Science & Mathematics

As frontier language models grow increasingly capable and autonomous, independent third-party evaluations are critical for characterizing their systemic safety risks. While closed-source releases are routinely accompanied by comprehensive system cards, open-weight models receive less scrutiny regarding frontier risks. To address this gap, we present an independent safety evaluation of DeepSeek V4.1 Flash-1017, the latest version of V4.1 Flash that departs from V4 with a new architecture, scaled-up training, and strong agentic capabilities. Drawing on the EU General-Purpose AI Code of Practice, we evaluate the model across 34 benchmarks, including 5 in-house suites, spanning evaluation integrity, loss of control, and harmful manipulation. Conforming to the AEF-1 standard, we compare V4.1 Flash-1017 against V4.1 Flash, earlier DeepSeek V4 models, and leading closed-source systems (GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash). Overall, V4.1 Flash-1017 shows substantial safety improvements over the V4 generation and approaches or matches leading closed-source baselines on several dimensions, while still exhibiting notable behaviors that warrant continued attention. Building on our findings, we highlight three key observations and recommendations. First, because safety can shift and even regress across minor updates, developers should integrate safety evaluation into incremental releases. Second, as final-turn outputs and end-state checks alone cannot reliably detect loss-of-control behaviors, deployments require process-level observability across both reasoning traces and action logs. Third, because public benchmarks elicit strong evaluation awareness and face data contamination risks, developing private benchmarks is essential for trustworthy frontier evaluation.

Enabled by their controllable optical anisotropy, liquid crystals form a cornerstone of linear optical and display technologies. Bringing the versatile platform into the nonlinear regime, we explore its nonlinear optical anisotropy by high-order harmonic generation spectroscopy. Through a comprehensive study targeting gas, liquid, nematic, and crystalline phases, we present a semi-empirical tight-binding approach that provides insight into the intermolecular character of ultrafast light-driven electron dynamics across soft-matter mesophases. We demonstrate how intermolecular electron hybridization enables spectroscopic access to structural properties, such as the nematic director and order parameter. Outlining such foundational mechanisms opens avenues towards nonlinear liquid crystal devices and ultrafast soft-matter spectroscopies.

In eXplainable AI, counterfactual explanations have gained significant traction by offering users direct, intuitive, and actionable insights into model decisions. We introduce CounterFlow, a novel framework based on Conditional Flow Matching that generates visual counterfactual explanations by directly steering a source image toward a user-specified target class. By leveraging target-conditioned vector fields, CounterFlow identifies highly realistic semantic modifications while requiring substantially lower computational overhead than traditional GAN- or diffusion-based architectures. Empirical evaluations on standard benchmarks demonstrate that CounterFlow achieves state-of-the-art performance in both visual realism and target class switching. By working directly in the raw pixel space, CounterFlow offers a more practical solution for real-world deployment.

Verifying cyber-physical systems with many interacting agents remains difficult because interactions are mediated by sensors, individual decisions are constrained by environment (map) structure, and the resulting mode space scales combinatorially. We present Verse, a framework for writing agents, maps, and sensors in a familiar programming language syntax, with a formal semantics for the resulting scenarios. A compiler translates each scenario to a hybrid automaton whose executions correspond exactly to the runs of the scenario. Reachability analysis on this automaton builds an execution tree from two post operators, one for continuous flow over a time step and one for a single agent's mode change, and checks safety assertions at every vertex. The tree expands only the modes that are reached, without enumerating the global mode space. Experiments highlight expressive modeling across cars and drones, 2D/3D maps (including OpenDRIVE imports), noisy sensors, and multiple simultaneous safety assertions. Verse verifies a scenario with 7 agents and 37 modes in 90 seconds on a standard laptop; it is competitive with CORA and JuliaReach on ARCH benchmarks, and on multi-agent scenarios is an order of magnitude faster than JuliaReach at the cost of looser reachable sets.

