Effective machine learning depends not only on how we model data, but also on what data we choose to collect. For enterprises deploying AI at scale, the cost of gathering labeled data often dominates project budgets. A new theoretical study from researchers at multiple institutions examines whether the emergent in-context learning (ICL) capabilities of large sequence models can streamline this data selection process—an ability akin to "intrinsic curiosity."
The paper, titled "Can In-Context Learning Support Intrinsic Curiosity?" and posted on arXiv, tackles a longstanding challenge: how to automate data collection so that an AI agent actively seeks out the most informative observations. Classic approaches reward an agent based on its "learning progress," which measures how much a new observation improves a world model's predictive accuracy. However, calculating these rewards traditionally requires expensive inner loops of gradient descent updates within each trajectory, rendering them computationally impractical at scale.
The authors—Eric Elmoznino, Sangnie Bhardwaj, Johannes Oswald, Rajai Nasser, Blaise Agüera y Arcas, João Sacramento, Rif A. Saurous, and Guillaume Lajoie—ask whether ICL can serve as an immediate, update-free world model, thus eliminating the gradient-descent bottleneck. Specifically, they evaluate whether an exploration policy can be trained to maximize learning progress using only the prediction errors and counterfactual context manipulations of an in-context learner.
Theoretical Results: Possibility and Impossibility
The study proves two contrasting results. First, for general Markov decision processes (MDPs)—which model sequential decision-making with states, actions, and rewards—the researchers demonstrate that using ICL-derived intrinsic rewards is impossible in an unbiased way. The resulting rewards either suffer from nuisance terms that bias their estimation of true learning progress, or they cannot be implemented using an in-context learner's prediction errors.
Conversely, for a broad subclass of non-temporal settings—including active learning and Bayesian experimental design—ICL-derived rewards do work. Here, the authors prove that such rewards successfully bound and asymptotically converge to the true learning progress.
| Setting | ICL-Derived Reward Validity | Bias | Convergence to True Learning Progress |
|---|---|---|---|
| General Markov Decision Processes (temporal) | Impossible | Unavoidable bias from nuisance terms | Not achieved |
| Non-temporal settings (active learning, Bayesian experimental design) | Feasible | No bias | Asymptotic convergence achieved |
Experimental Corroboration
The theoretical findings are backed by controlled experiments across both continuous and symbolic environments. According to the paper, the ICL-driven framework successfully trains curious data-collection policies that explore optimally. These experiments demonstrate that, where theory permits, the method can replace gradient-based learning progress signals with in-context learning prediction errors, dramatically reducing computational overhead.
Implications for Enterprise AI Development
For CTOs and technology leaders, this research clarifies the boundaries of using in-context learning for automated data selection. While ICL cannot replace all exploration strategies—especially in temporal tasks like robotics or sequential supply-chain optimization—it offers a computationally efficient solution for non-sequential data collection. Applications such as active learning for document classification, Bayesian optimization for experiment design, or data augmentation for computer vision could benefit from this approach. The key insight is that in-context learners can serve as instant world models, cutting out the backpropagation loops that slow down current methods. However, deploying this in enterprise systems would require careful tuning to ensure the selected data aligns with business objectives—an area that future applied research will need to address.