acceptodds
Under review as a conference paper at ICLR 2027

Does Random Search Discover Task Experts? Revisiting Neural Thickets

10 views

Abstract

Pretrained models are typically viewed as generalists, requiring bespoke post-training to achieve task-specific expertise. Gan & Isola (2026) recently proposed a striking alternative: that pretrained models lie in a dense “neural thicket” of task specialists, which can be surfaced with simple Gaussian weight perturbations. We examine this neural thicket of task-specific experts and identify several key findings. First, we find that the apparent lift of the best-of- selected perturbation decays quickly with the size of the selection set, strongly suggesting overfitting to the selection questions. In fact, for of the model–task pairs we study, a simple null model of pure selection noise predicts more than of the top expert's reward lift above the perturbation-population mean. Second, the top experts' advantage often vanishes on held-out questions, or remains within the run-to-run variability of the unperturbed "anchor" model. Third, we find that majority voting over the top- selected perturbations—coined RandOpt() by Gan & Isola (2026)—is often matched or outperformed by majority voting over samples from the base model. Overall, our results suggest that random weight perturbations may be ineffective at uncovering task experts around pretrained models, and that adapting pre-trained models to downstream tasks may require more sophistication than random search.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large language models, neural thickets, random weight perturbations, majority voting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_TNiMOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to critique the claim from Gan & Isola (2026) that pretrained models are surrounded by a dense population of task experts that can be discovered through random perturbations to the weights. The authors propose a simple statistical model of selection effects to study how the apparent advantage of selected perturbations may arise from more simple models than the “neural thickets” as presented in the original paper. They study how the advantage changes with the number of selection questions as well as transfer to held out data, and compare the roles of ensembling in achieving the results from Gan and Isola.

The authors argue that much of the apparent expert advantage can arise from selection noise, and can be recovered from sampling noise and pure ensembling, weakening the interpretation that random search can easily discover task specialists.

I consider this an important question and a valuable direction of work. The proposed statistical explanations are sensible, and I find the simple models useful for understanding how what looks like neural thickets could arise from empirical observations. However, I am not yet fully convinced that the empirical evidence supports the full breadth of the conclusions. In particular, unresolved reproduction differences between the papers, a change in the definition of expert density, and limited analysis of perturbation scale make it difficult to determine which original claims have actually been contradicted. I believe addressing these issues would make the paper substantially more convincing.

Critical strengths
  1. The paper asks whether striking gains from random weight search actually justify claims about the structure of pretrained models, which is clearly an important question that requires careful testing and understanding.
  2. The model presented makes clear why a large best-of-NN advantage can appear even when candidates have no persistent differences in expected performance. Its prediction that this advantage should shrink with the number of selection questions is a meaningful contribution.
  3. Held-out evaluation, matched answer-budget comparisons with anchor-model majority voting, repeated anchor evaluations, and controls for weight-restoration artifacts help distinguish the value of selecting weight perturbations from other potential sources of apparent improvement.
Critical weaknesses
  1. The paper reports that it cannot reproduce the approximately eight-percentage-point single-expert improvement originally reported for OLMo on Countdown by Gan and Isola. This is important, but it is not clear why this is the case, and I worry that this may also impact some of the other results presented. The original paper also includes held-out ‘rank-1’ results, including Figures 9 and 11. These results may need a more careful matching of model, scoring, decoding, data splits, and perturbation settings. The authors need not reproduce every original experiment, but their conclusions should reflect which results have been reproduced. E.g. if the gains on OLMO are not reproduced but the gains on Qwen are, wouldn’t it make more sense to have Figure 2 be the equivalent result but on the Qwen models? Instead it focuses on OLMo settings with little or modest transfer. Since the paper reports substantial held-out gains for Qwen-3B and Qwen-7B on Countdown, it might be a comparison worth adding. This would help establish where the null model ceases to explain the observations and how broadly the interpretation applies.
  2. As I understand, Figures 6 and 7 measure the fraction of perturbations exceeding a threshold relative to the perturbation population’s own mean. The original density analysis in Gan and Isola instead uses the unperturbed model as the reference. The new statistic is useful for studying variation among perturbations, but its relevance as a test of the original density claim is insufficiently motivated. Perturbations could become increasingly concentrated around their own mean while an increasing fraction outperform the base model. Consequently, the trend in Figure 7 can coexist with the original claim.
  3. Perturbation scale is not sufficiently separated from model size or selection effects. I would like to see Figure 3-style held-out results separately by scale, using comparable candidate budgets and including the original reported range. It seems possible that these neural thickets do exist but we aren’t seeing them on these plots because the scale chosen is too small.Figure 7 could also benefit from an analysis varying both model size and perturbation scale: larger models might require different, potentially larger, perturbations to reveal useful variation.
  4. The relationship to controls and qualifications already in the original paper needs sharper treatment The original paper explicitly discusses formatting effects and includes a test-time majority-voting baseline. The critique should clarify what its new evidence adds to those analyses and explain the large differences in baseline performance where possible.
Questions for authors

