Language-model gains may come from diversity, not hidden experts

Small changes can make a language model give different answers. The open question is what those changes actually reveal.

What the study found

Dufour and colleagues revisit the Neural Thickets hypothesis: the idea that a pretrained language model sits near many task-specific experts in weight space, which random weight perturbations can uncover. The authors say they reproduce the reported gains, but find that changing a single decoder block recovers much of the improvement.

They also report that leaving the model frozen and adding random biases to output logits matches or exceeds full-weight perturbations on most reasoning tasks they studied. Their proposed Perturb-Select-Vote procedure generates varied predictors, selects strong candidates using a small validation set, and combines their predictions.

Why it matters

The result offers an alternative explanation: diversity, selection and voting may account for gains previously attributed to nearby experts. The authors connect that interpretation to existing inference-time methods. They also report improved vision performance under domain shift, including Earth observation.

Limits and unknowns

This is an arXiv preprint, not a settled finding. This briefing uses the abstract alone: it does not identify every model, dataset or evaluation condition, or quantify the reported gains. The results should not be assumed to apply to every task or model.

Read the research

Authors: Nicolas Dufour, Elouan Gardès, Amil Dravid, Patrick Pérez, Loic Landrieu

Published: 2026-10-29

Source basis: Based on the abstract only

Original abstract

The Neural Thickets hypothesis posits that pretrained language models lie in dense neighborhoods of task-specific experts in weight space, which random Gaussian perturbations can uncover. We reproduce the reported gains, but show that they do not require this geometric explanation. Perturbing a single decoder block already recovers much of the improvement, while keeping the model entirely frozen and adding a random bias to its output logits matches or exceeds full-weight perturbations on most reasoning tasks. This suggests an alternative explanation based on output diversity, validation-based selection, and voting. We formalize this view as Perturb-Select-Vote (PSV), a general procedure that generates perturbed predictors, selects high-performing members on a small validation set, and aggregates their predictions. This interpretation connects Neural Thickets to established inference-time methods, including sampling, self-consistency, and verifier-guided selection. A first-order analysis further shows that small weight perturbations act primarily through their induced logit shifts. We extend this analysis to vision, where logit PSV substantially improves performance under domain shift, including on Earth observation, while full-weight perturbations provide little additional benefit. Together, these results suggest that the gains attributed to nearby weight-space experts can instead be obtained through output diversity, selection, and voting. Logit PSV realizes these benefits with a fully frozen model, providing a simple, training-free alternative to full-weight perturbation ensembles across language and vision.

This article is an AI-generated adaptation, not a substitute for the original research. Preprints may not have been peer reviewed; AI can make mistakes.