Safety & Evaluation A

Finite-dimensional diffusion policies suffer temporal drift from discretization that degrades long-horizon performance. The paper introduces a backward Kolmogorov equation that lifts diffusion policies into a Cameron-Martin space to make them more robust.

The classical continuous-time description of temporal-difference learning with linear features is an ODE capturing asymptotic mean dynamics but neglecting stochasticity. This work provides a diffusion approximation for TD learning under Markovian noise to capture those fluctuations.

For the forward solution of nonlinear PDEs, the method uses Bellman-Kalaba quasilinearization to reduce the nonlinear problem to a sequence of linear subproblems, each discretized by collocation and solved with physics-informed neural networks.

The paper evaluates open-source LLMs on multi-label classification of cyber threat intelligence (CTI) reports using MITRE ATT&CK techniques. Summary is title-based and neutral; details and figures are as presented by the source and not independently verified.

LLMs produce legal text of at least median quality, yet no benchmark evaluates doctrinal legal reasoning, the interpretive core of legal work. The paper benchmarks doctrinal reasoning under the EU AI Act and discusses the measurement gap in legal automation.

The paper proposes WEQA, a framework for question answering over wearable health sensor data using query-adaptive agentic reasoning, arguing that diverse sensor modalities and user intents cannot be handled by a fixed reasoning workflow.

AI agents are shifting from advisors to actors that book travel and run procurement. Existing animal-welfare benchmarks grade only text answers, so this work introduces an agentic benchmark testing whether implicit animal-welfare reasoning transfers to agent actions in frontier models.

The paper formalizes 'cognitive atrophy,' a process-level behavioural measure of AI-mediated mental-health support, capturing whether interactions help users keep reflecting, coping, and deciding, a dimension distinct from safety and static response quality.

Conversational AI localizes responses using user metadata, yet the regional biases this hidden context introduces remain poorly understood. The paper analyzes the unintended effects of geographic conditioning on large language model outputs.

Embedded machine learning moves inference from the cloud to resource-constrained devices. This practice-oriented synthesis lays out data, feature, evaluation and deployment pipelines for an embedded ML workflow on microcontroller-class platforms.

LLM apps build prompts from templates, with Handlebars the default in Microsoft Semantic Kernel. While double-brace expressions HTML-escape values, triple-brace interpolation inserts them raw. The paper studies structural role injection and the limits of HTML auto-escaping.

RAG grounds model outputs in external evidence, but its dominant evaluations and defaults are oriented toward factual question answering. HistoRAG embeds historical methodology into retrieval-augmented generation through critical technical practice for interpretive historical studies.

The paper presents IsabeLLM, applying AI-based automated theorem proving to formally verify blockchain consensus, aiming to automate much of the expertise-intensive verification workload and make formal verification more accessible.

An NVIDIA technical post explains techniques for optimizing transformer-based models during low-precision training. The export raw_excerpt was blocked (cookie/query data), so this summary is based only on the title and source; specific methods and figures are unverified.

Structured state space models such as S4 and S4D capture long-range dependencies but are hard to deploy on constrained devices. S4oP introduces operator-level pruning to enable efficient deployment of SSMs on time- and resource-constrained hardware.

The paper presents EAGG, an embodiment-aligned grasp generator that represents each end-effector with a topology-aware graph and embodiment-specific conditioning, aiming to generalize grasp generation across objects and diverse robot embodiments.

Google DeepMind presents an AI Control Roadmap for securing the future of AI agents, combining traditional safeguards with real-time monitoring to protect internal systems. The framework lays out layered defenses against agent misuse and unsafe behavior as agents proliferate.

Gradient descent in deep learning can operate at the edge of stability, where the loss Hessian's top eigenvalue hovers near the stability threshold. Classical tools fail there, so Edge Flow offers a tractable, predictive continuous-time model of this regime.

The paper proposes a multi-agent framework for healthcare that mitigates premature diagnostic handoff and silent clinical hallucinations, replacing LLM-as-a-judge routing with deterministic orchestration constraints and adding two safety mechanisms.

NoiseTilt (NTRK) is a reward-guided diffusion sampler that injects reward gradients through the noise term, leaving the score kernel unchanged and needing only a single sample per step, improving reward alignment of pretrained diffusion models.

As LLM-based agents enter autonomous scientific research, resisting pseudoscience matters. PseudoBench is an adversarial benchmark measuring how such agents may rapidly generate plausible yet misleading studies that contaminate academic literature.

LLM agents rely on reusable tool specifications (skills), but real tasks require composing multiple skills. The paper formalizes compositional skill routing: decomposing a complex query into atomic sub-tasks, retrieving relevant skills, and composing them.

Vision-language-action models combine vision-language backbones with expressive generative action heads trained via flow matching on large robotic datasets. Despite strong performance, the paper studies uncertainty quantification for these flow-based VLA models.

NVIDIA announced that its Blackwell GPU architecture topped the MLPerf Training 6.0 benchmark with what it calls industry-leading scale and performance. Summarized neutrally from the title; the export excerpt was blocked (cookie/query data), so figures are vendor claims, not independently verified.

Tool-using LLM agents use the Model Context Protocol to answer from heterogeneous sources like search, APIs, databases and clinical records. ProvenanceGuard provides source-aware factuality verification to catch provenance-sensitive failure modes that standard metrics miss.

Looped transformers scale latent computation by repeating shared blocks, but sequential looping raises latency and KV-cache memory with loop count. Building on parallel loop transformers, LoopCoder-v2 makes loop count a practical knob for efficient test-time computation scaling.

Legal-AI systems hallucinate at aggregate rates near 52%, but averages hide where and how errors concentrate. LegalHalluLens is an auditing framework pairing typed hallucination auditing with calibrated multi-agent debate to give compliance officers actionable signals for trustworthy legal AI.

Conditional independence testing is fundamental to statistics and causal inference. The paper proposes a fast nonparametric conditional independence test based on two-stage regression, aiming to improve computational efficiency and power.

The paper introduces LLM Consumer Behavior Theory, a proposed field analyzing consumer behavior in agentic markets where LLMs make consumption decisions on behalf of users, drawing on classical and behavioral economics alongside NLP.

In masked diffusion language models, padding and semantic termination roles get entangled. VoidPadding introduces a [VOID] token to handle padding so that [EOS] can focus on signaling semantic termination, improving generation behavior.