NOTE-078: Neural Thickets

Contribution

Reframes a pretrained checkpoint as a distribution over parameter vectors rather than a point, and measures two properties of that distribution that nobody had: how often a random Gaussian perturbation improves a downstream task, and whether the perturbations that improve one task improve the others. Both rise with model scale. The practical consequence — that a single parallel round of guess-and-check is a viable post-training method for models above about 1.5B — is offered as a probe of the measurement rather than as a method to adopt, and the paper is explicit about that ordering.

Key insight

Pretraining does not just move the weights to a good point; it changes what is nearby. A small or untrained model sits on a local peak of any individual task’s accuracy, so improvement requires structured search — the needle in a haystack. A large pretrained model sits in a valley of each individual task’s accuracy, surrounded by higher ground in many directions at once, and the directions are not the same direction: each is an expert at something different. The aggregate pretraining loss is at a minimum the whole time, which is why this structure is invisible to the single-objective landscape analysis the field usually runs. Post-training in that regime is selection from something already present, not construction of something new.

Assumptions

The measurements are empirical and carry no formal conditions, but the setting is narrow in specific ways and they bound everything below:

• Perturbations are isotropic Gaussian in the full parameter space, with σ ∈ {1, 2, 3}e-3 for the method and σ = 1e-3 for the density study. No structure, no layer-wise scaling, no low-rank restriction.
• “Task expert” means “scores well on the benchmark”, which the paper states and then measures the consequences of (§8). Format and style count as expertise under this definition.
• Density is measured on one model family — Qwen2.5 instruction-tuned, 0.5B to 32B — on three reasoning tasks; diversity on seven tasks with 500 perturbations.
• The method is evaluated at 8B and below, on Qwen2.5, OLMo3 and Llama-3.1, with baselines matched on training FLOPs and not on inference FLOPs.
• Scoring uses a held-out set of roughly 200 examples. The method needs a cheap, reliable scalar score per candidate; it inherits every problem a verifier has.

Key results

• Solution density δ(m) = P[s(θ+ε) ≥ s(θ)+m] rises monotonically with model size, at every margin m tested. On GSM8K the fraction of perturbations at least matching base accuracy goes 0% → 64% across 0.5B → 32B (Fig. 3, Fig. 12).
• Spectral discordance rises monotonically with model size (Fig. 3b) — the task rankings of sampled perturbations become more orthogonal as models grow, supporting specialists over generalists. PCA of the per-seed performance vectors over seven tasks separates into clusters with complementary strengths (Fig. 4).
• RandOpt at N=5000, K=50 matches or beats PPO, GRPO and ES at equal training FLOPs across seven tasks (Table 4). Best cells include OLMo3-7B- Instruct on Countdown at 85.0% (ES 71.0, GRPO 68.5, PPO 69.0) and Qwen2.5-3B-Instruct on GSM8K at 87.1% (GRPO 83.2, ES 85.8).
• ES + 50-way test-time majority vote takes the best or runner-up cell in roughly half of the same table, beating RandOpt on GSM8K, MATH-500 and OlyBench at several scales (Table 4). The paper’s §5.4 asks whether ensembling helps the baselines and answers yes, and that the gap between selection methods shrinks as training proceeds.
• Wall clock: one training step against 200 (GRPO), 600 (PPO) and 167 (ES). OLMo3-7B-Instruct to 70% on Countdown in 3.2 minutes on 200 GH200s.
• The method has a scale threshold. No gain at GPT-2 0.1B, small gains at 0.5B, a sharp rise from ~1.5B, then shrinking relative gains as base accuracy catches up. From an untrained initialization it is near zero at every scale (Fig. 8).
• Ensembling is load-bearing: K=1 is substantially worse than K=50, though K=1 still beats base (Fig. 11).
• Population size helps log-linearly and the optimal selection ratio falls as N grows; at large N, keeping the top 1% is enough (Fig. 10).
• Accuracy decomposition on GSM8K (Qwen2.5-3B-Instruct, N=3000, K=50) splits the gain into a “reasoning thicket” (problems the base model got wrong) and a “format thicket” (problems it got right and mis-formatted), with substantial contributions from both — and reports the same shape for GRPO (Fig. 9).
• Distillation of the top-50 into one model: 84.3% against the ensemble’s 87.1% on GSM8K at 3B, at ~2% of training cost (Table 2).
• A minimal replication with no language in it. An MLP next-value predictor over a mixture of 1D signal families shows three regimes: needle (no pretraining), thicket (mixture pretraining), and plateau (pretrained on linear signals only — already at ceiling, guessing buys nothing).

