Overview
Evaluating and predicting the performance of large language models (LLMs) in multi-turn conversational settings is computationally intensive, particularly when key events, such as jailbreaks or successful task completion, manifest only after repeated interactions. These events may be infrequent and difficult to observe within typical computational budgets. Recent conformal survival frameworks offer reliable lower predictive bounds (LPBs) on the number of iterations required to trigger an event of interest but utilize static budget allocation, which is inefficient for multi-turn setups.
This research introduces Dynamic Allocation via PRojected Optimization (DAPRO), a dynamic budget allocation framework explicitly designed for bounding time-to-event in multi-turn LLM interactions. DAPRO is presented as the first theoretically valid framework of its kind.
Research Context
The evaluation of LLMs in multi-turn conversations presents a challenge due to the computational expense and the potential rarity of critical events. Observing these events often necessitates numerous interactions. Prior conformal survival frameworks generate LPBs for the number of iterations until an event. However, their reliance on static budget allocation limits their efficiency in multi-turn interaction scenarios. A key limitation of these earlier approaches is their assumption of conditional independence between censoring and event times.
Approach
DAPRO was developed to address the inefficiencies of static budget allocation in multi-turn LLM evaluation. The framework aims to dynamically allocate computational budget while providing guarantees on event time bounding. A core theoretical contribution of DAPRO is a novel coverage bound. This bound scales with the square root of the mean censoring weight, in contrast to prior work which scaled with the worst-case weight.
Findings
- DAPRO satisfies the budget constraint.
- The framework provides distribution-free, finite-sample coverage guarantees.
- DAPRO does not require the conditional independence assumption between censoring and event times, which was assumed by prior conformal survival approaches.
- The novel coverage bound yields provably tighter guarantees than previous work by scaling with the square root of the mean censoring weight.
- DAPRO can generate unbiased, low-variance estimates of population-level evaluation metrics, such as the jailbreak rate, even with limited computing resources.
- Comprehensive experiments were conducted across various evaluation scenarios, including agentic task success, adversarial jailbreaks, toxic content generation, and RAG hallucinations.
- LLMs tested included Llama 3.1 and Qwen 2.5.
- Experimental results demonstrated that DAPRO consistently achieved coverage closer to the nominal level.
- DAPRO exhibited lower variance compared to static baselines.
- The framework successfully satisfied the budget constraint during experiments.
Why This Matters
The development of DAPRO offers a method for more efficiently evaluating LLMs in multi-turn conversational settings. Its ability to provide tighter coverage guarantees and unbiased estimates of key metrics under budget constraints can contribute to more effective assessment of LLM performance, specifically concerning behaviors like jailbreaking or successful task execution, without the need for the conditional independence assumption.