A curated archive of frontier intelligence, operator-grade guides, and strategic analysis.
New research from arXiv reveals that In-Context Learning (ICL) task identity is encoded as distributed output format templates, not single-position activations, fundamentally reshaping understanding of LLM behavior.
Read the briefing
New research reveals homogeneous LLM multi-agent debate often fails due to sycophancy, fragility, and consensus collapse, with self-correction offering better...
SCOPE-FE, a new framework from arXiv, addresses the computational bottleneck of automatic feature engineering in high-dimensional datasets by intelligently reducing...
A new roundtrip verification and repair framework enhances LLM autoformalization fidelity from 45-61% to 83-85% without ground-truth annotations.
Frontier coding agents can now autonomously implement an AlphaZero-style machine learning pipeline for Connect Four, performing comparably to external solvers.
Verbal Process Supervision (VPS) uses structured natural-language critique to improve LLM reasoning performance, achieving 94.9% on GPQA Diamond without training.
MaSH Loops framework evaluates generative AI as pluralist sociotechnical systems, examining how models, users, and institutions recursively co-construct meaning and...
LLHKG framework enables lightweight language models to construct knowledge graphs with performance comparable to GPT-3.5, automating entity and relation extraction.
A complete guide to Hippo, the open-source memory framework for AI agents. Learn how its neuroscience-inspired architecture solves context loss...
