SOTA-212: Post-train by scoring many random weight perturbations in one parallel pass and majority-voting the best of them

Source, and who is recommending this

The evidence is Gan and Isola (2026), LIT-233 — ARXIV-2603.12228. The recommendation is not theirs. That paper measures the method and declines to promote it, in as many words: “our goal is not to promote RandOpt as superior to alternative methods. Rather, we use it as a probe.” The instruction — post-train by perturbing the weights and selecting among the results, rather than by following a gradient — was stated by Qiu et al. (LIT-211), and this is that instruction with the iteration count set to one and an ensemble on the end. introduced_by: says so, which is the distinction ADR-030 requires every practice to make.

The procedure, called RandOpt, is four lines:

1. Draw N seeds. Give each a noise scale σ drawn from a small set — the paper uses Σ = {1, 2, 3}e-3 — and form θ_i = θ + σ_i · ε(seed_i) with ε ~ N(0, I).
2. Score every θ_i on a small held-out set (~200 examples).
3. Keep the top K.
4. At inference, generate from all K and majority-vote the answers.

N = 5000, K = 50 throughout the paper. Nothing is sequential: the workers never communicate during training and the scores are gathered once, against O(T) rounds of communication for ES and O(T) optimizer steps for PPO and GRPO.

Why this is a recommendation and not just a result

The claim that carries it is the wall-clock and parallelism one, not the accuracy one. Matched on training FLOPs, RandOpt wins or ties most cells of a seven-task table across Qwen2.5 0.5B–3B, OLMo3-7B base and instruct, and Llama-3.1-8B-Instruct. But it does that in one step: OLMo3-7B-Instruct reaches 70% on Countdown in 3.2 minutes on 200 GH200s. Where nodes are cheap, interconnect is expensive and time-to-result is what binds — and in federated or otherwise decentralized settings, where communicating gradients is the problem — that is a different trade from the one every other post-training method offers, and it is the reason to reach for this rather than a margin on a benchmark.

Conditions, and they are severe

The base model has to be big enough. Below roughly 1.5B parameters the gains are small; at GPT-2 scale there are none; applied to un-pretrained weights the method returns near zero at every scale tested. This is not a tuning failure — THEORY-006 is the account of it, and the practice inherits its boundary. As base accuracy rises the relative gain also shrinks, so the window is real on both sides.

The answer has to be votable. Majority voting is the whole of the inference step. For a story, a molecule or an image there is no obvious vote; the paper offers mean-ensembling a diffusion model’s denoising steps as a proof of concept and claims nothing more.

Inference costs K forward passes. K = 1 is substantially worse than K = 50, so the ensemble is not optional. Distilling the top-K into one model recovers most of it — 84.3% against 87.1% on GSM8K for Qwen2.5-3B-Instruct, at about 2% of training cost — but that reintroduces the sequential training the method exists to avoid.

Part of what you are buying is formatting. The paper’s own decomposition of the GSM8K gain finds a substantial share is answers the base model already had, emitted in a form a strict checker rejects. It reports the same of GRPO, so this is not a reason to prefer the baseline — it is a reason to check what moved before believing a number.

Against the record’s post-training spine

SOTA-129 makes reinforcement learning with verifiable rewards the third stage of the reasoning recipe, SOTA-145 recommends the group baseline inside it and SOTA-146 corrects the objective. All three assume the paradigm. This is the second practice in the record to reach the same goal from outside it, after SOTA-154 — which it extends: rather than rivals. Both are gradient-free, both perturb the full parameter space with Gaussian noise, and this one is ES with the iteration count set to one and an ensemble on the end.

The honest comparison against SOTA-154 is a tie, not a win. The paper’s headline gives this method a 50-way ensemble and the RL baselines a single sample, which it says plainly “disadvantages the baseline but reflects current standard usage”. It also runs the matched version — ES plus 50-way test-time majority vote — and that arm takes the best or runner-up cell in roughly half the table, beating this one on GSM8K, MATH-500 and OlyBench at several scales. The paper’s own conclusion is that once you ensemble, the selection method matters less and less. What this practice is evidence for is the paradigm and the parallelism; a practitioner choosing between it and SOTA-154 on accuracy alone has no result to choose on.

That tie is also the most useful thing here for a reader evaluating anyone else’s post-training numbers: an ensembled method compared against a one-sample baseline is a comparison about ensembling.

Known implementations

• None in a released model. The authors publish code at github.com/sunrainyg/RandOpt.