Weekly Recap

An artificial intelligence‐powered One Health framework designed to transition from surveillance to predictive prevention and advances the concept of “Precision Environmental Health”, supporting anticipatory governance and promoting equitable interventions at a global scale.

Lateral movement in networks can be curtailed by binding user processes to fixed access scopes that persist across hosts, fundamentally changing how we secure enterprise environments.

Achieving nearly 99% precision in anomaly detection, HSMLog redefines how we approach security log analysis in hardware systems.

Action-aware reinforcement learning can boost code localization performance, achieving up to 8x better results than larger models.

AI agents not only accelerated the verification of liblzma's memory safety but also uncovered critical vulnerabilities in its zero-input handling.

An artificial intelligence‐powered One Health framework designed to transition from surveillance to predictive prevention and advances the concept of “Precision Environmental Health”, supporting anticipatory governance and promoting equitable interventions at a global scale.

Lateral movement in networks can be curtailed by binding user processes to fixed access scopes that persist across hosts, fundamentally changing how we secure enterprise environments.

Achieving nearly 99% precision in anomaly detection, HSMLog redefines how we approach security log analysis in hardware systems.

Action-aware reinforcement learning can boost code localization performance, achieving up to 8x better results than larger models.

AI agents not only accelerated the verification of liblzma's memory safety but also uncovered critical vulnerabilities in its zero-input handling.

Lower-cost models can achieve up to 94% of the localization quality of premium explorers while slashing agent time and token usage by up to 95%.

EMERGE-Policy reveals that a robot's intelligence can thrive through a collaborative network of specialized agents rather than a monolithic policy.

Temporal validity in patient memory design can significantly enhance clinical reasoning accuracy, challenging the notion that more historical data is always better.

Strong and weak edges in choice-derived item graphs encode qualitatively different relationships, fundamentally reshaping our understanding of collaborative filtering dynamics.

Function vectors can enhance emotion detection in multiple languages without the computational burden of traditional few-shot learning.

Query language alone can determine whether local suppliers are included in product recommendations, regardless of the user's location.

The Socioscope pipeline has generated a massive, analyzable dataset that redefines how qualitative research can scale in complex domains like the food system.

Unlocking insights from customer letters could revolutionize how life insurers tackle cancellation issues and enhance retention strategies.

AMINA not only streamlines nonprofit operations but also actively counters misinformation, transforming how marginalized immigrant organizations navigate politically sensitive environments.

LLMs struggle to mimic human belief updating in persuasion, showing only slight agreement and distinct preferences in persuasive cues.

Most AI fairness evaluations overlook critical uncertainty factors, risking misinterpretation of bias impacts in hiring systems.

Model decisions can reverse by over 50% when evaluated across different versions, challenging the reliability of post-hoc assessments in AI governance.

Agents can now prove compliance with governance policies without ever sharing sensitive data, fundamentally transforming trust in multi-agent AI systems.

Context closure could revolutionize WebPKI by ensuring that mutable authorization contexts are securely authenticated without the heavy data costs of traditional methods.

ActReal reveals that system-level mobile agents can successfully evade automation detection by generating synchronized touch and IMU signals, achieving over 77% attack success rates.

SkillGuard cuts attack success rates to as low as 4.8% by redefining agent capabilities in response to contamination from untrusted data.

A new algebraic algorithm for Search-LWE achieves polynomial complexity improvements, challenging previous assumptions in cryptographic problem-solving.

Benign actions can be misclassified as attacks due to causal signal shifts, leading to significant utility loss in tool-using LLMs.

Achieving 95.7% precision in parameter extraction, this open-weights model challenges the reliance on commercial systems for reproducible agentic workflows in materials science.

Native code bugs in Python applications can go undetected but are critical to software quality, with this study revealing their hidden patterns and effective resolution strategies.

Bluezz's Reactive Peripheral Modeling uncovers vulnerabilities in BLE firmware that traditional rehosting methods miss, achieving over 2.6 times the basic-block coverage of its predecessors.

Optimizing attacker sequences can dramatically enhance the effectiveness of hard-label black-box text attacks, outperforming traditional methods by leveraging a strategic combination of attack strategies.