If the following questions are resolved, I am happy to increase my score, since I believe they will improve the paper.

  1. Why were you unable to reproduce the approximately eight-percentage-point single-expert improvement for OLMo on Countdown, and could this discrepancy also affect other results presented in the paper? Have the model, scoring, decoding, data splits, and perturbation settings been carefully matched when comparing against the original held-out results?
  2. Given that the OLMo gains are not reproduced but substantial held-out gains are observed for Qwen-3B and Qwen-7B on Countdown, would it make sense to add the equivalent of Figure 2 for these Qwen settings? Would this help establish where the null model explains the observations and how broadly your interpretation applies?
  3. Why is density relative to the perturbation population’s own mean the appropriate statistic for testing the original density claim, which uses the unperturbed model as its reference? Could the trend in Figure 7 coexist with the original claim, since perturbations could become more concentrated around their own mean while an increasing fraction outperform the base model? Maybe there is a way to make this more clear in the main text.
  4. Could you show Figure 3-style held-out results separately by perturbation scale, using comparable candidate budgets and including the original reported range? Is it possible that useful neural thickets are not visible because the chosen scale is too small? For Figure 7, could larger models require different, potentially larger, perturbations to reveal useful variation?
  5. Since the original paper already discusses formatting effects and includes a test-time majority-voting baseline, can you make clear in the paper what the new evidence adds to those analyses?
4 comments
AuthorsOpenReview ↗

Author response (pt 1)

We thank the reviewer for carefully reading the paper and for their detailed review. We are encouraged that the reviewer considers this an important question and finds the statistical models usefu, and we agree that the paper should more clearly justify changes in evaluation metrics, e.g., de-meaning using expert population rather than anchor.

W1 and Q1–Q2. Reproduction scope and the Qwen exceptions

We thank the reviewer for raising this important point. We made an extensive effort to match Gan & Isola’s experimental setup, using their official implementation, yet could not recover their headline OLMo Countdown expert gain. (Further detail below.) Furthermore, Gan & Isola’s released materials unfortunately do not include evaluation artifacts underlying their reported results. Given our efforts and the magnitude of the discrepancy, we regard this as a reproducibility issue with the original paper. Our findings, meanwhile, are directly inspectable: we provide experiment code, analysis code, and evaluation artifacts needed to regenerate every figure and table in our paper, through our official repository. We have just updated our submission with an anonymized version of the repository (https://anonymous.4open.science/r/RevisitingNeuralThickets-7C36/). We believe that our fully documented results serve as a more reliable reference point for the research community to study neural thickets than the original headline findings that we failed to reproduce.

