robustness of the utilized solution. We discuss inference performance and address the class imbalance problem in short-term gaze prediction. These results identify a gap in standardized evaluation methods, necessitating a shift toward P95-focused, subject-specific metrics to develop reliable and perceptually stable gaze-contingent systems.
Exploring the Capabilities of Large Language Model Encoders for Image-Text Retrieval in Chest X-rays
arXiv:2509.15234v2 Announce Type: replace
Abstract: Multimodal learning from paired medical images and clinical text is a central challenge in medical data-driven informatics, where effective cross-modal alignment is critical for scalable analysis and retrieval. In chest radiography, vision-language pretraining is constrained by heterogeneous radiology reports that contain abbreviations, impression-only notes, and institution-specific writing styles. Unlike general-domain settings, naively aggregating large collections of noisy reports can plateau or even degrade multimodal learning when reporting styles differ substantially. We propose a domain-adapted bidirectional large language model text encoder for chest radiograph reports, trained with masked token prediction and supervised contrastive learning on stylistically diverse but clinically equivalent report variants to produce robust, generalizable text embeddings. We then integrate this encoder into a dual-tower contrastive vision-language framework using parameter-efficient adaptation to improve image-text alignment. Across 1.6 million paired studies from public datasets and a de-identified hospital cohort, the proposed models improve bidirectional retrieval accuracy and external generalization, achieving GREEN scores of 0.308 on MIMIC-CXR and 0.618 on Open-I, while reducing the degradation observed when abbreviation-rich, impression-only hospital reports are added to training.
arXiv:2509.15385v3 Announce Type: replace
Abstract: We propose a Variable-Preconditioned Transformed Primal-Dual (VPTPD) method for solving generalized Wasserstein gradient flows based on the structure-preserving JKO scheme. This is a nontrivial extension of the TPD method [Chen et al. (2025) SIAM J. Sci. Comput.] incorporating proximal splitting techniques to address the challenges arising from the nonsmoothness of the objective function. Our key contributions include: (i) a semi-implicit-explicit iterative scheme that combines proximal gradient steps with explicit gradient steps to treat the nonsmooth and smooth terms respectively; (ii) variable-dependent preconditioners constructed from the Hessian of a regularized objective to balance iteration count and per-iteration cost; (iii) a proof of existence and uniqueness of bounded solutions for the generalized proximal operator with the chosen preconditioner, along with a convergent and bound-preserving Newton solver; and (iv) an adaptive step-size strategy to improve robustness and accelerate convergence under poor Lipschitz conditions of the energy derivative. Comprehensive numerical experiments spanning from 1D to 3D settings demonstrate that our method achieves superior computational efficiency--achieving up to a 20$\times$ speedup over existing methods-thereby highlighting its broad applicability through several challenging simulations.
arXiv:2509.15394v3 Announce Type: replace
Abstract: Accurate electricity demand forecasting is challenging due to the strong multi-periodicity of real-world demand series, which makes effective modeling of recurrent temporal patterns crucial. Decomposition techniques make such structure explicit and thereby improve predictive performance. Variational Mode Decomposition (VMD) is a powerful signal-processing method for periodicity-aware decomposition and has seen growing adoption in recent years. However, existing studies often suffer from information leakage and rely on inappropriate hyperparameter tuning. To address these issues, we propose VMDNet, a causality-preserving framework that (i) applies sample-wise VMD to avoid temporal leakage; (ii) represents each decomposed mode with frequency-aware embeddings and decodes it using parallel temporal convolutional networks (TCNs), ensuring mode independence and efficient learning; and (iii) introduces a Stackelberg game inspired bilevel scheme to guide the selection of VMD's two key hyperparameters. Experiments on three widely used electricity demand datasets show that VMDNet consistently outperforms state-of-the-art baselines.
