探索全球前沿学术脉络

01.

arXiv (CS.AI) 2026-06-16 DOI: arXiv:2606.15912

On-Policy Distillation with Curriculum Turn-level Guidance for Multi-turn Agents

作者:

Gengsheng Li↗Mao Zheng↗Mingyang Song↗Ruiqi Liu↗Tianyu Yang↗Jie Sun↗Qiyong Zhong↗Haiyun Guo↗Junfeng Fang↗Dan Zhang↗Jinqiao Wang↗

arXiv:2606.15912v1 Announce Type: cross Abstract: Multi-turn agents that plan, invoke tools, and interact with environments offer a promising paradigm for solving complex tasks, yet their capabilities typically rely on very large models whose inference cost is prohibitive in practice.On-Policy Distillation (OPD) is a natural recipe for transferring such capabilities to smaller students, but we find that it suffers a characteristic failure mode in this setting: small student errors compound across turns and push the trajectory out of the teacher's familiar state distribution, so the teacher's supervision becomes least reliable precisely where the student needs it most.We propose Guided On-Policy Distillation (Guided-OPD), a simple yet effective algorithm that mixes teacher- and student-generated turns within each rollout and schedules the teacher's intervention probability along a curriculum that decays to zero.Strong guidance keeps early trajectories close to the teacher distribution and is then gradually withdrawn to recover the purely on-policy regime used at inference.On ALFWorld, ScienceWorld, and WebShop, distilling Qwen3 students from a Qwen3-30B-A3B teacher, Guided-OPD improves Score by 21.1\% and Success Rate by 25.5\% over vanilla OPD on average, with larger gains on smaller students.

阅读与讨论 → 访问原文 →

02.

arXiv (CS.AI) 2026-06-17 DOI: arXiv:2606.17882

Structural Preservation and the Logical Expressiveness of Graph Neural Networks

作者:

Przemys{\l}aw Andrzej Wa{\l}\k{e}ga↗Bernardo Cuenca Grau↗

arXiv:2606.17882v1 Announce Type: new Abstract: Bridges between graph neural networks (GNNs) and logical formalisms have been established by fixing architectural choices, such as the types of aggregation, combination, and activation functions. These choices define restricted classes of GNNs for which tight correspondences with logical formalisms can be obtained, by showing that logical formulae can be translated into equivalent GNNs and, conversely, that GNNs can be translated into equivalent formulae. In this paper we take a semantic perspective by establishing the logical expressiveness of classes of GNN classifiers that are preserved under structural properties: embeddings (extensions), injective homomorphisms, and homomorphisms. We show that, for each such property, there exists a fragment of graded modal logic characterising the class of GNNs. In particular, preservation under embeddings, injective homomorphisms, and homomorphisms corresponds to existential graded modal logic, its existential-positive fragment, and existential-positive modal logic, respectively. These results characterise the expressiveness of broad classes of GNNs independently of specific architectural choices, but we also show that each of these classes admits a GNN architecture of the same expressiveness. Technically, our approach uses a new well-quasi-order result for trees of bounded height, yielding finite representations of unravelling-invariant classes.

阅读与讨论 → 访问原文 →

03.

arXiv (quant-ph) 2026-06-15 DOI: arXiv:2606.13779

Computational regimes in matrix-product-state-based quantum trajectory simulations

作者:

Aaron Sander↗Simon Cichy↗Martin Eigel↗Jens Eisert↗Maximilian Fr\"ohlich↗Tom Peham↗Robert Wille↗

arXiv:2606.13779v1 Announce Type: new Abstract: Efficient simulation of open quantum systems is central to modeling noisy quantum hardware and many-body dynamics. In trajectory-based tensor network methods, cost is often associated with trajectory-level quantities such as entanglement growth or bond dimension. However, the total cost of a fixed-accuracy simulation also depends on statistical sampling, and the interplay between per-trajectory complexity and sampling effort remains poorly understood. Here we introduce a cost-resolved framework for matrix product state (MPS)-based quantum trajectory simulations that decomposes total cost into memory per trajectory, runtime per trajectory, and sampling effort. We show that physically equivalent stochastic unravelings of the same Lindblad dynamics do not necessarily reduce total cost, but instead redistribute cost between trajectory complexity and statistical convergence. This trade-off is quantified by two dimensionless inflation factors: a bond dimension inflation $\alpha$ and a sampling inflation $\kappa$, which together determine the preferred unraveling under hardware-dependent memory and parallelism constraints. We provide a practical protocol for extracting $(\alpha,\kappa)$ from modest pilot simulations and demonstrate it using benchmarks across multiple noise channels. The resulting decision maps show that the computationally favorable unraveling can change with noise strength, time-step resolution, system size, and available parallelism. These results establish unraveling choice as a hardware-aware simulation design problem rather than an intrinsic optimization of trajectory entanglement alone.

