Research map / Large language models / Agents, code and math
The Imitation Game: Turing Machine Imitator is Length Generalizable Reasoner
Zhouqi Hua, Wenwei Zhang, Chengqi Lyu, Yuzhe Gu, Songyang Gao, Kuikun Liu, Dahua Lin, Kai Chen
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 |
|---|---|---|---|
| add Atomic State (ablation) | TAIL · Greedy task, L range | pass@1 accuracy: 16.0 -> 62 | Improved |
| add Atomic State (ablation) | TAIL · Recursion task, L range | pass@1 accuracy: 32.0 -> 87.0 | Improved |
| add Linear Transition (ablation) | TAIL · Recursion task, L range | pass@1 accuracy: 30.8 -> 87.0 | Improved |
| add Linear Transition (ablation) | TAIL · Greedy task, L range | pass@1 accuracy: 11.2 -> 62 | Improved |
| add Memory Fetcher (ablation) | TAIL · Enumeration task, L range | pass@1 accuracy: 63.6 -> 92.6 | Improved |
| add Memory Fetcher (ablation) | TAIL · Simulation task, L range | pass@1 accuracy: 88.0 -> 90.0 | Improved |
| add Memory Fetcher (ablation) | TAIL · Iteration task, L range | pass@1 accuracy: 67.6 -> 84.4 | Improved |
| add Atomic State and Linear Transition and Memory Fetcher (ablation) | TAIL · DP task, L range | pass@1 accuracy: 74.2 -> 71.8 | Mixed |
| use TAIL instead of Index Hint | Qwen2.5-7B · Large-number addition, L range | pass@1: 24.0 -> 86.5 | Improved |
| use TAIL instead of Reversed Format | Qwen2.5-7B · Large-number addition, L range | pass@1: 35.0 -> 86.5 | Improved |
| use TAIL-CoT instead of DeepSeek-R1 distilled CoT for SFT | Qwen2.5-7B · Long range (L) | pass@1: 61.8 -> 90.0 | Improved |
| use TAIL-CoT-styled natural language instead of plain TAIL-CoT (ablation) | TAIL · length generalization | accuracy | No clear effect |
Closest papers by meaning
- Universal Length Generalization with Turing Programs (ICML 2025 · similarity 0.91)
- Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Trace (NeurIPS 2026 · similarity 0.89)
- Beyond Single-Task: Robust Multi-Task Length Generalization for LLMs (NeurIPS 2025 · similarity 0.89)
- Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision (NeurIPS 2026 · similarity 0.88)
- LIFEBENCH: Evaluating Length Instruction Following in Large Language Models (NeurIPS 2025 · similarity 0.88)
- Auto-Regressive Next-Token Predictors are Universal Learners (ICML 2024 · similarity 0.88)
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.
