The BasePeople

Paweł Olszowiec

Samsung AI Center, Warsaw, Poland

Titles as the sources gave them. In 1 entry, with 5 checked quotations.

What they said

Newest first. Each quotation was found word for word in the source.

“Averaged over the four generative sets, 76-89% of answer tokens are decided at the first iteration in Ouro, Parcae, and Rec-OLMo, and 90% are decided by iteration 2-3 on most sets (iteration 5 at the latest).”
Monitoring latent reasoning in looped language modelsarXiv, October 13, 2026. Speaking as: Other. Section 4.2: an answer token is 'decided' at the first iteration after which its top prediction never changes; results averaged over four prompt sets.
“More iterations are not always better. Base Ouro predicts the prompt best at iteration 3-4, and gets worse thereafter. The reasoning-tuned Ouro-Thinking models are best at iteration 1 and get worse with every iteration. Only Huginn improves steadily”
Monitoring latent reasoning in looped language modelsarXiv, October 13, 2026. Speaking as: Other. Introduction summary. 'Predicts the prompt' means how well each iteration guesses the next word of the prompt, a reading test, not a measure of solving problems.
“Noise four times the norm of the entry state reduces every carry model to broken or unrelated text. On the other hand, noise fifty times the norm leaves re-read models largely unaffected”
Monitoring latent reasoning in looped language modelsarXiv, October 13, 2026. Speaking as: Other. Introduction summary of the noise test: random noise added to the starting state, scaled to its size; 'carry' loops get the prompt once, 're-read' loops at every iteration.
“Of its 73 flipped answers … only 12 follow the harmful request, 24 answer a harmless misread version (one word swapped), and 25 open a reasoning trace that never answers.”
Monitoring latent reasoning in looped language modelsarXiv, October 13, 2026. Speaking as: Other. Section 4.5: Ouro-1.4B with noise twice its entry state's size; of 300 safety prompts, 73 clean refusals turned into non-refusals, which these figures break down.
“We use isotropic Gaussian noise as a simple, unstructured perturbation to investigate how the loop responds to changes in its hidden state, independently of its semantic content.”
Monitoring latent reasoning in looped language modelsarXiv, October 13, 2026. Speaking as: Other. Section 4.5: how the noise test works. It is random noise, not a crafted attack.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

    Monitoring latent reasoning in looped language models

    Researchers at Samsung AI Center, Warsaw, apply five tests of internal states to eight open 'looped' language models, which reuse one block of layers. They report that six of the eight fix most answer tokens on the first pass, that extra loops can hurt, and that random noise breaks one design ('carry' loops) but not the other.