Research map / Large language models / Instruction and fine-tuning
From Growing to Looping: A Unified View of Iterative Computation in LLMs
Ferdinand Kapl, Emmanouil Angelis, Kaitlin Maile, Johannes von Oswald, Stefan Bauer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| grow the model in depth by middle-layer duplication (LIDAS) | SmolLM 1.7B · Reasoning Primitives | accuracy: 34.62 -> 48.02 | Improved |
| use MIDAS instead of LIDAS growth operator (ablation) | SmolLM 1.7B · Reasoning Primitives | accuracy: 48.02 -> 40.88 | Hurt |
| loop a 12-layer block twice instead of standard 12-layer model (iso-param) | SmolLM 1.7B · Reasoning Primitives | accuracy: 35.42 -> 41.22 | Improved |
| fully loop a 4-layer block 6 times vs standard iso-inference baseline | SmolLM 1.7B · holdout NLL / knowledge | NLL: 1.96 -> 2.12 | Hurt |
| keep unique encoder and decoder layers and loop only middle block (Loop 4-4x4-4) (ablation) | looped 1.7B · layer-swap robustness | relative accuracy | Improved |
| repeat a 4-layer middle block at inference time without training to loop | MIDAS and LIDAS · Copy Real Words | accuracy: up to 2x | Improved |
| repeat a middle block more than twice at inference (ablation) | grown models · reasoning primitives | accuracy | No clear effect |
| more in-context examples for looped or grown models | LIDAS, Loop models · Reasoning Primitives few-shot | accuracy | Improved |
| use Nemotron-CC-Math-4+ instead of FineMath-4+ for 20% math cooldown (ablation) | LIDAS 360M · Math Word Problems | accuracy: 16.54 -> 25.64 | Improved |
| add 20% Nemotron-CC-Math-4+ cooldown | SmolLM 1.7B LIDAS · GSM8K | accuracy: 5.61 -> 13.34 | Improved |
| retrofit one extra loop of layers 8-11 during math cooldown | LIDAS 1.7B · GSM8K | accuracy: 13.34 -> 17.36 | Improved |
| retrofit loop of layers 12-15 instead of 8-11 (ablation) | LIDAS 1.7B · Reasoning Primitives | accuracy: 54.84 -> 56.60 on primitives; 17.36 -> 15.47 on GSM8K | Mixed |
Closest papers by meaning
- Reasoning with Latent Thoughts: On the Power of Looped Transformers (ICLR 2025 · similarity 0.89)
- LoopFormer: Elastic-Depth Looped Transformers for Latent Reasoning via Shortcut Modulation (ICLR 2026 · similarity 0.89)
- Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (NeurIPS 2025 · similarity 0.88)
- Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision (NeurIPS 2026 · similarity 0.88)
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (ICML 2026 · similarity 0.88)
- Looped Diffusion Language Models (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
