Research map / Large language models / Instruction and fine-tuning
Reasoning-preserved Efficient Distillation of Large Language Models via Activation-aware Initialization
Junlin He, Yihong Tang, Tong Nie, Guilong Li, Binyu Yang, Jinxiao Du, Lijun Sun, Wei Ma
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 activation-aware channel-selection initialization of projection matrices | LRC-1.5B (Llama3.2-3B teacher) · GSM8K | accuracy: 0.21 -> 0.44 | Improved |
| add activation-aware initialization | LRC-1.7B (Qwen2.5-3B teacher) · GSM8K | accuracy: 0.09 -> 0.42 | Improved |
| add activation-aware initialization | LRC-2.7B (Llama2-7B teacher) · Reasoning average | accuracy: 0.06 -> 0.10 | Improved |
| add activation-aware initialization | LRC-4B (Qwen2.5-7B teacher) · Reasoning average | accuracy: 0.18 -> 0.33 | Improved |
| add activation-aware initialization | LRC-4B · General ability average | accuracy: 0.69 -> 0.69 | No clear effect |
| add activation-aware initialization | LRC-1.5B (Llama3.2-3B teacher) · General ability average | accuracy: 0.59 -> 0.61 | No clear effect |
| add activation-aware initialization (ablation) | linear autoencoder proxy · reconstruction loss | loss: 1.24 -> 0.08 | Improved |
| add activation-aware initialization (ablation) | linear autoencoder proxy · hidden eRank | eRank: 81.30 -> 263.28 | Improved |
| use random orthogonal init instead of activation-aware init (ablation) | LRC distillation · GSM8K | accuracy: 0.44 -> 0.30 | Hurt |
| add orthogonal regularization (1e-4) to random init (ablation) | LRC distillation · GSM8K | accuracy: 0.21 -> 0.26 | Improved |
| raise distillation temperature from 1 to 40 (ablation) | RED-1.5B · GSM8K | accuracy: 0.44 -> 0.08 | Hurt |
| raise distillation temperature from 1 to 40 (ablation) | RED-1.5B · MMLU | accuracy: 0.50 -> 0.52 | Improved |
Closest papers by meaning
- Efficient Reasoning via Constrained Optimization in Latent Space (NeurIPS 2026 · similarity 0.89)
- T1: Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling (ICML 2025 · similarity 0.89)
- Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression (AAAI 2026 · similarity 0.88)
- LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models (NeurIPS 2026 · similarity 0.88)
- Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models (ICML 2026 · similarity 0.88)
- KaVa: Latent Reasoning via Compressed KV-Cache Distillation (ICLR 2026 · 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.
