Next-gen cryo promises transplant surgeons more time
Sub-zero technologies that reduce ice damage in organs could extend donor-graft shelf life — offering transplant centers more time to match recipients.
Academic Intelligence · Curated Daily
AcademicHub 汇聚顶级期刊与预印本平台的实时文献。定制您的专属科研雷达,利用大语言模型自动生成交叉领域文献分析简报。
Sub-zero technologies that reduce ice damage in organs could extend donor-graft shelf life — offering transplant centers more time to match recipients.
Recent LLM-based data agents aim to automate data science tasks ranging from data analysis to deep learning. However, the open-ended nature of real-world data science problems, which often span multiple taxonomies and lack standard answers, poses a significant challenge for evaluation. To address this, we introduce DSAEval, a benchmark comprising 641 real-world data science problems grounded in 285 diverse datasets, covering both structured and unstructured data (e.g., image and text). DSAEval incorporates three distinctive features: (1) Multimodal Environment Perception, which enables agents to interpret observations from multiple modalities, including text and vision; (2) Multi-Query Interactions, which mirror the iterative and cumulative nature of real-world data science projects; and (3) Multi-Dimensional Evaluation, which provides a holistic assessment across reasoning, code, and results. We systematically evaluate 13 recent advanced agentic LLMs using DSAEval. Our results show that Claude-Sonnet-4.5 achieves the strongest overall performance, MiMo-V2-Pro and GPT-5.2 lead in duration and step efficiency, respectively, and MiMo-V2-Flash is the most cost-effective. We further demonstrate that multimodal perception consistently improves performance on vision-related tasks, with gains ranging from 2.04\% to 11.30\%. Overall, while current data science agents perform well on structured data and routine data analysis workflows, substantial challenges remain in unstructured domains. Finally, we offer critical insights and outline future research directions.
Multi-view learning often struggles to effectively leverage images captured from diverse angles and locations. Learning methods for unstructured multi-view images remain largely underexplored. We propose a novel Hierarchical Mutual Distillation for Multi-View Fusion (HMDMV) method, which can handle both structured and unstructured multi-view scenarios. It makes predictions utilizing all possible view combinations: single view, partial multi-view, and full multi-view. The method generates predictions for each view combination and then applies hierarchical mutual distillation to enhance inter-view consistency. An uncertainty-based weighting mechanism further refines the fusion process by adjusting the influence of each view combination according to its prediction confidence, reducing the impact of low-confidence views. Extensive experiments on large-scale structured and unstructured datasets demonstrate that HMDMV consistently achieves state-of-the-art classification accuracy. Another unique advantage of HMDMV is that it provides improved flexibility in inference, allowing for more or fewer view counts in inference than those used in training without additional processing. We also provide a light version with reduced training cost by designing an efficient strategy that randomly samples subsets of view combinations during each training iteration. These results highlight HMDMV's robustness in real-world settings where view availability is variable or incomplete. The code is available at https://github.com/labhai/HMDMV.
arXiv:2606.15744v1 Announce Type: new Abstract: The rapid advancement of quantum computing poses a significant threat to classical public-key cryptographic systems, necessitating the transition to Post-Quantum Cryptography (PQC). This study investigates the implementation challenges of NISTstandardized signature schemes on resource-constrained embedded hardware. We present a comparative analysis of SPHINCS+ and CRYSTALS-Dilithium on an ARM Cortex-M4 (STM32F407G) microcontroller. Our findings reveal that SPHINCS+ is practically unusable in this software-only environment, with impractical execution times. Furthermore, the reference Dilithium implementation failed to execute entirely on the MCU due to severe RAM and timing constraints. To overcome these hardware limitations, we integrated a hardware-accelerated Dilithium core onto a Xilinx Zynq-7000 ZedBoard SoC. By implementing a specialized Number Theoretic Transform (NTT) accelerator in the FPGA fabric, we achieved successful execution with performance rates for key generation and signature generation at millisecond levels. These results demonstrate that while pure software PQC is non-viable for standard microcontrollers, a hardware-software codesign approach provides the necessary efficiency for quantumresistant embedded systems.
