Research map / Large language models / Instruction and fine-tuning
IA2: Alignment with ICL Activations improves Supervised Fine-Tuning
Aayush Mishra, Daniel Khashabi, Anqi Liu
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 IA2 activation-alignment priming before SFT (ablation) | SFT with LoRA · single-token tasks, N=4 | accuracy and ECE: SST2 acc 65.2 -> 90.4; ECE 0.22 -> 0.06 | Improved |
| Add IA2 priming before SFT on multi-token tasks (ablation) | SFT · GSM8K, GSM8Ks, MATH Algebra, SciQ, N=4 | accuracy: GSM8K 70.9 -> 73.6; HMathA 50.4 -> 55.3 | Improved |
| Add IA2 priming before SFT versus using ICL at inference | Qwen3-4B-Base math · GSM8K, HMathA | accuracy: GSM8K 76.4 -> 73.6 (ICL vs IA2->SFT); HMathA 60.4 -> 55.3 | Hurt |
| Train with IA2 only (no SFT) (ablation) | Qwen3-4B-Base · SciQ generation, N=4 | accuracy: 40.8 -> 06.9 | Hurt |
| Add IA2 loss jointly with SFT on ICL responses (ablation) | SFT on ICL responses · GSM8K, N=4 | accuracy: 66.4 -> 77.0 | Improved |
| Use soft-label KD on ICL responses instead of SFT only (ablation) | Llama SFT · GSM8K and SST2 | accuracy: GSM8K 28.2 -> 36.1; SST2 52.3 -> 55.9 | Mixed |
| Train only QK matrices instead of QKO in LoRA (ablation) | IA2 -> SFT, Llama-1B · FinS, PoemS, SST2 | accuracy: FinS 89.8 -> 50.4 | Hurt |
| Increase LoRA rank from 8 to 16 (ablation) | IA2 -> SFT · GSM8K, SST2 | accuracy: GSM8K 73.6 -> 74.1 | No clear effect |
| Swap LoRA target from IA2 to (IA)^3 adapters | IA2 -> SFT · SciQr/QASCr | accuracy: QASCr 32.8 -> 48.6 vs SFT only | Improved |
Closest papers by meaning
- Capturing In-Context Learning Dynamics with Task Operators (NeurIPS 2026 · similarity 0.88)
- The Missing Alignment Link of In-context Learning on Sequences (ICML 2025 · similarity 0.88)
- The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning (ICLR 2024 · similarity 0.88)
- Revisiting In-context Learning Inference Circuit in Large Language Models (ICLR 2025 · similarity 0.87)
- In-Context Learning through the Bayesian Prism (ICLR 2024 · similarity 0.87)
- Joint Localization and Activation Editing for Low-Resource Fine-Tuning (ICML 2025 · 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.
