AI models often need adaptation to new tasks, but retraining is costly. Researchers previously thought randomly tweaking a model's internal weights could unlock specialized experts. This paper shows that the gains come from output diversity, not weight tweaks. The authors applied random biases to the model's output scores (logits) while keeping the model frozen. They selected the best biased versions using a small validation set and aggregated their predictions through voting. This logit-only approach achieved the same benefits. Logit PSV improved accuracy on a math test from 63.3% to 81.3%, exceeding the weight-tweaking method. It is simpler, training-free, and works across language and vision domains.
Language models gain from varied answers, not necessarily altered weights
Dufour and colleagues show that changing output scores, selecting useful variants and combining their answers can match or exceed ensembles built by perturbing model weights.
Big Tech
Nicolas Dufour · Elouan Gardès · Amil Dravid · Patrick Pérez · Loic Landrieu
Kyutai · Meta FAIR · LIGM · CNRS · Univ Gustave Eiffel
Research Digest··2 min read
In one line
Output diversity, not weight-space experts, explains the gains from random perturbations in pretrained models.
Before you read this
An idea this paper builds onRead it as
Randomly changing a language model’s weights can improve reasoning when selected variants vote on an answer.
The authors compared perturbing an entire model’s weights, perturbing a single decoder block and leaving all weights unchanged while altering output logits, the scores used to choose the next token.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.