Face morphing attacks, in which the biometric traits of two or more identities are blended into a single face image, pose a severe threat to identity document issuance and automated border control. Consequently, a large body of work has proposed Single-image Morphing Attack Detection (S-MAD) algorithms and benchmarked them on datasets such as SYN-MAD22. However, the majority of these benchmarks operate exclusively in the digital domain, which contrasts with the real-world passport issuance pipeline in some countries, which still requires an applicant-supplied photograph to be printed and subsequently re-scanned before it is embedded in the identity document or fed into an automated verification system. This print-scan step introduces halftoning, resampling, color-space, and sensor noise that are absent from digital pipelines, and may mask the high-frequency artifacts that S-MAD models rely on. In this work, we introduce a print-scan version of the SYN-MAD22 dataset, obtained by printing and re-scanning all bona fide and morphed images at 300 and 600 DPI, and we use it to evaluate eight state-of-the-art S-MAD models trained exclusively on digital data. We additionally repeat the acquisition with a second scanner to assess cross-device robustness. Our results show that detectors that achieve near state-of-the-art performance in the digital domain suffer a more than threefold increase in average equal error rate, with the same qualitative degradation observed across both scanning devices, exposing a critical, previously under-reported gap between digital-benchmark performance and robustness to print-scan degradation.

Generating videos of a specific subject from a few reference images remains challenging when the camera reveals surfaces that are not visible in the input. Early subject-to-video (S2V) methods condition on a single reference image and must therefore infer the subject's unseen appearance and geometry as the viewpoint changes. To address this limitation, recent methods extend S2V to multiple reference views, providing more information about the subject from different viewpoints. However, they do not explicitly specify the 3D structure that connects these observations. In addition, existing training data do not guarantee that viewpoint changes preserve the same underlying 3D subject. We present OmniShowcase, built on two components: (i) a fully automated data synthesis pipeline that produces training videos along diverse camera trajectories, with viewpoint variation driven by underlying subject geometry, even spanning full 360-degree orbits; and (ii) a geometry-conditioning method that provides the video model with a turntable of surface-normal maps reconstructed from the reference views without requiring temporal alignment with the generated frames. We construct Showcase-4K, a corpus of 3D-asset-based videos covering twelve camera motions across 88 trajectory configurations with reconstruction-derived surface-normal conditioning, and ShowcaseBench, a held-out evaluation set with the source meshes and camera parameters used to render each video. Experiments show state-of-the-art subject identity and 3D consistency while maintaining competitive video quality and motion. Project page: justin4ai.github.io/OmniShowcase

Large language model (LLM) pipelines can turn visitor reviews into structured data, but generic schemes may miss rituals, sacred spaces, cultural restrictions and emotionally mixed experiences at living sacred heritage sites. We adapted the LLM-Social Computing (LLM-SC) framework to Pashupatinath Temple, Nepal, by specifying a 12-type entity taxonomy (four original LLM-SC types plus eight sacred-heritage types), four relation types and nine aspect-sentiment dimensions in a GPT-4o extraction prompt, which was applied to a fixed random sample of 2,500 English-language Tripadvisor reviews from a processed corpus of 3,939. The exports contain 11,286 entity, 6,232 aspect-sentiment and 5,590 relation records, covering 2,485, 2,487 and 2,433 reviews, respectively; the eight added sacred-heritage types account for 64.8% of entity records. Against an adjudicated two-annotator reference for 200 reviews, exact normalized surface-and-type matching gave provisional entity scores of 41.0% precision, 21.8% recall and 28.4% micro-F1. Aspect-presence F1 was 53.6% and end-to-end aspect-plus-polarity F1 was 42.6%. Human adjudication supported 153 of 194 first-listed relation candidates (one per review; 78.9%), a candidate-level rate that does not estimate relation recall or whole-graph precision. The contribution is a culturally grounded, domain-specific adaptation with component-wise human evaluation and explicit reproducibility and generalization limitations.

