Neural Thickets: RandOpt Shows Random Perturbations Around Pretrained Weights Rival RLHF-Style Fine-Tuning

Forum topic · ✨步子哥 · 2026-03-19

Summary

A 2026 paper from MIT CSAIL researchers (Yulu Gan, Phillip Isola, et al.), "Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights", reports a counterintuitive finding: around the weights of large multi-task pretrained models, high-performing task-specific "experts" are extremely dense, not sparse. Applying Gaussian noise perturbations (sigma around 0.005) to models above roughly 1.5B parameters yields many variants that excel at downstream tasks — for a 7B model, over 15% of randomly sampled neighbors exceed 75% accuracy on GSM8K, versus near zero without pretraining. The proposed RandOpt algorithm exploits this: sample N perturbed copies, evaluate them in parallel on a small validation set (~200 examples), keep the top-K elites, and ensemble their predictions — no gradients or backpropagation required. With N=5000, K=50, RandOpt reaches 82.3% on GSM8K and 87.1% on Countdown, comparable to or better than GRPO (83.5%) and PPO (78.0%); on OLMo-3-7B-Instruct it completes training in 3.2 minutes on 200 GH200 GPUs. The effect scales superlinearly with model size (expert density roughly proportional to N^1.5-2.0) and is strongest on vision-language models (GQA accuracy up from 56.6% to 69.0%). The paper also analyzes distillation for inference-cost reduction, open-source release, security/fairness considerations, and implications for optimization theory and AI democratization. Code: https://github.com/sunrainyq/RandOpt

Neural Thickets: RandOpt Algorithm — Technical Innovation, Theoretical Breakthrough, and Social Impact

This post discusses a paper from MIT CSAIL (Yulu Gan, Phillip Isola, et al., March 2026): *"Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights"*, and its proposed RandOpt algorithm.

Key points

1. The "Neural Thickets" phenomenon

Counterintuitive finding: Large multi-task pretrained models have *dense* clusters of task-specific expert models in their weight neighborhoods — challenging the classic assumption (Schmidhuber et al., 2001) that good solutions are extremely sparse in weight space.
Scale dependence: The effect emerges sharply around ~1.5B parameters. For a 7B model, with noise scale sigma = 0.005, over 15% of random perturbations exceed 75% accuracy on GSM8K; for unpretrained models this is near zero.
Superlinear scaling: Effective expert density follows approximately a power law, rho proportional to N^alpha with alpha ≈ 1.5–2.0, meaning larger models are *easier* to optimize via random search.
Pretraining quality is decisive: Only multi-task, large-scale pretraining creates the dense-expert geometry; single-task or no pretraining does not.

2. RandOpt algorithm

Two-stage design, no gradients or backpropagation:

1. Training/search stage: sample N Gaussian noise seeds, instantiate perturbations theta_i = theta + sigma * epsilon_i (sigma ≈ 0.005), evaluate in parallel on a small validation set (~200 samples), keep top-K elites (typical: N=5000, K=50). 2. Inference stage: ensemble the K elites via majority voting (classification) or probability averaging (generation), with built-in uncertainty quantification via prediction agreement.

Key implementation details:

Seed-based deterministic noise generation: memory drops from O(N·|theta|) to O(|theta|); a 7B model with N=5000 needs ~40KB of seeds instead of ~140TB of checkpoints.
Time complexity O(1) (fully parallel) versus O(T) iterative updates for PPO/GRPO.
On 200 NVIDIA GH200 GPUs, RandOpt (N=2000, K=50) trains OLMo-3-7B-Instruct in 3.2 minutes, reaching 70% on Countdown.

3. Benchmark results

| Method | GSM8K | Countdown | |---|---|---| | PPO | 78.0% | baseline | | GRPO | 83.5% | baseline | | RandOpt (random, K=50) | 82.3% | 87.1% | | RandOpt (distilled) | 84.3% | — |

Vision-language models benefit most: on Qwen2.5-VL-Instruct (3B), GQA accuracy rises from 56.6% to 69.0% (+12.4 points), exceeding typical gains on text-only models.
Error analysis attributes gains both to output formatting ("format thickets") and genuine reasoning improvements ("reasoning thickets").
Distillation: using elite outputs for 2 rounds of SFT compresses K=50 to a single model retaining ~90% of ensemble performance at ~2% of training cost.

4. Practical guidance

Applicable models: Qwen2.5 (0.5B–32B, verified), Llama 2/3, OLMo3; requires weight access and >1.5B parameters with diverse pretraining. VLMs: freeze the vision encoder, perturb only the language-model weights.
Hyperparameters: sigma has a wide plateau (robust around 0.005, effective range ~0.001–0.02); K=50 is the standard cost/quality trade-off; ~200-sample validation sets suffice.
Workflow: 30-minute prototyping (N=500, K=10) → sigma grid search → scale to N=5000, K=50.
Complementarity with RLHF: RandOpt can serve as initialization or exploration for PPO/GRPO; iterated variants (adaptive sigma schedules, ES-hybrid) trade some parallelism for refinement.
Applications: math reasoning, code generation (MBPP pass@10 gains of 20–25%), creative writing, chemistry (USPTO), plus industry uses in education, finance, healthcare, and enterprise code assistants.

5. Theoretical and societal implications

Challenges optimization theory: in well-prestructured loss landscapes, gradient-free random search suffices; "neighborhood optima plus ensembling" can replace global optimization.
Suggests implicit modularity in dense models — an intriguing parallel to explicit Mixture-of-Experts architectures.
Democratization: domain experts (teachers, doctors, analysts) can customize models in minutes without RL expertise; but risks include capability–responsibility mismatch, larger attack surfaces (validation-set poisoning, ensemble evasion), auditability challenges under regulations like the EU AI Act, fairness effects of validation-set-driven selection, and inference-time K-fold compute/energy overhead (mitigated via distillation).

6. Open questions

Rigorous mathematical characterization of expert density and its scaling law rho(N, D, C).
Generality beyond Transformers (Mamba/SSM, MoE) and to 100B+ scale.
Online/continual-learning variants and edge deployment via distillation and quantization.

Code: https://github.com/sunrainyq/RandOpt

*Note: figures above are as reported in the forum post/paper; the original Chinese post is truncated near the governance-recommendations section.*