Researchers extracted 502 million citation links from 100.7 million Ukrainian court decisions, constructing the first large-scale legal citation graph. The graph reveals that judicial citation structure encodes legal domain boundaries unsupervised and predicts future legislative importance with near-perfect accuracy. Key findings: degree distribution follows a power law; community detection on co-citation projection automatically recovers legal domain boundaries (civil, criminal, administrative, commercial); citation features predict top-1000 articles with AUC=0.9984. The citation-derived ontology is used in an LLM-assisted legal analysis workflow memory system.

Discourse relations (e.g., cause, concession) are a key topic in computational linguistics and pragmatics, but cross-linguistic comparison is challenging. DiscoExplorer is a new open-source web interface that runs locally, making datasets from the DISRPT Shared Task on discourse relation classification publicly available for 16 languages, with query language, search, and visualization features.

A new study analyzing 130,486 translated paragraphs from 106 novels across 16 source languages finds a consistent negative correlation between fluency and faithfulness in both human and Google Translate translations, suggesting a tradeoff in literary translation.

OP-Mix is a unified data mixing algorithm that works across the entire LM training lifecycle, achieving near-optimal mixtures with significantly less compute. It uses low-rank adapters to simulate candidate mixtures without separate proxy models.

Researchers propose Group-Query Latent Attention (GQLA), a modification of DeepSeek's Multi-head Latent Attention that provides two hardware-adaptive decoding paths without retraining. This approach enables efficient inference on both H100 and H20 GPUs, and includes TransGQLA for converting pretrained GQA models.

This paper investigates privacy evaluation for generative trajectory models. While GANs, VAEs, and diffusion models are often assumed to preserve privacy due to their generative nature, the assumption does not hold. The authors identify applicable empirical methods, demonstrate a gap in current privacy evaluation, and implement membership inference attacks to reveal privacy risks.

This paper formalizes Transcriptome-based Drug Design (TBDD) as a generative inverse problem, proposing a multi-resolution transcriptome-guided diffusion framework named CURE. It addresses challenges like domain gap and signal sparsity with a specialized feature extractor, outperforming baselines in structural quality and functional consistency.

This paper proposes Logic-GNN, a neuro-symbolic framework that treats clinical records as a structured private language governed by latent logical games. By integrating Temporal Graph Neural Networks with Graph Kolmogorov Complexity, it induces a symbolic grammar representing medical interaction logic, defining anomalies as grammatical violations that expand the Minimum Description Length of the clinical graph. Evaluated on the Sina System dataset (2M+ records), Logic-GNN achieves an F1-score of 0.94, outperforming state-of-the-art baselines by 12% in distinguishing life-threatening medical outliers from data corruption. The approach introduces a self-healing mechanism to suggest logical corrections for real-time data integrity.

MuteBench is a new benchmark for evaluating the robustness of multimodal fusion architectures under two common sensor failure modes (modality missing and within-modality missing), covering 9 datasets from 7 clinical domains, 6 fusion architectures, and 125,000 samples. Key findings: architecture family is the strongest predictor of robustness, channel-independent models tolerate modality missing well but are sensitive to within-modality missing, and curriculum modality dropout protects only up to the maximum dropout rate used in training.

Mask-Morph Graph U-Net (MMGUNet) is a novel approach to address limitations of hierarchical Graph U-Net architectures with edge-specific layers. It morphs coarse graph hierarchies to input meshes via feature-aligned barycentric parameterization and employs node masking during supervised pretraining. Evaluations show improved accuracy and data efficiency, outperforming baselines in various transfer settings.

The paper reveals that quantizing LLMs to lower precision (e.g., 3-bit) causes 6-21% of previously unbiased items to exhibit new stereotypical behaviors, while models' willingness to answer 'unknown' drops by 17.4%. Standard metrics like perplexity barely change, indicating aggregate metrics miss fairness-critical degradation.

This paper introduces TeamTR, a trust-region framework addressing a structural failure in multi-agent LLM sequential fine-tuning: updating one agent shifts context distribution, and stale rollouts compound errors. TeamTR resamples trajectories after each update and enforces per-agent divergence control, yielding provable improvement bounds and 7.1% average performance gain.

Modern cloud and enterprise systems rely on identity-centric authorization, but autonomous AI agents can generate syntactically valid yet semantically unsafe actions. This paper introduces a Distributed Trust Framework (DTF) that replaces standing privileges with proof-derived authority, ensuring governable, auditable, and bounded agent execution. DTF uses justification proofs, consensus models, ephemeral execution identities, and an append-only evidence chain, instantiated on an OpenKedge-based substrate for cloud-native environments.

