cs.LGMar 12, 2026

Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights

Authors: Yulu Gan, Phillip Isola

Organizations: MIT CSAIL

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Pretraining produces a learned parameter vector that is typically treated as a starting point for further iterative adaptation. In this work, we instead view the outcome of pretraining as a distribution over parameter vectors, whose support already contains task-specific experts. We show that in small models such expert solutions occupy a negligible fraction of the volume of this distribution, making their discovery reliant on structured optimization methods such as gradient descent. In contrast, in large, well-pretrained models the density of task-experts increases dramatically, so that diverse, task-improving specialists populate a substantial fraction of the neighborhood around the pretrained weights. Motivated by this perspective, we explore a simple, fully parallel post-training method that samples NN parameter perturbations at random, selects the top KK, and ensembles predictions via majority vote. Despite its simplicity, this approach is competitive with standard post-training methods such as PPO, GRPO, and ES for contemporary large-scale models.

Figures & tables

Figure 1: (a) Schematic of the main effects we observe (see Fig 2 for a version with real data). Left: Small models live in a needle in a haystack regime, where good solutions to downstream tasks occupy a tiny fraction of the surrounding weights. In this regime, it is important to have a smart search algorithm, such as gradient descent or other forms of iterative optimization. Right: Large models are surrounded by a veritable thicket of task-specific solutions. In this regime, random sampling is sufficient to quickly land on promising adaptations, which can then be ensembled to yield strong behavior, an approach we call RandOpt . (b) Solution density – i.e. density of task-improving weights in a Gaussian neighborhood of the pretrained weights – scales with model size. (c) RandOpt is O(1)\mathcal{O}(1) in training steps, FLOP-efficient, and competitive in converged accuracy with GRPO and ES. Results are shown on the Countdown task with Olmo-3-7B-Instruct ; RandOpt uses 5000 random weight guesses and ensembles the top KK ; K-pass baselines use Test-time Majority Vote (TT-MV). More results are shown in Fig. 6 and Table 4 .
Figure 2: Accuracy landscapes in weight space across model scales and reasoning tasks. We perturb the pretrained Qwen2.5 models (from 0.5B to 32B) with 1000 random weight perturbations and project the perturbed models into 2D using random projection. Colors show relative accuracy change (acc−base)/base×100(\mathrm{acc}-\mathrm{base})/\mathrm{base}\times 100 ( blue : degraded, white : equivalent, red : improved.) Dashed circles indicate the mean perturbation distance and stars mark the best-performing perturbations. Larger models have warmer landscapes, with more high-performing neighborhoods. The last column shows an RGB visualization where GSM8K, Olympiad and Countdown accuracies are mapped to R,G,B channels; richer colors indicate more task experts.
Figure 3: Scaling laws of solution density and diversity (using Qwen-2.5 instruction tuned models). (a) Solution density increases with model scale, showing that larger models have a higher fraction of good solutions. (b) Spectral discordance across model scales, measuring solution diversity. Together, these results demonstrate that larger models have both denser and more diverse solution landscapes in the neighborhood around their pretrained weights.