arXiv:2605.05057v4 Announce Type: replace
Abstract: Open-vocabulary human-object interaction (HOI) detection requires recognizing interaction phrases that may not appear as annotated categories during training. Recent vision-language HOI detectors improve semantic transfer by matching human-object features with text embeddings, but their predictions are often dominated by object affordance and phrase-level co-occurrence. As a result, a model may predict \textit{cut cake} from the presence of a knife and a cake without verifying whether the hand, tool, target, contact pattern, and object state jointly support the action. We propose \textbf{ScriptHOI}, a structured framework that represents each interaction phrase as a soft scripted state transition. Rather than treating a phrase as a single class token, ScriptHOI decomposes it into body-role, contact, geometry, affordance, motion, and object-state slots. A visual state tokenizer parses each detected human-object pair into corresponding state tokens, and a slot-wise matcher estimates both script coverage and script conflict. These two quantities calibrate HOI logits, expose missing visual evidence, and provide training constraints for incomplete annotations. To avoid suppressing valid but unannotated interactions, we further introduce interval partial-label learning, which constrains unannotated candidates with script-derived lower and upper probability bounds instead of assigning closed-world negatives. A counterfactual script contrast loss swaps individual script slots to discourage object-only shortcuts. Experiments on HICO-DET, V-COCO, and open-vocabulary HOI splits show that ScriptHOI improves rare and unseen interaction recognition while substantially reducing affordance-conflict false positives.
arXiv:2510.03690v4 Announce Type: replace
Abstract: Real-world graph datasets often arise from mixtures of populations, where graphs are generated by multiple distinct underlying distributions. In this work, we propose a unified framework that explicitly models graph data as a mixture of probabilistic graph generative models represented by graphons. To characterize and estimate these graphons, we leverage graph moments (motif densities) to cluster graphs generated from the same underlying model. We establish a novel theoretical guarantee, deriving a tighter bound showing that graphs sampled from structurally similar graphons exhibit similar motif densities with high probability. This result enables principled estimation of graphon mixture components. We show how incorporating estimated graphon mixture components enhances two widely used downstream paradigms: graph data augmentation via mixup and graph contrastive learning. By conditioning these methods on the underlying generative models, we develop graphon-mixture-aware mixup (GMAM) and model-aware graph contrastive learning (MGCL). Extensive experiments on both simulated and real-world datasets demonstrate strong empirical performance. In supervised learning, GMAM outperforms existing augmentation strategies, achieving new state-of-the-art accuracy on 6 out of 7 datasets. In unsupervised learning, MGCL performs competitively across seven benchmark datasets and achieves the lowest average rank overall.
arXiv:2605.04534v2 Announce Type: replace
Abstract: Large language models (LLMs) are becoming increasingly embedded in students' learning practices, yet much of what is known about how students use LLMs and how this usage impacts learning comes from problem-solving domains or constrained experimental settings. We present an analysis of data on LLM usage collected during two offerings of a research-oriented course where students learn to read, reason about, and critique academic papers. Without restrictions on whether or how to use LLMs, students reported their LLM usage practices when asked to do these activities as a series of homework assignments during the course. This paper extends prior work done on data from a single offering of the same course by presenting a refined bottom-up categorization of LLM usage types, cross-labeled by the extent of student initiative these usages entail. Furthermore, we examine how LLM use impacts student learning, measured by performance on three midterms, looking at factors such as frequency and type of usage.
arXiv:2510.09222v3 Announce Type: replace
Abstract: Flow Matching (FM) has shown remarkable ability in modeling complex distributions and achieves strong performance in offline imitation learning for cloning expert behaviors. However, despite its behavioral cloning expressiveness, FM-based policies are inherently limited by their lack of environmental interaction and exploration. This leads to poor generalization in unseen scenarios beyond the expert demonstrations, underscoring the necessity of online interaction with environment. Unfortunately, optimizing FM policies via online interaction is challenging and inefficient due to instability in gradient computation and high inference costs. To address these issues, we propose to let a student policy with simple MLP structure explore the environment and be online updated via RL algorithm with a reward model. This reward model is associated with a teacher FM model, containing rich information of expert data distribution. Furthermore, the same teacher FM model is utilized to regularize the student policy's behavior to stabilize policy learning. Due to the student's simple architecture, we avoid the gradient instability of FM policies and enable efficient online exploration, while still leveraging the expressiveness of the teacher FM model. Extensive experiments show that our approach significantly enhances learning efficiency, generalization, and robustness, especially when learning from suboptimal expert data.
arXiv:2510.09260v2 Announce Type: replace
Abstract: Recent work has shown that RLHF is highly susceptible to backdoor attacks. However, existing methods often rely on rare tokens or fixed triggers, limiting their impact in realistic scenarios. In this work, we develop GREAT, a novel framework for crafting natural distributional backdoors in RLHF. Specifically, GREAT targets harmful response generation for a vulnerable user subpopulation featured by semantically violent requests paired with emotionally angry triggers. At the core of our framework is a trigger identification pipeline that operates in the model's latent embedding space, leveraging dimensionality reduction and clustering techniques to identify representative triggers. To enable this, we introduce a hierarchical and diversity-driven prompting strategy to construct Erinyes, a high-quality dataset of over 5,000 angry triggers curated from GPT-4.1. Our experiments show that GREAT significantly outperforms baselines in attack generalization to unseen triggers, while preserving standard utility and maintaining stealth under defenses.
