Information Gain: Two Gates for Research Agents
How a two-gate scorer splits information gain into standalone signal and marginal gain over an archive, so research agents stop resubmitting known rules.

How a two-gate scorer splits information gain into standalone signal and marginal gain over an archive, so research agents stop resubmitting known rules.

RLHF alternatives through Ashby's law of requisite variety: why a rich environment lets a simple update rule replace a reward model, and when RLVR wins.

How agent harnesses add a controller turn that picks explore, plan or execute each step, parses the choice with fallbacks and keeps the budget in view.

Test-time adaptation for LLM agents runs at two timescales: scratchpads and episodic memory in context, or LoRA weight updates. When each is the right call.

Why t-tests fail when an agent proposes hundreds of rules, and how frozen permutation nulls and cluster bootstrap lower bounds keep benchmark validity.

What mode collapse is in LLMs, how to detect it with pass@k, and why a sharp drop after peak-only data selection behaved differently in our run.

Why Schmidhuber's Gödel machine was never built, and how agent self-improvement swaps its proof gate for empirical gates: benchmarks, archives, survival.

Procedural generation gives an agent endless tasks; Schmidhuber's PowerPlay and OOPS let it invent its own. What each gate costs, and how to mix the two.

Meta-learning, or learning to learn, from Schmidhuber's 1987 self-referential learners to the inner and outer loops of today's self-training LLM agents.

Why strategy diversity falls as self-training agents improve: selection prunes failing strategies into a few basins, and how to tell pruning from collapse.