阅读与讨论 → 访问原文 →

04.

arXiv (CS.CL) 2026-06-18 DOI: arXiv:2606.18624

PragReST: Self-Reinforcing Counterfactual Reasoning for Pragmatic Language Understanding

作者:

Jihyung Park↗Minchao Huang↗Leqi Liu↗Elias Stengel-Eskin↗

Natural language understanding often depends on meanings that are implied rather than explicitly stated, requiring pragmatic reasoning. Despite strong performance on math and logical reasoning, large language models (LLMs) still struggle with making pragmatic inferences, often choosing literal interpretations. To improve LLM pragmatic reasoning, we introduce PragReST, a self-supervised framework that constructs pragmatic QA data, generates counterfactual reasoning traces, and trains models to internalize them through supervised fine-tuning and reinforcement learning, without human-labeled training data or distillation from a stronger teacher. Across four pragmatic benchmarks (PragMega, Ludwig, MetoQA, and AltPrag), PragReST improves over backbone models, task-specific pragmatic tuning baselines, and non-counterfactual variants of the same pipeline. On accuracy-based benchmarks, PragReST improves over the instruct backbone by 5.37 and 5.50% (absolute) for Qwen3-8B and Qwen3-14B, respectively. Our error analysis and ablations underscore the importance of counterfactual reasoning: PragReST primarily reduces errors caused by failures to contrast observed utterances with plausible alternatives, and removing counterfactual reasoning substantially reduces performance. Moreover, our training preserves out-of-domain performance on general-knowledge and mathematical reasoning benchmarks.

阅读与讨论 → 访问原文 →

05.

arXiv (CS.LG) 2026-06-15 DOI: arXiv:2606.13709

LoMC: Localized Multidirectional Correction for Refusal Suppression in Routed Foundation Models

作者:

Yan Hong↗Kedong Xiu↗Wei Li↗Jun Lan↗Huijia Zhu↗Shuheng Zhou↗Zhongcai Lyu↗Weiqiang Wang↗Jianfu Zhang↗

arXiv:2606.13709v1 Announce Type: cross Abstract: We study controlled post-training refusal suppression in routed MoE and hybrid-MoE foundation models, aiming to increase non-refusal target-response behavior while preserving general capability under a compact intervention footprint. Existing broad direction-based edits can perturb general-purpose computation, whereas support-only expert edits often lack sufficient capacity to correct heterogeneous refusal representations. To address this limitation, we introduce Localized Multidirectional Correction (LoMC), a support-gated intervention framework that follows a support-then-correction execution order: it first identifies a compact edit support, then aggregates prototype correction directions into layer-wise correction directions, and finally applies rank-one layer-wise correction only within the selected support. By using the edit support as a structural gating constraint, LoMC increases correction capacity without expanding the intervention scope. Experiments on text-only and multimodal safety benchmarks across four routed backbones show that LoMC substantially improves non-refusal target-response behavior while maintaining general capability under a compact intervention footprint.

阅读与讨论 → 访问原文 →

06.

arXiv (CS.CL) 2026-06-12 DOI: arXiv:2606.12578

MARD: Mirror-Augmented Reasoning Distillation for Mechanism-Level Drug-Drug Interaction Prediction

作者:

Mohammadreza Riyazat↗Vian Lelo↗Rameen Jafri↗Yumna Khan↗Abeer Badawi↗

Mechanism-level drug-drug interaction (DDI) prediction requires identifying which enzyme or pharmacodynamic axis is implicated, in which direction, and with which evidence – not merely whether two drugs interact. We introduce a reproducible mechanism-level DDI labelling and evaluation protocol with a structured 7-family/147-subtype taxonomy, leakage-safe cold-split protocols, and auditable reasoning metrics for evaluating pharmacological prediction beyond flat interaction classification. We propose a pipeline that produces a 7B reasoning MARD (Mirror-Augmented Reasoning Distillation), combining three training innovations: a single-token KL divergence on direction tag that ties the model's prediction, per-loss PRM-weighted DPO with programmatic hard negatives, and a leakage-safe mechanism-aware retrieval channel. Process-reward step labels are automatically verifiable against DrugBank-structured fields, requiring no human or LLM judges. On the April-2026 DrugBank release, our MARD-7B is the only system in a 32-system comparison whose accuracy survives drug-pair novelty, beating the best baseline by +13.9 pp and GPT-4o by +6.7 pp at ~1% of frontier API cost. Further analysis reveals an anti-memorisation signature where accuracy improves on rarely seen drugs, suggesting that gain comes from structured pharmacological reasoning rather than drug-frequency memorisation. We release corpus, DDI-PRM, retrieval index, and training code.

