AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Meta-in-context learning in large language models

Julian Coda-Forno, Marcel Binz, Zeynep Akata, Matthew Botvinick, Jane X Wang, Eric Schulz

NeurIPS 2023 · Large language models › Language, safety and interpretability · approach: In-context learning analysis · 6 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Present examples from previous tasks in the context before a new regression taskGPT-3 · 1D regressionMSE: -13.26 ± 3.08 task-number effect on MSEImproved
Present examples from previous tasks in the context before a new bandit taskGPT-3 · Two-armed banditregret: -0.031 ± 0.012 task-number effect on regretImproved
Use sequential task experience to adapt the model before real-world regressionGPT-3 · Real-world regression benchmarkRMSEImproved
Apply meta-in-context learning to the real-world regression benchmarkGPT-4 · Real-world regression benchmarkRMSEImproved
Use GPT-4 rather than GPT-3 for real-world regression, before or after meta-in-context learningReal-world regression · Real-world regression benchmarkRMSEHurt
Encounter more similar previous tasksGPT-3 after meta-in-context learning · Real-world regression benchmarkRMSE: -0.147 ± 0.020Improved
Encounter more similar previous tasksGPT-4 after meta-in-context learning · Real-world regression benchmarkRMSE: -0.196 ± 0.024Improved
Accumulate additional observations within a taskGPT-3 · 1D regressionMSE: -30.614 ± 3.08 trial-number effectImproved
Accumulate additional observations within a taskGPT-3 · Two-armed banditregret: -0.221 ± 0.012 trial-number effectImproved
Accumulate additional observations within a taskGPT-3 · Real-world regression benchmarkRMSE: -0.089 ± 0.020 trial-number effectImproved
Meta-in-context learning shifts prior expectations toward encountered task statisticsGPT-3 · Two-armed banditprior expected reward mean: 545.22 -> 6.49Improved
Meta-in-context learning constrains predictions to plausible target valuesGPT-3 · Real-world regression benchmarkproportion of predictions outside target rangeImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy