Depth-gated abstention in a looped language model

Version 2 (adds the analyses asked for in review: what "None" replaced, the unreadable share in training, precision and recall, d' by depth, a bootstrap over original questions and seeds, a per-seed split, and a figure). We trained a looped language model, Ouro-1.4B, which applies its 24 layers four times per token, and a plain model of similar size, Qwen3-1.7B-Base, with identical reinforcement learning from checked answers (GRPO, LoRA, 100 steps of GSM8K practice, three seeds), and tested both on 400 GSM-Plus twin pairs, where a pair counts only when both answers are right. Ouro started ahead: 228 pairs against 180, and 13 against 3 answers of "None" on the 50 questions that have a needed fact removed. After training, Ouro reached 283-287 pairs and Qwen 209-224; Ouro's extra gain, 4.7 points, is not separable from zero (95% hierarchical bootstrap over seeds and questions, -0.9 to 10.3). In every seed, 13-14 pairs of the gap come from the questions with a missing fact, where trained Ouro answered "None" on 28, 28 and 34 of 50 and trained Qwen on 7, 5 and 8, although no practice question was unanswerable. The switch is not a reward artefact: everywhere trained Ouro says "None", the untrained model had given a readable wrong number, which the reward scores the same as "None". In Ouro the behaviour depends on depth. At one loop no adapter answered "None" at all, even on questions whose answerable version it still solved; at two loops the same adapters did so on 18-25 of the 50; and the ability to tell unanswerable from answerable questions (d') peaks at the trained depth of four and falls at eight. Ouro's head start means our data cannot separate looping from pretraining; the note names the controls that would. Released: the code exactly as it produced the results, the per-question outputs, the training logs, and stats.py and figure.py, which recompute every number and the figure (code-and-results-v2.tar.gz).

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24240092
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Depth-gated abstention in a looped language model

Mark Hutto
Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
preprint

Depth-gated abstention in a looped language model

Mark Hutto
preprint en

Abstract

Version 2 (adds the analyses asked for in review: what "None" replaced, the unreadable share in training, precision and recall, d' by depth, a bootstrap over original questions and seeds, a per-seed split, and a figure). We trained a looped language model, Ouro-1.4B, which applies its 24 layers four times per token, and a plain model of similar size, Qwen3-1.7B-Base, with identical reinforcement learning from checked answers (GRPO, LoRA, 100 steps of GSM8K practice, three seeds), and tested both on 400 GSM-Plus twin pairs, where a pair counts only when both answers are right. Ouro started ahead: 228 pairs against 180, and 13 against 3 answers of "None" on the 50 questions that have a needed fact removed. After training, Ouro reached 283-287 pairs and Qwen 209-224; Ouro's extra gain, 4.7 points, is not separable from zero (95% hierarchical bootstrap over seeds and questions, -0.9 to 10.3). In every seed, 13-14 pairs of the gap come from the questions with a missing fact, where trained Ouro answered "None" on 28, 28 and 34 of 50 and trained Qwen on 7, 5 and 8, although no practice question was unanswerable. The switch is not a reward artefact: everywhere trained Ouro says "None", the untrained model had given a readable wrong number, which the reward scores the same as "None". In Ouro the behaviour depends on depth. At one loop no adapter answered "None" at all, even on questions whose answerable version it still solved; at two loops the same adapters did so on 18-25 of the 50; and the ability to tell unanswerable from answerable questions (d') peaks at the trained depth of four and falls at eight. Ouro's head start means our data cannot separate looping from pretraining; the note names the controls that would. Released: the code exactly as it produced the results, the per-question outputs, the training logs, and stats.py and figure.py, which recompute every number and the figure (code-and-results-v2.tar.gz).

Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Depth-gated abstention in a looped language model — Mark Hutto · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS