Research map / Large language models / Language, safety and interpretability
Rethinking Label Consistency of In-Context Learning: An Implicit Transductive Label Propagation Perspective
Haoyang Chen, Richong Zhang, Junfan 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 |
|---|---|---|---|
| select demonstrations with TopK on label-interpolated synthetic embeddings (TopK-SD) instead of original embeddings | TopK demonstration selection · 9 datasets avg, LLaMA3-8B | average accuracy: 79.8 -> 81.1 | Improved |
| use TopK-SD with fixed lambda 0.7 instead of TopK | TopK demonstration selection · 9 datasets avg, DeepSeek-7B | average accuracy: 76.6 -> 76.9 | No clear effect |
| use TopK-SD with fixed lambda 0.7 instead of TopK | TopK demonstration selection · GPT-J-6B and LLaMA2-7B, 9 datasets avg | average accuracy: 72.5 -> 74.0; 75.9 -> 77.1 | Improved |
| use TopK-SD on MNLI instead of TopK | TopK demonstration selection · MNLI, DeepSeek-7B and LLaMA2-7B | accuracy: 50.6 -> 48.4 (DeepSeek); 52.2 -> 51.3 (LLaMA2) | Hurt |
| replace stage-1 TopK with TopK-SD in two-stage selection methods (MDL, ConE, DPP) (ablation) | MDL, ConE, DPP · 9 datasets avg, LLaMA3-8B | average accuracy: +0.9, +0.4, +0.6 | Improved |
| increase label consistency of demonstrations (share of demos with the query label) (ablation) | ICL demonstration set · SST-2 and AGNews | ICL accuracy | Improved |
| let LLM infer on selected demonstrations instead of simple label voting (ablation) | TopK-SD demonstrations · SST-2, SST-5, AGNews, TREC, LLaMA3-8B | accuracy: TREC 84.6 -> 92.6 at lambda 0.6 | Improved |
| set lambda to 0.0 (embedding fully replaced by label center) (ablation) | TopK-SD · SST-5, LLaMA3-8B | accuracy: 18.4 vs 52.7 at lambda 1.0 | Hurt |
| tune lambda per dataset instead of fixed 0.7 (ablation) | TopK-SD · 9 datasets, LLaMA3-8B | average accuracy | Improved |
| use TopK-SD instead of TopK with few demonstrations (3 or fewer) (ablation) | TopK demonstration selection · TREC, varying number of demonstrations | accuracy | Improved |
| random demonstration selection with few shots (ablation) | ICL · TREC | accuracy: 47.0 -> 23.2 at 1 demonstration | Hurt |
Closest papers by meaning
- In-Context Learning Learns Label Relationships but Is Not Conventional Learning (ICLR 2024 · similarity 0.90)
- Disentangling Latent Shifts of In-Context Learning with Weak Supervision (NeurIPS 2025 · similarity 0.88)
- MAPLE: Many-Shot Adaptive Pseudo-Labeling for In-Context Learning (ICML 2025 · similarity 0.88)
- Revisiting In-context Learning Inference Circuit in Large Language Models (ICLR 2025 · similarity 0.88)
- Mixture of Demonstrations for In-Context Learning (NeurIPS 2024 · similarity 0.87)
- Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization (ICLR 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.
