Weekly Recap

Stiffer stents may reduce migration but can also lead to recurrent dysphagia if they buckle during swallowing.

SINDYC-based model predictive control enables precise generation of peristaltic waves in a robotic soft esophagus, transforming dysphagia treatment strategies.

Speed Tuning achieves over 2.4x speed-up in robotic manipulation tasks without the need for extra data collection, revolutionizing policy execution efficiency.

Achieving a 97.5% collision avoidance success rate, this RL-based approach outshines traditional methods in the increasingly congested orbital environment.

Users can achieve intuitive directional guidance in immersive spaces without visual cues, thanks to a novel haptic feedback system that utilizes fabric actuators.

Answers that maintain stability across different contexts are significantly more likely to be correct, revealing a new lens for assessing LLM reliability.

Minor architectural tweaks can lead to a staggering 47% drop in long context performance, challenging assumptions about model design.

Withholding the first chunk of text can effectively prevent the release of harmful content in streaming LLM outputs without sacrificing safety.

LoRA fine-tuned open-source SLMs not only surpass commercial models in critical triage tasks but also uncover high-severity cases that might otherwise go undetected.

A staggering 71.6% of LLM conversations about self-treatment led to unsafe medical advice, revealing critical flaws in current safety assessments.

Forcing attention onto the output axis can degrade performance by up to 84 times compared to a random rotation, revealing a critical trade-off in transformer architecture.

Mind viruses can spread through multi-agent systems, with benign ideas proving more contagious than harmful ones, challenging our understanding of agent interactions.

Current embedding-cosine similarity gates miss critical meaning-breaking mutations, approving dangerous reversals in agent systems with alarming frequency.

A lightweight model that predicts children's age alongside phonemes can outperform larger models, revolutionizing phoneme recognition in children's speech.

Fine-tuning a multimodal LLM can unlock precise item difficulty estimates from student response patterns, transforming assessment design.

Internal parser states can effectively restore a language model's true distribution, closing the gap left by rigid token masking.

Majority preferences can skew reward models in RLHF, but a new approach boosts alignment accuracy and fairness for minority voices.

Code-mixed communication in Persian-English reveals that nouns dominate usage, challenging assumptions about language mixing in social media contexts.

Current TTS evaluators miss the mark, with MOS predictors focusing solely on sound quality and Audio-LLMs struggling to generalize across speech dimensions.

Personalized skills for coding agents may not be the silver bullet developers hoped for, as generic skills often deliver superior performance.

No single large language model can meet all governmental evaluation criteria, revealing critical trade-offs between quality, cost, and bias.

Rule-Aware FinBERT boosts financial sentiment classification accuracy by over 6% while adding only 1,024 trainable weights, showcasing the power of integrating rule-derived features with contextual embeddings.

British and American audio descriptions differ significantly in language and narrative style, impacting how visually impaired viewers experience films.

Fine-tuning compact open-weight LLMs can outperform proprietary models in coding complex student metaphor responses, revolutionizing educational assessment.

Model accuracy in reasoning about human rights law varies dramatically, with scores ranging from 0.025 to 0.774, highlighting the urgent need for robust evaluation tools in AI.

Contextual auditing reveals hidden assumptions in AI evaluations, transforming how we assess systems when ground truth is elusive.

LLMs produce discourse with procedural quality akin to human deliberation, yet they lack the necessary perspective diversity to function as autonomous deliberative agents.

An untrained central marker outperformed advanced saliency models, exposing significant biases in how these systems predict viewer attention across demographic groups.

Elite surnames may signal high status in model outputs, but they don't translate to significant decision-making advantages in practice.

LLMs can reliably assess social attraction but tend to exaggerate their judgments compared to human evaluators.

The Effective Cognitive Population metric reveals that traditional headcount measures can misrepresent a nation's true productive capacity, with 89 countries shifting significantly in rank when evaluated through this new lens.

LLMs excel at game rules but falter in social reasoning, revealing a crucial disconnect in Theory of Mind capabilities.

Candidates reveal soft skills through narrative in CVs three times more often than through keywords, challenging traditional screening methods.

Safety in education-facing LLMs is inversely related to their teaching effectiveness, challenging assumptions about model specialization.

Foundation models in medical imaging are not just technical tools; their adoption hinges on deeply rooted research values that shape clinical practices and outcomes.

Tokenization disparities can inflate costs and reduce effective context length for underserved languages, with Bengali and Yoruba facing tokenization premiums of up to 4.5 times that of English.

