EGGROLL: low-rank evolution strategies make gradient-free LLM fine-tuning fast

What happened#

A paper posted to arXiv on 20 November 2025 introduces EGGROLL — Evolution Guided GeneRal Optimisation via Low-rank Learning — a variant of evolution strategies (ES) designed to make gradient-free optimisation cheap enough to use on billion-parameter language models. The authors are Bidipta Sarkar, Jakob Nicolaus Foerster, Shimon Whiteson and seventeen collaborators. The work reached us via bycloud's video coverage, published 27 April 2026; the paper itself, "Evolution Strategies at the Hyperscale" (arXiv:2511.16652), is what this article is about.

The motivation: evolution strategies are black-box optimisers that are highly parallelisable and handle non-differentiable and noisy objectives, but naïve ES becomes prohibitively expensive at scale on GPUs. The bottleneck, per the paper, is the low arithmetic intensity of batched matrix multiplications with unstructured random perturbations — the hardware spends its time moving random matrices around rather than computing.

What does the EGGROLL paper actually claim?#

The abstract makes four substantive claims:

1. Speed. Structuring individual perturbations as rank-$r$ matrices instead of full unstructured random matrices improves arithmetic intensity, producing "a hundredfold increase in training speed for billion-parameter models at large population sizes," and "up to 91% of the throughput of pure batch inference."
2. Theory. The paper gives a theoretical analysis of Gaussian ES for high-dimensional parameter objectives, identifying a "linearising effect" and proving consistency between EGGROLL and standard ES as parameter dimension increases.
3. Integer-only pretraining. EGGROLL enables stable pretraining of nonlinear recurrent language models that operate purely in integer datatypes — a setting where backpropagation through the model is not available in the usual floating-point form.
4. LLM post-training and RL. EGGROLL is "competitive with GRPO for post-training LLMs on reasoning tasks" and "does not compromise performance compared to ES in tabula rasa RL settings, despite being faster."

Note the precision of those verbs: competitive with, does not compromise. The paper does not claim to beat GRPO.

Chart drawn from the source's reported figures. Data: https://arxiv.org/abs/2511.16652.

How does the low-rank trick speed up evolution strategies?#

Standard ES perturbs the model's parameters with a full random matrix per population member, evaluates each perturbed copy, and takes a fitness-weighted average of the perturbations as the update. On GPUs, applying a different unstructured random matrix to each population member is a batch of small, irregular matrix multiplications — exactly the workload GPUs are worst at.

EGGROLL's fix is to parameterise each perturbation as a rank-$r$ product rather than a dense random matrix. A rank-$r$ perturbation can be applied as a sequence of two thin matrix multiplications, which restores the arithmetic intensity that makes batched GPU execution efficient. The paper's theoretical contribution argues this restriction is not just a speed hack: it identifies a linearising effect in high dimensions and proves that EGGROLL's updates become consistent with full ES updates as the parameter dimension grows — the regime LLMs live in.

What is genuinely new here?#

Evolution strategies themselves are decades old, and applying ES to neural networks is not new either. The novel contributions, as the paper states them, are the rank-$r$ structuring of perturbations as the fix for ES's GPU inefficiency, the convergence analysis showing when and why low-rank ES matches full ES in high dimensions, and the three demonstrations that the result is practically useful: integer-only recurrent LM pretraining, GRPO-competitive reasoning post-training, and tabula rasa RL.

The broader framing — that evolution strategies are "back from the museum" — is the reporter's. bycloud's coverage points to a lineage including OpenAI's 2017 ES paper (arXiv:1703.03864) and an earlier 2025 paper titled "Evolution Strategies at Scale" (arXiv:2509.24372) as prior steps in this revival [unverified — neither source is resolved in our evidence bundle]. What the EGGROLL paper itself claims is narrower and more defensible: ES was too slow on GPUs, low-rank perturbations fix that, and the resulting method is competitive — not dominant — on the tested tasks.

Where does the coverage differ from the paper?#

Only one reporter covered this story, so there is no cross-source disagreement to resolve. There is, however, a framing gap worth naming. The video's title — "A new way to fine-tune LLMs just dropped" — presents EGGROLL primarily as a fine-tuning method. The paper presents three equal contributions, two of which (integer-only pretraining and tabula rasa RL) are not fine-tuning at all, and the fine-tuning claim is "competitive with GRPO," not a replacement for it.

Because no transcript was available for the video, we cannot verify or attribute any more specific claims from the coverage; everything stated as fact above comes from the paper's abstract.

What are the limitations?#

Several caveats follow from the paper's own wording. The headline throughput figure is "up to" 91% of pure batch inference — a best case, at large population sizes, not a guaranteed operating point. The hundredfold speedup is against naïve ES, the straw-man baseline the method is designed to beat, not against gradient-based training. "Competitive with GRPO" on reasoning post-training means the method holds its own in the authors' experiments; the abstract gives no benchmark numbers, so the margin and the evaluation suite cannot be checked from what we have.

The theoretical result is also conditional: the paper investigates "conditions needed for ES updates to converge in high dimensions," which implies those conditions are not automatic. And this article is written from the abstract alone — the full text would be needed to assess the experimental design, model sizes, and whether the integer-datatype pretraining results generalise beyond the recurrent architectures tested.

FAQ#

What is EGGROLL?#

EGGROLL (Evolution Guided GeneRal Optimisation via Low-rank Learning) is a variant of evolution strategies introduced in arXiv:2511.16652 by Sarkar, Foerster, Whiteson and colleagues. It structures the random weight perturbations ES uses as rank-$r$ matrices, which makes population-based gradient-free training fast enough on GPUs to apply to billion-parameter language models.

Is EGGROLL better than GRPO for fine-tuning LLMs?#

The paper does not claim that. It reports that EGGROLL is "competitive with GRPO for post-training LLMs on reasoning tasks" — meaning comparable performance in the authors' experiments, not superior performance. The abstract gives no benchmark numbers, so the exact margin is not publicly stated there.

How much faster is EGGROLL than standard evolution strategies?#

The authors report a hundredfold increase in training speed for billion-parameter models at large population sizes, compared with naïve ES using full unstructured random perturbations. At its best, EGGROLL reaches up to 91% of the throughput of pure batch inference — meaning the training loop runs almost as fast as simply running the models forward.

Why would you train a model without gradients?#

Gradient-free methods like evolution strategies handle objectives that are non-differentiable or noisy, and they parallelise trivially because each population member is evaluated independently. The EGGROLL paper demonstrates one concrete case where gradients are unavailable in the usual form: pretraining nonlinear recurrent language models that operate purely in integer datatypes.

Where can I read the EGGROLL paper?#

The paper, "Evolution Strategies at the Hyperscale," is on arXiv as arXiv:2511.16652, posted 20 November 2025.

Sources#

Primary sources#

• Evolution Strategies at the Hyperscale — Bidipta Sarkar, Mattie Fellows, Juan Agustin Duque et al.; 2025-11-20; arXiv:2511.16652

Reporting that surfaced this#

• bycloud — A new way to fine-tune LLMs just dropped (2026-04-27)