bayesian inference in icl

**Bayesian inference in ICL** is the **theoretical view that in-context learning approximates Bayesian updating over latent task hypotheses using prompt evidence** - it models prompt demonstrations as observations that update internal belief over possible tasks. **What Is Bayesian inference in ICL?** - **Definition**: Model behavior is interpreted as selecting predictions by posterior-weighted task hypotheses. - **Prompt Role**: Examples in context serve as evidence that shifts internal task belief state. - **Approximation**: Transformers may implement heuristic Bayesian-like updates rather than exact inference. - **Scope**: Useful for explaining calibration shifts and few-shot adaptation dynamics. **Why Bayesian inference in ICL Matters** - **Theory**: Provides principled framework for analyzing few-shot generalization behavior. - **Prompt Design**: Guides construction of demonstrations that disambiguate latent tasks. - **Robustness**: Helps explain failure under ambiguous or conflicting evidence. - **Evaluation**: Supports prediction of confidence and uncertainty behavior in ICL settings. - **Research Direction**: Connects transformer behavior to probabilistic inference models. **How It Is Used in Practice** - **Hypothesis Sets**: Design tasks where latent hypotheses are explicit and measurable. - **Evidence Control**: Vary demonstration quality and quantity to test posterior-shift predictions. - **Mechanistic Link**: Map Bayesian-like behavior to concrete circuits with causal tracing. Bayesian inference in ICL is **a probabilistic framework for interpreting few-shot adaptation in prompts** - bayesian inference in ICL is most convincing when theoretical predictions align with both behavior and circuit-level evidence.