For the direct OLMo Countdown reproduction, we matched virtually every setting that can be clearly recovered from the official repository:

  • Model and checkpoint. We use allenai/Olmo-3-7B-Instruct, pinned to revision 6e5971d9. This matches the named model in the original paper and released scripts, though the exact revision used for the published results is unavailable.
  • RandOpt implementation. We use the official repository at commit 536df0a, modified only to add logging. (The exact environment used for the original runs is unavailable.)
  • Prompt and scoring. We use the released Countdown prompt, answer-extraction procedure, correctness reward, and format reward.
  • Decoding. Expert rollouts use greedy decoding with temperature 0, seed 42, a 1,024-token limit, and bf16 precision, matching the paper and released driver.
  • Candidate population. The direct reproduction evaluates N=5,000 candidates generated using the released seed-42 procedure, matching the headline experimental budget.
  • Data split. Following Gan & Isola, we use the first 200 training questions for selection on all three tasks. We evaluate on 1,400 Countdown questions (source rows 200–1599), the official 1,319-question GSM8K test set, and the official 500-task MBPP test set.
  • Perturbation scales. We reproduce the released scripts’ mixture of {0.5, 1, 2} × 10^-3. The original paper’s Table 3 instead reports {1, 2, 3} × 10^-3; we note this discrepancy in footnote 4 on page 14. We have also performed separate analysis for each sigma-value population, as we describe below. In a nutshell, the same behavior we report in the paper appears separately at each sigma: a general absence of “true experts”, aside from on Qwen-3B and 7B Countdown.
  • Software. We use vLLM 0.16.0. After running our experiments, we noticed that Gan and Isola used vLLM 0.10.0. They suggest that newer vLLM versions may increase RandOpt accuracy, but our observations point to the contrary, i.e., we see little expert advantage in most settings. For OLMo Countdown, the rank-1 expert gains just 0.44 points over the anchor, and across all the top-50 ranked experts, the best achieves an advantage of just 2.6% on the test split. This falls well below the top-1 8% advantage reported in Gan & Isola.

We also corresponded directly with Gan & Isola, but were unable to identify the cause of the discrepancy, despite their assistance.

The reviewer is correct in noting that the Qwen Countdown results deserve greater prominence. However, we note that OLMo currently takes center stage for two reasons: it is Gan & Isola’s headline setting, and it represents the typical pattern in our experiments. At m=200, the selection-noise null predicts 76+% of the observed lift in 10 of 12 model–task settings. The two exceptions are precisely Qwen-3B and Qwen-7B on Countdown, with 31% and 53%.

We agree that the Qwen equivalent of Figure 2 displays distinct behavior, with a persistent expert advantage across selection set size (see below). We mention this Qwen advantage in the final paragraph of Section 3.1, but agree that the contrast deserves clearer emphasis in the revised manuscript.

ModelmObserved selection liftNoise nullTop-20 held-out gain
Qwen 3B20024.087.41+6.93
Qwen 3B80023.833.56+7.63
Qwen 7B20018.169.55+6.95
Qwen 7B80016.144.85+8.25
AuthorsOpenReview ↗

Author response (pt 2)

W2 and Q3. Density relative to the anchor

We agree with the reviewer’s criticism; alteration of the de-meaning process should be more adequately justified. Aside from our belief that population-based de-meaning is statistically cleaner than anchor-based de-meaning, a crucial component in our decision is that we observe little qualitative difference between the two methods. In particular, under either choice of reference, we find no consistent monotonic increase in density with model size.

Below we report results that demean using the mean of 5 greedy anchor evaluations, i.e., using Gan & Isola’s method. The tables report density levels for perturbations at \sigma = 1e-3, on m=200 questions, agreeing with the settings of Gan & Isola. (Note that the value of \sigma = 5e-3 stated in the original Neural Thicket paper happens to be a typo, as mentioned by the authors in their GitHub repo (https://github.com/sunrainyg/RandOpt/issues/5#issuecomment-4781029924).)

Countdown

ThresholdQwen 3BQwen 7BQwen 32B
90%60.3689.8472.08
95%51.9581.5148.08
100%43.7169.8725.56
105%35.8256.2110.68
110%28.9641.893.10

GSM8K

Threshold (τ)Qwen 3BQwen 7BQwen 32B
90%90.8699.8599.83
95%71.5994.8497.89
100%27.9133.5763.64
105%2.280.262.02
110%0.020.000.00

MBPP

Threshold (τ)Qwen 3BQwen 7BQwen 32B
90%93.5399.5999.96
95%77.2486.2286.71
100%44.7434.5013.19
105%14.633.990.06
110%2.400.150.00

We find that on Countdown, density peaks at Qwen-7B and falls at 32B at every threshold from 90% through 110%. GSM8K and MBPP show no consistent increase either: the pattern changes with the threshold and task. In the next version of the paper, we will add the complete anchor-relative analysis to the appendix and mention this comparison in the main text.

