A new arXiv preprint (2607.20486v1) released on July 24 describes OPTScientist, a system where four AI agents collaborate to discover new optimizer algorithms for transformer pretraining S¹. The system uncovered an optimizer named RS-MR, which the researchers claim outperforms strong baselines. What makes this different from every other "AI discovers something" paper is the mechanism: when the agents hit a wall, they rewrite their own programming language to break through it.

My read: This is the first optimizer-discovery system I've seen that treats the search space itself as mutable. Most automated discovery work, including OR-Agent's approach to algorithm discovery P⁴ and the evolutionary multi-task optimization framework from Gozeten et al. P², searches within a fixed space. OPTScientist's second stage, where it proposes DSL extensions to escape representational bottlenecks, is the genuinely novel idea. I don't buy the RS-MR superiority claim yet, because the paper offers no quantified improvements and uses what the authors call their "native evaluation protocol" S¹, which is a custom benchmark. Until I see numbers on standard benchmarks, this is a promising framework with an unverified result.

How four agents divide the labour

The OPTScientist framework operates using four distinct agents within one orchestration loop: Theorist, Designer, Engineer, and Reviewer S¹. Each has a distinct job. The Theorist leverages mathematical concepts to suggest new optimizer architectures. The Designer converts these concepts into code using a typed domain-specific language (DSL), which is a limited programming language designed solely for defining optimizer update rules. The Engineer executes training trials for the proposed optimizers. The Reviewer assesses the outcomes and provides feedback into the system.

This division mirrors how a research team operates, but with one key difference: the loop runs automatically, and the agents can iterate far faster than humans clicking through training runs.

The trick that makes the search space grow

The first stage of OPTScientist is evolutionary search. The agents generate candidate optimizer programs, test them, and mutate the best performers, standard genetic-algorithm territory. The same approach appears in related work like AlphaOPT, which creates a library of self-improving experiences to develop optimization programs P⁵, and OR-Agent, which integrates evolutionary search techniques with organized hypothesis tracking P⁴.

The second stage is where OPTScientist diverges. If the agents consistently fail to surpass current optimizers, the system views this as an indication that the DSL is too restrictive, creating a representational bottleneck. It then suggests minor additions to the language, introducing new operators or functions that enable the agents to articulate optimizer concepts that were previously inexpressible S¹.

Think of it like a mechanic who, after failing to fix an engine with the tools on hand, invents a new wrench. Most systems just keep trying different wrenches from the same box.

What RS-MR actually is

The newly discovered optimizer is dubbed RS-MR, standing for reduced-state matrix optimizer S¹. The study characterizes it as a matrix-based optimizer, similar to Lion and Sophia, which utilize matrices to monitor gradient data during training instead of depending only on per-parameter metrics like Adam.

The authors report that RS-MR improves transformer pretraining over strong baselines S¹. But the abstract provides no specific percentage improvements, no named baselines, and no reference to standard benchmarks. The improvement is qualified as occurring "under our native evaluation protocol" S¹, meaning the authors ran their own custom training comparisons rather than using the established evaluation setups the broader community uses to compare optimizers.

Where this sits in the field

Automated algorithm discovery is an active research area. The SciAgentsDiscovery GitHub repository, which has 605 stars, uses multi-agent setups for broader scientific discovery tasks P³. Research by Gozeten et al. from the University of Michigan and UCSD investigates LLM-driven program discovery that handles several tasks at once through an evolutionary multi-task optimization approach P². OR-Agent integrates evolutionary search techniques with organized hypothesis tracking P⁴. AlphaOPT creates a library of self-improving experiences to develop optimization programs P⁵.

OPTScientist's contribution is narrowing this approach to optimizer discovery specifically, and adding the DSL-extension mechanism. The pattern across these systems is the same: multiple specialised agents, each handling a slice of the problem, orchestrated in a loop.

What to do about it

For a machine learning team training transformers, the practical question is whether RS-MR or any OPTScientist-discovered optimizer is worth swapping in for AdamW or Lion. Right now, the answer is no, not from this paper alone. The absence of standard benchmark numbers means there is no way to compare RS-MR against established optimizers on equal footing.

A team running pretraining experiments could, however, use the OPTScientist framework itself as a research tool. If you have the compute budget to run many short training experiments, the four-agent loop could search your own optimizer space, including domain-specific constraints your team cares about that standard optimizers ignore. A language model pretraining team, for instance, might let the agents explore optimizers tuned for their specific architecture and data mixture rather than relying on general-purpose defaults.

The concrete thing to check this week: read the paper's description of the typed DSL and the DSL-extension mechanism. If your team has an internal optimizer search pipeline, the idea of expanding the search space on failure is directly portable, even without adopting the full multi-agent framework.

What we don't know yet

The paper is an unreviewed arXiv preprint. All empirical claims are self-reported by the authors with no independent verification S¹. The abstract contains no quantitative results or named baselines, and makes no reference to standard benchmarks. The qualifier "under our native evaluation protocol" means the comparisons may not translate to the evaluation settings other researchers use.

We don't know the specific training configurations tested, the model sizes used, or whether RS-MR generalises beyond the authors' experimental setup. We don't know whether the DSL extensions the system proposed are genuinely novel mathematical operators or recombinations of existing ideas. We don't know how the four agents' outputs compare to a single-agent baseline doing the same evolutionary search.

The next signal: the full paper, which should contain experimental details and benchmark numbers beyond the abstract. We'll check RS-MR against standard optimizer benchmarks when the authors release complete results or when independent researchers attempt replication. If you want that follow-up in your inbox, subscribe and we'll send it the day it drops.

Sources: S1 — OPTScientist: Multi-Agent Discovery of Typed Optimizer Programs for Tr · P2 — Evolutionary Multi-Task Optimization for LLM-Guided Program Discovery · P3 — lamm-mit/SciAgentsDiscovery · P4 — OR-Agent: Bridging Evolutionary Search and Structured Research for Aut · P5 — Paper page - AlphaOPT: Formulating Optimization Programs with Self-Imp

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.