The Archive

Mesh Memory Protocol enables cross-session cognitive collaboration and state sharing for multi-agent LLM teams.

Agentic AI with anomaly detection for proactive fall risk management and elderly activity monitoring.

Google DeepMind announces partnerships with global consultancies to commercialize frontier AI for enterprise adoption.

Cyber Defense Benchmark evaluates LLM-agent threat hunting on 106 real Windows attack scenarios with MITRE ATT&CK coverage.

Benchmark: Haiku 4.5 with agent skill (84.3%) outperforms baseline Opus 4.7 (80.5%) across 880 evals at 5x lower cost.

Google Ads Advisor adds three agentic safety features for policy compliance and account optimization.

Claude Code UX improvement: double-paste expands truncated pasted content, fixing long-standing issue.

OpenAI teases Image V2 model ahead of public release.

EVPO addresses critic noise in sparse-reward LLM RL by casting baseline selection as Kalman filtering, challenging PPO/GRPO design tradeoffs.

GPT-Image-2 implements self-review iteration mechanism; ~11 min generation time for multi-pass refinement.

Dual-Glob applies supervised contrastive learning to pitch accent classification in Seoul Korean speech; narrow linguistic domain.

Replica-based fairness audit of deployed Early Warning System at Centennial College reveals gender/age/residency disparities across ML pipeline.

Neural operator framework discovers stability and receptivity in complex systems from data alone, without governing equations.

LePREC dataset (769 Malaysian Contract Act cases) reveals LLM limitations in legal issue identification despite diverse generation capability.

Reddit speculation that Claude 4.7 performance degradation stems from cost optimization and OpenClaw re-enablement; unverified claim.

Less than a year ago, Apple made headlines for a lack of AI announcements at its annual WWDC event. Ten months later, the company has announced that hardware executive John Ternus will succeed longtime CEO Tim Cook as chief executive - and the official release doesn't mention…

GPT-5 and DeepSeek-R1 exploit formalization-faithfulness gap in Lean 4 proofs despite valid logical reasoning; evaluates on FOLIO and Multi-LogiEval.

Four-axis alignment framework (factual, reasoning, compliance, regulatory) for evaluating long-horizon enterprise AI agents in loan/claims/clinical domains.

ZC-Swish activation stabilizes BN-free deep networks in micro-batch and federated regimes by centering activations to prevent mean-shift.

Personalized federated learning for industrial failure prediction accommodates heterogeneous degradation across clients while preserving data privacy.

Dual-track LLM-assisted stylometry (embedding + register) detects biblical allusions in Cormac McCarthy via KJV vocabulary and semantic matching.

Unsupervised method calibrates confidence estimates in reasoning LLMs using self-consistency proxy targets and distillation, enabling reliable deployment with single generation.

Large-scale study across 15 LLMs and 8 tasks reveals zero-shot ability poorly predicts optimization performance in LLM-guided evolutionary search, identifying hidden trajectory factors.

CTLF branching-time logic framework formalizes bias detection in generative AI with counting worlds semantics and probability distribution verification.

AI music startup GRAI says fans want to remix tracks, not generate songs from scratch.

CAST model captures fine-grained item semantics for sequential recommendation by distinguishing true complementary relations from spurious popularity correlations.

Visual Contrastive Editing (VCE) mitigates object hallucination in large vision-language models by counteracting language priors without inference cost.

GOLD-BEV framework learns dense bird's-eye-view semantic maps for autonomous driving using time-synchronized aerial imagery as weak supervision.

HP-Edit applies RLHF to diffusion-based image editing via human-preference post-training framework, addressing scalability gaps in preference dataset collection.

MM-JudgeBench—first multilingual multimodal evaluation benchmark with 60K+ preference instances across 25 languages—reveals LVLM judge generalization failures.