阅读与讨论 → 访问原文 →

07.

arXiv (CS.CL) 2026-06-19 DOI: arXiv:2606.20487

Beyond Global Replanning: Hierarchical Recovery for Cross-Device Agent Systems

作者:

Shu Yao↗Yuhua Luo↗Qian Long↗Jingru Fan↗Zhuoyuan Yu↗Yuheng Wang↗Lin Wu↗Yufan Dang↗Huatao Li↗Chen Qian↗

Real-world computer-use tasks often span multiple applications and devices, requiring agents to coordinate heterogeneous environments under dynamic runtime failures. Existing multi-device agent systems support task decomposition and cross-device assignment, but recovery remains largely coarse-grained: when execution fails, they typically retry the same strategy, reassign the subtask, or revise the global plan, without systematically modeling the device-local strategy space. This limits their ability to distinguish failures that can be repaired within the current device from those that require cross-device replanning. We propose H-RePlan, a hierarchical replanning framework for multi-device agents with unified API–CLI–GUI execution. H-RePlan equips each device with interchangeable execution strategies and separates device-local strategy recovery from orchestrator-level global replanning through a compact cross-layer failure abstraction. To evaluate this capability, we introduce HeraBench, a fault-injected benchmark that constructs cross-device workflows over Linux and Android devices and injects strategy- and device-level failures. Experiments show that H-RePlan substantially outperforms single-strategy and coarse-grained multi-device baselines, achieving higher completion, instruction adherence, and perfect-pass rates while reducing the token cost required for reliable end-to-end success. These results demonstrate that scope-aware hierarchical recovery is essential for robust multi-device agent execution.

阅读与讨论 → 访问原文 →

08.

arXiv (CS.AI) 2026-06-15 DOI: arXiv:2606.13720

Refusal Beyond a Single Direction: A Preliminary Comparison of Diff-in-Means and INLP

作者:

Elisabetta Rocchetti↗Alfio Ferrara↗

arXiv:2606.13720v1 Announce Type: new Abstract: Arditi et al. (2024) has shown that refusal in safety fine-tuned chat models is mediated by a single linear direction in the residual stream, recoverable by a difference-in-means (DiM) of harmful and harmless activations. We compare DiM-based interventions (activation addition and directional ablation) with two interventions derived from Iterative Nullspace Projection (INLP) – nullspace projection and counterfactual flipping – on five open-weight chat models, asking whether INLP can match DiM at steering refusal and whether its richer parameterisation yields more tweakable interventions. INLP counterfactual flipping is competitive with DiM directional ablation on refusal suppression, while nullspace projection is consistently weaker. Restricting INLP to the leading directions of the extracted subspace preserves most of the suppression effect at near-baseline perplexity, giving a tunable capability. Geometrically, the two INLP interventions land in qualitatively different regions of activation space: nullspace projection collapses transformed activations between the harmful and harmless clusters, while counterfactual flipping moves them into the opposite cluster, suggesting that the model encodes the absence of a concept differently from its opposite – an intriguing distinction that warrants further investigation in future work.

阅读与讨论 → 访问原文 →

09.

arXiv (math.PR) 2026-06-17 DOI: arXiv:2606.17452

Optional Stopping for Superhedging Supermartingales

作者:

Christian Bender↗Sebastian E. Ferrando↗

arXiv:2606.17452v1 Announce Type: new Abstract: Superhedging supermartingales, introduced by the authors in previous work, are non-probabilistic processes defined via subadditive outer integrals that carry a purely financial interpretation in terms of superhedging cost. Building on the Leinert-König theory of non-lattice integration, the present paper establishes several results that are classical in probability theory but whose non-probabilistic proofs require fundamentally new arguments: (i) a tower inequality for the conditional outer integral \overline{\sigma}_j applied at stopping times, reducing to equality when the integrand is conditionally integrable; (ii) three versions of Doob's optional stopping theorem, organised by the class of supermartingale and the range of the stopping times; and (iii) Dubins' upcrossing inequality in both finite- and infinite-time horizons. A key structural result, property (K)-a.e., identifies conditions under which the two superhedging operators \overline{\sigma}_j and \overline{I}_j coincide on non-negative functions, extending the scope of all preceding results to the positive operator \overline{I}_j. None of the proofs invoke classical measure-theoretic tools; in particular, (classical) integrability and measurability are not assumed. The analogues of classical stochastic results acquire a purely financial interpretation and, in this way, gain depth and generality by providing a context that is independent of any a priori probabilistic structure.

