Researchers frame agentic in-context learning as a learning algorithm problem
An arXiv paper proposes Agentic In-Context Learning, a framework for studying how fixed-parameter Transformers learn through their own interaction, with expressivity and training results.
In plain words
In-context learning is when a language model adapts to examples placed in its prompt without its weights changing. The authors study what happens when the model's own actions decide which examples it sees next, a setting they call Agentic In-Context Learning. They analyze when Transformers can imitate such learning algorithms and how different training rewards shape the behavior that results.
What was announced
The paper, posted on arXiv, argues that existing work either supplies contexts from outside or relies on interaction histories carrying rewards, while research on general tool-using agents lacks this learning algorithm perspective. The authors describe an interactive learner through three components: its state, its update rule, and its decision rule, a framing that covers active learning, bandits, and online reinforcement learning.
On the theory side, the authors construct causal Transformers that approximate these algorithms over a finite horizon and bound the resulting loss in return; they also characterize when optimizing process rewards preserves behavior that maximizes the final outcome. In experiments, distilled models learned to execute three representative recursions. Models trained from outcome or process rewards, without oracle action or state targets, developed distinct decision policies depending on the reward type, and shifting process supervision toward outcome-optimal action preferences shifted the learned decisions. The authors report that such behavioral differences persisted in a pretrained language model across textual paraphrases.
What it does not show
- The source is the paper's abstract only, so experimental details, baselines, and peer review status are not available in the supplied evidence.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.