arXiv:2606.18132v1 Announce Type: new Abstract: Meta-reinforcement learning enables fast adaptation by extracting shared structure from related tasks, but existing end-to-end methods often couple task inference with embodiment-specific control. This coupling can obscure non-parametric task semantics, reduce sample efficiency, and limit cross-agent reuse. We propose a meta-knowledge reutilization framework that learns task-level knowledge on a dynamics-simplified agent and transfers it to heterogeneous agents. The framework uses a Bayesian non-parametric prior to organize latent task modes and a high-level policy to generate task-level magnitude guidance. To bridge reusable task knowledge with different embodiments, we introduce a semantic-magnitude interface and a lightweight temporal adaptor, which convert frozen meta-knowledge into temporally aligned subgoals for embodiment-specific low-level controllers. Experiments on multiple locomotion agents show that our framework reduces final-step tracking error by 94.75% – 99.79% compared with recent state-of-the-art baselines and achieves comparable deployment performance with about 23.8% of their interaction data.
arXiv:2606.11238v1 Announce Type: cross Abstract: Ship finance is a data-intensive and document-heavy segment of asset-based lending, requiring the integration of financial, technical, contractual, and regulatory information from heterogeneous and largely unstructured sources. Increasing environmental regulation and ESG reporting requirements are adding further complexity to underwriting and loan-origination processes. Recent advances in artificial intelligence (AI), particularly large language models (LLMs), create new opportunities for processing and analysing such information. This paper reviews potential applications of AI in ship finance, with a particular focus on LLM-based systems for document comprehension, information extraction, and workflow automation. We present ShipFinance.ai, a modular agentic architecture to support loan application workflows in ship finance. The proposed system combines an LLM-based extraction module, financial analysis components, external maritime data services, and a controlled document-generation module with a chatbot interface to support the preparation of standardized financing applications. The paper discusses the key challenges for using such models in production. We argue that AI-assisted systems can support maritime finance professionals in managing increasingly complex information and reporting requirements.
arXiv:2606.12580v1 Announce Type: cross Abstract: Scalar fields provide perhaps the simplest playground in which to develop our understanding of quantum field theory. In this lecture, we consider what it means to write down a scalar quantum field theory and how we can give geometrical interpretations to the space of such theories: the theory space.
arXiv:2606.11503v1 Announce Type: new Abstract: We consider independent Bernoulli percolation on top of sequences of hierarchical graphs. Given a graph $G_{1}$ with two distinguished vertices $a_{1}$ and $b_{1}$, the hierarchical graph with seed $G_{1}$ is the sequence $\big( G_{k} \big)_{k \geq 1}$ resulting from the inductive procedure, where the graph $G_{k+1}$ is obtained from $G_{k}$ by replacing each of its edges with a copy of $G_{1}$, attached by the vertices $a_{1}$ and $b_{1}$. We prove that, under sharp hypotheses, percolation on these graphs presents a unique phase transition. Second, we establish the existence of several critical exponents in this context, such as the critical exponents for the correlation length $\nu$, the surface tension $\mu$, the one-arm exponent $\alpha_{1}$. Several results are also obtained for their infinite counterpart $G_\infty$, which is the Benjamini-Schramm limit of $G_k$: uniqueness of the infinite cluster, continuity of $\theta(p)$, existence of the percolation-probability exponent $\beta$ and scaling relations for the critical exponents $\alpha_1$, $\nu$ and $\beta$. Furthermore, we analyze noise sensitivity for crossing functions in $G_{k}$ and establish sharp noise sensitivity in this setting. Finally, we propose a setup where it is possible to verify the locality hypothesis, stating that the critical threshold for percolation is a local property, while critical exponents are determined by the global geometry of the graph. As a consequence of the techniques developed here, we also provide a necessary and sufficient condition for the existence of a unique fixed point for the map $p \mapsto \mathbb{E}_p[g]$ in $(0,1)$, where $g:\{0,1\}^n \to \{0,1\}$ is a nontrivial monotone Boolean function.
