What is In-Context Learning (ICL), and how does it differ from fine-tuning (4)?
In-Context Learning (ICL) enables a model to perform a task by conditioning on input-output examples without updating weights.
Key Differences from Fine-Tuning:
No weight updates → The model remains unchanged.
Works immediately without extra training.
Practically useful → Avoids costly data labeling and fine-tuning complexity.
Intellectual intriguing -> Genralization from few examples as test for “intelligent behaviour”
Why is In-Context Learning (ICL) practically useful?
ICL avoids costly data labeling & fine-tuning:
Labelling data is costly
Requires domain expertise.
Emergent, time-sensitive scenarios.
Fine-tuning is challenging
Training is sensitive to hyperparameters.
Not enough validation data.
Expensive in memory and time.
We dont quite understand how fine-tuning works
What are 3 biases and sensitivities in In-Context Learning?
Key biases:
Encoding Sensitivity → The format of prompts affects performance (e.g., stars, sentiment labels).
Majority Label Bias → Frequent training answers dominate predictions.
Recency Bias → Examples near the end of the prompt influence predictions more.
What are the challenges of reasoning in Large Language Models (LLMs)?
Types of reasoning problems:
Arithmetic Reasoning
Symbolic Reasoning (Letters)
Commonsense Reasoning
What methods improve reasoning abilities in LLMs?
Name 4 techniques
Approaches to enhance reasoning:
Train with 100x more data.
Increase model size for better knowledge retention.
Specific techniques:
Chain of Thought (CoT) → Step-by-step problem solving.
Few-Shot CoT → Step-by-step demonstration and response.
Zero-Shot CoT → Two-stage prompting + step-by-step answer.
Self-Consistency → Majority voting for answer selection.
What is alignment, and why is it necessary in LLMs?
Problem: There is a mismatch between pre-training and user intent.
Solution: Alignment → Adjusting models to ensure they do what we want.
Goal: Make the pre-trained model follow human instructions reliably.
What is Instruction-Tuning, and what are the 3 steps?
Instruction-Tuning fine-tunes LLMs on a collection of datasets that involve mapping language instructions to their corresponding desirable generations.
Steps:
Collect example (instruction, output) pairs across multiple tasks.
Fine-tune the model on these examples.
Test on unseen tasks to check generalization.
What are common tasks used in instruction-tuning?
Frequent tasks in instruction-tuning:
Sentiment classification
Text summarization
Question answering
Machine translation
Textual entailment
What are the main challenges in instruction-tuning?
Challenges:
Expensive → Requires ground-truth data for tasks.
Hard for open-ended tasks → No clear right answer for creative outputs.
Prone to hallucinations → Generates incorrect but confident responses.
What are the advantages of instruction-tuning for LLMs?
Advantages:
Improves zero-shot performance.
Scaling up instruction data improves accuracy.
Minor costs compared to pretraining
What are the limitations of instruction-tuning?
Limitations:
Hard to collect diverse labeled data.
Limited creativity → Generates token-by-token, not holistic ideas.
Encourages hallucinations → Model does not verify its outputs.
Last changed3 months ago