Few-shot in-context learning appears in six data domains, but not uniformly

single source· 1 articles · confidence: medium · first seen 2026-09-11 20:00 UTC

What this means for you

Nothing to build on yet. No model, weights or code are mentioned, and the abstract gives no scores, so the claim cannot be checked against a number. Worth noting only if you train sequence models outside text — protein or genomic — where it hints in-context learning may transfer, but that is unverified.

A preprint posted to arXiv on 11 September reports that few-shot in-context learning — where a model infers a pattern from examples in its prompt and applies it to new inputs — emerges across six kinds of data, not only text. The authors instantiated one task suite in language, genome, integer sequences, time series, images and proteins. Paired-mapping tasks beat controlled baselines in all six, and per-task effects correlated in five. That is partial support for their Convergent Emergence Hypothesis: that tasks helped by in-context learning in one domain tend to be helped in others. The posted text gives no numeric scores and no evaluation date.

Key facts

  • ·The preprint was posted on arXiv on 11 September 2026. source
  • ·Six modalities were tested on one shared task suite: language, genome, integer sequences, time series, images and proteins. source
  • ·Paired-mapping in-context learning surpassed controlled baselines in all six modalities. source
  • ·Per-task ICL effects correlated across five of the six modalities. source
  • ·The authors name the claim under test the Convergent Emergence Hypothesis and report support for it in some modalities but not all. source
  • ·The posted abstract gives no numeric benchmark scores and no evaluation date. source

What the sources say

  • Hugging Face Daily Papers — Cross-modality test of whether in-context learning shares a difficulty profile; reports partial support.

Sources

The original reporting. Follow these — they did the work.

← the wire