Skip to content
AI-generated research idea🟠Medium🔬Significant ResearchNo close match found

Beyond Majority Voting: Ensembling Continuous Structured Predictions

Implement a representation-space ensembling method using hidden-state interpolation or a reward model verifier to execute Best-of-K selection for continuous structured predictions.

Time

8 hours

Compute level

Large cluster

Resources

2 nodes of 8x NVIDIA GH200 GPUs for 4 hours to evaluate representation interpolation and reward-model scoring over thousands of perturbations.

What the paper used

Stable Diffusion XL model evaluated on text-to-image generation tasks using mean ensembling across K models at each denoising step.

Prerequisites

basic ML knowledge

Scope

new model architecture

The Gap

In Section 11, the authors note that the majority-vote ensembling in Algorithm 1 works well for discrete classes or integers, but does not support complex structured predictions.

“We have focused primarily on problems where the answer is a single discrete class (or an integer), in which case majority-vote ensembling is straightforward to apply.”
From the paper, Section 11, page 14. Matched against the PDF text.

Limitation

The RandOpt algorithm cannot natively ensemble answers for open-ended text, molecular SMILES strings, or complex Python code where identical output strings are rarely generated by independent perturbations.

Proposed Solution

Implement a generalized representation-space ensembling method, such as interpolating hidden states or utilizing a verifier/reward model to execute a Best-of-K selection at inference time, replacing the strict majority voting mechanism.

Why This Matters

Extending RandOpt to successfully handle structured and continuous outputs unlocks scalable, gradient-free post-training for code generation, conversational AI, and computational biology.

How to Get Started

How to test it: Hypothesis: Generalized representation-space ensembling (such as hidden state interpolation or reward-model-guided Best-of-K) outperforms strict majority voting on open-ended text and code generation tasks. Baseline: Standard majority voting and base model generation. Metric: Task-specific pass rate and BLEU/CodeBLEU score. Stopping condition: If representation interpolation fails to surpass the base model score on open-ended generation benchmarks.

  1. Modify the inference pipeline to retain multiple candidate continuous structured outputs from the base model instead of collapsing via majority vote.
  2. Train or integrate a verifier or reward model capable of scoring continuous structured predictions.
  3. Implement hidden-state interpolation and Best-of-K selection algorithms to aggregate candidate predictions.
  4. Benchmark the generalized ensembling method against strict majority voting on tasks with continuous outputs.

Code linked to this paper on Hugging Face: github.com/sunrainyg/RandOpt

Suggested by AI. Adapt it to your setup.

Has This Been Done?

No close match in 10 results

The listed works such as Cluster Ensembles, Majority Vote, Voter Eligibility and Privileged Voters and Soft voting ensemble classifier for liquefaction prediction discuss traditional voting methods, classification ensembles, and specific domain applications. None of the provided results implement a generalized representation-space ensembling method or Best-of-K selection to replace majority voting for continuous structured predictions.

We searched OpenAlex and arXiv for “Implement a representation-space ensembling method using hidden-state interpolation or a reward model verifier to execute Best-of-K selection for continuous structured predictions.” and “Beyond Majority Voting: Ensembling Continuous Structured Predictions” and an AI compared the idea with the top 10 results. A small search can miss relevant work.

Source Paper

Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights

Yulu Gan, Phillip Isola · 2026

This paper reveals that large pretrained models exist in a 'thicket' regime, where the local neighborhood around their weights is densely populated with diverse, task-specific experts. Based on this finding, the authors introduce RandOpt, a gradient-free post-training method that samples random weight perturbations and ensembles the best ones via majority voting to achieve competitive performance against standard RL methods.

Related Ideas

Is this a good idea?

Share onXLinkedIn