# Paper Argues Evolution Strategies Beat RL at Keeping LLM Answer Sets Diverse > Source: > Published: 2026-08-15 05:41:29+00:00 ## What happened Seven researchers posted a preprint on arXiv arguing that reinforcement learning post-training collapses the diversity of a language model's answers, and that evolution strategies — optimizing directly in weight space through random perturbations — preserve that diversity and raise pass@k. The paper was submitted on 13 August 2026 and has not been peer reviewed. The publicly visible abstract contains no benchmark names, model sizes, or measured figures. A preprint titled "Beyond the Best Guess: Improving LLM Solution Coverage with Evolution Strategies" was submitted to arXiv on 13 August 2026 and catalogued under artificial intelligence (cs.AI) with a secondary listing in neural and evolutionary computing (cs.NE). It is credited to seven authors — Conor F. Hayes, Elliot Meyerson, Kajetan Schweighofer, Roberto Dailey, Babak Hodjat, Risto Miikkulainen and Xin Qiu. The arXiv record does not list institutional affiliations, and this report does not attribute the work to any organization. The paper's framing starts from how language models are typically used in discovery settings such as mathematics and the sciences: a problem is presented, and the model's single answer is taken as the proposed solution. The authors argue that this "best guess" mode leaves value on the table, because additional test-time compute can be spent generating many candidate solutions instead of one. That regime is usually measured with pass@k, a metric that counts a problem as solved if at least one of k sampled attempts is correct. The central claim is a diagnosis of a side effect in current practice. Post-training a model with reinforcement learning, the authors write, narrows the model's output distribution around high-reward outputs — and that narrowing causes solution coverage to collapse. In other words, RL can make the first answer better while making the set of distinct answers smaller, which specifically penalizes the pass@k regime the authors care about. As an alternative, the paper proposes evolution strategies: a population-based, gradient-free post-training method that optimizes directly in weight space by applying random perturbations to model parameters and selecting on the results, rather than backpropagating a reward signal. The abstract states that ES achieves consistently higher pass@k than RL, produces a broader output distribution with greater solution coverage, and that this coverage in turn translates into better results on standard math benchmarks. What the abstract does not supply is most of the evidence. It names no model families or parameter counts, no specific benchmarks, no values of k, no baseline RL algorithm, and no numerical results — the words "consistently higher" and "better results" are the only characterizations offered. The submission is 449 KB and full text is available as PDF and experimental HTML, so those details may well be in the paper; they are simply absent from the record evaluated here. It is a version-one preprint with no indication of peer review, no visible code or data link, and no independent replication. [Read the primary source: arxiv.org ↗](https://arxiv.org/abs/2608.12679) ## Why it matters In domains where a candidate answer can be checked — proofs, code, scientific hypotheses — the useful ceiling is not whether a model's first guess is right but whether a correct answer appears anywhere in a batch of tries. If RL post-training systematically shrinks that pool, the industry's dominant alignment recipe may be trading away exactly the property that discovery and agentic search depend on. The distinction the paper draws — between a model's best single answer and the breadth of what it can produce across many samples — is not academic. A growing share of high-value AI work runs in a generate-then-verify loop: propose many candidate programs and run the test suite, propose many proof steps and check them mechanically, propose many hypotheses and screen them. In all of those settings, a verifier decides which candidate is right, so the binding constraint is whether a correct candidate was ever generated. Coverage is the ceiling, and accuracy on the first try is only one point beneath it. That makes the diagnosis potentially consequential for the dominant post-training recipe. Reinforcement learning from reward signals is how most current frontier models are shaped after pretraining, and sharpening the output distribution is close to the point of the exercise. If that sharpening reliably costs coverage, then the standard pipeline may be optimizing for benchmark scores measured at k=1 while quietly degrading performance in the regime where models are increasingly deployed. The concern that RL narrows model diversity has been raised before in the research literature; the contribution claimed here is a concrete alternative rather than a fresh diagnosis. The proposed remedy carries its own well-known trade-offs. Evolution strategies avoid gradients entirely, which sidesteps some of the instability of RL fine-tuning, but they historically pay for that with sample efficiency: estimating a useful update direction from random weight perturbations typically requires many forward passes across a population, which parallelizes well but consumes compute. Whether that arithmetic works at the scale of modern language models is precisely the question a reader would want answered, and the abstract does not address cost at all. For the public and for practitioners, nothing changes today. This is a research claim about training methodology, not a product, a model release, or a safety finding. Its practical significance would arrive indirectly — through models tuned to explore rather than to commit, or through providers exposing knobs that trade first-answer accuracy for breadth. There is no evidence in the source that any deployed system uses this method, and no vendor has been described as adopting it. It is also worth stating the limits of what can be concluded from a preprint abstract. The claim that ES beats RL on pass@k is the authors' report of their own experiments, not an independently established fact. Comparisons between optimization methods are notoriously sensitive to tuning effort, compute budget, and choice of baseline, and a result that holds for one model scale or one benchmark family often does not generalize. ## What to watch next The load-bearing details are in the full PDF and not the abstract: which models were trained, which benchmarks were used, what values of k, and crucially whether ES and RL were compared at matched compute. Also worth watching is whether code is released, whether the result reproduces outside mathematics, and whether any frontier lab adopts the method. The first thing to check is the full paper's experimental setup, and specifically whether the ES and RL runs were compute-matched. A gradient-free method that consumes substantially more training compute than its RL baseline can look better for reasons unrelated to the mechanism the authors propose. Readers should look for the model sizes trained, the RL algorithm used as a comparison, the number of perturbations per generation, and the values of k at which pass@k was measured — coverage curves often cross, and a method that wins at large k can lose at k=1. Second, watch for scope. The abstract's concrete claim about downstream gains is limited to "standard math benchmarks," a domain with cheap automatic verification and heavy prior optimization. Whether the coverage advantage transfers to code generation, scientific hypothesis generation, or open-ended agentic tasks — where verification is noisier and correctness is not binary — is unestablished by the material available. Third, watch for artifacts and replication. The arXiv listing shows no associated code or data release. A released implementation, or a reproduction by a group unaffiliated with the authors, would move this from a claim to a result. Absent that, the appropriate posture is interest rather than confidence, and any adoption signal should come with its own numbers. Fourth, the verifier question bounds how much any coverage gain is worth. Higher pass@k only converts into useful output when something can identify the correct candidate among many. In mathematics and software, checkers exist; in most commercial applications they do not, and selecting the right answer from a broader pool becomes its own unsolved problem. Follow-up work pairing ES-trained models with selection mechanisms would be the natural next step. Finally, watch the publication track. The paper is a v1 preprint with no stated venue or review status. Revisions, a conference submission, or public critique from other researchers working on RL post-training diversity would all sharpen how seriously the central claim should be taken.