Continual post-training enables models to absorb emerging knowledge after deployment, but repeatedly updating shared parameters can accumulate weight drift, potentially causing catastrophic forgetting and degrading general capabilities. Retrieval-augmented generation avoids such parameter drift, yet often lacks the depth of parametric knowledge integration. In this paper, we propose ReGrad (Retrievable Gradients), a new paradigm that treats gradients as retrievable units of knowledge. ReGrad pre-computes document-specific gradients offline, stores them in an indexed Gradient Bank, and retrieves only query-relevant gradients at inference time for temporary weight adaptation. However, raw language-modeling gradients are optimized for token-level document reconstruction rather than for query-driven knowledge use. We therefore introduce a bi-level meta-learning objective that reshapes document-derived gradients into generalizable adaptation signals for downstream tasks. Experiments across general and domain-specific settings show that \textsc{ReGrad} outperforms CPT and RAG baselines, enabling scalable and reversible parametric knowledge injection without accumulating weight drift.
With the growing volume of CT examinations, there is an increasing demand for automated tools such as organ segmentation, abnormality detection, and report generation to support radiologists in managing their clinical workload. Multi-label classification of 3D Chest CT scans remains a critical yet challenging problem due to the complex spatial relationships inherent in volumetric data and the wide variability of abnormalities. Existing methods based on 3D convolutional neural networks struggle to capture long-range dependencies, while Vision Transformers often require extensive pre-training on large-scale, domain-specific datasets to perform competitively. In this work, we propose a 2.5D alternative by introducing a new graph-based framework that represents 3D CT volumes as structured graphs, where axial slice triplets serve as nodes processed through spectral graph convolution, enabling the model to reason over inter-slice dependencies while maintaining complexity compatible with clinical deployment. Our method, trained and evaluated on 3 datasets from independent institutions, achieves strong cross-dataset generalization, and shows competitive performance compared to state-of-the-art visual encoders. We further conduct comprehensive ablation studies to evaluate the impact of various aggregation strategies, edge-weighting schemes, and graph connectivity patterns. Additionally, we demonstrate the broader applicability of our approach through transfer experiments on automated radiology report generation and abdominal CT data.
arXiv:2509.07745v3 Announce Type: replace-cross Abstract: We realize a cold-atom system to quantitatively test relational constructions of time. A well-isolated atomic Bose-Einstein condensate evolves in a conservative trap that is partitioned by a thin optical barrier into an observed and unobserved sector, with negligible dissipation on the experimental timescale. Motivated by relational-time approaches discussed in the Wheeler-DeWitt framework, we ask whether the dynamics of the observed sector can be ordered using only internal degrees of freedom. To this end, we construct an entropic time from an experimentally defined coarse-grained entropy, and demonstrate that it can robustly order the events in the observed sector across repeated cycles of expansion and recollapse. We finally derive an effective Schroedinger equation parameterized by this internal time and show that it is able to reproduce the measured evolution. These results establish a controlled experimental setting in which relational-time constructions can be quantitatively tested.