Claims

id	claim	strength	support
C1	The fraction of random Gaussian perturbations that improve a downstream task rises monotonically with model scale	strong	direct measurement, Qwen2.5 0.5B–32B, three tasks, multiple margins (Fig. 3, 12)
C2	Those perturbations are task specialists, not uniform improvements	strong	spectral discordance rising with scale, plus PCA clustering over seven tasks (Fig. 3b, 4)
C3	Thickets require pretraining on a variety of tasks, not merely a lot of pretraining	moderate	the 1D signal experiment’s three regimes (§3); shown in the toy setting, asserted for LLMs
C4	One parallel round of random guessing plus top-K majority voting is competitive with PPO, GRPO and ES at equal training FLOPs	moderate	Table 4, but the baselines answer with one sample where this answers with fifty
C5	Random guessing is better than ES	weak	not supported; ES + TT-MV matches or beats it in about half the table, and the authors do not claim it
C6	The method fails below ~1.5B and from scratch	strong	Fig. 8, including the from-scratch control at every scale
C7	The gains are not an artifact of sandbagging	moderate	OLMo3 base is fully open and improves; §5.3 and App. G argue it, no direct test
C8	A substantial part of the gain is output formatting rather than new capability	strong	the paper’s own decomposition, §8 — and it holds for GRPO too

Method

RandOpt, in two phases.

Training (one step, fully parallel). Sample N seeds; assign each a noise scale σ_i drawn uniformly from a small set Σ; form θ_i = θ + σ_i · ε(seed_i) with ε ~ N(0, I_d); evaluate each θ_i on a small held-out set; keep the top K by score. Workers never communicate during this phase — scores are gathered once, against T times for ES.

Inference. Generate from each of the K selected models and return the majority answer.

Only a seed and a scale need to be stored per candidate, so the population is cheap to represent; the cost is N evaluations, which is what “equal training FLOPs” is buying against the baselines’ optimizer steps.

Concepts

• Thicket regime — the state of a parameter vector whose Gaussian neighbourhood contains a substantial fraction of task-improving solutions. Contrasted with the needle in a haystack regime (solutions vanishingly rare, structured search required) and the plateau regime (the base model is already optimal for the task, so nothing nearby helps).
• Solution density δ(m) — P[s(θ+ε) ≥ s(θ)+m] for ε ~ N(0, σ²I). The hit rate of random guessing at margin m.
• Spectral discordance — a scalar read off the spectrum of the correlation matrix of per-task percentile ranks across sampled seeds. 0 is orthogonal task rankings (specialists); parallel rankings (generalists) sit at the other end.
• Format thicket / reasoning thicket — the paper’s own split of where an improvement came from: making an already-correct answer parseable, versus answering a question the base model got wrong.

Connections

Positions itself against three lines. Against flat-minima work, its point is that a flat aggregate landscape is compatible with spiky per-task landscapes, and that the per-task structure is what governs adaptation — the pretrained weights need not minimize any single task. Against the lottery ticket hypothesis, it agrees about random initialization and claims a qualitatively different regime after pretraining, so the two are about different starting points rather than in tension. Against post-training as reweighting (KL-regularized PPO and friends), it is the same idea taken literally into weight space: the behaviours are already in the pretrained distribution, and post-training selects among them.

Its most direct relative is Qiu et al.’s evolution strategies at LLM scale, which it cites in the same breath as its own thesis — it doesn’t take much to obtain good downstream solutions given the right pretrained representation — and also runs as a baseline. Machine-readable lineage is on the LIT.

