Research map / Architectures and efficiency / Pruning and efficient tuning
Turbo Connection: Reasoning as Information Flow from Higher to Lower Layers
Tang Mohan, Sidi Lu
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 dense multi-layer downward connections (TurboConn) | Llama 3.2 1B fine-tuned · Parity | accuracy (%): 92.87 -> 100.0 | Improved |
| add TurboConn | Qwen3 1.7B fine-tuned · Parity | accuracy (%): 53.78 -> 100.0 | Improved |
| add TurboConn | Llama 3.1 8B fine-tuned · Multi-step Arithmetic | accuracy (%): 48.32 -> 51.86 | Improved |
| add TurboConn | Llama 3.1 8B fine-tuned · GSM8K (No CoT) | accuracy (%): 23.92 -> 24.82 | Improved |
| replace dense downward connections by single soft-token feedback (ablation) | TurboConn on Llama 3.2 1B · Multi-step Arithmetic | accuracy (%): 42.66 -> 38.00 | Hurt |
| replace dense downward connections by single soft-token feedback (ablation) | TurboConn on Llama 3.2 1B · Parity | accuracy (%): 100.0 -> 96.51 | Hurt |
| increase group size from 4 to 8 at alpha=100 (ablation) | TurboConn on Llama 3.2 1B · Parity | accuracy (%): 100.00 -> 51.42 | Hurt |
| increase group size from 4 to 16 at alpha=100 (ablation) | TurboConn on Llama 3.2 1B · GSM8K (No CoT) | accuracy (%): 8.32 -> 6.94 | Hurt |
| use alpha=1 instead of alpha=100 at group size 16 (ablation) | TurboConn on Llama 3.2 1B · Parity | accuracy (%): 51.40 -> 98.85 | Improved |
| use alpha=100 instead of alpha=1 at group size 4 (ablation) | TurboConn on Llama 3.2 1B · Parity | accuracy (%): 98.94 -> 100.00 | Improved |
| use alpha=100 instead of alpha=1 at group size 4 (ablation) | TurboConn on Llama 3.2 1B · Multi-step Arithmetic | accuracy (%): 40.13 -> 42.66 | Improved |
| add TurboConn to a CoT-conditioned answer model | Llama 3.2 1B · NuminaMath-CoT | accuracy (%): 30.96 -> 33.98 | Improved |
Closest papers by meaning
- Transformers Provably Solve Parity Efficiently with Chain of Thought (ICLR 2025 · similarity 0.89)
- Grokking of Implicit Reasoning in Transformers: A Mechanistic Journey to the Edge of Generalization (NeurIPS 2024 · similarity 0.89)
- Transformers Provably Learn Chain-of-Thought Reasoning with Length Generalization (NeurIPS 2025 · similarity 0.89)
- Multi-head Transformers Provably Learn Symbolic Multi-step Reasoning via Gradient Descent (NeurIPS 2025 · similarity 0.88)
- Efficient Reasoning via Constrained Optimization in Latent Space (NeurIPS 2026 · similarity 0.88)
- Self-Verifying Reflection Helps Transformers with CoT Reasoning (NeurIPS 2025 · 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.