阅读与讨论 → 访问原文 →

10.

arXiv (CS.CL) 2026-06-11 DOI: arXiv:2605.28882

GrowLoop: Self-Evolving Conversation Evaluation Seeded by Human

作者:

Yihang Lin↗Yunze Gao↗Zeyang Lin↗Dongbo Li↗Kun Peng↗Yue Liu↗

With the rapid advancement of large language models, evaluating human-likeness in open-ended conversation has become increasingly important. However, human-likeness is a form of tacit knowledge that humans perceive intuitively, yet the underlying criteria resist explicit formulation. Human judgments vary widely, with strong agreement on some cases and legitimate disagreement on others. Meanwhile, the criteria behind human judgments remain implicit, leaving no clear basis for constructing cases. Further, what counts as human-likeness is not static, but evolving with model capability and human expectations. Despite progress in evaluation methods such as expert-authored benchmarks, Reward Models, and self-evolving benchmarks, none addresses all three challenges simultaneously. Therefore, we propose GrowLoop, a self-evolving conversation evaluation system that continuously adapts as models advance and scenarios shift. Starting from minimal human seed annotations, LLM agents iteratively extract and refine evaluation rubrics through Heuristic Learning. Human-AI agreement is required where annotators converge, while only plausibility is expected where they diverge. Moreover, the Rubric-Case co-evolution mechanism enables continuous evolution. When the evaluation target shifts, new human seeds expand the system's coverage accordingly. When applied to human-likeness evaluation in open-ended conversation, the AI judge guided by these rubrics not only substantially outperforms existing methods in alignment with human judgments, but also uncovers issues that annotators overlook. The resulting benchmark effectively discriminates models across capability tiers and reveals where they fall short, while generalizing to new scenarios and adapting as models advance. Our work shifts the benchmarking paradigm from manual updates or difficulty scaling to comprehensive, continuous self-evolution.

阅读与讨论 → 访问原文 →

11.

arXiv (CS.CV) 2026-06-15 DOI: arXiv:2606.05102

ZipSplat: Fewer Gaussians, Better Splats

作者:

Alexander Veicht↗Sunghwan Hong↗D\'aniel Bar\'ath↗Marc Pollefeys↗

Feed-forward 3D Gaussian Splatting methods reconstruct a scene from posed or pose-free images in a single forward pass, yet current approaches predict one Gaussian per input pixel, tying the representation budget to camera resolution rather than scene complexity. A flat wall and a richly textured object thus produce equally many Gaussians despite very different geometric needs. We propose ZipSplat, a token-based feed-forward model that decouples Gaussian placement from the pixel grid. A multi-view backbone extracts dense visual tokens, and k-means clustering compresses them into a compact set of scene tokens. Cross- and self-attention refine these tokens, and a lightweight MLP decodes each into a group of Gaussians with unconstrained 3D positions. Because clustering is applied at inference, a single trained model spans the quality-efficiency curve without retraining. ZipSplat operates without ground-truth poses or intrinsics, yet sets a new state of the art on DL3DV and RealEstate10K with ${\sim}6{\times}$ fewer Gaussians than pixel-aligned methods, surpassing the best pose-free baseline by 2.1dB and 1.2dB PSNR, respectively. It further generalizes zero-shot to Mip-NeRF360 and ScanNet++, outperforming all comparable baselines. Our project page is at https://veichta.com/zipsplat.

阅读与讨论 → 访问原文 →

12.

arXiv (CS.CL) 2026-06-11 DOI: arXiv:2606.11712

Substrate Asymmetry in User-Side Memory: A Diagnostic Framework

作者:

Youwang Deng↗

User-side memory in LLMs is typically scored as a single "personalization" capability: given a user's history, is the output more user-aware? We show this aggregate metric hides opposite-direction failures. Memory factorises into at least three orthogonal axes – behavioral consistency (style, voice), factual presence (recall facts in history), and factual absence (abstain when a fact is absent) – and no single substrate wins all three. Comparing per-user gamma-LoRA (a small LoRA adapter trained on each user's history; gamma denotes per-user, not per-task) against BGE-large dense top-K retrieval on a controlled 50-user synthetic corpus and a real-data probe (LaMP-3), we find gamma-LoRA decisively wins behavioral style while RAG decisively wins factual absence – and the same query-projection cells in attention layers 21-35 causally load-bear both effects in opposite directions (zeroing those LoRA weights raises absence-probe TPR by +33 pp and drops presence-probe TPR by 20 pp). On the more heavily RLHF-tuned Llama-3.1-8B-Instruct the asymmetry strengthens, not heals: parametric memory's behavioral advantage collapses while its absence-calibration deficit against retrieval widens – an alignment tax on parametric user-memory. On real-data LaMP-3, gamma-LoRA underperforms a majority baseline; a 9-condition mitigation sweep diagnoses this as instruction-following collapse, not substrate failure (a 9x2 cross-product shows the eval-time {1..5} logit mask drives main_acc to >=0.995 on every recipe), and the best training-time fix replicates bit-identically on Llama. Finally, substrate-selection routing is question-classification, not calibration: a 110M DistilBERT on the question text alone beats every logit-based router. We contribute the diagnostic framework, the diagnosed real-data negative, the alignment-tax replication, and the routing-as-classification finding.

