Research & Deep Dives
Research papers, reports, scenarios, and technical findings translated into practical implications for builders and operators.
Open Research & Deep DivesRecursive Language Models argue that very long prompts should become an external environment the model can inspect, decompose, and recursively call itself over. Recursion converts context management from passive storage into an active reasoning process.
An agent architecture argues that the append-only event log should be the runtime's source of truth, not an observability layer bolted on afterward. That choice makes recovery, replay, observability, and policy enforcement part of one architecture.
FinePhrase shows that high-quality synthetic pretraining data comes from prompt recipes, mix-in strategy, output diversity, and data infrastructure, not simply larger generator models. The result is a repeatable engineering discipline for controlling quality, coverage, and cost.
A survey argues that agent reliability depends as much on the execution harness as on the model and gives builders a vocabulary for the infrastructure surrounding agents. It organizes execution, context, state, tools, recovery, and evaluation into one reliability model.
AgingBench argues that deployed agents can degrade as their memory state changes, so reliability needs lifespan testing rather than only day-one benchmarks. Its evaluation tests how accumulated state and environmental change affect performance over time.
Language Models Need Sleep argues that long-context systems may require offline consolidation time, not merely larger memory, to reason over context they can no longer attend to. The analogy motivates architectures that compress experience between active reasoning cycles.
A legal history argues that AI authorship is not a new copyright problem but the latest form of mediated authorship. Earlier disputes over photography and mechanical creation reveal recurring legal tests for control and originality.
Polar shows how to train coding agents through the harnesses they already use instead of rebuilding those environments for reinforcement learning. The approach preserves realistic tools and feedback while making training operationally tractable.