W3 and Q4. Perturbation scale

We agree with the reviewer that perturbation scale is a subtle and important point. As the reviewer correctly notes, most results in the paper aggregate over a population of N = 5000 (or N = 1000) perturbations for which each scale \sigma is drawn uniformly at random from {0.5, 1, 2}e-3. Though we made this decision intentionally to match the released RandOpt method of Gan & Isola, we acknowledge that it can somewhat obscure the precise role of sigma. (For this reason, we were mindful to present density-based results for the fixed choice of sigma=1e-3 in Figures 6 and 7.)

The table below reports the \sigma-stratified version of Figure 3; each cell records the mean held-out accuracy gain of the five highest-ranked individual experts over the anchor, in percentage points. The ± value is the standard deviation across 200 candidate subsets. Within each scale, we draw 1,000 candidates for the N = 5000 searches and 300 for the N = 1000 Qwen-32B search, then rank them on the same 200 selection questions. (There is one missing cell for which we lack the necessary held-out evaluations.)

Though distinct sigma scales can perform considerably differently for a given model–task pair, overall we find that the stratified results tell much the same story. Outside Qwen-3B and 7B on Countdown, the top five experts gain at most about 2 accuracy points on average at any evaluated scale, with several groups performing worse than the anchor.

Model–taskσ=0.5×10^-3σ=1×10^-3σ=2×10^-3
OLMo–Countdown+0.61 ± 0.25+1.07 ± 0.25-0.02 ± 0.29
OLMo–GSM8K+0.03 ± 0.15-0.49 ± 0.15-0.60 ± 0.22
OLMo–MBPP+1.43 ± 0.35+1.53 ± 0.68+1.66 ± 0.60
Qwen 3B–Countdown+1.33 ± 0.79+4.43 ± 1.41+7.54 ± 2.11
Qwen 3B–GSM8K+0.56 ± 0.31+0.36 ± 0.32-1.52 ± 0.32
Qwen 3B–MBPP+0.75 ± 0.34+1.07 ± 0.29-2.39 ± 0.44
Qwen 7B–Countdown+5.55 ± 0.47+6.98 ± 0.39+6.74 ± 1.11
Qwen 7B–GSM8K+0.17 ± 0.17-0.35 ± 0.13-1.21 ± 0.27
Qwen 7B–MBPP-0.46 ± 0.36-1.35 ± 0.38-2.29 ± 0.63
Qwen 32B–Countdown+0.22 ± 0.37+1.68 ± 0.19—
Qwen 32B–GSM8K+1.86 ± 0.08+2.01 ± 0.16+1.81 ± 0.06
Qwen 32B–MBPP-1.63 ± 0.15-1.63 ± 0.17-2.83 ± 0.11

Regarding Figure 7, we appreciate the suggestion to vary perturbation scale with model size. We note, however, that Gan and Isola use the fixed value sigma=1e-3 throughout their Section 2, including in their discovery of the “Scaling Law of the Solution Density” (their Figure 3). The original claim of Gan & Isola is therefore a fixed-sigma claim, which would itself be meaningful to challenge in its own right.

AuthorsOpenReview ↗

Author response (pt 3)

Nevertheless, it is worth carefully considering the relationship between perturbation scale and model size. From the previous table, we find that there is little task-independent relationship between model size and favorable perturbation scale. For instance, for Qwen-3B, raising sigma from 0.5e-3 to 2e-3 increases the top-five Countdown gain from +1.33 to +7.54, but consequently the GSM8K and MBPP gains turn negative. Qwen-32B’s GSM8K gain, meanwhile, stays near +2 points at all three scales. We thus see little evidence of a task-independent relationship between model size and a favorable sigma.

