Machine learning models are trained before deployment, yet the context they are deployed into—the data distribution they will serve, or the environment they will act in—is often unavailable during training. This thesis studies test-time context mismatch: the setting in which information needed for reliable model behavior is only revealed at deployment time. It investigates how two kinds of test-time information, limited unannotated real data and environment interaction, can be converted into signals for adaptation.

The first contribution addresses mismatch in the training data. We introduce and formalize Synthetic Dataset Quality Estimation (SynQuE), the problem of ranking synthetic datasets by their expected real-world task performance using only limited unannotated real data. We establish the first comprehensive benchmark for this problem by adapting distribution- and diversity-based distance measures as proxy metrics, and we propose Lens, a novel proxy that leverages large language model (LLM) reasoning to characterize the differences between synthetic and real data through natural-language rubrics. Across sentiment analysis, text-to-SQL parsing, image classification, and web navigation, SynQuE proxies correlate with real task performance; on text-to-SQL, selecting the top-3 synthetic datasets by proxy score raises accuracy from 30.4% to 38.4% on average over indiscriminate selection, and Lens consistently outperforms other proxies on complex, long-horizon tasks.

The second contribution addresses mismatch in the deployment environment. We identify two failure modes of LLM agents in novel environments—syntactic misunderstanding of environment-specific formats and semantic misunderstanding of state-transition dynamics—and propose an annotation-free adaptation strategy for each. Online syntactic alignment learns a lightweight adaptation vector during deployment that aligns the agent’s output distribution with the environment’s syntax at roughly 3% latency overhead. Deployment-time dynamics grounding uses persona-driven exploration to build an in-context world model of the environment’s causal dynamics before task execution. Both strategies improve performance across function-calling and web-navigation benchmarks; on the WebArena multi-site split, dynamics grounding raises the agent’s success rate from 2% to 23%.

Finally, the thesis develops a unified view of these two adaptation routes, comparing the signals they consume and the timescales they operate on, and showing that both replace labeled supervision with structure available at test time. This view positions test-time context as a practical resource for adaptation precisely in the privacy-sensitive and low-resource settings where labels, demonstrations, and retraining are unavailable.