arXiv:2606.02328v1 Announce Type: new
Abstract: We explore Riemannian optimization techniques for rank-factored matrix parameters, targeting contemporary deep learning applications. We examine ten points in the algorithm design space: two geometries for rank-$r$ matrices, three geometries for rank-$r$ partial isometries, and block-matrix variants of these five, where factors are shared across block-rows and block-columns. We apply our methods to the multihead attention parameters in small language models. After tuning learning rates, our methods do not conclusively outperform an AdamW baseline. Our implementations are available online.
arXiv:2510.10541v2 Announce Type: replace
Abstract: Current benchmarks are inadequate for evaluating progress in reinforcement learning (RL) for large language models (LLMs).Despite recent benchmark gains reported for RL, we find that training on these benchmarks' training sets achieves nearly the same performance as training directly on the test sets, suggesting that the benchmarks cannot reliably separate further progress.To study this phenomenon, we introduce a diagnostic suite and the Oracle Performance Gap (OPG) metric that quantifies the performance difference between training on the train split versus the test split of a benchmark. We further analyze this phenomenon with stress tests and find that, despite strong benchmark scores, existing RL methods struggle to generalize across distribution shifts, varying levels of difficulty, and counterfactual scenarios: shortcomings that current benchmarks fail to reveal.We conclude that current benchmarks are insufficient for evaluating generalization and propose three core principles for designing more faithful benchmarks: sufficient difficulty, balanced evaluation, and distributional robustness.
arXiv:2510.10943v2 Announce Type: replace
Abstract: Bias in large language models (LLMs) remains a persistent challenge, often leading to stereotyping and unfair treatment across social groups. While prior work has mainly focused on individual LLMs, the emergence of multi-agent systems (MAS), where multiple LLMs collaborate and communicate, introduces new and underexplored dynamics in how bias emerges, propagates, and amplifies. To systematically investigate these dynamics, we propose a simple evaluation framework with three agent-level metrics that quantify bias emergence, propagation, and amplification throughout multi-agent interaction. We evaluate MAS across three bias benchmarks under varying LLM backbones, social-group configurations, communication behaviors, and adversarial settings. Our results show that communication can trigger up to 70\% new bias emergence, propagate bias across over 80\% of agents, and amplify stereotypes by more than 3$\times$. We further find that denser and competitive communication generally increases bias. Finally, we demonstrate that MAS are highly vulnerable to simple bias injection attacks, and existing defense strategies provide only limited protection. Our findings provide important insights into the fairness and robustness of multi-agent LLM systems.
arXiv:2311.00260v2 Announce Type: replace
Abstract: In collaborative active learning, where multiple agents try to learn labels from a common hypothesis, we introduce an innovative framework for incentivized collaboration. Here, rational agents aim to obtain labels for their data sets while keeping label complexity at a minimum. We focus on designing (strict) individually rational (IR) collaboration protocols, ensuring that agents cannot reduce their expected label complexity by acting individually. We first show that given any optimal active learning algorithm, the collaboration protocol that runs the algorithm as is over the entire data is already IR. However, computing the optimal algorithm is NP-hard. We therefore provide collaboration protocols that achieve (strict) IR and are comparable with the best known tractable approximation algorithm in terms of label complexity.
arXiv:2506.07342v2 Announce Type: replace
Abstract: We study sketching trimmed statistics of a frequency vector, including the $F_p$ moment of the top-$k$ coordinates and of the trimmed-$k$ vector. Despite their natural role in robust analytics, this is the first time these problems have been studied in any sublinear space setting. For $p \in [0,2]$, we obtain $poly(\log n/\varepsilon)$-space algorithms for both tasks when $k$ is moderately large, and for general $k$ we identify a sharp structural threshold that characterizes exactly when sublinear space is possible: in particular, it is actually determined by the ratio between $a_k^2$ and $\|x_{-k}\|_2^2/k$. We extend these results to $p > 2$ and present several applications including algorithms for thresholded $F_p$ estimation and generalized impact indices. Notably, we improve the space bounds of Govindan, Monemizadeh, and Muthukrishnan (PODS 2017) for computing the $h$-index.
