TACIT-Switch: Cost-Aware Model Escalation for LLM Agents from Censored Supervision

The researchers proposed the TACIT-Switch method, which uses Teacher-Annotated Censored Intervention Times (TACIT) data marked as interval-truncated observations on a cumulative risk scale. This method learns permanent transition strategies through a hybrid-cure threshold model, achieving cost-perception improvement of language model agents without teacher involvement in deployment. Experiments show that this method treats annotations as interval-truncated observations on a cumulative risk scale and estimates the probability of strong rolling success and conditional transition thresholds. In mechanism-based multi-step simulations, TACIT-Switch achieves 7.4–11.1 percentage points higher success rate and comparable cost compared to task-level, step-level, and fixed-prefix routing baselines. Ablation experiments reveal that task characteristics and cumulative trajectory risk provide complementary information. After selecting working points based on development data, this method achieves 48.5% for the 4B Cheap model and 45.5% for the 9B Cheap model in ALFWorld.