Call for Papers

LLMs are inherently multitask systems, capable of performing a wide range of tasks using a single set of parameters. These models are adapted to specific downstream tasks and usage contexts through post-training, for example via fine-tuning on task-specific data, preference-based updates, or parameter-efficient adaptation methods.

In practice, downstream tasks addressed through post-training often do not remain fixed after deployment. Their definitions, data, and evaluation criteria can evolve over time as models are used in changing environments and usage contexts. This non-stationarity leads to repeated post-deployment modifications, such as continual or incremental updates. These practices are largely guided by heuristics and empirical validation, with limited theoretical understanding of when adaptation succeeds, when it fails, or how it affects previously learned capabilities.

This gap is particularly consequential in safety-critical settings. The International AI Safety Report, published in 2025 and 2026 and endorsed by several leading experts, identifies the lack of theoretical understanding as a key limitation of current AI safety approaches. As a result, empirical improvements alone are insufficient to reason about robustness under task evolution, unintended regressions, and the limits of safe adaptation. These challenges motivate the need for theoretical and statistical foundations that can support principled post-training in dynamic, real-world environments.