arXiv:2506.09035v2 Announce Type: replace
Abstract: We introduce Princeton365, a large-scale diverse dataset of 365 videos with accurate camera pose. Our dataset bridges the gap between accuracy and data diversity in current SLAM benchmarks by introducing a novel ground truth collection framework that leverages calibration boards and a 360-camera. We collect indoor, outdoor, and object scanning videos with synchronized monocular and stereo RGB video outputs as well as IMU. We further propose a new scene scale-aware evaluation metric for SLAM based on the optical flow induced by the camera pose estimation error. In contrast to the current metrics, our new metric allows for comparison between the performance of SLAM methods across scenes as opposed to existing metrics such as Average Trajectory Error (ATE), allowing researchers to analyze the failure modes of their methods. We also propose a challenging Novel View Synthesis benchmark that covers cases not covered by current NVS benchmarks, such as fully non-Lambertian scenes with 360-degree camera trajectories. Please visit https://princeton365.cs.princeton.edu for the dataset, code, videos, and submission.
arXiv:2510.11423v4 Announce Type: replace
Abstract: Community Notes, the crowd-sourced misinformation governance system on X (formerly Twitter), allows users to flag misleading posts, attach contextual notes, and rate the notes' helpfulness. However, our empirical analysis of 30.8K health-related notes reveals substantial latency, with a median delay of 17.6 hours before notes receive a helpfulness status. To improve responsiveness during real-world misinformation surges, we propose CrowdNotes+, a unified LLM-based framework that augments Community Notes for faster and more reliable health misinformation governance. CrowdNotes+ integrates two modes: (1) evidence-grounded note augmentation and (2) utility-guided note automation, supported by a hierarchical three-stage evaluation of relevance, correctness, and helpfulness. We instantiate the framework with HealthNotes, a benchmark of 1.2K health notes annotated for helpfulness, and a fine-tuned helpfulness judge. Our analysis first uncovers a key loophole in current crowd-sourced governance: voters frequently conflate stylistic fluency with factual accuracy. Addressing this via our hierarchical evaluation, experiments across 15 representative LLMs demonstrate that CrowdNotes+ significantly outperforms human contributors in note correctness, helpfulness, and evidence utility.
arXiv:2510.11560v2 Announce Type: replace
Abstract: The advent of LLMs has given rise to generative search, a new search paradigm in which LLMs retrieve information from the web related to a query and synthesize it into a single, coherent response. This paradigm differs fundamentally from traditional web search, where results are returned as a ranked list of independent web pages. In this paper, we ask: Along what dimensions does generative search differ from traditional search? We conduct a systematic comparison between Google organic search and five generative search systems from three providers: Google, OpenAI, and Perplexity. Our analysis reveals substantial variation among engines in their reliance on internal v.s. external knowledge, source diversity, and stability. While generative systems often achieve topical coverage comparable to traditional search, they do so using markedly different retrieval footprints and synthesis strategies. We further show that the outputs of generative search can vary across time and executions, raising new challenges for robustness. Our findings demonstrate that generative search introduces new dimensions that are not captured by existing evaluation paradigms, motivating the development of evaluations that explicitly account for retrieval behavior, synthesis, and stability in generative search systems.
arXiv:2510.11713v4 Announce Type: replace
Abstract: Real-world applications of Large Reasoning Models (LRMs) often require reasoning about changing prompts or environments. In this work, we challenge the frozen world assumption and evaluate LRM robustness under two realistic dynamic scenarios: interruptions, which test the accuracy of model responses under budget-constrained outputs, and dynamic context, which tests model adaptation to in-flight changes. Across mathematics and programming benchmarks that require long-form reasoning, static evaluations consistently overestimate robustness: even state-of-the-art LRMs, which achieve high accuracy in static settings, can fail unpredictably when interrupted or exposed to changing context, with performance dropping by up to 60% when updates are introduced late in the reasoning process. Our analysis further reveals several novel failure modes, including reasoning leakage, where models fold the reasoning into their final answer when interrupted; panic, where under time pressure models abandon reasoning entirely and return incorrect answers; and self-doubt, where performance degrades when trying to incorporate updated information. Project Page: http://dynamic-lm.github.io/