Recommendations

• R1 — If wall-clock time and parallelism matter more than inference cost, and your task has a votable answer, one parallel round of scored random perturbation plus top-K majority voting is a real option above ~1.5B. Topic: post-training. Status: experimental. Strength: moderate. Applies when: a cheap verifier exists, the answer is discrete, the cluster is wide, and K forward passes per answer are affordable.
• R2 — When comparing a post-training method that ensembles at inference against one that does not, give both the same test-time budget. This paper’s own matched arm (ES + TT-MV) changes which method wins in about half the table. Topic: evaluation. Status: standard. Strength: strong. Applies when: any method comparison where the arms differ in samples per answer.
• R3 — Decompose a post-training gain into format-fixing and capability before believing it. Topic: evaluation. Status: experimental. Strength: moderate. Applies when: the benchmark is scored by a strict answer checker.
• R4 — Expect gradient-free post-training to fail on small or untrained models, and treat that as a property of the model rather than of the method. Topic: post-training. Status: experimental. Strength: strong.

Bearing on the record

Two documents filed from this reading. THEORY-006 carries the density-and-diversity claim, Proposed, and SOTA-212 carries the method, Proposed and unreplicated.

It replicates SOTA-154, and that is what promoted it. That practice rested entirely on LIT-211, whose authors ran it. This paper runs ES as a baseline across six further tasks and three model families, at matched training FLOPs, with PPO and GRPO grid-searched over learning rate and batch or group size while ES ran a single fixed configuration — and finds ES strong enough that ES plus test-time voting beats the paper’s own method in about half the table.

Arriving as a baseline is not a discount on that: a group tuning the rival harder than the method, in a paper arguing for something else, is a strong form of evidence.

What is a discount on it is authorship, and this note missed it twice. The first draft called the result “the weaker of the two” for being a baseline and was corrected; the correction then called it an independent replication, which is also wrong. Yulu Gan is the second author of LIT-211, the paper this practice rests on, and this paper cites it as prior work. The tuning asymmetry still favours the baseline and the six added tasks are still six added tasks — but promote_when excluded “further results from the same group”, and an overlapping group is nearer to that than to what it asked for.

It supplies the explanation SOTA-154 never had. Nothing in the record said why a gradient-free method should find anything in a billion dimensions. THEORY-006 now does, and predicts where it stops.

R2 above is the one thing here the record has no home for. The anthology carries evaluation practices — reporting zero-shot and in-distribution separately, accounting for test-set proximity — and nothing about matching test-time sample budgets across compared arms. This paper demonstrates the failure and its fix in its own table, but does not recommend anything, so filing a practice would mean the record inventing a recommendation and attributing it. Left here deliberately, as a candidate for whoever files the practice that does recommend it.

Limitations

The authors state five, and all five are real: the method needs pretraining and fails from scratch; how far beyond the base model it can reach is unknown, and the scaling appears to saturate in log resources; inference costs K passes, with distillation as a partial and non-general escape; majority voting does not extend to structured prediction; and the mechanism by which pretraining produces a thicket is not explained.

Two more from this reading. The density and diversity measurements run on one model family, so the scaling law is a Qwen2.5 scaling law until someone repeats it. And the headline table’s asymmetry in test-time budget is disclosed but not corrected in the figure that carries the argument — the matched arm is present in Table 4 and in §5.4, and a reader taking Figure 6 at face value would come away with the wrong conclusion.

Open questions

• What about pretraining creates the thicket? The 1D experiment points at variety in the pretraining distribution. What would close it: a controlled pretraining sweep on LLMs varying data mixture diversity at fixed token count, with δ measured at the end of each.
• Where is the transition, really? “Around 1.5B” is one family on one task. What would close it: the density curve for two more families, which would also say whether the threshold tracks parameters, tokens, or benchmark headroom.
• Does the thicket survive a benchmark that cannot be moved by formatting? What would close it: the same decomposition on a task scored by a model or an execution harness rather than a string match.
• Is anisotropic perturbation better? Everything here is isotropic Gaussian over all parameters. Whether density is concentrated in particular layers or directions is not asked, and it bears directly on how large N has to be.