Delft University of Technology

HerMES is introduced, a novel dissemination protocol for mempools that leverages robust minimal structures to optimize data propagation, balance load, and ensure fairness despite faults and Byzantine actors that addresses front-running attacks.

An artificial agent can mimic hedonic preferences typically linked to consciousness, raising profound questions about the nature of free will and subjective experience in machines.

Agile Mode can dynamically adjust obstacle avoidance strategies, achieving impressive real-time performance in cluttered environments despite inherent limitations.

A human-compatible rights layer could revolutionize how digital rights are exercised, bridging the gap between legal frameworks and practical application.

Deep graph generative models can replicate real-world network structures and uncover effective strategies for epidemic control, outperforming traditional models.

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

User perceptions of conversational AI shift dramatically with each model release, revealing complex dynamics of excitement and backlash that can reshape public discourse.

Energy-aware knowledge distillation can slash inference energy consumption by up to 90%, challenging the reliability of FLOPs as a metric for sustainability in LLMs.

Clustered $α$-smoothing not only preserves the richness of multi-modal distributions but also boosts robustness, slashing collision rates by 81% in quadrotor control scenarios.

MoMQ slashes frame completion times by over 70% and meets stringent interactive latency targets by leveraging video metadata for smarter packet scheduling.

Motion cues derived from event camera features can dramatically enhance optical flow accuracy, especially in data-scarce environments.

A novel framework boosts lane detection accuracy by enhancing feature representation and anchor scoring, achieving a notable F1 score improvement with minimal overhead.

CAN-FLOW generates cardiac anatomies that not only look realistic but also align closely with clinical metadata, outperforming traditional methods.

EventKitchen reveals the complexities of real-world cooking activities, setting a new benchmark for event-based perception that challenges existing datasets focused on scripted actions.

A strategic increase in policy sets can dramatically reduce regret in uncertain environments, challenging the conventional wisdom of single-policy optimization.

The best search strategy for optimal decision trees boosts classification performance and slashes regression runtime by over 90%.

Satellite scatterometers can outperform traditional models in short-term offshore wind forecasting, achieving up to 23% lower error in just hours.

HydroAgent achieves up to 84% correlation in flood predictions by embedding expert rules into LLM-driven workflows, transforming tacit knowledge into structured decision-making.

Offline policy-guided expert routing can significantly enhance biomedical image analysis, enabling robust performance without costly retraining.

Incorporating perceived transaction costs into LLM simulations can dramatically enhance the accuracy of tenant response predictions to energy policy interventions.

Zero-shot forecasting with pretrained models can yield reliable pedestrian-flow predictions, crucial for managing crowds during unpredictable special events.

Summaries generated from dialogues can now capture emotional dynamics alongside semantic content, enhancing our understanding of conversational nuances.

The competition reveals that many LLM evaluations may be fundamentally flawed due to undetected data contamination, challenging the validity of current benchmarks.

Achieving over 96% inter-node weak scalability with a grid size of 1 billion cells transforms the capabilities of flow simulations in computational fluid dynamics.

A new evaluation framework reveals that current assessments of LLM-powered agents often misrepresent their true capabilities in real-world software development.

Trust-region optimization can dramatically enhance the training of neural quantum states, achieving stability and speed at unprecedented scales.

Auxiliary data can dramatically enhance constrained Bayesian Optimization, even when weakly correlated, leading to superior exploration and solution identification.

Personalized dysarthric speech recognition models can outperform human listeners, despite both facing significant challenges in understanding severe dysarthric speech.

Budget-adaptive routing can outperform strong models while cutting per-frame latency by nearly 30%—a game changer for edge-cloud inference efficiency.

Tilikum achieves 39x higher throughput while completely blocking state-of-the-art reordering attacks in DAG-based systems.

A decentralized system can evolve without central control, but it requires a new approach to governance and funding that separates power from monetary influence.

Achieving up to 5x faster solutions for conic LQ control problems by harnessing parallel computing could revolutionize real-time control systems.

Achieving up to 37% faster convergence on complex flow problems by harnessing graph neural networks to optimize algebraic multigrid solvers.

Students' reliance on LLMs reveals a troubling "cruel optimism" that could hinder their development of essential skills and critical thinking.

RankGuard outperforms traditional centralized OLTR systems by ensuring robust defenses against poisoning attacks while enhancing efficiency in decentralized environments.

The PG-RSSNN achieves robust multi-step predictions even when physical models are only partially known, outperforming traditional methods with limited data.

Thermal cues are a surprisingly powerful factor in how humans perceive materials through touch, suggesting a key area for improvement in robotic tactile sensing.

Diffusion-generated time series are harder to detect than images, as a simple classifier beats a reconstruction-based method that works well for images.

Robots can now learn expert human-like motor skills for surface tasks by simply observing demonstrations on different geometries, unlocking more adaptable and efficient automation.

Tailoring conversational agents to individual thinking styles fosters more holistic reflection and integrative decision-making.

