Research map / Large language models / Instruction and fine-tuning
Implicit meta-learning may lead language models to trust more reliable sources
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 fine-tuning LLMs (and ConvNets) on synthetic definitions tagged as reliable or unreliable induces implicit meta-learning, where later fine-tuning internalizes reliable-tagged content more, and studies factors and mechanisms behind it.
Problem
Whether standard SGD-based fine-tuning lets models learn indicators of source reliability and modulate how they internalize information from later training data, without explicit meta-learning algorithms.
Approach
Random-string define tags act as synthetic reliability indicators in a two-stage fine-tuning setup; the model learns during stage 1 to internalize consistent-tagged definitions more, and this transfers to stage 2 where tags are the only signal, termed implicit meta-learning (IML) and linked to gradient alignment and selective retrieval.
Key results
- Pythia-2.8B shows IML: after stage 1, stage-2 exact match on consistent-tagged (ineone) definitions exceeds that on inconsistent-tagged ones, with 20-seed confidence intervals.
- Larger models show more IML and IML first becomes noticeable at 1B parameters; smaller batch sizes give stronger IML.
- IML also appears in a from-scratch Pythia-70M set-inclusion task and in a ConvNeXt V2 MNIST task, though the Pythia-70M effect is weak.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune on Stage1 data where ineone definitions are consistent and inetwo definitions are inconsistent, then fine-tune on new definitions | Pythia-2.8B · CVDB, Stage2 validation QA | exact match | Helped |
| Use consistent definitions tagged ineone instead of no definitions | Pythia-2.8B · CVDB, Stage1 validation QA | exact match | Helped |
| Use inconsistent definitions tagged inetwo instead of no definitions | Pythia-2.8B · CVDB, Stage1 validation QA | exact match | Hurt |
| Increase the correspondence between define tag and definition consistency (ablation) | Pythia-2.8B · CVDB, entity attribution | accuracy | Helped |
| Use α=0.5, so the define tag is not predictive of consistency (ablation) | Pythia-2.8B · CVDB, entity attribution | accuracy | No clear effect |
| Use α=0.95, making ineone a highly predictive tag (ablation) | Pythia-2.8B · CVDB, entity attribution | accuracy | Helped |
| Change definition word order from TAE to EAT, TEA, or ETA (ablation) | Pythia-2.8B · CVDB, entity attribution | accuracy | Hurt |
| Change definition word order from TAE to AET (ablation) | Pythia-2.8B · CVDB, entity attribution | accuracy | Hurt |
| Increase model size (ablation) | Pythia · CVDB, in-distribution QA | exact match | Helped |
| Increase batch size during fine-tuning (ablation) | Pythia-2.8B · CVDB, validation QA | exact match | Hurt |
| Train on X1 and X2 jointly rather than in two stages (ablation) | Pythia-2.8B · CVDB, validation QA | exact match | No clear effect |
| Prepend definitions to questions as context instead of training on definitions as separate datapoints (ablation) | Pythia-2.8B · CVDB, validation QA | exact match | No clear effect |
Surprising findings
- Linear probes cannot decode the define tag of an alias's definition under the standard TAE ordering, although they succeed for ATE ordering.
- Out-of-distribution, definitions tagged with a new unseen tag are internalized better than those with the reliable tag learned in stage 1.
Limitations
- No conclusive mechanistic explanation for IML; gradient alignment evidence is suggestive only.
- Internalization is operationalized but not formally defined.
- Experiments use toy synthetic datasets rather than real LLM training data.
Experimental setup
Datasets: CVDB T-REx MNIST
Benchmarks: CVDB-based closed-book QA (in-distribution validation) entity attribution test set T-REx-based QA MNIST QA masked accuracy
Metrics: exact match masked accuracy cosine similarity gradient alignment linear probe accuracy
Baselines: QA-only baseline (QA3) unseen-variables baseline (QA7) inethree definitions baseline in-context learning
Builds on: MAML Reptile reversal curse Pythia
Closest papers by meaning
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.86)
- metaTextGrad: Automatically optimizing language model optimizers (NeurIPS 2025 · similarity 0.86)
- Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence (ICML 2025 · similarity 0.86)
- Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples (ICML 2025 · similarity 0.86)
- URLs Help, Topics Guide: Understanding Metadata Utility in LLM Training (NeurIPS 2025 · similarity 0.85)
- The Rise of Parameter Specialization for Knowledge Storage in Large Language Models (NeurIPS 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.