Figure 4: Performance spectra and clustering of random seeds. Sampled vectors possess diverse areas of expertise, with individual seeds specializing in specific tasks. (Left) Performance of 100 random seeds across seven evaluation datasets. Each line represents a specific seed, with four lines highlighted as examples. (Right) PCA visualization of these performance vectors, where seeds with similar behavior cluster together into different groups.
Figure 5: Pretraining a model of 1D signals, then probing the local neighborhood around the pretrained weights by random guessing N=1000N=1000 Gaussian perturbations. The plot shows the autoregressive predictions of a particular linear function (dashed blue line), given an observed context (solid blue line). Gray lines: random fθf_{\theta} ’s; Red lines: top-K fθf_{\theta} ’s. The figure shows three regimes: (a) No pretraining leads to needle-in-the-haystack search, (b) pretraining on several signal types leads to a thicket, (c) pretraining on just linear functions achieves nearly perfect predictions at pretraining time, hence post-training is at ceiling.
Figure 6: RandOpt vs. baselines on post-training LLMs. Marker size represents model scale (0.5-8B), shape indicates task family, and transparency distinguishes benchmarks within each task family. RandOpt matches or exceeds baselines in most settings. RandOpt is run with K=50 and ES+TT-MV also uses 50 test-time samples ensembled via majority vote. 1-pass PPO/GRPO use a single test-time sample; this setting disadvantages the baseline but reflects current standard usage of these models, in which there is no test-time ensembling. Full experimental results can be found in Appendix Table 4 .
ModelMethodGQA
Qwen2.5-VL-3B-InstBase56.6
RandOpt69.0
Table 1: RandOpt improves Accuracy (%) on GQA.
Figure 8
ModelMethodGSM8K
Qwen2.5-1.5B-Inst.Base58.8
Distill74.9
RandOpt76.4
Qwen2.5-3B-Inst.Base79.8
Distill84.3
RandOpt87.1
Table 2: Distilling the top-K RandOpt ensemble into a single model achieves performance comparable to the ensemble on GSM8K.
Figure 9: Accuracy decomposition on GSM8k using Qwen2.5-3B-Instruct (N=3000, K=50). Grey denotes strictly correct answers (both format and answer correct), light blue denotes the questions originally answered incorrectly that are corrected after training, and purple denotes the gains are just from fixing format. For both GRPO and RandOpt, a large portion of the gains come from format correction, while another portion is from solving the problem correctly.
Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10: Scaling curves for population size KK and ensemble size NN on Countdown task.
Figure 11: Model performance comparison on Countdown, GSM8k, MATH-500, and OlympiadBench using the Qwen2.5-1.5B-Instruct model. More results for additional models and baselines can be found in Appendix Table 4 .
Figure 12: Performance distributions of 500 randomly perturbed models relative to base model accuracy on the math reasoning task GSM8k (left) and Countdown (right). The fraction of perturbations matching or exceeding base performance (shaded region) increases with model size: 0%→64%0\%\to 64\% on GSM8K and 8%→60%8\%\to 60\% on Countdown as model size grows from 0.5B to 32B parameters. Note: x-axis shows relative performance; longer tails in smaller models do not equate to higher absolute accuracy.
