Large language models (LLMs) are often used in high-stakes environments where providing an incorrect answer can have serious consequences. While standard methods exist to help models "abstain" from answering when they are uncertain, these methods often focus on the average performance across an entire population. This can lead to a "group coverage gap," where a model appears reliable overall but consistently fails specific subgroups—such as those dealing with harder topics or different linguistic styles. This paper introduces HG-CRC, a framework designed to ensure that risk guarantees are met not just for the whole population, but for every specific subgroup within a defined hierarchy.
Addressing the Group Coverage Gap
Standard Conformal Risk Control (CRC) provides a "marginal" guarantee, meaning it ensures the average error rate across all users stays below a chosen budget. However, this approach does not account for the diversity of real-world populations. If a model is tested on a mix of easy and hard questions, it might satisfy the overall budget while failing significantly on the harder subset. The authors demonstrate that under realistic shifts in group composition, standard CRC can violate its risk budget in up to 47% of trials. HG-CRC solves this by enforcing simultaneous risk guarantees across all nodes of a user-defined hierarchy, ensuring that no specific group is sacrificed to maintain the global average.
How HG-CRC Works
HG-CRC organizes groups into a tree-like hierarchy, ranging from a broad global category down to specific, granular subgroups. The framework uses two key mechanisms to maintain control:
Bonferroni Correction: To ensure that all nodes in the hierarchy are certified simultaneously, the method applies a statistical correction that adjusts the confidence level for each node.
Leaf-First Selection Policy: When the model encounters a query, it checks the most specific (leaf) node first. If that node is certified and the query meets its threshold, the model provides an answer. If not, it "falls back" to broader, coarser nodes in the hierarchy.
This process requires only a held-out calibration set and does not require any model retraining, making it a practical post-hoc solution for existing systems.
Performance and Results
The researchers tested HG-CRC on three models (Qwen3-4B, Llama-3.1-8B-Instruct, and Gemma-3-4B) using two benchmarks, ARC Challenge and MMLU-Pro. On the ARC Challenge, the method achieved an empirical 0% violation rate for high-accuracy models. The study also found that the hierarchical structure is essential; when the "difficulty level" was removed from the hierarchy, the error rate returned to the 11% violation level seen in standard global CRC. While the method does come with a "participation cost"—meaning the model abstains more frequently to ensure safety—the authors argue this is a necessary trade-off for achieving equitable, principled reliability.
Important Considerations
While HG-CRC provides a robust framework for safety, it is not a "free" solution. The primary trade-off is reduced availability, as the model must abstain more often to guarantee that it stays within the risk budget for every subgroup. Additionally, the method relies on the quality of the uncertainty scores provided by the model; if a model is poorly calibrated, the system will naturally become more conservative and abstain more frequently. Finally, while the Bonferroni correction is mathematically necessary for the theoretical guarantee, the authors note that its impact is most significant when the hierarchy contains a large number of nodes.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!