Nov 17, 2025

• Heretic automates the removal of censorship from transformer-based language models using directional ablation and a TPE-based optimizer, achieving results comparable to manual methods without requiring deep technical knowledge.

• WeatherNext 2 is Google's latest AI model, delivering faster, more accurate, and higher-resolution weather forecasts, capable of generating hundreds of scenarios in under a minute using a single TPU.

• Supercookie leverages favicons to create a persistent tracking ID for users, which remains intact even in incognito mode and is resistant to common privacy measures like cache flushing or VPNs.

• Virtual Width Networks (VWN) enable wider representations without the quadratic cost of increased hidden size, achieving significant optimization speedups—over 2 times for next-token and 3 times for next-2-token predictions in large-scale experiments.

• MarsRL enhances multi-agent reasoning systems by utilizing reinforcement learning with agentic pipeline parallelism, allowing for improved solution refinement through multiple specialized agents.

• OpenGuardrails is the first fully open-source platform designed to ensure the safety, robustness, and privacy compliance of large language models (LLMs) by addressing content-safety violations, model-manipulation attacks, and data leakage risks.

• The Vector Space Mapping (VSM) protocol reveals that quantifying the specialization of attention heads in Transformer models can significantly enhance training efficiency, achieving over 160% faster structural reorganization with optimized hyperparameters.

• The I.S.A.O. framework enables individual researchers to achieve institutional-grade outcomes using consumer hardware, demonstrating its effectiveness through verified results and resilience during a nation-state cyberattack on November 11, 2025.

• The Prelim Attention Score (PAS) effectively detects object hallucinations in large vision-language models (LVLMs) by analyzing attention weights over prelim tokens, revealing a strong correlation between weak image dependence and hallucination occurrences.

• iMAD introduces a token-efficient framework that selectively engages Multi-Agent Debate (MAD) only when it can correct an initially wrong answer, enhancing reasoning without unnecessary computational costs.

• The Multi-agent Undercover Gaming (MUG) protocol enhances the Multi-Agent Debate (MAD) paradigm by employing multimodal counterfactual tests to identify hallucinating agents, thus improving the reliability of large language models (LLMs) in reasoning tasks.