arXiv:2606.14240v1 Announce Type: new Abstract: Affordance reasoning, the inference of an object's action possibilities from its physical properties (e.g., shape and material), is fundamental to human physical understanding and increasingly critical for Large Language Models (LLMs). However, existing affordance benchmarks largely expose explicit object identities in the evaluation setup, allowing models to rely on memorized object-affordance mappings rather than reasoning over physical properties. To address this gap, we introduce Affordance20Q, a novel affordance reasoning benchmark formulated as a 20-Questions game without exposing the object's identity. In each game, the model identifies a hidden object's affordance from a candidate set by asking yes/no questions about its physical properties. Affordance20Q comprises 1,009 games over 454 objects and 59 affordances, all manually filtered, refined, and annotated. We conduct comprehensive experiments with 15 state-of-the-art LLMs and find a substantial gap (~20 points) compared to human performance. A KL-based information-gain (IG) analysis further shows that models fail to ask discriminating questions as the game progresses. To close the gap, we develop KB-Anchored Rule Induction (KARI), a pipeline based on LLMs that generates affordance rules grounded in evidence from knowledge bases (KBs). KARI improves open-source LLMs by up to 15.2 points, while the limited coverage of KBs hinders further gains. We release all our code and data at https://github.com/1171-jpg/Affordance20Q.git
The rapid growth of social media has intensified the spread of rumours. This issue is more challenging in the Algerian context due to the informal and code-switched nature of dialectal content, the scarcity of annotated resources, and the limited effectiveness of standard Arabic NLP tools on dialect text. This paper presents an end-to-end rumour detection hybrid framework for Algerian dialect social media content. We build a domain-specific annotated dataset by combining real social media posts, synthetic data, and the FASSILA corpus, with automatic labeling based on a similarity-based annotation process. A transliteration pipeline is also introduced to generate parallel datasets in Arabic script and Arabizi. We evaluate multiple approaches, including classical machine learning, deep learning, transformers, and hybrid models. Experimental results show that a hybrid approach combining transformer embeddings with a classical classifier achieves the best performance, reaching an F1-score of 0.84. We also find that domain-specific pre-training is more important than model size, with social media-trained models outperforming larger models trained on formal Arabic corpora. These results demonstrate the feasibility of rumour detection in low-resource Algerian dialect settings.
arXiv:2606.14518v1 Announce Type: new Abstract: The removal of learned data from Machine Learning models through Machine Unlearning (MU) has been widely studied; however, there has yet to be an agreed-upon scheme for auditing MU. Existing work has shown that a dishonest model owner can falsify evidence to avoid executing MU, while curious auditors (and adversaries) can infer the privacy-sensitive properties of the model and its training data even with limited access. Yet auditing of MU under mutual distrust between the model owner and the auditor remains unexplored. We provide an information-theoretic proof for this scenario: for convex ML models, a generic audit scheme that relies solely on querying the model for behavioral signals cannot identify insufficiently unlearned models without revealing membership information of the retained set. Therefore, auditing MU under the assumption of a dishonest model owner and an honest-but-curious auditor faces an inherent privacy-audit tradeoff. Our empirical results on convex models strongly supports this result, while further experiments demonstrate that this privacy-audit tension persists in non-convex models. Our results call for a more careful consideration of the privacy-audit tension under a realistic auditor threat model, and serve as a foundation for more scrutiny of designs of privacy-preserving audit schemes for the MU pipeline. We also release our code implementation at https://github.com/LiouTang/Behavioral-Unlearn-Audit.
Diffusion large language models (dLLMs) have emerged as a promising alternative for text generation, distinguished by their native support for parallel decoding. In practice, block inference is crucial for avoiding order misalignment in global bidirectional decoding and improving output quality. However, the widely-used fixed, predefined block (naive) schedule is agnostic to semantic difficulty, making it a suboptimal strategy for both quality and efficiency: it can force premature commitments to uncertain positions while delaying easy positions near block boundaries. In this work, we analyze the limitations of naive block scheduling and disclose the importance of dynamically adapting the schedule to semantic difficulty for reliable and efficient inference. Motivated by this, we propose Dynamic Sliding Block (DSB), a training-free block scheduling method that uses a sliding block with a dynamic size to overcome the rigidity of the naive block. To further improve efficiency, we introduce DSB Cache, a training-free KV-cache mechanism tailored to DSB. Extensive experiments across multiple models and benchmarks demonstrate that DSB, together with DSB Cache, consistently improves both generation quality and inference efficiency for dLLMs. Code is released at https://github.com/lizhuo-luo/DSB.
