Korotyshova et al. (2025) — ARXIV-2507.04453
Key takeaways
It predates the full-parameter line and answers the same question differently. Posted two months before LIT-211, it accepts the premise that premise rejects — that ES cannot search a billion dimensions — and reduces the dimension instead. Three nested reductions: LoRA rather than full parameters; SVD of each LoRA adapter; and CMA-ES over only the top fraction of singular values. In the headline comparison only the top 40% of singular values are trained.
SFT initialization is load-bearing and stated as such. “The ES search does not begin with random LoRA adapters.” Adapters are supervised-fine-tuned on a small labelled set first, so the evolutionary search starts inside a region already adapted to the task format. This is the same structural move LIT-231 makes a year later with GRPO steps in place of SFT — seed the subspace with gradients, then search it without them.
The efficiency results are the strongest part. Against GRPO on GSM8K with Qwen2.5-Math-7B on 8 GPUs and identical initialization, ESSA passes 0.80 accuracy in under 25 minutes where GRPO needs more than 50 — and it does so from 100 training examples against GRPO’s full 5,978-example fold. Sample efficiency, not just wall clock. The baseline is LIT-127’s GRPO started from the same SFT-trained point, so what is compared is a gradient-based RL update of the model against a gradient-free search over the adapters’ top singular values.
It runs where backpropagation cannot. Qwen-32B aligned at BF16, INT8 and INT4 with a fixed population of 192, and the quantized runs are slightly better than BF16 on both convergence speed and final accuracy. Since ES needs only forward passes, the training precision can be the inference precision — which is the same argument LIT-229 later makes for integer-only recurrent pretraining, reached from the opposite end of the method space.
Rank has an interior optimum. LoRA ranks 4–32 converge fastest and reach the highest accuracy; rank 64 is slower and worse. The paper’s reading is the straightforward one — a larger search space is harder for ES to explore — and it is a useful counterweight to reading “reduce the dimension” as “reduce it as far as possible”.
What the evidence does not cover
One benchmark family. Mathematical reasoning with verifiable accuracy-based rewards, which the authors name as the limit: ESSA’s reliance on a scalar fitness “may limit its applicability to subjective or sparse feedback scenarios, such as aligning human preference in open dialogue” — which is what the title’s word alignment would normally mean.
The GRPO comparison differs in more than the optimizer. Different runtimes (vLLM against VERL), different training-set sizes by design, and a wall-clock x-axis. It is a convincing efficiency demonstration and not a controlled comparison of optimizers.
No forgetting or drift measurement, which is the axis the later papers in this line turn out to care about — though searching a few hundred singular values bounds the drift by construction, which is worth noticing given what LIT-235 says drift scales with.
Standing in the anthology
The line’s head, chronologically, and the record filed it last. Everything
else here descends from or argues with LIT-211’s full-parameter claim; this
came first and took the dimensionality objection seriously rather than
overturning it. Filing it last is an artifact of having followed citations
backward from a 2026 paper, and the corrective is in the curation entry rather
than in a relation — it is not extends: or corrects: anything here,
because the later work did not build on it.
It makes the subspace approach a line rather than a one-off. LIT-231 (Hyper-ES) argues from lemmas that full-parameter ES cannot work and searches a gradient-seeded subspace instead. This did the same thing eighteen months earlier without the argument, and reported it working at 7B and 32B. So the record now holds two independent instances of seed a low-dimensional subspace with gradients, then search it gradient-free — and the fact that they were developed independently is a better argument for the approach than either paper’s own results.
It does not support SOTA-154, and the distinction matters: that practice recommends evolution strategies over the full parameter space in place of policy-gradient RL. This is ES over a few hundred coefficients, initialized by supervised fine-tuning. Same family, different recommendation.