UpLoop: Training Looped Transformers via Upcycling at No Extra Pretraining Compute
Organizations: Shanghai Jiaotong University · University of Science and Technology of China
Abstract
Looped language models can strengthen reasoning through repeated computation with shared weights, increasing execution depth without adding independent parameters. However, each additional pass increases training floating-point operations (FLOPs), making these gains more expensive to obtain. To address this cost, we propose UpLoop, a pretraining schedule that uses upcycling to train looped Transformers while exactly matching a standard dense baseline in final parameter count, total training tokens, and training FLOPs. Specifically, UpLoop begins with a shallow Transformer, progressively grows its depth by copying learned layers, and uses the computation saved early to support repeated execution of a shared middle block later in training. Experiments across model scales show improvements in language modeling and downstream performance, with the average score across 22 downstream tasks increasing from 22.99 to 27.52 at 1.3B parameters. These gains also extend to mathematical reasoning after supervised fine-tuning.
Figures & tables
| Scale | Model | OBQA | WinoG. | ARC-c | ARC-e | HellaSwag | SIQA | PIQA | BoolQ | Average |
|---|---|---|---|---|---|---|---|---|---|---|
| 370M | Vanilla | 20.80 † | 52.01 † | 28.84 | 62.75 | 35.02 | 38.69 | 67.90 | 57.37 | 48.43 |
| UpLoop | 23.80 † | 53.28 † | 29.01 | 63.22 | 36.02 | 40.17 | 68.66 | 61.62 | 49.78 | |
| 760M | Vanilla | 26.80 † | 55.01 | 32.17 | 68.06 | 39.67 | 39.66 | 70.78 | 58.35 | 51.96 |
| UpLoop | 26.20 † | 55.17 | 32.85 | 69.57 | 41.58 | 42.07 | 70.89 | 62.51 | 53.52 | |
| 1.3B | Vanilla | 28.40 † | 56.51 | 37.54 | 73.36 | 43.21 | 40.84 | 73.12 | 59.76 | 54.91 |
| UpLoop | 30.40 † | 57.85 | 37.20 | 73.23 | 44.84 | 42.37 | 73.72 | 66.06 | 56.47 |
| 370M | 760M | 1.3B | ||||
| Task | Vanilla | UpLoop | Vanilla | UpLoop | Vanilla | UpLoop |
| Closed-book question answering | ||||||
| TriviaQA | 13.36 | 13.72 | 20.75 | 20.99 | 22.24 | 20.79 |
| TyDiQA-NoContext | 8.66 | 8.81 | 13.36 | 12.39 | 12.71 | 12.84 |
| NaturalQuestions | 5.10 | 5.53 | 7.93 | 8.69 | 9.43 | 9.04 |
| WebQuestions | 10.85 | 11.17 | 14.67 | 16.76 | 17.13 | 15.00 |
| Model | Params. | Tokens | FLOPs | NLL |
|---|---|---|---|---|
| Vanilla | 371M | 2.78 | ||
| Loop from scratch | 371M | 2.76 | ||
| Loop with dense initialization | 371M | 2.76 | ||
| Dense-20 (Iso-FLOPs) | 438M | 2.76 | ||
| Dense-20 (Iso-Tokens) | 438M | 2.74 | ||
| LIDAS (Iso-FLOPs) ( Kapl et al., 2025 ) | 371M | 2.75 |
| Model | Params. | Zero-shot (%) | Five-shot (%) |
|---|---|---|---|
| Vanilla | 1.3B | 8.72 | 5.46 |
| UpLoop | 1.3B | 12.05 | 8.11 |
| Model | Normal 2 calls | Extended 8 calls |
|---|---|---|
| UpLoop | 0.867 | 1.906 |
| Loop with dense initialization | 0.933 | 2.965 |
| Loop from scratch | 0.923 | 2.644 |
| Scale | Native | Control | Target | Drop (pp) |
|---|---|---|---|---|
| 760M | 84.8 | 86.3 | 77.7 | 8.6 |
| 1.3B | 89.1 | 89.5 | 75.0 | 14.5 |
| Scale | Vanilla | UpLoop | |
|---|---|---|---|
| 370M | 2.7770 | 2.7430 | |
| 760M | 2.5770 | 2.5381 | |
| 1.3B | 2.4332 | 2.3982 |
| Scale | Vanilla | UpLoop | Ratio |
|---|---|---|---|
| 370M | |||
| 760M | |||
| 1.3B |
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Category ( valid) | Vanilla | UpLoop | |
|---|---|---|---|
| Closed-book Q&A (4) | 9.49 | 9.81 | +0.32 |
| Open-book Q&A (4) | 18.20 | 19.51 | +1.32 |
| LAMBADA (1) | 34.02 | 34.58 | +0.56 |
| HellaSwag (1) | 42.60 | 44.93 | +2.33 |
| Math word (1) | 1.97 | 2.52 | +0.55 |
| Reas. primitives (2) | 28.50 | 36.20 | +7.70 |
| Benchmark | Vanilla | UpLoop | |
|---|---|---|---|
| Closed-book question answering | |||
| TriviaQA | 13.36 | 13.72 | +0.36 |
| TyDiQA-NoContext | 8.66 | 8.81 | +0.15 |
| NaturalQuestions | 5.10 | 5.53 | +0.43 |
| WebQuestions | 10.85 | 11.17 | +0.32 |
| Open-book question answering | |||
| Model | Attention | MLP |
|---|---|---|
| UpLoop | 0.00919 | 0.01929 |
| Loop with dense initialization | 0.01427 | 0.03426 |
| Loop from scratch | 0.02840 | 0.05341 |
| Scale | Task | Vanilla | UpLoop |
|---|---|---|---|
| 760M | Math | 39.8 | 74.2 |
| Code | 45.3 | 81.2 | |
| Average | 42.6 | 77.7 | |
| 1.3B | Math | 34.4 | 81.2 |
| Code | 57.0 | 85.2 | |
| Average | 45.7 | 83.2 |
| Task | Core 1 (pp) | Core 2 (pp) |
|---|---|---|
| Math | 0.0 | 14.1 |
| Code | 0.0 | 14.8 |
| Average | 0.0 | 14.5 |
Efficient Language Model Training
Efficient language model training studies how to reduce the data, compute, memory, time, energy, and cost needed to build capable language models.
See all 172 papers