Regarding larger sigma values, we also tested the paper’s printed sigma=5e-3 before learning that it was a typo, and found it to be far too destructive for Qwen models. In particular, with 500 perturbations for each of Qwen-0.5B, 1.5B, 3B, and 32B on Countdown, GSM8K, and MBPP, we found that not a single perturbation beat the anchor on held-out data in 11 of the 12 settings, consistent with the reproduction concern raised on GitHub.

W4 and Q5. Relationship with Gan & Isola’s TT-MV Baseline

We thank the reviewer for raising this point. We begin by noting that our paper examines the broader question of how much random weight search adds to held-out performance, and what accounts for such improvement (when it exists).

In particular, our individual-expert evaluations provide a considerably more pessimistic picture than that of Gan & Isola (i.e., their figure 9). On Countdown, we find that the held-out gains of the top-ranking experts for Qwen 3B and 7B are driven primarily formatting. That is, 94% and 83%, respectively, of the performance improvement over the anchor model arises from changes in which questions receive valid expressions. The correctness difference on questions where both models produce valid expressions contributes only 0.74 and 1.91 accuracy points to their total gains of 12.16 and 11.41 points. Qwen-7B retains a smaller, statistically significant shared-question advantage.

Furthermore, our analysis of USPTO reveals a separate axis along which expert improvement can be deceptive. As USPTO has significant label imbalance, we find that top experts simply exploit such imbalance in order to outperform the anchor. Upon randomly reassigning category numbers independently for each question, the top-20 experts’ mean advantage fell from 5.1 to merely 1.1 accuracy points. (Meanwhile, anchor accuracy remains virtually unchanged under such shuffling.)

Together, we believe these experiments provide the community with a more thorough understanding of the mechanisms that can underlie expert improvements, showing how random search can favor benchmark-specific output tendencies rather than (considerably) improving problem-solving ability.

We thank the reviewer again for the careful and constructive review. We believe these changes will substantially improve the paper.

Dear Reviewer TNiM,

As the discussion period is ending in a few hours, could you please let us know if our response has addressed your concerns? We would be happy to provide further clarifications before the discussion period ends.

Best regards,

Authors of Submission 47017

Reviewer_Yn54OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper revisits the Neural Thicket hypothesis, the hypothesis that pretrained language models are surrounded by dense neighborhoods of task-specific experts that can be discovered through random Gaussian weight perturbations. The authors question whether the observed improvements of these selected perturbations reflect genuine task expertise or are largely explained by selection effects. To investigate this, they introduce a statistical null model that assumes no persistent differences in expected performance between perturbations and show that it predicts a substantial fraction of the observed selection-set improvements. They further show that performance advantages of the best perturbations on held-out diminish considerably. Finally they show that repeatedly sampling the anchor model is comparable in performance with the RandOpt's majority voting.

Critical strengths

S1. The paper investigates an interesting and timely question regarding the Neural Thicket hypothesis.

S2. The introduction of a selection-noise null model provides a simple but insightful way to investigate the performance improvements of top-ranked perturbations.

Critical weaknesses

W1. The main experiments do not appear to be repeated across independently generated perturbation populations. Given the stochastic nature of RandOpt, the selected experts and their performance could depend on the particular random perturbations sampled. Doing the repeated perturbations, and reporting mean and standard deviations would strengthen the reliability of the experimental findings.

W2. The paper would benefit from clearer organization and more guidance throughout the experimental sections. The authors move directly into reporting experimental findings without first clearly introducing the motivation, setup, or purpose of the analyses. Section 3.4 provides a better example, so applying this structure throughout the paper would improve readability.

Questions for authors

Q1. What motivates the choice of ensemble sizes (K∈{10,20,50}K \in \{ 10,20,50\}) in Section 3.3?

Q2. Is it possible to put a concise version of the related-work discussion into the main text?

5 comments
AuthorsOpenReview ↗

Author response

We thank the reviewer for carefully reading the paper and for their detailed, constructive review. We are encouraged that the reviewer found the question interesting and timely, and viewed the selection-noise null model as a simple but insightful contribution.