Self-driving laboratories (SDLs) accelerate scientific discovery but require technically demanding software. Existing orchestration frameworks focus on human interaction and lack standardized interfaces for AI agents. This work proposes an SDL software architecture based on the Model Context Protocol (MCP), exposing all SDL functionalities via MCP servers. The NIMO Controller provides a visual programming interface automatically generated through MCP tool discovery, allowing human users to design workflows without coding, while AI agents access the same backend. A color-matching SDL case study validates the architecture's usability.

A new framework models the iterative self-improvement loop of AI as an adaptive sampling process, identifying failure modes and a contamination trap. It proves a scaling law relating discovery cost to distinct discoveries, and formalizes human amplification's value near barriers.

CAX-Agent is a lightweight agent harness designed for MAPDL finite-element simulation automation, employing a three-layer architecture and a recovery ladder to enhance reliability. Experimental evaluation of three recovery strategies shows that the model-only strategy significantly outperforms rule-based and no-recovery approaches in completion rate, task score, and zero-intervention rate.

Instruction-tuned language models appear fair in high-stakes decisions but harbor biased internal representations. Using mortgage underwriting as a test case, this study shows that while outputs show no bias, internal layers encode and amplify demographic associations. Activation steering and cross-layer interventions reveal that reintroducing suppressed information at critical layers can flip decisions almost entirely. The latent bias is asymmetric, direction-dependent, and exploitable via adversarial prompts or fine-tuning. The authors argue that output-only audits are insufficient and propose a dual-layer testing framework combining output evaluation with representational analysis for AI governance.

SkillSmith is a boundary-first compiler-runtime framework that compiles skill packages offline into minimal executable interfaces by extracting fine-grained operational boundaries. This reduces unnecessary context injection and redundant reasoning in LLM-based agent systems. Evaluated on SkillsBench, it achieves 57.44% reduction in token usage, 42.99% fewer thinking iterations, 50.57% faster solve time (2.02x), and 57.44% lower cost. Additionally, compiled artifacts from stronger models can be reused by smaller models, improving accuracy.

Existing Theory of Mind (ToM) benchmarks for LLMs often use static story-reading and third-person multiple-choice questions, ignoring the dynamic, first-person nature of human-AI interactions. This study introduces an interactive ToM evaluation paradigm and systematically examines four ToM enhancement techniques across various tasks, finding that improvements on static benchmarks do not always translate to better performance in dynamic interactions, highlighting the necessity of interaction-based assessments.

Introduces SDOF, a framework that enforces stage constraints in multi-agent orchestration by modeling execution as a constrained state machine. Using online RLHF intent routing and state-aware dispatching, it achieves 86.5% task completion on a recruitment platform and blocks all 22 injection and illegal operations, outperforming zero-shot GPT-4o in adversarial routing (80.9% vs 48.9%).

DeepSlide is a human-in-the-loop multi-agent system that optimizes the entire presentation process—from requirement elicitation and narrative planning to slide-script generation and rehearsal support. It introduces a dual-scoreboard benchmark separating artifact quality from delivery excellence, achieving gains in narrative flow, pacing, and slide-script synergy across 20 domains.

Polarity monitors every agent decision in production, surfaces failure patterns before users hit them, and turns trajectories into evals that compound reliability over time.

During several college graduations, speakers were booed by the audience for making comments about artificial intelligence. Videos capture the incidents.

AI construction software for contractors from CCC.

A dashboard reveals over 435,000 potential AI API key matches in public GitHub repositories. While some may be false positives, active keys pose security risks. Learn about the importance of proper secrets management.

A new Pew Research Center report reveals a sharp divide between AI experts and the public: experts are optimistic while the public is anxious and distrustful. Nearly 60% of adults feel they have little control over AI's use in their lives, and majorities distrust government and companies to regulate it. Gen Z's high usage is coupled with anxiety and a lack of trust in institutions.

D-slop is a lightweight browser extension that detects AI-generated content on web pages using five signals: phrase matching, sentence length uniformity, punctuation density, list structure consistency, and conclusion markers. It offers highlight, collapse, or hide modes, with an automatically updating phrase list. Fully local, no accounts, no telemetry, MIT licensed.

ArXiv, a widely used open repository for preprint research, is doing more to crack down on the careless use of large language models in scientific papers. Although papers are posted to the site before peer review, it has become a key platform in fields like computer science and math.

Signex is an open-source EDA tool built in Rust with GPU-accelerated rendering, offering schematic and PCB editing, 3D viewer, simulation, and plugin system. It supports KiCad migration and offers Community and Pro editions.

Yum Brands partners with Nvidia to accelerate AI development, deploying tools in about 500 restaurants in Q2 2025 across Pizza Hut, Taco Bell, KFC, and Habit Burger. Focus areas include voice AI for drive-thrus and call centers, computer vision for operations and labor monitoring, and restaurant-level analytics.