You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$15,000 left to reach the funding goal.
Looped models (such as Ouro from ByteDance) repeat layers for each token. This is referred to as depth. Reasoning models will also use compute to produce longer reasoning chains (referred to as length).
Given a fixed compute budget, which is optimal to invest in more loops or tokens? This could be problem specific.
Smoke tests are complete, full study is currently running on free GPU's within the notebook. Funding is required to run this study to completion reliably.
What I found so far (Ouro-2.6B on AIME problems, small samples):
Depth outperformed breadth when the total number of tokens was the same. A depth-4 with a token limit of 16k outperformed a depth-2 with a token limit of 32k by +0.20 (95% CI: +0.08 to +0.33) on 30 problems using two samples each.
Halving the depth caused the model to be verbose. Reducing the depth to 2 from 4 during an answer increased the proportion of answers which hit their maximum token limit nearly 2 fold (23% to 43%).
The choice of depth is made upfront. Increasing the depth after 1k tokens improved performance (+0.20), but increasing the depth after 4k tokens barely improves performance.
There is scope to reduce computation by varying the depth based on the problem. An ideal selector would have the same accuracy as a constant deep setting but with approximately 25% less computations.
Plan:
Finish the full diagnostic: 7 models (Ouro 1.4B/2.6B, TaH2 variants, Qwen3 baseline), 6 benchmarks (AIME, HMMT, MATH-500, GPQA, LiveCodeBench, HumanEval+).
A controlled study of the “decide early” effect, with same-depth controls.
A method that picks both depth and token budget per problem, tested against depth-only and length-only baselines at equal compute.
Release all reasoning traces and code publicly, post an arXiv preprint, and submit to A* conference.
Why it matters for AI safety: if extra loops replace written reasoning, part of the model’s thinking moves somewhere humans can’t read. My data measures how many visible tokens each loop replaces.
100% GPU rental: RTX PRO 6000 at about $1.69/hour, the same GPU my results so far come from and pay small amount of money from my pocket to finish the smoke test.
$1,500 minimum (~880 GPU-hours): finish the core diagnostic (~400 h), the “decide early” study (~100 h), method training data (~150 h), and method evaluation (~200 h).
Up to $15,000 (~5,900 GPU-hours): all of the above, plus:
a second looped architecture, Huginn-3.5B (~1,500 h)
full 32k-token limits and more samples for tighter results (~1,000 h)
method training and ablations on H100s (~1,500 h)
a 20% buffer for experiments reviewers ask for
Just me: MD Azizul Hakim, independent researcher in Bangladesh, working without a lab or supervisor.
NeurIPS 2026 main-track poster (paper #2592), on training quality and test-time reasoning efficiency.
Website: https://logicsame.github.io/md-hakim.github.io/
LinkedIn: https://www.linkedin.com/in/md-azizul-hakim-0b2635359/
The per-problem method doesn’t beat depth-only routing. Recent work (RecurTrace) routes depth per problem, but only for short answers. Then the paper becomes a diagnostic paper, which is still publishable.
The findings only hold for Ouro, not other looped models. I would report that honestly as a limitation.
Either way, the full trace dataset (thousands of reasoning traces at depths 1–4) gets released and is useful to other researchers.
There are no bids on this project.