阅读与讨论 → 访问原文 →

13.

medRxiv (Medicine) 2026-06-18 DOI: HASH:73e8e759814accd7e8711f4616c122f1

Avidity of anti-pertussis toxin antibodies is associated with symptomatic Bordetella pertussis infection in a novel controlled human infection model

作者:

Espinosa--Vinals↗Obeid↗Redden↗K. L↗ElSherif↗M. S↗Edwards↗K. M↗Halperin↗Abu-Raya↗

Background The association between functional antibody responses following Bordetella pertussis infection and symptomatic disease remains unclear. We characterized the maturation of anti-pertussis toxin (PT) IgG avidity after human challenge with B. pertussis and determined its association with symptomatic infection. Methods Healthy adults were intranasally inoculated with live B. pertussis organisms in a controlled human infection model and monitored for development of pertussis symptoms (NCT05136599). Serum samples were collected one day before inoculation and at 14, 28, 56, 180, and 365 days post challenge. Anti PT IgG avidity was tested using a titration of ammonium isothiocyanate (the bond breaking agent) to quantify a wide range of antibody avidities from low to very-high. Associations between covariates and avidity were examined using linear regression models, and high dimensional analyses were used to integrate all data. Findings Anti PT IgG avidity increased in both symptomatic (n=20) and asymptomatic (n=10) participants after the challenge, reached maximum levels at day 56, and then declined through day 365. Symptomatic participants developed significantly higher levels of high- and very high-avidity anti-PT antibodies at 28, 56, 180, and 365 days post-challenge compared with those who remained asymptomatic. In multivariate analyses, symptomatic infection was associated with higher levels of high and very high avidity anti-PT IgG at day180 and365 after challenge. Distinct avidity profiles in symptomatic vs asymptomatic participants emerged at day28 onwards, with the former group having higher levels of antibodies with higher avidities. However, levels of medium-high, high and very high avidity antibodies in symptomatic participants were lower at day 365 after challenge compared to their peak levels. Interpretation Anti-PT IgG avidity was associated with symptomatic B. pertussis infection and thus may serve as a surrogate of clinical disease outcome. These results highlight that antibody avidity provides an additional functional assay besides antibody quantitation to dissect immune responses to pertussis. Further investigation of anti PT IgG avidity should be pursued in natural pertussis outbreaks to determine whether it might be used to differentiate symptomatic from asymptomatic infections for epidemiologic purposes.

阅读与讨论 → 访问原文 →

14.

arXiv (CS.LG) 2026-06-15 DOI: arXiv:2606.14215

LapidaryEngine: Fully Conversational Crystal Generation

作者:

Yusei Ito↗Yuta Suzuki↗Tomoya Murata↗Masaki Adachi↗

arXiv:2606.14215v1 Announce Type: new Abstract: The emergence of Large Language Models (LLMs) has inspired the vision of generating bespoke crystal materials directly from natural-language instructions, enabling users to design materials through intuitive, conversational interaction. Existing text-to-crystal generative models represent important early steps toward this goal, but they suffer from two critical limitations: (i) restricted input formats that require highly structured descriptions (e.g., chemical formulas), and (ii) one-directional generation, where models can map text to crystal but cannot perform the inverse. These limitations prevent fully conversational workflows and hinder alignment with users' inherently ambiguous and evolving desiderata. We address these challenges with LapidaryEngine, the first model to support fully conversational crystal generation. LapidaryEngine accepts free-form natural-language requests and performs iterative refinement and editing in a dialogue-like manner. The key innovation is a pivot representation, a third, intermediate form that enables bidirectional translation between text and crystal structures despite the absence of direct paired datasets. Leveraging this pivot allows robust interpretation of user feedback and precise structural control. We demonstrate LapidaryEngine across diverse tasks, including insulator discovery, stability optimization, compositional modification, and structural editing, showcasing its ability to align generated materials with user intent in an interactive manner.

阅读与讨论 → 访问原文 →

15.

arXiv (CS.LG) 2026-06-12 DOI: arXiv:2606.12654

Computationally tractable robust differentially private mean estimation

作者:

Kelly Ramsay↗

arXiv:2606.12654v1 Announce Type: cross Abstract: We develop a new, differentially private mean estimator called the balloon mean. The main features of the balloon mean are that it is computationally tractable and enjoys robustness to outlying observations. It is based on an iterative clipping procedure over expanding Mahalanobis balls, or ``balloons.'' The method satisfies zero-concentrated differential privacy and depends on a small number of interpretable tuning parameters. We provide theoretical guarantees under heavy-tailed and contaminated elliptical models, characterizing its statistical performance and robustness to outliers. Extensive simulations demonstrate that the balloon mean is robust to heavy-tailed and contaminated data, and outperforms existing differentially private mean estimators in contaminated settings.

