Research map / Large language models / Language, safety and interpretability
On the Noise Robustness of In-Context Learning for Text Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that noisy demonstration labels significantly hurt in-context learning on text generation tasks, and proposes Local Perplexity Ranking (LPR), a training-free wrapper that swaps low-ranked demonstration candidates for semantic neighbors with lower perplexity.
Problem
In-context learning for text generation is sensitive to noisy annotations in demonstrations, unlike the reported robustness on classification tasks; existing selection methods (Random, TopK, DPP) do not account for label noise.
Approach
Decompose perplexity into inherent and matching components; approximate inherent perplexity via semantic neighbors so that candidates with higher perplexity than their local neighborhood are likely mislabeled and get replaced by a cleaner neighbor. The method is training-free and plugs into existing selection methods.
Key results
- On SCIQ with 60% irrelevant noise, LPR raises TopK exact match from 29.31 to 48.06 (+18.75).
- With Mistral-7B and 60% irrelevant noise, LPR improves random selection average from 19.24 to 27.07 across six datasets.
- LPR takes about 0.57 hours versus 3.05 hours for global perplexity ranking in the efficiency comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LPR to TopK selection | TopK ICL on SCIQ · SCIQ, 60% irrelevant label noise | EM: 29.31 -> 48.06 | Helped |
| Add LPR to Random selection | Random ICL on NQ · NQ, 60% irrelevant label noise | EM: 4.23$\pm$0.46 -> 8.87$\pm$0.74 | Helped |
| Use noisy annotations in demonstrations | ICL for text generation · Text generation tasks, increasing noise rates | EM | Hurt |
| Increase demonstration set size under noisy annotations | ICL for text generation · Noisy settings, K=2 versus K=8 | EM | Hurt |
| Add LPR to ICL demonstrations | ICL on clean SCIQ with TopK selection · SCIQ, 0% noise | EM: 68.62$\pm$1.13 -> 70.06$\pm$0.32 | No clear effect |
| Lower the LPR threshold γ (ablation) | LPR · Generation tasks, especially 60% noise | EM | Helped |
| Set the threshold γ to 75% (ablation) | LPR · Generation tasks with noise | EM | Helped |
| Increase local-neighbor count from 0 to 2, 4, or 6 (ablation) | LPR · Generation tasks with noise | EM | Helped |
| Increase local-neighbor count beyond the saturation point (ablation) | LPR · Generation tasks with noise | EM | No clear effect |
| Use global rather than local perplexity ranking (ablation) | Perplexity-based ICL selection · Generation tasks, most settings | EM | Hurt |
| Use global rather than local perplexity ranking (ablation) | Perplexity-based ICL selection · Generation tasks | time (h): 3.05/0.57 | Hurt |
| Use cosine similarity instead of BM25 to find local neighbors (ablation) | LPR · Generation tasks with noisy annotations | EM | No clear effect |
Surprising findings
- Irrelevant noise hurts more than relevant noise, and larger demonstration sets under noise can perform worse than smaller ones.
- Under clean conditions LPR sometimes slightly lowers performance (e.g., TopK on NQ 20.25 to 19.19).
Limitations
- Assumes clean annotations are the majority; performance gains shrink at 70-90% noise.
- Assumes semantically similar examples share inherent perplexity, which weakens with smaller LMs (e.g., OPT-1.3B).
- No theoretical analysis of how noisy annotations affect ICL.
Experimental setup
Datasets: NQ WebQ SQuAD SCIQ GeoQuery NL2Bash SST2 AGNews
Benchmarks: NQ WebQ SQuAD SCIQ GeoQuery NL2Bash SST2 AGNews MT-bench Arena-Hard
Metrics: exact match BLEU accuracy GPT-4 score
Baselines: Random TopK DPP Zero-Shot Zero-Shot-CoT Manual-CoT Auto-CoT Global perplexity ranking
Builds on: TopK demonstration retrieval DPP demonstration selection OpenICL
Closest papers by meaning
- Perplexed by Perplexity: Perplexity-Based Data Pruning With Small Reference Models (ICLR 2025 · similarity 0.87)
- What is Wrong with Perplexity for Long-context Language Modeling? (ICLR 2025 · similarity 0.87)
- Understanding In-Context Learning from Repetitions (ICLR 2024 · similarity 0.86)
- Making Text Embedders Few-Shot Learners (ICLR 2025 · similarity 0.86)
- In-Context Language Learning: Architectures and Algorithms (ICML 2024 · similarity 0.85)
- InstructRAG: Instructing Retrieval-Augmented Generation via Self-Synthesized Rationales (ICLR 2025 · similarity 0.85)
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.