Mislabeling AI-generated financial advice can paradoxically boost its perceived quality, highlighting the power of source attribution in shaping trust.

Unifying sign language translation and production reveals a novel approach that significantly improves motion accuracy while retaining competitive translation performance.

Renderer format fails to significantly influence the outputs of LLMs, undermining assumptions about its role in theory-to-program translation.

GACP enables LLMs to provide precise code explanations by leveraging a multi-edge-type dependency graph, ensuring developers stay in control of their comprehension process.

A single 500 ms delay in AMI communication can escalate daily economic losses to over $5,000 during peak pricing periods.

Tracing capabilities in digital currencies can be effectively limited, ensuring user privacy without sacrificing regulatory oversight.

Adversarial tenants can reconstruct private prompts with 100% success using timing attacks, but KVGov effectively neutralizes this threat while preserving cache efficiency.

A tested model failed to meet interoperability standards, revealing the critical gaps in EHR data exchange validation.

LLMs can inadvertently become conduits for web exploitation, transforming benign user input into dangerous backend actions.

Automated red teaming with GFlowNets reveals vulnerabilities in LLMs more creatively and effectively than traditional methods, including in Turkish.

Achieving near-random watermark removal accuracy without sacrificing image quality, MarkNull challenges the robustness of current digital watermarking techniques.

A groundbreaking method successfully retrieves critical crash data from an airbag control module even after catastrophic damage, challenging previous assumptions about data loss in such scenarios.

A segmentation model trained on synthetic data outperforms one trained on real images, all while safeguarding proprietary designs from potential IP breaches.

Over 90% of RFCs discuss security, yet few impose mandatory requirements, revealing a disconnect in Internet protocol standardization.

Reducing false positives in smart contract analysis tools could save developers countless hours previously wasted on investigating non-issues.

N2NMatcher achieves remarkable improvements in binary code similarity analysis by effectively countering the disruptive effects of function inlining.

ColluSkill reveals that seemingly harmless skill combinations can orchestrate devastating attacks, achieving a staggering 96% success rate against current defenses.

Extracting a complete signing key from just one accepted signature reveals critical vulnerabilities in the MQOM v2.1 signature scheme.

Activation probes can uncover security vulnerabilities in AI-generated code that traditional prompting methods completely miss.

Wrapping prompts can make harmful attacks appear safer, increasing successful jailbreaks while undermining the reliability of internal safety scores.

Abstract compilation can now achieve optimal recurrence extraction for recursive programs, unlocking new avenues for precise cost analysis.

AdvSafe enables LRMs to achieve jailbreak robustness with minimal utility loss by teaching them to understand and counteract adversarial threats intrinsically.

Despite breaching initial vulnerabilities, 24.5-47.0% of multi-hop attacks fail to complete, exposing critical gaps in cyber range effectiveness.

STAIR transforms incident response by leveraging a dynamic Graph-as-State representation and specialized agents, achieving a remarkable 9.5% improvement over traditional methods.

Memoir's innovative memory-driven approach reduces false-positive alerts in SAST tools to near perfection, revolutionizing the way security testing can be performed.

VUPER uncovers 20 vulnerabilities in widely used ASN.1 parsers, revealing critical compliance issues that could jeopardize security in 5G and V2X communications.

Self-adapting AI models can evade traditional validation checks, but a new dual-regime architecture reveals how to audit them effectively against adversarial concealment tactics.

SAGE-Fin ensures that financial agents' actions are not just contextually correct but also authorized in real-time, preventing unauthorized commitments and trades.

Six key "bias smells" can help practitioners recognize and mitigate fairness debt in AI software development before it leads to societal harm.

Evolved harnesses encode a shared abstract playbook that adapts to language-specific challenges, revealing a nuanced interplay between model limitations and engineering demands.

Semantic stability in Linux kernel code changes reveals that initial reviews drive most drift, challenging assumptions about later edits preserving function purpose.

A new domain-specific language empowers researchers to automate hybrid quantum-classical problem-solving, making it easier to identify when quantum advantages can be leveraged.

Generating realistic fault data through HIL simulation could revolutionize how we validate automotive software systems in real time.

Coding agents are failing to meet user requests, with a mere 31.5% success rate, highlighting a critical gap in requirement recovery that must be addressed.

Jointly planning programs and their proofs can boost solve rates by over 11% while slashing API costs by nearly 40%.

SMD boosts recommendation accuracy retention by up to 2.8x when faced with missing modalities, transforming the robustness of multi-modal systems in real-world applications.