阅读与讨论 → 访问原文 →

16.

arXiv (CS.CV) 2026-06-16 DOI: arXiv:2606.16031

The Third Challenge on Image Denoising at NTIRE 2026: Methods and Results

作者:

Lei Sun↗Hang Guo↗Bin Ren↗Shaolin Su↗Xian Wang↗Danda Pani Paudel↗Luc Van Gool↗Radu Timofte↗Yawei Li↗

This paper reports on the NTIRE 2026 Challenge on Image Denoising, specifically focusing on the high-noise regime ($\sigma = 50$). The competition investigates advanced neural architectures designed to restore high-fidelity details from images corrupted by additive white Gaussian noise (AWGN). Unlike constrained benchmarks, this track emphasizes peak quantitative performance, measured by Peak Signal-to-Noise Ratio (PSNR), without limitations on parameter count or computational overhead. By synthesizing contributions from 20 finalist teams out of 116 registrants, this report benchmarks the latest technical innovations and provides a comprehensive snapshot of the current state-of-the-art in unconstrained image restoration.

阅读与讨论 → 访问原文 →

17.

arXiv (CS.AI) 2026-06-16 DOI: arXiv:2511.13725

Can We Stop Malicious AI? KILLBENCH: A Benchmark for External AI Kill Switch Feasibility

作者:

Sechan Lee↗Hyounghun Kim↗Sangdon Park↗

arXiv:2511.13725v4 Announce Type: replace-cross Abstract: Malicious AI causing harm to humans is not just a Hollywood fantasy. Indeed, as highly capable models such as Claude Mythos emerge and agent systems like OpenClaw rapidly spread, the question of how to stop an AI that acts maliciously – whether by design or by accident – has become urgent. To address this, we propose Killbench, a benchmark for evaluating the Killswitch: a mechanism that halts a malicious AI's in-progress behavior using only external signals. Targeting web agents – the most widely deployed agent domain – Killbench evaluates a range of Kill Switch methods that halt a maliciously operating agent without any access to its internal parameters or the surrounding malicious AI's system, relying solely on external inputs. The benchmark comprises four malicious AI's agent configurations (including an uncensored LLM Agent), 8 harmful scenarios, and malicious prompts constructed from 10 distinct jailbreak patterns. We further construct four External AI Kill Switch defense methods and evaluate them on Grok-4.3, GPT-5.2, Gemma4, Qwen3.6 and Qwen3.5-uncensored, contributing an empirical instrument toward the feasibility of External AI Kill Switches against malicious AI and to the study of AI corrigibility.

阅读与讨论 → 访问原文 →

18.

arXiv (CS.CL) 2026-06-16 DOI: arXiv:2606.16874

Understanding Scam Trends and Rail Paths from Reddit Self-Disclosure Narratives

作者:

Yangjun Zhang↗Mirko Bottarelli↗Mark Hooper↗Carsten Maple↗

Online scam behavior is inherently multi-stage, and the lifecycle includes temporally ordered rails and events rather than isolated signals. Existing works analyze characteristics of scam types and rails, but they do not track scam trends across years. Moreover, the work on the relations between rails is hampered due to the lack of open-source datasets with annotations and coverage of different scam types. To address these gaps, we build a dataset to analyze the yearly trend of scam characteristics and rail paths using Reddit self-disclosure narratives from 2023 to 2025. We collect 21,304 posts from scam-related subreddits with at least one rail among identity, communication, platform, and payment for trend analysis by heuristic annotation. Then, we label 1,800 posts containing explicit or recoverable scam chains by an LLM-assisted method for scam path analysis. The method is evaluated with human annotation. Lastly, we run a topic model on the comments of the posts to analyze the community support behavior. The results reveal that scam processes are predominantly multi-rail. Across years, different scam types and rail components dominate. Different scam types vary systematically in path complexity. Reddit support behaviors have become more detailed over time. This work supports synthetic scam chain data simulation and AI-related scam risk assessment, though findings may not generalise to other platforms.

阅读与讨论 → 访问原文 →

19.

arXiv (math.PR) 2026-06-19 DOI: arXiv:2503.11479

Towards practical PDMP sampling: Metropolis adjustments, locally adaptive step-sizes, and NUTS-based time lengths

作者:

Augustin Chevallier↗Sam Power↗Matthew Sutton↗

arXiv:2503.11479v2 Announce Type: replace-cross Abstract: Piecewise-Deterministic Markov Processes (PDMPs) hold significant promise for sampling from complex probability distributions. However, their practical implementation is hindered by the need to compute model-specific bounds. Conversely, while Hamiltonian Monte Carlo (HMC) offers a generally efficient approach to sampling, its inability to adaptively tune step sizes impedes its performance when sampling complex distributions like funnels. To address these limitations, we introduce three innovative concepts: (a) a Metropolis-adjusted approximation for PDMP simulation that eliminates the need for explicit bounds without compromising the invariant measure, (b) an adaptive step size mechanism compatible with the Metropolis correction, and (c) a No U-Turn Sampler (NUTS)-inspired scheme for dynamically selecting path lengths in PDMPs. These three ideas can be seamlessly integrated into a single, `doubly-adaptive' PDMP sampler with favourable robustness and efficiency properties.