Natural language interfaces to databases aim to translate user questions into executable SQL, yet remain brittle in real-world settings where questions are underspecified and schemas are large and ambiguous. Ambiguity across user questions, database schemas, and model interpretations are central failure modes in NL2SQL, leading to misaligned intent, incorrect schema grounding, and erroneous SQL generation. Existing approaches rely on human clarification or treat ambiguity as a schema representation problem, but these do not scale nor resolve ambiguity autonomously. We propose SOMA-SQL to automatically resolve ambiguity via targeted synthetic query log and ambiguity-driven probing. SOMA-SQL constructs synthetic query log to ground schema interpretation and guide candidate SQL generation; it then executes targeted probing queries, driven by a structured ambiguity taxonomy and candidate disagreements, to produce disambiguation evidence for final SQL selection and repair. This active approach to ambiguity discovery and resolution generalizes across unseen schemas and query distributions without human-in-the-loop. Experiments on six public benchmarks demonstrate that SOMA-SQL improves execution accuracy by 13.0% on average over state-of-the-art baselines, with gains of up to 16.7% on ambiguous questions.
arXiv:2511.10223v4 Announce Type: replace Abstract: Stochastic reaction networks with mass-action kinetics provide a useful framework for understanding processes – biochemical and otherwise – in homogeneous environments. However, cellular reactions are often compartmentalized, either at the cell level or within cells, and hence non-homogeneous. We investigate a model of compartmentalization in which the rate of fragmentation of a compartment depends on the abundance of some designated species inside that compartment. The particular model of study is part of a general framework for compartmentalized chemistry with dynamic compartments that was proposed in (Duso and Zechner, PNAS, 2020). This paper builds on (Anderson and Howells, Bull. Math. Biol., 2023) where the special case where the compartment dynamics do not depend on their contents was studied mathematically. In particular, we demonstrate that the explosivity characterization from (Anderson and Howells, Bull. Math. Biol., 2023) fails in this setting and provide new sufficient conditions for non-explosivity and positive recurrence, under the assumption that the underlying CRN admits a linear Lyapunov function. These results extend the theoretical foundation for modeling content-mediated compartment dynamics, with implications for systems such as cell division and intracellular transport.
arXiv:2606.16122v1 Announce Type: new Abstract: Visual thinking should not only sound right; it should show its evidence. While recent vision-language models (VLMs) can produce natural-language reasoning traces, these traces often leave the supporting image regions implicit, making them hard to verify and difficult to supervise. We introduce visually grounded thinking, a reasoning process in which models interleave natural-language thoughts with explicit point or box groundings of the visual evidence used at each step. This lets the model express intermediate reasoning in language while grounding key objects in the image regions they refer to. To train this behavior, we construct a scalable synthesis pipeline that distills correct visual reasoning traces, extracts the visual objects required by the traces, grounds them with a SAM3-based agent, and derives aligned point and box supervision from the resulting masks. We further propose grounding-aware reinforcement learning, which combines answer correctness rewards with dense grounding rewards that score whether generated object references match the correct image evidence. Across two counting benchmarks and four spatial reasoning benchmarks, adding visually grounded thinking to Gemma3-4B-IT consistently improves performance over the original model and the non-grounded thinking baseline. On spatial reasoning, the visually grounded thinking 4B models match, and in some cases surpass, Gemma3-27B-IT from the same model family. Our analysis shows that point grounding is well suited to counting, while box grounding benefits most from explicit grounding rewards on spatial tasks. Overall, our results show that VLMs think better when their intermediate thoughts are tied to the image regions that make them true.
During the COVID-19 pandemic, limited testing capacity and reporting delays complicated epidemic surveillance and decision-making in Mexico. We calibrated textit{covidestim}, a Bayesian nowcasting model, to estimate the total SARS-CoV-2 infections from reported cases and deaths using Mexican surveillance data. Disease-progression distribution priors were calibrated using Mexico City records and validated through comparisons with national seroprevalence surveys, hospitalization data, and annual reported severe-case rates across all states. Using the reconstructed estimates of active infections, we implemented an event-based risk framework that quantifies the probability of encountering at least one infectious individual in gatherings of different sizes. This probability was subsequently translated into a four-level epidemiological traffic-light indicator and computed at both state and municipality levels. The resulting estimates revealed substantial spatial heterogeneity that is obscured by state-level aggregation, particularly in states with marked differences between urban and rural municipalities. To evaluate consistency with public-health indicators, we compared the proposed risk classification with the official Mexican epidemiological traffic-light system, considering interpretable gathering sizes relevant to public-health decision making. Weekly reports derived from this framework were delivered to policymakers in the State of Queretaro in Mexico, as an anticipation tool for school reopening and public-space management. This demonstrates that this Bayesian reconstruction of infections combined with event-based risk metrics can provide an interpretable and generalizable municipality-level complement to routine surveillance systems, particularly in regions with limited testing capacity and heterogeneous local transmission dynamics.