PolyFlow uncovers hidden cross-language vulnerabilities that traditional static analyzers miss, thanks to its innovative use of LLMs in a neuro-symbolic framework.

JITterFlip reveals that targeting the JIT serving control plane can amplify attack effects by over 7 million times, exposing a major vulnerability in LLM inference systems.

Current MLLMs can only identify and explain environmental risks in secure authentication scenarios with limited effectiveness, revealing a significant vulnerability in their design.

LOOM achieves robust kernel monitoring with negligible overhead by leveraging TrustZone's hardware isolation, effectively countering sophisticated rootkit threats.

Users preferred a context-based authentication method that was more memorable and less burdensome than traditional PINs, even when it took longer to enter.

ARMOR boosts aerial object detection robustness by up to 32% against adversarial attacks, even in low-data environments.

One-third of Chain-of-Thoughts in VLA models for autonomous vehicles are unreliable, highlighting a critical gap in safety validation.

SKILLFORGE achieves rapid formal verification of Dafny programs by intelligently orchestrating a library of specialized skills, outperforming existing methods in both speed and accuracy.

A single linear context-free language exists that is not a parsing expression grammar, revealing fundamental limitations in PEGs' closure properties.

LLMs struggle with interactive programming tasks, revealing a significant gap in their algorithmic reasoning abilities under constrained information conditions.

By framing geo-localization as a human-like reasoning challenge, GeoPAVE significantly reduces perceptual errors and enhances decision accuracy in ambiguous visual contexts.

PhysWave achieves spatially consistent audio generation by combining physics-based priors with a unified control framework, revolutionizing text-to-FOA applications.

A single natural language correction can restore near-perfect performance in VLA models that would otherwise fail under environmental shifts.

Event-triggered control can maintain performance levels of time-triggered systems while effectively managing computational delays in networked environments.

Achieving a 97.5% success rate in zero-shot robotic ultrasound scanning shows that task-relevant feature dynamics can outpace traditional simulation realism.

A sliding window filter can drastically enhance the consistency of inertial navigation estimates by effectively managing unknown measurement delays.

Achieving under 0.64 N and 0.042 Nm RMS errors in handle estimation could revolutionize the accuracy of haptic feedback for everyday objects.

Agents can now effectively collaborate in multi-robot systems even when facing unpredictable computational delays, leading to improved learning and control performance.

Agents can now learn from their mistakes in real-time, improving safety and performance in autonomous driving by recognizing when to seek expert intervention.

MiBOT achieves significant cardiovascular benefits comparable to human massage, all while being a silent, head-worn robotic solution.

A new planner achieves zero invalid trajectories in complex bimanual tasks, revealing a staggering 62% error rate in traditional methods.

Humanoid robots can now traverse complex 3D structures with remarkable agility and precision, completing intricate maneuvers in real-world trials.

LARC achieves a staggering 10.28x speedup in trajectory certification without sacrificing accuracy, revolutionizing how we approach robot motion planning.

SmoothRL achieves real-time policy optimization in robotics, effectively bridging the gap between asynchronous execution and online reinforcement learning.

Transforming open-loop action into a reliable closed-loop system, AGM achieves superior task performance while keeping VLA policies frozen and lightweight.

Achieving a 3.36-fold speedup in action generation, DriftingVLA eliminates the need for iterative refinement while maintaining high performance in robot control tasks.

Explicitly teaching models about molecular scaffolds can reshape how embeddings are organized, leading to improved property predictions in self-supervised learning.

Charged mixtures reveal unexpected ion-specific thermodynamic effects that could redefine our understanding of electrolyte behavior.

VeloBins achieves unprecedented accuracy in aerial inertial odometry by transforming velocity estimation into a classification task, eliminating the pitfalls of traditional regression methods.

Tactile representations can significantly enhance robotic manipulation by focusing on physical contact rather than device-specific tactile appearances.

Single-module control learning enables scalable shape control in soft robots, eliminating the need for configuration-specific data collection.

Achieving both generalizability and scalability in multi-agent planning, this diffusion approach reduces collision risks while adapting to diverse task specifications.

Robots can achieve impressive manipulation skills using only proprioception, revealing the untapped potential of joint encoder signals in dexterous tasks.