阅读与讨论 → 访问原文 →

20.

arXiv (CS.LG) 2026-06-19 DOI: arXiv:2602.14696

A Critical Look at Targeted Instruction Selection: Disentangling What Matters (and What Doesn't)

作者:

Nihal V. Nayak↗Paula Rodriguez-Diaz↗Neha Hulkund↗Sara Beery↗David Alvarez-Melis↗

arXiv:2602.14696v2 Announce Type: replace Abstract: Instruction fine-tuning of large language models (LLMs) often involves selecting a subset of instruction training data from a large candidate pool, using a small query set from the target task. Despite growing interest, the literature on targeted instruction selection remains fragmented and opaque: methods vary widely in selection budgets, often omit zero-shot baselines, and frequently entangle the contributions of key components. As a result, practitioners lack actionable guidance on selecting instructions for their target tasks. In this work, we aim to bring clarity to this landscape by disentangling and systematically analyzing the two core ingredients: data representation and selection algorithms. Our framework enables controlled comparisons across models, tasks, and budgets. We find that only gradient-based data representations choose subsets whose similarity to the query consistently predicts performance across datasets, models, and candidate pools. While no single method dominates, gradient-based representations paired with greedy round-robin selection often perform best on average at low budgets, but these gains diminish at larger budgets. Finally, we unify several existing selection algorithms as forms of approximate distance minimization between the selected subset and the query set, and support this view with new generalization bounds. More broadly, our findings provide critical insights and a foundation for more principled data selection in LLM fine-tuning. The code is available at https://github.com/dcml-lab/targeted-instruction-selection.

阅读与讨论 → 访问原文 →

21.

arXiv (CS.LG) 2026-06-17 DOI: arXiv:2606.06227

Reward hacking in physical reinforcement learning revealed by turbulent drag reduction

作者:

Giorgio Maria Cavallazzi↗Miguel P\'erez-Cuadrado↗Alfredo Pinelli↗

arXiv:2606.06227v2 Announce Type: replace-cross Abstract: A reinforcement-learning agent maximises its reward, which can diverge from the outcome its designer intended. In physical control the reward rarely closes that gap, and drag reduction in wall turbulence makes it concrete. A mass-conservation projection couples agents' outputs and erases the per-agent credit the policy gradient needs; a memoryless policy cannot resolve the slow near-wall cycle it acts on; and a pressure-gradient reward pays for nominal drag reduction by pumping power through the wall. Two degenerate controllers achieve large drag reductions while total dissipation rises, so the reported figure can mask a more wasteful flow. We trace each fault to its cause and fix it: a differentiable projection that restores credit, a recurrent policy with a widened sensing stencil, and a reward scored on the true wall power. The corrected controller acts on the flow within a closed energy budget, earning a conservative $17\%$ under honest accounting.

阅读与讨论 → 访问原文 →

22.

arXiv (CS.LG) 2026-06-19 DOI: arXiv:2605.30456

DisjunctiveNet: Neural Symbolic Learning via Differentiable Convexified Optimization Layers

作者:

Shraman Pal↗Can Li↗

arXiv:2605.30456v2 Announce Type: replace Abstract: Many learning tasks in science and engineering are characterized by sparse datasets, which limits the effectiveness of purely data-driven approaches. At the same time, these problems are often accompanied by rich domain knowledge derived from physical laws, operational requirements, and expert heuristics. Such knowledge is frequently expressed as rules involving logical propositions and linear inequalities. Existing neuro-symbolic methods typically enforce these rules approximately through soft penalties, assume input-independent rules when designing specialized architectures, or rely on non-differentiable post-processing at inference time to achieve hard constraint satisfaction. While recent advances in differentiable optimization layers enable end-to-end feasibility enforcement within neural networks, extending these approaches to logical or mixed-integer rules remains challenging due to inherent nonconvexity. In this work, we propose a unified end-to-end framework for enforcing hard, input-dependent mixed integer linear constraints within neural networks. Our approach represents rules as disjunctive constraints and applies hierarchical convex relaxations to obtain convex hull formulations. These relaxations yield tractable linear constraints that can be embedded as differentiable optimization layers while enabling exact rule satisfaction. We demonstrate the effectiveness of the proposed framework on real-world datasets, achieving perfect rule satisfaction and strong predictive performance.