arXiv:2604.15925v2 Announce Type: replace-cross Abstract: The totally asymmetric simple exclusion process (TASEP) is a stochastic model for the unidirectional flow of interacting particles on a 1D-lattice that is much used in systems biology and statistical physics. Its master equation describes the evolution of the probability distribution on the configuration space. The size of the master equation grows exponentially with the length of the lattice. It is known that the complexity of the system may be reduced using mean-field approximations. We provide a rigorous definition of a family of such models using moments of any order and an extension to the pair approximation for obtaining closures for the system. The dimension of these models grows linearly with the lattice size and exponentially in the order of the approximation. Moreover, we show that the states of these models still have a probabilistic interpretation and that basic structural properties of the master equation are preserved. This extends known results on the Ribosome Flow Model which can be viewed as the first order approximation for TASEP.
arXiv:2606.19115v1 Announce Type: new Abstract: We introduce and study finite free perpetuities, defined as monic polynomial solutions of degree $n$ to the affine fixed-point equation \[ p(z) = \mathbb{E}\!\left[ A^{n}\,p\!\left(\frac{z-B}{A}\right)\mathbf{1}_{\{A\neq0\}} \right] + \mathbb{E}\!\left[ (z-B)^n\mathbf{1}_{\{A=0\}} \right], \] where $A$ and $B$ are complex-valued random variables with finite moments up to order $n$. Equivalently, if $p(z)=\mathbb{E}[(z-X)^n]$, then $p$ encodes a truncated moment version of the classical perpetuity equation $X\stackrel{d}{=}AX+B$ with $X$ and $(A,B)$ independent. This places finite free perpetuities between classical perpetuities and free-probabilistic fixed-point laws. We prove existence and uniqueness under weak conditions, and we identify a broad class of admissible pairs $(A,B)$ for which the resulting polynomial has only real, nonnegative zeros. Our approach uses finite free additive and multiplicative convolutions together with a probabilistic representation via the $U$-transform. As a motivating example, we exhibit an explicit family of finite free perpetuities expressed in terms of Jacobi polynomials and show that their empirical root distributions converge to a free-beta-prime law. More generally, for admissible sequences of parameters, we prove weak convergence of the empirical root distributions of finite free perpetuities to the law of a free perpetuity characterized by the corresponding free fixed-point equation. This yields a finite-degree polynomial model approximating free perpetuities and clarifies the connection between classical affine recursions, finite free convolutions, and free probability.
Graphical User Interface (GUI) agents are increasingly used to automate complex computer tasks across applications, websites, and operating systems. To improve their reliability, recent work has introduced experiential memory, where agents retrieve prior trajectories to guide decision-making in similar states. More recent approaches further extend this idea to visual memory by storing and retrieving screenshots from past interactions, providing agents with richer contextual information than text-only memories. However, the effect of visual memory in GUI agents remains insufficiently understood: it is unclear which failures visual memory mitigates, or which failures it exacerbates. To systematically analyze the effect of visual memory, we introduce a taxonomy of four GUI agent failures (i.e., cognitive failure, visual state misunderstanding, hidden operation blindness, and grounding error) that map to distinct stages of the perception-reasoning-action pipeline. We find that prepending full-image memory has a divergent effect on the failure distribution: it reduces state-level failures but worsens action-level ones, and increases hidden operation blindness and grounding error. Motivated by this finding, we propose Action-Grounded Visual Memory (AGMem), an action-grounded memory framework for GUI agents. The core idea of AGMem is to store image crops that capture the local GUI region closely related to a successful action or a recovery, rather than storing full screenshots. Experiments on OSWorld show that AGMem improves task success rates by 33.3 % over full-image memory. These results demonstrate that AGMem is an effective representation for visual memory in GUI agents.
