Research map / Large language models / Language, safety and interpretability
On the Power of Context-Enhanced Learning in LLMs
Xingyu Zhu, Abhishek Panigrahi, Sanjeev Arora
Read the paperCheck your idea against 137,217 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.
Formalizes context-enhanced learning (gradient training where curriculum text is in context but receives no loss) and shows on a synthetic multi-level translation task that it can be exponentially more sample-efficient than SFT, with the benefit tied to more accurate gradients.
Problem
Whether gradient-based LLM training can benefit from helpful material placed in context without auto-regressive loss on that material, and whether such learning is more sample-efficient than standard SFT and leaks less recoverable information.
Approach
Train with standard auto-regressive loss on targets while conditioning on curriculum text (phrasebook excerpts with dropout) that receives no loss; this is a gradient-based analog of in-context learning. Theory uses a surrogate transformer model on a multi-level translation task to show exponential sample-complexity gap, attributed to more accurate gradient signals.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use annealed rule dropout with correct phrasebook context instead of no context | Llama 3.2-3B ICL-capable model on multi-level translation · Multi-level translation, n=8 and n=10 | training samples needed: 10x less samples | Improved |
| Train with subsets of phrasebook rules and 20% dropout | Llama 3.2-3B ICL-capable model · Multi-level translation heldout test | next-token prediction accuracy: perfect heldout test-time accuracy with 100% dropout rate | Improved |
| Use fixed rule dropout instead of annealed rule dropout | Llama 3.2-3B ICL-capable model · Multi-level translation, n=8 and n=10 | sample efficiency | Hurt |
| Remove dropout while providing the relevant phrasebook rules (ablation) | Context-enhanced learning · Multi-level translation | test accuracy without context | Hurt |
| Replace correct phrasebook rules with incorrect rules (ablation) | Annealed-dropout context-enhanced learning · Multi-level translation | test accuracy | Hurt |
| Start from a non-ICL-capable 3B base model (ablation) | Annealed-dropout context-enhanced learning · Multi-level translation | test accuracy | Hurt |
| Include unused phrasebook rules in context during training (ablation) | Annealed-dropout context-enhanced model · Multi-level translation, heldout rule split | translation accuracy on unused rules | No clear effect |
| Filter generation to tokens in alphabet A_{i+1} (ablation) | Context-enhanced learning model queried for phrasebook rules · Phrasebook rule recovery, d=5, n=10 | rule recovery success rate: 0.00% -> 1.66%; 0.00% -> 1.28%; 0.89% -> 0.94% | Mixed |
| Increase the rule dropout rate | ICL-capable surrogate model · Gradient prediction accuracy | gradient prediction accuracy | Hurt |
| Drop rules from multiple phrasebooks rather than fewer phrasebooks | ICL-capable surrogate model · Gradient prediction accuracy | gradient prediction accuracy | Hurt |
| Increase the gradient-computation batch size | ICL-capable surrogate model · Gradient prediction accuracy | gradient prediction accuracy | Improved |
Surprising findings
- Phrasebook rules are hard to recover from output probabilities after context-enhanced learning, even under adversarial token filtering.
Limitations
- Experiments use a synthetic translation task rather than real math or coding tasks.
- Theory relies on a simplified surrogate model; extension to actual transformers is open.
- Only a limited set of recovery attacks was explored.
Experimental setup
Datasets: synthetic multi-level translation (MLT) samples
Benchmarks: held-out synthetic MLT translation tasks
Metrics: next-token accuracy on final answer tokens recovery success rate for phrasebook rules gradient prediction accuracy l2 norm of hidden representation change
Baselines: vanilla SFT no-dropout context ablation wrong-context ablation non-ICL-capable base model
Builds on: learning using privileged information CoT internalization statistical query framework
Closest papers by meaning
- Let's (not) just put things in Context: Test-time Training for Long-context LLMs (ICLR 2026 · similarity 0.89)
- Understanding In-Context Learning in Transformers and LLMs by Learning to Learn Discrete Functions (ICLR 2024 · similarity 0.88)
- Understanding Synthetic Context Extension via Retrieval Heads (ICML 2025 · similarity 0.87)
- Is In-Context Learning Sufficient for Instruction Following in LLMs? (ICLR 2025 · similarity 0.87)
- In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering (ICML 2024 · similarity 0.87)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · 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.