Figure 13: Batch/group size vs. accuracy under one-step training. Scaling baseline parallelism does not get RandOpt’s performance. Task: GSM8K, Model: Qwen2.5-3B-Instruct, N=5000, K=50)
CategoryParameterGRPOPPOESRandOpt
Budget & ScaleBatch Size1024128––
Group/Population Size81305000
Iterations2006001671
Optimization ComplexityActor Learning Rate1e-61e-65e-4 ( α\alpha )–
Critic Learning Rate–1e-5––
OptimizerAdamWAdamW-–
Appendix
Table 3: Hyperparameters. ’ – ’ indicates that the hyperparameter is not required for the corresponding algorithm.
ModelMethodCountdownGSM8kMATH-500OlyBenchMBPPROCStoriesUSPTO
  (math) ↑\uparrow(math) ↑\uparrow(math) ↑\uparrow(math) ↑\uparrow(prog.) ↑\uparrow(writing) ↑\uparrow(chemistry) ↑\uparrow
Qwen2.5- 0.5B-InstBase0.1 ±0.1{}_{\pm\text{0.1}}39.9 ±0.0{}_{\pm\text{0.0}}19.8 ±0.0{}_{\pm\text{0.0}}4.2 ±0.4{}_{\pm\text{0.4}}30.9 ±0.2{}_{\pm\text{0.2}}22.0 ±0.0{}_{\pm\text{0.0}}29.9 ±0.3{}_{\pm\text{0.3}}
TT-MV †0.3 ±0.3{}_{\pm\text{0.3}}41.0 ±0.1{}_{\pm\text{0.1}}29.2 ±0.5{}_{\pm\text{0.5}}6.7 ±0.2{}_{\pm\text{0.2}}37.0 ±0.3{}_{\pm\text{0.3}}19.5 ±0.2{}_{\pm\text{0.2}}30.2 ±0.2{}_{\pm\text{0.2}}
Best-of-N ‡5.9 ±2.0{}_{\pm\text{2.0}}40.5 ±1.4{}_{\pm\text{1.4}}34.5 ±0.5{}_{\pm\text{0.5}}10.5 ±0.3{}_{\pm\text{0.3}}34.9 ±1.3{}_{\pm\text{1.3}}20.3 ±0.3{}_{\pm\text{0.3}}26.8 ±1.2{}_{\pm\text{1.2}}
PPO14.8 ±2.8{}_{\pm\text{2.8}}43.2 ±1.2{}_{\pm\text{1.2}}33.4 ±1.7{}_{\pm\text{1.7}}16.1 ±0.0{}_{\pm\text{0.0}}37.8 ±3.0{}_{\pm\text{3.0}}19.1 ±0.8{}_{\pm\text{0.8}}31.2 ±0.6{}_{\pm\text{0.6}}
GRPO13.0 ±0.0{}_{\pm\text{0.0}}48.4 ±0.8{}_{\pm\text{0.8}}33.7 ±0.9{}_{\pm\text{0.9}}6.9 ±0.1{}_{\pm\text{0.1}}42.8 ±3.7{}_{\pm\text{3.7}}30.9 ±1.1{}_{\pm\text{1.1}}31.7 ±0.2{}_{\pm\text{0.2}}
Appendix
Table 4: Experimental results on reasoning benchmarks. We compare RandOpt against RL-based methods (PPO, GRPO), Evolution Strategies (ES), Best-of-N and test-time majority voting (TT-MV) across model scales. Results are averaged over 3 runs with standard deviation shown in gray . Bold indicates best performance, underlined indicates runner-up. Details of the hyperparameters for all methods are provided in Appendix E.3 .
Table 5: RandOpt on 1D signals (approximation). Each cell corresponds to a pretraining–post-training pair. In this experiment we plot results on the same function that we fit to during post-training; this is a test of apprpoximation only, not generalization to new functions. We also compare to Xavier initialization ( Glorot & Bengio, 2010 ) and Kaiming initialization He et al. (2015) . For each pretraining method, we pick a perturbation noise scale, σ\sigma that is large enough to show functional variation; this value is shown underneath the pretraining method name.
Table 6: RandOpt on 1D signals (generalization). Each row corresponds to a pretraining–post-training pair. For all rows, the test set functions are of the post-trained type, and we show three random test examples. Rightmost column shows the average mean squared error over the entire test set, for each method.
Table 7: RandOpt on Text-to-Image models (Train Set). We use the Stable Diffusion XL model ( Podell et al., 2023 ) . Images are generated from a text prompt. RandOpt selects the top-K models by scoring generated images with a target text (e.g., “blue”) using GPT-5.2, and performs mean ensembling over the K models at each denoising step.
Table 8: RandOpt on Text-to-Image models (Test Set). Top-K selected on the training set and evaluated on the test set.
More in this topic

