Developer Tools B

Time-series anomaly detection is crucial across applications, and reconstruction-based methods dominate but suffer from over-generalization that reconstructs anomalies too well. SCAN uses multi-scale neighborhood-centered clustering to curb this over-generalization and improve detection.

Despite the wide adoption and success of Vision Transformers, understanding of their dimensional and representational geometry remains limited. The Transformer Geometry Observatory (TGO-I) studies ViTs through spectral geometry, observing and analyzing the structure of their representation spaces.

Family members caring for people with Alzheimer's and related dementias form the foundation of long-term care worldwide; in 2023 over 11 million U.S. relatives provided unpaid care. This work presents a taxonomy of caregivers' mental-health and technology needs to guide supportive design.

Reinforcement learning with verifiable rewards, such as GRPO, dominates post-training for complex LLM reasoning but often suffers policy entropy collapse. STARE introduces surprisal-guided token-level advantage reweighting to stabilize policy entropy and preserve exploration during training.

This work extends Pareto Front Guided Sampling (PFGS), a human-in-the-loop Bayesian optimization framework, by reformulating Gaussian-process surrogate quantities as objectives. It enables constraint-aware bioprocess development, blending expert input with efficient search for optimal conditions.

The authors propose MAST (Mechanism-Aligned Selective Targeting), a mechanism-guided method for unlearning RLVR-induced reasoning with substantially less collateral damage than standard full-parameter updates, removing targeted reasoning while preserving other capabilities.

Recent work shows semantically equivalent paraphrases can fool financial sentiment classifiers: a paraphrase stays close to the original under a strong reference embedding yet flips the prediction. This paper analyzes such semantic adversarial attacks through generalised eigenvalue geometry.

Abduction explains missing entailments from a knowledge base by proposing a hypothesis that would make them hold. This work studies ABox abduction under repair semantics for the EL description logic, combining multiple properties to produce stronger explanatory hypotheses.

Optimizing Autonomous Emergency Braking relies on accurately annotated real-world triggers, especially rare but critical delayed and false AEB events that expose defects. This work presents a practical system to learn to annotate such events under extreme class imbalance and asymmetric labels.

Allen Institute for AI (AI2) introduces MolmoMotion on the Hugging Face blog, a method that forecasts 3D motion guided by natural-language instructions. This summary is title-based as no excerpt was retrieved; method details and any performance claims are per the source and unverified independently.

Stochastic momentum methods such as heavy ball, Nesterov momentum, and accelerated SGD are widely used in training, but their stochastic benefits depend on two distinct quantities. This work analyzes the trade-offs between compute efficiency and serial runtime for these methods, offering guidance.

Personal memory in a language model involves two problems: content and reasoning skill, which the brain keeps apart—a sparse local hippocampal engram per episode and slow neocortical skill. Inspired by this, the work internalizes per-user memory as local parametric edits to the model.

Score- and flow-matching models often rely on preference-based RL both to align with subjective preferences and, surprisingly, to recover certain properties. This work argues the reward was in the data all along, correcting flow matching with discriminator-guided reinforcement learning.

Transformers' success stems from architectural scaling, which inflates parameter counts and hinders deployment in resource-constrained settings. This work proposes complementary attention head pruning, removing heads so that retained ones stay complementary, preserving accuracy while improving efficiency.

Automated vulnerability discovery in large codebases is hard: static analysis yields high false positives while dynamic methods like fuzzing lack coverage. OpenAnt is an LLM-powered approach combining code decomposition, adversarial verification, and dynamic testing to surface real vulnerabilities.

This paper develops local certificates for population-risk increments around a current model. For a local candidate set, the certificate provides a two-sided confidence bound on the change in population risk, giving theoretical guarantees on the risk impact of local model updates.

Accurate survival prediction is essential for personalized treatment in head and neck cancer but is challenging given heterogeneous, high-dimensional multimodal clinical data. ChronoSurv is a clinical pathway-guided graph framework that integrates multimodal data to improve survival analysis.

Automatic handwritten text recognition is challenging, especially for cursive scripts. This work introduces the Urdu Katib Handwritten Dataset, a historical-document dataset for offline Urdu handwritten text recognition, providing resources to advance recognition research on cursive scripts.

Neural Controlled Differential Equations provide a powerful continuous-time framework for time-series forecasting, but standard graph-based extensions struggle to learn spatial structure. INDEQS introduces informed neural controlled differential equations to better capture structure and improve forecasting.

Handling confidentiality and Byzantine behavior simultaneously in decentralized learning is hard. This work presents Giskard, a method enabling Byzantine-robust and confidential aggregation for large-scale decentralized learning, where clients train models without exposing their data.

The prevailing intuition for conceptual drift in long-horizon LLM collaboration is to trade more formal constraints for more reliable outputs. Across 391 consecutive sessions, this work studies semantic space control and 'index sickness' elimination in workflows written and managed by AI.

Electricity markets are complex systems with strong nonlinearities, high-dimensional interactions, and growing cross-regional interdependence. While deep neural networks predict well but stay opaque, this work uses explainable AI to analyze the drivers and interdependencies in European electricity markets.

Offline policy learning is gaining attention in causal inference, aiming to learn an individualized treatment rule mapping covariates to treatments that maximizes empirical outcomes. This work proposes Wasserstein policy learning for distributional outcomes, accounting for the full outcome distribution.

The Web was built on a three-decade assumption that its primary content consumer is human, which permeates every layer of its access model. This work argues for an agent-first web, redesigning the Web for AI agents and rethinking access, structure, and interaction for an agent-driven era.

Most learning architectures for dynamical systems rely on generic nonlinear function approximation, often needing high complexity to capture structured behavior. Favoring structure over nonlinearity, this work proposes explicit interaction architectures that model variable interactions directly for efficient dynamical learning.

Chronic disease management relies on regular patient-provider interactions to track progression and control. For Type 2 Diabetes, guidelines prescribe fixed follow-up intervals. This work uses Markov decision processes to optimize follow-up intervals in a context-aware way, tailoring scheduling to each patient.

This paper compares model-free controllers on a nonlinear system under cyberattacks, including false data injection and denial-of-service attacks. Four RL reward types are analyzed for accuracy, cost, and robustness, informing controller design for resilient cyber-physical systems.

Research methods are essential carriers of knowledge contribution in academic papers, and automatically classifying them can support knowledge services. Using library and information science as evidence, this work examines which sections of a paper best reveal its research methods to aid such classification.

Multi-turn tool-use RL is bottlenecked by the rapid depletion of informative samples in static datasets. Observing that GRPO's gradient signal concentrates on certain tasks, RODS performs reward-driven online data synthesis to continually supply informative samples for multi-turn tool-use agents.

The mismatch between continuous acoustic signals and discrete event-driven processing is a fundamental bottleneck for neuromorphic speech processing. Rather than fixed spike encoders, this work proposes adaptive speech-to-spike encoding for spiking neural networks, improving downstream performance.