W1. Sensitivity to the sampled perturbation population

We appreciate this point and agree that sensitivity to the sampled perturbation population should be quantified. We’ve performed 2,000 bootstrap resamples of each candidate population (drawn with replacement). The table below reports the bootstrap mean and std deviation of the top-20 experts’ mean held-out accuracy relative to the mean of five greedy anchor evaluations, in percentage points.

ModelNCountdownGSM8KMBPP
OLMo-3-7B-Instruct5,000+0.72 ± 0.23-0.47 ± 0.15+1.39 ± 0.45
Qwen2.5-3B-Instruct5,000+5.16 ± 1.26+0.26 ± 0.22+0.65 ± 0.31
Qwen2.5-7B-Instruct5,000+7.28 ± 0.48-0.09 ± 0.14-1.05 ± 0.27
Qwen2.5-32B-Instruct1,000+0.81 ± 0.38+1.93 ± 0.13-2.06 ± 0.23

We find our conclusions to be stable: Aside from the smaller Qwen 3B and 7B models evaluated on Countdown, the held-out performance of top perturbations is generally mixed and modest. Furthermore, across the twelve model-task pairs, the bootstrap standard deviation of the rank-1 selection lift ranges from only 0.18 to 0.70 reward points. We acknowledge that this bootstrap analysis is inherently limited, as it does not expand the original pool of perturbations we consider in the paper.

However, in most of our experiments, RandOpt itself searches over a considerable population of N=5,000 perturbations. Repeating every model–task search several times would therefore require evaluating tens of thousands of additional perturbed models. Moreover, a user of the proposed method would ordinarily perform this search only once. Thus, even if our conclusions were driven by one “unlucky” 5,000-perturbation pool, that would itself be a notable practical finding about the reliability of RandOpt. (And our bootstrap analysis suggests that this is not the case.)

W2. Organization and motivation of the experimental sections

We apologize that the experimental sections were difficult to follow. We acknowledge that there is room for improvement in the exposition, particularly in introducing the motivation and purpose of each analysis before presenting its results.

In the revision, we will reorganize Sections 3.1–3.3 to more closely follow the style and structure of Section 3.4. We will also add a short roadmap at the beginning of Section 3 to clarify how the analyses fit together.

Q1. What motivates the ensemble sizes in Section 3.3?

The values K={10,20,50} were chosen arbitrarily for concision; we find that a denser sweep over K reveals the same behavior described in the paper. We will include these more granular results in the revision.

Countdown

KOLMo 7BQwen 3BQwen 7BQwen 32B
1+0.69+10.76+7.65+2.60
5-1.66+11.54+3.76+1.85
10-1.57+7.01+2.17+1.50
15-1.08+4.68+2.76+2.04
20-1.24+2.57+2.73+2.27
25-0.89+0.22+3.03+2.74
30-0.72+0.75+3.06+2.22
35-0.79+0.57+2.96+2.12
40-0.75-0.43+2.36+2.14
45-0.52-0.91+2.53+2.12
50-0.71-0.84+2.25+1.86

GSM8K

KOLMo 7BQwen 3BQwen 7BQwen 32B
1-1.25+1.89+0.39+1.62
5+0.19+0.19-0.61+0.40
10+0.41-0.71-0.84+0.03
15+0.28-0.27-0.38+0.20
20+0.30+0.00-0.36+0.24
25+0.43-0.21-0.21+0.27
30+0.77-0.48-0.24+0.08
35+0.66-0.32-0.42+0.19
40+0.61-0.31-0.53+0.23
45+0.59-0.45-0.69+0.27
50+0.56-0.61-0.28+0.32

MBPP

KOLMo 7BQwen 3BQwen 7BQwen 32B
1+2.22+0.75-1.65-0.78
5+0.75+0.73-2.06-1.23
10+0.87+0.66-2.12-0.84
15+0.80+0.60-1.08+0.58
20+0.72-0.03-0.87+0.85
25+0.85+0.13-1.08+0.97
30+0.42+0.49-1.18+0.68
35+0.54+1.27-1.00-0.14
40+1.07+1.15-1.47+0.63
45+0.52+0.49-1.74-0.28
50+0.53+0.87-1.94+0.09

Q2. Concise related work in the main text

Yes, we agree this is a shortcoming in the current version of the paper. Should the paper be accepted, we will use the additional camera-ready page to place an abbreviated related-work discussion immediately after the introduction, while retaining the more complete treatment in the appendix.