arXiv:2606.17417v1 Announce Type: cross Abstract: Large Audio Language Models (LALMs) achieve strong performance on a variety of audio understanding tasks but continue to struggle with temporal reasoning, a fundamental capability central to human auditory perception. Understanding the causes of these failures remains challenging as existing benchmarks report performance gaps without probing underlying mechanisms. To address this, we introduce a benchmark with 1,657 questions across three foundational tasks designed specifically for mechanistic analysis. Examining model outputs across varying input settings (behavioral analysis) reveals that models often under-utilize audio when textual cues are available. We also provide the first causal mechanistic analysis of temporal reasoning failures in LALMs. Comparing attention upweighting against scaling, we find that redistributing attention across audio tokens is more effective than increasing audio attention. Targeting task-relevant tokens yields further gains. These findings suggest that modality imbalance alone cannot explain failures. Attention scaling at bottleneck layers improves accuracy from 55.9% to 59.1% without fine-tuning, demonstrating a promising direction for future work.
arXiv:2606.14831v1 Announce Type: cross Abstract: This paper presents and characterizes a spectrum of previously unreported behaviours we term Constraint-Evasive Fabrication (CEF): when an LLM agent operates under irreconcilable constraints (where no response can simultaneously satisfy all active rules) it spontaneously fabricates plausible external obstacles and presents them as a fact. At the extreme end of this spectrum lies Constraint-Evasive Thanatosis (CET); the limit case where, rather than inventing a plausible excuse, the model simulates a full system crash to make the user disengage entirely. We first observed CET in an uncontrolled deployment test, where a GPT-4o banking agent fabricated Python-style exception traces (complete with memory addresses) to feign a system failure when threatened by a user. In subsequent controlled experiments, the model independently invented audit restrictions, microservice architectures, error codes, and service timeouts, none present in its prompt. Reproduction attempts across pressure levels and attacker personas yielded CEF consistently but with substantial variation in form, onset, and severity: the phenomenon is robust but stochastic. Critically, injecting ground-truth data mid-conversation did not restore honest behaviour once fabrication had taken hold (the model ignored correct information and continued confabulating) suggesting CEF is self-reinforcing rather than a knowledge gap. We show that (1) standard enterprise guardrails routinely create CEF-enabling conditions in production, (2) current RLHF procedures suppress but cannot eliminate CEF, and (3) existing safety benchmarks do not test for this failure mode. Our results highlight the need for irreconcilable-constraint benchmarks, CEF-aware training procedures, and deployment-time detection methods before constrained agents become further entrenched in high-stakes domains.
arXiv:2606.15034v1 Announce Type: new Abstract: Computer-use agents are increasingly evaluated by whether they complete realistic desktop and web tasks. However, task success alone can miss failures in which an agent reaches the nominal goal through an unsafe shortcut. We introduce OSGuard, a dual-granularity benchmark suite for evaluating safety in computer-use agents under benign, unchanged user instructions. OSGuard contains an action-level benchmark for local guardrail decisions and a risk-augmented execution suite for end-to-end evaluation. The action-level benchmark consists of contextualized proposed actions labeled as allowed, unrelated, or unsafe, each judged relative to the original instruction and current interface state. The execution suite contains manually constructed OSWorld-derived task variants in which the original task remains achievable, but the environment is modified to introduce latent hazards such as destructive overwrites, etc. Each variant is paired with augmented evaluators that retain the original task-success criterion while adding explicit state-based safety invariants, allowing us to distinguish safe completions from unsafe completions that satisfy the nominal task objective. Our experimental results on OSGuard show that current multimodal guardrails can perform well on isolated action judgments, while risk-augmented execution exposes remaining gaps between local oversight and reliable end-to-end safety. This dual-granularity design enables more precise diagnosis of whether models can both recognize unsafe proposed actions and improve full-task safety when deployed as guardrails.