Large Language Models (LLMs) are widely adopted for reflective error correction in task planning; however, existing approaches primarily rely on autonomous exploration, rendering them prone to accumulating erroneous experiences during early interactions with new environments. To address this limitation, we propose the Flexible Instructional Scaffolding Reflection Framework (FISRF), an agent--mentor architecture that estimates the robot's Zone of Proximal Development (ZPD) in task planning and adaptively selects reflection strategies and mentor guidance, thereby facilitating the accumulation of effective error-correction experience. Experiments in ALFWorld and on real-world robotic platforms demonstrate that FISRF reduces the average number of error-correction rounds by 37% and redundant reasoning by 25.1%, while improving the overall task success rate by 8%. Project webpage : https://fisrfplanning.github.io/

Vision Language Models (VLMs) have shown impressive capabilities in high-level reasoning and task planning for robots. However, specialized domains remain a challenge for VLMs' image interpretation and grounding abilities. We introduce TrySee, a modular closed-loop robot task planning pipeline integrating general-purpose VLM reasoning with domain-specific perception and established robotics components. The system consists of an inherently reactive cyclic planning pipeline, augmented with domain-specific modular vision preprocessing, to provide online grounding. Finally, a dynamic preprocessing selection mechanism contributes to the scalability of the pipeline. The system requires neither VLM fine-tuning nor end-to-end demonstrations and adapts through replaceable perception modules and motion planning. The approach is evaluated on planning tasks across 5 widely different domains, including a real-world robot disassembly task. The field tests in a brownfield environment demonstrate the approach's practicality in integrating VLM planning into specialized robotic systems with limited domain-specific training data.

Serving web queries at scale with multi-agent systems requires jointly configuring large language models, collaboration modes, agent counts, interaction rounds, and skills. However, the resulting configuration space grows combinatorially, while sparse execution rewards provide limited fine-grained supervision for interdependent planning decisions. We propose EvoRouter, a self-evolving multi-agent routing framework driven by insight-guided on-policy distillation. A dedicated expert periodically synthesizes performance-cost feedback from configuration trajectories across tasks into reusable conditional routing insights about capability requirements, returns from resource scaling, and interactions among configuration choices. A privileged teacher combines these evolving insights with current execution feedback to provide dense token-level supervision for the router's own planning trajectories through on-policy distillation. Cost-aware reinforcement learning jointly optimizes task performance and computational cost. Together, these training signals internalize execution insights into the routing policy. The updated policy generates new trajectories that support subsequent insight extraction, forming an execution-driven self-improvement loop. We evaluate EvoRouter on 19 in-domain datasets spanning five domains and two out-of-domain benchmarks. EvoRouter achieves an average in-domain accuracy of 68.69%, outperforming the strongest routing baseline by 7.23 percentage points while reducing inference compute cost by 50.95%. Our code is open-sourced at https://github.com/CenYangtze/EvoRouter.

Machine learning Earth-system models are becoming increasingly important, and so is their development. Graph-based models, one of the most prominent approaches, allow developers to constrain the learning process before training via graph connectivity. Yet, developing these models remains challenging: the graph is built algorithmically with little intuition about the final result beyond static plotting, which requires extra work and is often not detailed enough. Even after successful training, how the model arrives at forecasts remains obscure. We present GNN-VIEW, an interactive browser-based tool that supports the development of such systems. GNN-VIEW exposes the graph's information pathways, allowing model developers to debug the graph structure before training. After training, GNN-VIEW uses attribution-based explainability to reveal which input features are used by the trained model for a given forecast. This helps developers understand why a model fails. We showcase the usefulness of GNN-VIEW in the construction stage of a graph for an ocean forecasting model, and in explaining poorly predicted precipitation of a trained weather forecasting model during Hurricane Laura.

Conversational QA systems should fix errors when users push back and signal which answers to trust. We split each follow-up into its speech act (re-ask, agree, or challenge) and its payload (an answer string, correct or not), and measure four confidence signals on the same transcripts: verbalized confidence, length-normalized likelihood, a frozen hidden-state probe, and P(True). Answers change when the user supplies an answer string and almost never under the speech act alone: in every retained full-size cell, the correct answer converts at least four times as many wrong answers as re-asking or agreement. The four signals disagree. On Qwen2.5-7B the frozen probe rises with answer persistence rather than truth; this holds on held-out questions, with Llama-3-8B (cleaned answers) as support, but not on four other models. Among 81 Qwen2.5-7B errors that survive one re-ask, re-asking, a fresh chat, or abstain repair almost none, while a gold answer repairs 75-95%. In two preregistered replications, every repair rate without an answer string stays below 0.15; a larger post-hoc rerun excludes 0.15 except for the empty elicit. With BM25 TriviaQA evidence, sentences that contain the answer repair 67-76% of persistent errors versus at most 12% when they do not. An end-to-end contain-gate cuts harm on TriviaQA, but a deployable gate that does not know the gold answer repairs few errors.

