MoA vs. Moat: Agentic LLMs for Drug Competitor Mapping Cut Diligence Time 20×
A mechanism-first look at how scaffolded web agents and conservative validation turn messy biotech diligence into faster, more measurable competitor mapping.
A mechanism-first look at how scaffolded web agents and conservative validation turn messy biotech diligence into faster, more measurable competitor mapping.
A mechanism-first reading of Preference Chain, a Graph RAG method that uses small behavioural samples to make LLM mobility agents less generic and more locally plausible.
GLARE shows that legal AI improves when it broadens candidate charges, learns from precedent reasoning paths, and searches only for missing legal premises.
AgentScope 1.0 shows how tool-using agents become more operationally credible when ReAct is wrapped in disciplined abstractions, tracing, evaluation, and runtime isolation.
A spectral reading of SFT and RL fine-tuning shows why RL often restores lost generalization rather than manufacturing new capability from scratch.
Deep-DxSearch shows that diagnostic RAG becomes more useful when retrieval behaviour is trained as a policy, not scripted as a prompt.
A mechanism-first reading of why external rivalry can make LLM agent teams cooperate more internally, and why that lesson is useful but not yet operational proof.
Google’s Gemini serving paper is less interesting as a tiny per-prompt footprint claim than as a practical accounting template for measuring AI inference.
aiXiv shows that the hard part of AI-generated science is not generation, but building the review, revision, security, and governance machinery around it.
BusiAgent shows how multi-agent LLMs can turn broad business requests into governed workflows, but its real value is orchestration discipline, not artificial executive genius.