Introduction

1. In-Context Learning (ICL):
   • Language models perform tasks by conditioning on examples provided in the prompt.
   • No parameter updates; the model uses the context to make the predictions.
2. Objective of the Paper:
   • Investigate the role of demonstrations in ICL.
   • Understand what aspects of demonstrations are crucial for model performance.

Introduction

1. How important are correct input-label mappings in demonstrations?
2. Which aspects of demonstrations contribute most to performance gains?
3. How does meta-training with an ICL objective affect these findings?

Main Research Questions

Experiments

• 12 Models:

  • GPT-2 Large (774M parameters)

  • GPT-J (6B parameters)

  • Fairseq LMs (6.7B & 13B parameters)

  • GPT-3 (175B parameters)

  • MetaICL (Meta-trained GPT-2 Large)
     

• ​​Inference Methods:

  • ​Direct Method: Predict output directly from input and context.

  • Channel Method: Predict input given output and context.

Experimental Setup

• Evaluation Data:

  • 26 Datasets covering:

    • ​Sentiment analysis, paraphrase detection, Hate speech detection, etc.

  • ​All tasks are classification or multi-choice tasks
     

• Experimental Details:

  • ​Demonstrations: K = 16 examples per prompt.

  • Repetitions: 5 random seeds, experiments run 5 times.

    • ​For Fairseq 13B and GPT-3: subset of 6 datasets and 3 seeds

  • ​Metrics:

    • ​Macro-F1 for classification tasks

    • Accuracy for multi-choice tasks

• Method: Replace gold labels with random labels in demonstrations.
  C = small discrete set of possible labels => labels are randomly sampled at uniform from C
• Findings: Performance drops only marginally compared to using correct labels.

Importance of Correct Input-Label Mappings

• Objective: Assess the effect of using out-of-distribution (OOD) inputs in demonstrations.
• Method: Replace in-distribution inputs x1,x2,…,xkx_1, x_2, \dots, x_kx1​,x2​,…,xk​ with random sentences from an external corpus (e.g., CC-News).
• Findings:
  • ​Significant performance drop (3–16% absolute decrease) for most models.
  • In some cases, worse than no demonstrations.

Impact of Input Distribution

• Objective: Examine the role of the label space in model performance.
• Method: Replace task-specific labels     with random English words.
• Findings:
  • Significant performance drop (5–16% absolute decrease).
  • Indicates that knowledge of the correct label space is important.
  • Direct models have significant gaps, but channel models:
    • Minimal impact (0–2% drop or slight increase).
    • Channel models are less affected by changes in label space.

Impact of Label Space

• Objective: Assess the importance of the input-label pairing format in demonstrations.
• Method:
  • ​No Labels: Use inputs only in demonstrations.
  • Labels Only: Use labels only without inputs.
• Findings:
  • Removing the input-label pairing format leads to performance worse than or close to no demonstrations.
  • Keeping the format allows models to retain a significant portion of performance gains.
• Key Insight: Format is crucial for triggering models to mimic and complete tasks.

Impact of Input-Label Pairing (Format)

• MetaICL Model:
  
  • GPT-2 Large meta-trained with an in-context learning objective.
• Observations:
  • Patterns are more pronounced with MetaICL.
  • Ground truth input-label mappings matter even less.
  • Format becomes even more important.
• Hypothesis:
  • Meta-training encourages models to exploit simpler aspects (like format) and ignore complex ones (like input-label mappings).
    1. The input-label mapping is likely harder to exploit
    2. The format is likely easier to exploit
    3. The space of the text that the model is trained to generate is likely easier to exploit than the space of the text that the model conditions on

Impact of Meta-Training

Conclusion

• Main Findings:
  • Correct input-label mappings have little impact on performance.
  • Performance gains come from independent specification of inputs and labels.
  • Format of demonstrations is key to performance.
  • Meta-training magnifies reliance on superficial aspects.
• Implications:
  • Models may not learn new tasks but exploit patterns and priors from pertaining.
  • In-context learning might not work for tasks not captured during pretraining.

Conclusion