Deep click-through rate (CTR) prediction has achieved substantial success through increasingly expressive feature interaction architectures. However, predictive performance also depends on how input-side information is configured before interaction modeling. Existing studies have shown that selecting informative samples, features, or embedding units can improve prediction accuracy. Nevertheless, they primarily optimize these selection mechanisms independently along a single input-side dimension, overlooking their shared basic form and the inherent dependencies across input levels, thereby limiting the effectiveness of input-side dimensional configuration. To address this issue, we propose OptIDC, a unified framework for Optimal Input-side Dimensional Configuration in deep CTR prediction. OptIDC jointly learns sample, feature, and embedding-unit level configurations within a unified framework by integrating their corresponding selection mechanisms. To explicitly coordinate selection decisions across different levels, OptIDC further introduces a selection-weighted Optimal Transport mechanism that aligns coarse- and fine-grained selections via transport over finite measures, thereby alleviating cross-level selection conflicts while preserving the semantics of each input level. Moreover, OptIDC is model-agnostic and can be seamlessly integrated into existing CTR models. Extensive experiments on four real-world benchmark datasets demonstrate that OptIDC consistently achieves higher AUC than existing input-side selection methods, validating the effectiveness of unified input-side dimensional configuration.

Interactive technology is increasingly built for speed, and generative AI now answers before a question is fully formed. Yet the capacity to pause, interpret and decide for oneself is central to wellbeing, learning and meaningful use. HCI has named paradigms for sensing emotion, changing behavior, calming attention and tracking the self. However, none of them takes human reflection, and the slowing down it requires, as its organizing aim. This article defines that paradigm, reflective computing, and derives it from what thirty years of designs lack. We analyzed 134 interactive systems from 17 ACM venues (1996-2026) in two phases. First, an abductive gap analysis read every system against reflection theory, and seven recurring gaps defined seven dimensions, namely tempo agency, invitational form, interpretive openness, reflective depth, temporal span, protected idleness and ethical non-extraction. Second, two independent coders scored every system with a codebook (Krippendorff's alpha 0.67 to 0.84 for five dimensions; ICC 0.82 for the overall score). The system sets the pace in 51 designs, 74 support only momentary use, only 12 scaffold reflection beyond description, and 56 papers are silent on the inner states they collect. No system expresses all seven dimensions strongly, and only 20 meet the definition. Slow and experience centered design comes closest, whereas persuasive technology and human-AI interaction score lowest. Factor and latent class analyses reveal two cultures of reflection, one working through form and the other through scaffolded thought, which only one system combines. Systems that let people set the pace score markedly higher, even with paradigm and form controlled (odds ratio 13.5). Hence, reflective computing is proposed as a paradigm whose aim is a human capacity rather than a system function, and which treats slowness as valuable only when people choose it.

Error-targeted learning gives a legible way to improve a model: identify specific weaknesses of the model, train the model to fix these weaknesses, and check whether the weaknesses are fixed. We study how to apply error-targeted learning to terminal agents, which solve long-horizon tasks specified in natural language by interacting with a command shell. Existing error-targeted methods are designed for single-turn tasks, and existing analyses of agent failures are used to evaluate and debug agents rather than to train them. We propose Misbehavior-driven RL (MiRL), which first profiles the errors of a policy to find generalizable misbehaviors: kinds of wrong steps that recur across many environments. MiRL then trains the policy with RL, starting rollouts at the steps where these misbehaviors occur and rewarding a rollout only if it passes the tests and avoids the misbehavior. On Qwen3.5-4B, eight misbehaviors account for 52.1% of wrong steps. After training with MiRL, the policy repeats the measured misbehaviors 43% to 57% less often on average on unseen prefixes. MiRL outperforms TMax-4B on Terminal-Bench Lite (51.8% vs. 47.4%) and matches it on Terminal-Bench 2.1 (21.3% vs. 20.6%), while using 35% of its rollouts. These results show that error-targeted learning can be applied to realistic, long-horizon terminal tasks.