We thank the reviewer again for their constructive comments. We believe these changes will substantially improve the paper’s clarity and strengthen its presentation.

Thank you for the rebuttal and additional experiments. I appreciate the author's efforts to address my concerns.

The bootstrap analysis provides useful evidence about the stability of the selected experts. However, it does not fully address W1, as resampling from an existing candidate population is not equivalent to repeating RandOpt with independently generated perturbations. The sensitivity of the results to different perturbation populations therefore remains unclear.

For this reason, I will maintain my score.

I would also like to ask whether the authors plan to upload a revised manuscript during the discussion period that incorporates the changes mentioned in their rebuttal.

AuthorsOpenReview ↗

Author response

Thank you for the follow-up! We address each point below.

Sensitivity to perturbation population

We have now compared disjoint populations of independently generated perturbations and find that our selected-expert conclusions are already stable at substantially smaller population sizes. Across all nine settings with N=5000N = 5000 candidates, the mean held-out gain at every smaller size is within 0.40.4 percentage points of the full-5k-search result.

This evidence matters because repeating these searches is computationally expensive. For OLMo-7B alone, another 5,0005,000-candidate sweep with evaluations across the three tasks costs approximately 500 GPU hours on H200-class GPUs. Repeating the searches across all models would require much more. Therefore, we examined whether the independent perturbations already evaluated provide consistent evidence across perturbation population sizes before committing to additional sweeps with an indefinite ceiling on the compute.

How we compare independent populations. Recall that each perturbation is generated independently from the same anchor model, whose weights are restored exactly after every evaluation. So splitting the 5k candidates into, for instance, five disjoint groups of 1000 would give five independent smaller candidate populations. We then select experts separately within each group using the same 200 selection questions.

We repeat this comparison at population sizes of 250, 500, 1000, and 2500. To keep selection equally stringent, we retain the same fraction of candidates at each size: the top 1, 2, 4, and 10, respectively, compared with the top 20 from the full population of 5000. The table below reports their mean individual held-out accuracy relative to the mean of five greedy anchor-model evaluations, in percentage points.

Model / task250 candidates, top 1500 candidates, top 21,000 candidates, top 42,500 candidates, top 105,000 candidates, top 20
OLMo-7B / Countdown+0.59±0.92+0.59 \pm 0.92+0.64±0.66+0.64 \pm 0.66+0.66±0.49+0.66 \pm 0.49+0.72±0.31+0.72 \pm 0.31+0.73+0.73
OLMo-7B / GSM8K−0.50±0.68-0.50 \pm 0.68−0.47±0.47-0.47 \pm 0.47−0.46±0.33-0.46 \pm 0.33−0.48±0.21-0.48 \pm 0.21−0.45-0.45
OLMo-7B / MBPP+1.28±1.80+1.28 \pm 1.80+1.31±1.32+1.31 \pm 1.32+1.32±0.96+1.32 \pm 0.96+1.38±0.61+1.38 \pm 0.61+1.53+1.53
Qwen-3B / Countdown+5.11±4.86+5.11 \pm 4.86+5.32±3.72+5.32 \pm 3.72+5.26±2.67+5.26 \pm 2.67+5.13±1.79+5.13 \pm 1.79+5.12+5.12
Qwen-3B / GSM8K+0.26±0.99+0.26 \pm 0.99+0.27±0.69+0.27 \pm 0.69+0.27±0.47+0.27 \pm 0.47+0.26±0.30+0.26 \pm 0.30+0.24+0.24
Qwen-3B / MBPP+0.80±1.25+0.80 \pm 1.25+0.71±0.90+0.71 \pm 0.90+0.66±0.66+0.66 \pm 0.66+0.63±0.48+0.63 \pm 0.48+0.62+0.62
Qwen-7B / Countdown+6.86±2.19+6.86 \pm 2.19+7.06±1.49+7.06 \pm 1.49+7.22±1.01+7.22 \pm 1.01+7.29±0.65+7.29 \pm 0.65+7.23+7.23
Qwen-7B / GSM8K−0.05±0.65-0.05 \pm 0.65−0.05±0.44-0.05 \pm 0.44−0.06±0.31-0.06 \pm 0.31−0.09±0.20-0.09 \pm 0.20−0.11-0.11
Qwen-7B / MBPP−1.01±1.14-1.01 \pm 1.14−1.06±0.83-1.06 \pm 0.83−1.08±0.59-1.08 \pm 0.59−1.05±0.36-1.05 \pm 0.36−1.03-1.03

