Research & Deep Dives
Research papers, reports, scenarios, and technical findings translated into practical implications for builders and operators.
Open Research & Deep DivesA NeurIPS position paper argues that LLM social simulations need traceable belief models, not merely fluent personas and plausible survey answers. Explicit belief state, update rules, and validation make simulated populations more scientifically useful.
A systematization paper argues that reusable agent skills are becoming the procedural layer of agent systems, with meaningful performance gains and supply-chain risk. The survey maps benefits alongside provenance, permissions, poisoning, and dependency management.
A preprint argues that agent reliability is governed by the runtime harness around the model: execution, tools, context, state, lifecycle policy, and evaluation. The framework gives teams a way to reason about reliability beyond benchmark scores.
A paper argues that AI automation depends less on isolated task exposure and more on whether AI-capable steps sit next to each other in a workflow. The workflow graph becomes the right unit for estimating exposure, redesign, and value.
A paper argues that enterprise AI should use language models as bounded interfaces while knowledge, rules, and repeatable computation live in explicit systems. The design keeps outputs auditable, deterministic where needed, and easier to govern.
A paper argues that repeated agent workflows should be compiled into reusable MCP blueprints instead of re-reasoned from scratch every time. Compilation reduces repeated inference while preserving validation, portability, and a path back to source intent.
Diffusion language models promise faster generation, but a paper shows why speed does not yet translate into reliable planning, tool use, or agent behavior. The gap comes from weak sequential control, state management, and verification under iterative work.
AlphaFund's whitepaper reframes recursive self-improvement as a measurable capital-allocation loop: create more edge than you lose, then reinvest the gains. The framework turns improvement into an investment decision governed by measurable returns and downside.