LLM Post-Training

Large language model post-training adapts pretrained models into more useful, reliable, and specialized assistants through supervised and preference-based learning.

See all 246 papers

Explore similar work

CardsList
Oct 7, 2026

Sequential Pretraining Favors Large Models

Large neural networks often acquire capabilities that small models fail to learn. Does this stem from large models learning more representative features, or from being more robust to unaccounted-for adverse effects introduced during training? We define and quantify one such adverse effect, primacy bias, as the extent to which exposure to early data distributions impairs later learning. We show that small models can allocate learning capacity inefficiently toward early distributions, whereas sufficiently overparameterized models are robust to this effect. This inefficiency is particularly consequential in pretraining, where foundation models often encounter heterogeneous data distributions sequentially rather than jointly. As a result, small foundation models can struggle to learn distributions encountered late in training, which is particularly harmful when later data emphasizes desirable capabilities such as code, mathematics, and reasoning. Motivated by these findings, we introduce Exposure Therapy (ET), a simple regularization that promotes more efficient allocation of learning capacity during sequential pretraining. We demonstrate that ET improves foundation models' performance on late data distributions as well as overall capability in models up to the billion-parameter scale. Overall, our results suggest that some benefits of large foundation models may arise from greater robustness to adverse training effects, rather than from learning more representative features, and that improved training algorithms can recover some of these advantages in smaller models.

Mohnish Harwani, Yujia Zheng

Purdue University · University of Illinois Urbana–Champaign

Language Model PretrainingContinued Pretraining
AskPDF
May 12, 2026

Early Data Exposure Improves Robustness to Subsequent Fine-Tuning

How can we train models whose post-trained capabilities survive subsequent fine-tuning? Rather than focusing on downstream interventions to mitigate forgetting of upstream capabilities, we study how upstream training choices - that is, the manner in which a capability is acquired - shape how robustly that capability is retained. We investigate this question in a controlled three-stage language-model pipeline: pretraining, post-training to acquire a target capability, and downstream fine-tuning on a new objective. Across 135M and 1B models, two post-training domains, and two downstream fine-tuning tasks, we find that immediate post-training performance does not reliably predict retention after subsequent fine-tuning: training recipes that look equivalent immediately after post-training can retain the target capability very differently after subsequent fine-tuning. In particular, early exposure - mixing post-training data into pretraining - consistently improves the frontier between retained upstream performance and downstream performance. In compute-matched experiments, where the target data must be allocated between pretraining and post-training, we find that the optimum lies at neither extreme. Together with our other empirical and theoretical findings, this supports the view that post-training drives immediate specialization while early exposure improves robustness to later forgetting. Replay and dropout, typically used to mitigate forgetting as it occurs during fine-tuning, provide complementary gains to early exposure when applied during post-training. Our findings suggest that robustness to subsequent fine-tuning should be treated as a first-class objective of upstream training, addressed preventatively through choices like early exposure rather than reactively during fine-tuning itself.

Lawrence Feng, Gaurav R. Ghosal, Jacob Mitchell Springer +2

Carnegie Mellon University

Language Model PretrainingContinual Learning for LLMsLLM Fine-TuningLanguage Model RobustnessCatastrophic Forgetting
AskPDF
Jun 3, 2026

TENP: Trapezoidal Expert Neuron Pruning For Mixture-of-Experts

Mixture-of-Experts large language models (LLMs) scale efficiently through sparse activation, yet their deployment is fundamentally constrained by the large static parameter footprint of experts. Existing compression approaches either remove entire experts, disrupting routing topology and harming performance, or rely on unstructured weight pruning with limited practical efficiency. To address the limitations, we propose TENP, a structured Trapezoidal ExpertNeuron Pruning framework. Using a few samples, we identify and retain important experts, while applying expert neuron pruning (ENP) to less important experts, reserving model parameters in a trapezoidal pattern from shallow to deep layers. When evaluating expert importance, we jointly consider both the magnitude of the expert output and its ability to change the direction of the input vector. For ENP, we measure each neuron's projected contribution to the expert output to identify and retain important neurons. We conduct extensive experiments on the Qwen and DeepSeek models. Under a routing expert sparsity of 40% and an average of 63.76% activated expert parameters, the DeepSeek model suffers only a 1-point drop in accuracy compared to the full-parameter model. Moreover, it outperforms the full-parameter model by 10% on code generation tasks.

Jiangyang He, Shaolin Zhu, Deyi Xiong

TJUNLP Lab, School of Computer Science and Technology, Tianjin University, China

Mixture-of-Experts PruningLLM PruningMixture-of-Experts Language ModelsLLM CompressionStructured SparsityStructured Pruning+1
AskPDF