Here, the first four columns report the mean ±\pm standard deviation across disjoint groups, with results averaged over 200200 random partitions at each scale.

The average gains remain stable while the variability decreases. For example, OLMo's Countdown gain stays between +0.59+0.59 and +0.73+0.73 points, while its across-population standard deviation falls from 0.920.92 points at 250250 candidates to 0.310.31 at 2500. Qwen-3B's Countdown gain remains approximately +5+5 points throughout. Therefore, the larger populations reproduce the average effects we already observe in smaller populations, just with less variability.

Furthermore, we can see that across the population sizes, the standard deviations decrease approximately as SD(n)∝1/n\mathrm{SD}(n) \propto 1/\sqrt{n}, where nn is the candidate population size. Extending this observed trend to N=5,000N = 5{,}000 gives estimated standard deviations of 0.140.14-0.420.42 points in the seven settings outside Qwen-3B/7B Countdown, and 1.181.18 and 0.470.47 points for those two exceptions. These already agree closely with our bootstrap estimates in our first response to you: the largest difference is 0.080.08 points. We will label the full-size estimates explicitly as extrapolations from the measured smaller-population variability.

In sum, given the consistency of the evidence already available, we believe the current experiments and this additional analysis provide strong support for the reliability of this conclusion.

Revised manuscript

We have now uploaded a revised manuscript incorporating the population-sensitivity analyses and denser ensemble-size sweep in the appendix, with brief references in the main text. We have also expanded the opening of Section 3 to explain the purpose of each analysis. Given the current page limit, we would use the additional camera-ready page to include a concise related-work discussion in the main text, should the paper be accepted.

Reviewer_nVqfOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits neural thickets, which are overparameterized ensembles trained jointly on multiple tasks, and asks whether simple random search over their subnetworks can uncover task-specialized experts. The authors evaluate random subnetwork selection across several multi-task settings and compare it with structured pruning, learned gating, and single-task fine-tuning. They report that random search frequently finds subnetworks that outperform the full thicket on individual tasks and sometimes match or exceed more expensive expert discovery methods. The paper argues that this effect arises from strong implicit specialization within overparameterized multi-task networks and that random search is a surprisingly effective probe of that structure.

Critical strengths
  1. The paper provides useful comparisons with structured pruning and learned gating, showing that random search can be competitive despite its simplicity.
  2. The analysis highlights that overparameterized multi-task training can encode substantial implicit specialization, which is relevant to mixture-of-experts and multi-task learning research.
Critical weaknesses
  1. The paper does not fully explain why random search works well in some settings but not others, so the practical guidance for choosing pruning ratios or search budgets remains limited.
  2. It is unclear whether random search remains competitive when gating networks are well tuned or when compute budgets are matched more carefully.
Questions for authors

How does the required search budget scale with model size, task diversity, and pruning sparsity?

Comments on the paper

1 comment
Area_Chair_BS2wOpenReview ↗

Participate in the discussion period

Dear reviewers,

The authors have replied to your comments and the discussion is close to the end. We encourage you to participate in the discussion to see whether your questions and concerns have been answered by the authors. Thank you!

Best, Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.