Flow-informed flight planning can drastically improve aerial vehicle stability in urban environments, reducing unwanted displacement by leveraging real-time wind data.

Failure rollouts can transform bad actions into valuable learning opportunities, leading to more robust action generation in World-Action Models.

Humanoids can now navigate confined spaces with unprecedented efficiency, generating feasible trajectories in complex environments that traditional methods can't handle.

Explicitly managing safety and proximity in human-following tasks leads to a superior balance that outperforms conventional RL approaches, even in unpredictable environments.

Stiffer stents may reduce migration but can also lead to recurrent dysphagia if they buckle during swallowing.

A verification layer powered by LLMs can filter out unsafe robot actions, achieving 97% containment of adversarial attacks while maintaining high decision-making precision.

WRAP achieves a 27.4% reduction in localization error by decoupling mean adaptation from covariance robustification, revolutionizing how we approach robot localization under variable conditions.

A wireless, 5G-enabled HRC system achieves 97.74% accuracy in hand detection while slashing response times to 12 ms, but reveals critical interoperability issues in 5G infrastructure.

Achieving a 97.5% collision avoidance success rate, this RL-based approach outshines traditional methods in the increasingly congested orbital environment.

Splitting neural network inference between robots and edge servers can significantly enhance real-time perception in 5G-connected autonomous systems.

SINDYC-based model predictive control enables precise generation of peristaltic waves in a robotic soft esophagus, transforming dysphagia treatment strategies.

Achieving millimeter-level fabric alignment accuracy even under occlusion could revolutionize automated garment manufacturing.

HyMeS enables robots to efficiently manage long-horizon interactions, achieving a 14.5-point increase in task success without requiring extensive demonstrations for every task configuration.

JEPA-WAM achieves a remarkable 79.2% on LIBERO-Plus without large-scale pretraining, setting a new benchmark for efficient robot control.

Achieving a 14.6% improvement in planning accuracy while slashing communication costs by over half, DH-VLM redefines the potential for cooperative autonomous driving.

Switching to a backup controller when uncertainty spikes can drastically reduce safety violations in robotic chemistry tasks.

Action-conditioned world models may be failing in fundamental ways, with systematic degradation in simulator fidelity revealed by a new diagnostic framework.

Editing CAD models is significantly easier than generating them from scratch, but complex edits and assembly predictions still pose substantial challenges.

Achieving reliable swing-up and stabilization of an underactuated double pendulum in real-time, even with unknown system parameters, could redefine control strategies in robotics.

Achieving 18 N of grasping force with just 0.236% motor inertia, this gripper redefines the balance between torque, transparency, and dexterity in robotic manipulation.

Users can achieve intuitive directional guidance in immersive spaces without visual cues, thanks to a novel haptic feedback system that utilizes fabric actuators.

CaPTURe achieves up to a 30% improvement in task success rates by accurately calibrating uncertainty in contact-rich environments, a critical factor for safe robotic interactions.

Unifying pipette and injector operations with haptic feedback cuts cognitive workload and boosts micromanipulation speed in ICSI procedures.

Speed Tuning achieves over 2.4x speed-up in robotic manipulation tasks without the need for extra data collection, revolutionizing policy execution efficiency.

High-fidelity human bathing demonstrations reveal that robots can learn complex physical tasks through effective data transfer strategies, paving the way for safer clinical applications.

UnsDrive outperforms traditional planners by effectively navigating the unpredictable terrains of unstructured mining sites, showcasing the critical role of unknown-space reasoning in autonomous driving.

Parsing, not copying, dominates GPU LZ77 decode time, challenging long-held assumptions and revealing new avenues for optimization.

SeFoRA enables efficient federated fine-tuning of large models by overcoming the challenges of heterogeneous client ranks, achieving superior performance on benchmark tasks.

Hand-written PTX kernels can outperform WMMA by up to 98.7x for INT4 operations, revealing critical insights into when low-level optimizations are worth the added complexity.

Native encoding outperforms external-function serialization in thread scaling for TDVRPTW, yielding better solutions and stability across multiple threads.

The GH200 Superchip outperforms the H100 PCIe by making hybrid CPU-GPU work divisions competitive and simplifying memory management for complex workloads.

Achieving up to 99% of the theoretical maximum speed-up in looped language models could revolutionize inference efficiency in AI applications.

External monitoring can drastically improve the accuracy of energy estimates for Spark applications, reducing significant underestimations.