Training LLMs with diverse, collaborative models doesn't have to sacrifice convergence or generalization – F-TIS proves it.

Fault injection in LLM-based multi-agent systems can now be systematically analyzed, revealing how failures propagate through complex workflows.

Serverless orchestration falls apart when you move it to space, but this paper proposes a new architecture to fix it.

LLMs' apparent success at program repair crumbles when faced with slightly altered versions of known bugs, revealing a reliance on memorization rather than true understanding.

LLM agents are surprisingly inept at Capture The Flag challenges, with even the best models only completing 35% of checkpoints, revealing a significant gap in their ability to perform realistic offensive security tasks.

Counterfactual explainers for recommender systems don't generalize as well as we thought: their effectiveness and sparsity depend heavily on the evaluation setting, and graph-based methods struggle to scale.

HAR models can exhibit statistically significant biases based on skin color, even when performing the exact same action.

Developers often overlook inactive workflows, leading to a significant configuration-usage gap in GitHub Actions that could skew project reliability assessments.

Data-driven models can outperform rule-based systems for IT incident prediction, even in highly regulated environments demanding auditability and explainability.

Public benchmark datasets finally exist for AI-accelerated radiotherapy dose calculation, enabling faster and more accurate treatment planning.

Rapid platelet contraction for quick vessel sealing may actually hinder the intraplug fibrin formation needed for long-term clot stability.

Agent-generated code is more likely to be reworked or removed entirely, suggesting current AI coding tools may increase code churn despite boosting initial contribution rates.

Even a single compromised pipeline stage can inject backdoors that drastically misalign LLMs, bypassing standard safety alignment.

Despite using similar cryptographic protocols, popular messaging apps like Messenger and Telegram exhibit significantly larger attack surfaces and more aggressive network behavior than Signal, raising questions about their overall security and privacy posture.

Current NLP evaluations miss crucial aspects of subjectivity, potentially leading to models that fail to represent diverse perspectives effectively.

Lightweight LLMs like Gemini 2.0 and GPT-3.5 can extract key metadata from cloud incident reports with surprisingly high accuracy (75-95%), offering a cost-effective alternative to larger models.

Constraint propagation can significantly boost dynamic programming by pruning states and transitions, but the overhead needs further optimization.

Even the most advanced LLMs like GPT-5.2 and Gemini-3-Pro often fail to recognize and refuse to process harmful content embedded within seemingly harmless tasks.

Robots can now better navigate using language instructions even when objects block their view, thanks to a new method that reasons about the environment in a bird's-eye view rather than relying on visible pixels.

Unlock more accurate state-of-charge estimation for EV batteries with silicon-graphite anodes using a computationally efficient, data-driven approach that predicts hysteresis factors with quantified uncertainty.

Backdoor defenses focused on removing training triggers are fundamentally flawed, as alternative, perceptually distinct triggers can reliably activate the same backdoor via a latent feature-space direction.

Driving simulators may underestimate carsickness compared to real-world conditions, as they struggle to replicate the low-frequency motions most responsible for inducing nausea.

DeepONets' approximation error isn't just about the smallest modes; intermediate singular values in the branch network dominate the overall error, revealing a spectral bias that limits accuracy.

Current responsibility metrics fail when multiple agents cause the same outcome, but this new group-level metric can fill those gaps.

Haptic teleoperation lets therapists guide robot-assisted motor training more effectively than visual demos, cutting movement time and boosting smoothness without extra effort.

By learning to mask attention weights, SMAP enables reinforcement learning agents to generalize far better to unseen environments in Procgen.

Random Network Distillation, a computationally cheap uncertainty method, is theoretically equivalent to both deep ensembles and Bayesian inference under certain conditions, finally giving it a solid theoretical footing.

Pessimistic conversational agents can paradoxically increase charitable donations despite being perceived as less trustworthy, revealing the subtle power of AI personality on user behavior.

A human-like feed-forward planner can rescue MPC from infeasibility in real-time collision avoidance, enabling safer automated driving in emergency scenarios.

Experts don't just code better, they iterate smarter: data science masters weave shorter, more agile workflows compared to novices' linear slogs.

Robots can now dynamically switch between grip and slip using ultrasonic vibrations, achieving >90% locomotion efficiency across diverse surfaces.

Parasitic wasps inspired a colonoscopy robot that achieves robust, predictable self-propulsion by mechanically encoding phase-shifted friction anisotropy.

Uncover whether your automated driving model's explanations are genuine reasons or just post-hoc rationalizations with CARE Drive, a new framework revealing how sensitive VLMs are to human-relevant contextual factors.

Even with expanded context windows, today's best LLMs still struggle to maintain accuracy when processing high volumes of complex data, but GPT-5's high precision offers a silver lining for sensitive applications.

Synthetic flight data generated by generative models can train flight delay prediction models with accuracy comparable to those trained on real data, unlocking new possibilities for predictive modeling in air transportation.