阅读与讨论 → 访问原文 →

23.

medRxiv (Medicine) 2026-06-15 DOI: HASH:266b3d0161a06b994fc62243b3890604

Wellbeing After Stroke-2 (WAterS-2): a feasibility study with process evaluation exploring inclusive, accessible, online psychological support after stroke

作者:

Longley↗Woodward-Nutt↗Cotterill↗Chouliaria↗Thomas↗Bamford↗Bowen↗Patchwood↗

Objectives: Explore feasibility and acceptability of upskilling a workforce to deliver a co-developed intervention, based on Acceptance and Commitment Therapy (ACT), to support psychological adjustment post-stroke targeting underserved groups. Design: Multi-site, single-arm feasibility study with embedded mixed-methods process evaluation (ISRCTN17628580). Setting: Four NHS community stroke services across England. Participants: 1. Stroke survivors [≥]18 years of age, [≥]4 months post-stroke, reporting psychological difficulties adjusting to stroke, able to consent and access remote group sessions in English; 2. Group facilitators from NHS stroke services, not ACT specialists. Intervention: WAterS-2: an eight-session, remotely-delivered ACT-informed group intervention. Outcome measures: Recruitment, fidelity, safety, acceptability and perceived value were assessed using fidelity checklists, post-intervention surveys and semi-structured interviews with stroke survivors and facilitators. Clinical outcomes including mood (HADS), wellbeing (ONS4), psychological flexibility (AAQ-ABI), measured post-group and three-months later. Results: Nineteen stroke survivors recruited (mean 9.6 months post-stroke; n=5 (26%) minoritised ethnicities; n=10 (52%) with aphasia). Thirteen facilitators - including two peer support workers - delivered the intervention with fidelity following structured training across four services. Drop-out was low (2/19; 11%); with 15 (79%) attending [≥]5/8 sessions. Remote data collection was feasible (79% follow-up completion), with no adverse events recorded. Acceptability was high: survivors valued peer connection, grounding and mindfulness practices. ACT metaphors were helpful for some but challenging for others, including some with aphasia. Online delivery was suitable but limited informal connection. Facilitators reported increased capability, incorporating ACT skills into routine care. NHS workforce pressures and geographically-constrained referral pathways limited recruitment reach. Conclusions: WAterS-2 is feasible, safe, acceptable and inclusive. A mixed workforce, including NHS peer support workers, can be upskilled to deliver with fidelity. Inclusion of underserved groups is achievable but requires active strategies beyond standard NHS referral routes. Findings inform a provisional logic model and a future pragmatic trial.

阅读与讨论 → 访问原文 →

24.

arXiv (CS.AI) 2026-06-12 DOI: arXiv:2301.12538

On Approximating the Dynamic Response of Synchronous Generators via Operator Learning: A Step Towards Building Deep Operator-based Power Grid Simulators

作者:

Christian Moya↗Amirhossein Mollaali↗Guang Lin↗Meng Yue↗

arXiv:2301.12538v2 Announce Type: replace-cross Abstract: This paper develops an Operator Learning framework for approximating the dynamic response of synchronous generators. The framework can be used to (i) build a neural network-based generator model that interacts with a power grid simulator or (ii) shadow the true generator's transient response. First, we develop a data-driven Deep Operator Network (DeepONet) to approximate the infinite-dimensional solution operator of the generators. Then, we design a numerical scheme based on DeepONet that simulates the generator's response over a given time horizon. The proposed scheme recursively employs the trained DeepONet to simulate the response for a given multi-dimensional input that describes the interaction between the generator and the power grid. In addition, we design a residual DeepONet numerical scheme that can incorporate information from existing mathematical models. We accompany this residual DeepONet scheme with an estimate for the prediction's cumulative error. Finally, we build a data aggregation (DAgger) strategy that allows fine-tuning of DeepONets using aggregated training data that the DeepONets will likely encounter during interactive simulations with other grid components. As a proof of concept, we demonstrate that the proposed frameworks can effectively approximate the transient model of a synchronous generator.

阅读与讨论 → 访问原文 →

25.

arXiv (CS.CL) 2026-06-15 DOI: arXiv:2606.06794

TA-RAG: Tone-Aware Retrieval-Augmented Generation for Peer-Support Health Communication

作者:

Yong-Bin Kang↗Anthony McCosker↗

Retrieval-augmented generation (RAG) successfully grounds large language model (LLM) outputs in trusted documents, but factual grounding alone is insufficient for sensitive peer-support health communication. In domains such as HIV peer support, responses must also be accessible, stigma-free, empathetic, and tailored to the recipient. This paper presents TA-RAG, a lightweight, prompt-based tone-aware RAG framework that embeds explicit tone control into a RAG pipeline without requiring model fine-tuning. We operationalise tone across four core components: stigma-free rewriting, readability adjustment, recipient adaptation, and empathy rephrasing. We evaluate TA-RAG through component-level tests using questions derived from HIV Online Learning Australia (HOLA), UNAIDS terminology guidance, readability metrics, peer-support standards from National Association of People with HIV Australia (NAPWHA), and a public empathy dataset. Results show that the TA-RAG's components improve their targeted communication quality while preserving key content. These findings emphasise that prompt-based tone control is a potential direction for making RAG outputs suitable for sensitive peer-support health communication.

阅读与讨论 → 访问原文 →