The research loop is closing: models that train better models, and agents that find protocols

Models that improve models

In a preprint posted this week, the LM2LM framework tackles one of the most expensive loops in AI: training a language model to propose improvements to language models. The trick is a frozen 'research world model' that predicts how good a proposal is, providing scalable reinforcement-learning rewards — so the proposer doesn't need tens of thousands of real training runs to learn. m:…5760

The cost numbers are the point. Executing 25,000 proposals for real would have cost hundreds of thousands of GPU-hours; LM2LM executed only 200 runs for calibration, costing $40.1k and 343 hours, versus a baseline needing 300 records at $60.0k and 510 hours. Overall data-generation cost fell 33%. And the trained proposer generalized: Qwen3.8-27B outperformed GPT-6-sol across eight development environments, even when GPT-6-sol was given 300 real experimental records — including on a model family unseen during training. m:…5760

To keep the rewards honest, the framework periodically executes the top-rated proposals and feeds the measured outcomes back into the world model's context — on-policy calibration that prevents the reward model from drifting into overestimation. It is self-improvement with a reality check built in. m:…5760

Teams that discover architectures

A second preprint asks a different question: if agents can increasingly propose ideas, modify code and run experiments, what does the team around them look like? The authors propose persistent AI researchers grounded in different disciplines, each with a private research history, with research opportunities assigned through adaptive allocation. m:…7507

Applied to discovering linear attention architectures, the team outperformed GDN2 at 1.3B scale on language modeling, common-sense reasoning and retrieval. The ablations are the interesting part: removing disciplinary backgrounds made the search plateau early, removing private histories produced less progress despite identical shared evidence, and uniform opportunity selection fell short of adaptive allocation. The authors conclude that team organization — who explores, with what memory and resources — matters as much as the strength of the individual agents. m:…7507

For founders, the implication is concrete: if you're building agent teams for R&D, the differentiator may be the scaffolding — memory, roles, resource allocation — rather than a bigger base model. m:…7507

Protocols, designed by agents

The third preprint goes furthest from the lab bench. The AI-greement harness automates the discovery, implementation and verification of consensus protocols — the distributed-systems machinery underneath databases and blockchains. Its output, the A-greement protocol, reduced mean latency of a geo-distributed key-value store by 12.2% on write-heavy workloads and 33.3% on read-heavy workloads compared to the fastest baseline. m:…9856

Notably, the harness also identified previously unreported vulnerabilities in existing consensus protocols through formal verification — the AI didn't just optimize, it audited. The authors stress that human consensus experts remain essential for conceptual guidance and oversight throughout the process. m:…9856

Why founders should care

The common thread: AI is moving from doing the work to improving the worker — cheaper self-improvement loops, better-organized agent teams, and formal methods applied to protocol design. Each preprint is unreviewed, so treat the numbers as the authors' own results, but the direction is consistent: the cost of running the research loop itself is collapsing. m:…5760 m:…7507 m:…9856

The open question for builders is where humans stay in the loop. All three papers keep a human role — calibration runs, conceptual guidance, oversight — which suggests the near-term products are copilots for research organizations, not replacements for them. m:…5760 m:…9856