SNF-ICON cuts average waiting times in HPC job scheduling while optimizing energy use, proving that smarter scheduling can significantly enhance performance.

SemSpot enables LLM agents to dynamically optimize cloud inference costs and performance by leveraging short-lived offers tailored to task semantics.

Achieving up to 42% power reduction in high-speed adiabatic logic circuits could redefine energy efficiency benchmarks in advanced chip design.

Site-specific vibrational probes can reveal collective protein motions, challenging the assumption that local measurements only capture immediate dynamics.

Cesium ions cluster in FLiBe, disrupting the fluoroberyllate network and preventing crystallization at room temperature, a key insight for fission reactor design.

Text-conditioned generative music models outperform their audio-conditioned counterparts in faithfully following emotional cues, revealing critical insights into genre-specific emotion-following dynamics.

ColSNAP allows for flexible, multi-level compression in visual document retrieval without sacrificing retrieval performance, adapting seamlessly across different architectures.

REIGN achieves state-of-the-art long-document retrieval efficiency while slashing training costs by four orders of magnitude compared to traditional methods.

TEMPO not only achieves superior timestamping accuracy in audio-language models but also integrates innovative techniques like atomic timestamp tokens and reinforcement learning for enhanced performance.

Current audio-language models struggle with temporal music grounding, but targeted training can lead to significant performance improvements.

PersonaGen-1M reveals that 78.3% of buyer queries are informational, providing a crucial lens into how generative engines prioritize brand recommendations.

Transforming retrieval failures into hard pairs can boost model performance while using 26% to 75% fewer synthetic samples than conventional methods.

Current agents struggle with indirect references in conversations, achieving only a 67% success rate in resolving them, revealing significant gaps in their contextual understanding.

Adaptive Doubly Robust evaluation reduces variance and improves accuracy in ranking policy assessments, even as the complexity of user behavior increases.

Image corruptions can drastically impair retrieval performance, revealing that clean image-only queries often outperform joint retrieval in snap-and-ask scenarios.

Integrating query intent throughout the generative retrieval pipeline led to a remarkable 21.7% boost in recall and a 3.52% increase in click-through rates in real-world applications.

Compiling scientific knowledge through LLMs and graph structures reveals a stable, source-traceable portrait of research evolution across multiple domains.

Current GUI world models often produce visually appealing screens but lack the contextual consistency needed for effective multi-step agent interactions.

Social media users exploit multimodal strategies in misinformation, with AI-generated content dominating tech narratives and news images lending credibility to vaccine falsehoods.

A hybrid model that combines generative and discriminative approaches achieves a remarkable 70.7% detection accuracy while significantly improving calibration in hallucination detection for LVLMs.

Role-linked visual biases in text-to-image models persist and even intensify when placed in unrelated contexts, challenging assumptions about context-free evaluations.

Achieving minute-long real-time interactive world generation while maintaining geometric consistency and dynamic subject identity is now a reality with Matrix-Game 3.5.

IDA-OPD not only boosts diversity in distilled models but does so while cutting computational costs, challenging the notion that more data always leads to better performance.

Smaller language models can efficiently replace larger ones in rubric-based reinforcement learning, achieving competitive performance with significantly reduced computational costs.

Firm-level characteristics can yield reliable risk bounds without the notoriously difficult-to-estimate cross-asset return covariances, reshaping portfolio management strategies.

The study highlights the importance of treatment planning decisions in reducing unnecessary radiation exposures to the surrounding normal tissues and supports the need for individualized, safety-conscious radiotherapy strategies.

This work reveals YTHDC1 as a pivotal coordinator of the glucose‐glutamine‐cystine metabolic network and provides ATF4 mRNA nanomedicine‐based combination strategies to improve TNBC therapeutic efficacy.

With longer follow-up, most participants receiving talquetamab continued to experience improvement in their multiple myeloma, and although side effects were common, very few participants stopped talquetamab treatment because of side effects.

To test the potential of so-called quantum physics-informed neural network technology for solving evolutionary partial differential equations, the problem of multiple wave propagation in a weakly nonlinear medium described by the Korteweg-de Vries equation with periodic boundary conditions is considered.