Tag
The article proposes a Multi-Key Attention idea where a single attention head uses multiple Key projections combined non-compensatorily before softmax to better handle multi-constraint retrieval tasks, and seeks feedback on potential flaws and existing benchmarks.
Anthropic introduces TASTE, a benchmark for evaluating AI models' ability to judge AI safety research proposals, finding that models like Fable 5 perform worse than human researchers.
A collection of Claude Code skills for academic research workflows, including literature reviews, PhD proposals, and fidelity-first academic slides with vector equations and citations.
Proposes treating semantic compression as a diffusion noise function for handling massive context beyond model windows, using multi-pass reading at decreasing compression levels. Untrained-model experiments show components work in isolation but the full chain needs training to resolve binding bottleneck.
The author proposes a novel experimental framework to study identity formation in LLMs as hypergraph evolution through multi-instance interaction, distinguishing it from standard multi-agent debate by focusing on structural divergence in activation space rather than task performance.