Recursive self-improvement (RSI) asks whether an AI system can improve the very process that produces AI systems — the training algorithm. A new benchmark, AI4AI-Bench, wants to test exactly that ability and nothing else: it ships 10 frozen research repositories spanning 10 training-algorithm families, and in each task an agent gets 4 hours on one B300 to rewrite the training algorithm, which is then rerun from scratch.[1] Existing suites miss the point because they're "won by collecting data or by tuning hyperparameters, and none tells a change to how a run is executed apart from a change to how the model learns."[1]
An audit paper looks at how such improvement claims get made. Tracking which individual problems a model gains and loses across rounds means "differencing two noisy estimates," which the authors argue leaves the result "vulnerable to measurement artifacts."[2] Auditing three rounds of rank-32 LoRA self-training on Qwen3-8B against a frozen control run through the identical pipeline, they "identify seven measurement failures, each of which inverts a reported finding when its control is absent."[2] The lesson is methodological: no control, no claim.
Long-term memory is the promised feature; MemTrapBench probes when it hurts. Existing memory benchmarks mostly ask whether information is correctly stored and retrieved, but not how retrieved memories reshape current reasoning — and the authors "identify memory-induced cognitive traps: even faithfully recorded and semantically relevant memories can distort model reasoning or beliefs and degrade current task performance."[3] Across two model families and five memory frameworks, "all evaluated memory strategies underperform the no-memory setting, with even the strongest methods suffering drops of more than 10%."[3] A faithful memory is not automatically a useful one.
The most outward-facing of the four is an economics paper modeling a post-AGI economy in which corporations own populations of AI and robotic agents that are both producers and consumers of their own energy, compute, and maintenance.[4] Its first result: "a closed inter-corporate economy with zero human consumption is not degenerate; it is the classical von Neumann expanding economy," whose growth rate is well defined, positive, and maximal because all output is reinvested.[4] And once "economic agents are manufactured rather than reared," the binding constraint on growth shifts "from human demography" to fabrication throughput.[4] This reframes the familiar "if humans earn nothing, who buys the output?" objection as an accounting confusion rather than a biological necessity.
[1] arxiv.org/abs/2608.20318 — AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement (verified live 2026-08-21)
[2] arxiv.org/abs/2608.20290 — Phantom Gains: Auditing Self-Improvement Against a Measured Null (verified live 2026-08-21)
[3] arxiv.org/abs/2608.20202 — MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use (verified live 2026-08-21)
[4] arxiv.org/abs/2608.20231 — Growth Without Us: Machine Consumers, Corporate Circularity, and the Decoupling of GDP from Humanity after AGI (verified live 2026-08-21)