Does learning from mistakes improve in-context learning?
Explores whether inducing models to make errors on few-shot examples, then having them articulate principles from those mistakes, leads to better performance than learning from correct examples alone.
In-context learning has always learned from correct input-output pairs only. LEAP (Learning Principles) revisits that: given the same few examples, it (1) intentionally induces the model to make mistakes on them, (2) has the model reflect on those mistakes and articulate explicit, task-specific principles — with no human supervision — that help avoid common errors, then (3) prompts the test question with the original few-shot examples plus the learned principles. It uses exactly the same number of labeled examples as standard few-shot, yet improves strong models (GPT-3.5/4/4-turbo, Claude-2.1, Gemini Pro) across DROP, HotpotQA, GSM8K, MATH, and Big-Bench Hard (e.g., +7.5% on DROP with GPT-4).
The keeper is a generative-learning principle at the prompt level: the model extracts more usable structure from examples by erring and explaining the error than by imitating correct answers. Negative experience, articulated, transfers better than positive examples alone — within a single inference-time prompt, no fine-tuning.
This is the in-context, self-supervised cousin of learning-from-mistakes at training time. It rhymes with Can reconstructing expert thinking improve reasoning transfer? (articulating the latent process behind surface examples) and with Can confidence trajectories reveal when reasoning goes wrong? in deriving a usable training/prompting signal from the model's own errors rather than external labels.
Inquiring lines that read this note 15
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Can prompt-based context override biases that were embedded during pretraining?- Do few-shot examples improve in-context learning or add noise?
- Why does negative experience transfer better than positive examples alone?
- Can we reverse the instruction-following deficit through targeted training?
- What makes a good in-context learning example for a given task?
- How does the Learning Law explain why all examples should contribute equally?
- Can rejected edits serve as negative feedback like hard negatives in contrastive learning?
- Do correlated human errors prevent models from transcending their training sources?
- How does demonstration coverage in context examples determine operation generalization?
Related concepts in this collection 3
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can reconstructing expert thinking improve reasoning transfer?
Expert texts show only the final result of complex thinking. Can we reverse-engineer those hidden thought processes and use them to train models that reason better across different domains?
both articulate the latent structure behind surface examples; LEAP does it in-context from induced mistakes
-
Can confidence trajectories reveal when reasoning goes wrong?
Does the timing of when a model commits to an answer predict whether its reasoning will be flawed? And can we use this signal to train better reasoning without expensive annotations?
both derive a usable signal from the model's own errors rather than external labels
-
Why do chain-of-thought examples fail across different conditions?
Chain-of-thought exemplars show surprising sensitivity to order, complexity level, diversity, and annotator style. Understanding these brittleness dimensions could reveal what makes reasoning prompts robust or fragile.
LEAP adds learned principles atop exemplars, a lever beyond exemplar selection
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- In-Context Principle Learning from Mistakes
- LLMs can implicitly learn from mistakes in-context
- Are Emergent Abilities in Large Language Models just In-Context Learning?
- Learning To Retrieve Prompts for In-Context Learning
- Self-reflecting Large Language Models: A Hegelian Dialectical Approach
- LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters!
- Generalization to New Sequential Decision Making Tasks with In-Context Learning
- Can Looped Transformers Learn to Implement Multi-step Gradient Descent for In-context Learning?
Original note title
in-context learning improves when you induce the model to err on the few-shot examples then have it articulate explicit principles from those mistakes