COMPLETED Research

Event-Triggered LLM Invocation — Uncertainty-Aware Sequential Decision Rules

When should a streaming pipeline pay for an LLM call? Cast as a risk-based sequential stopping problem, with six proved guarantees under which the classical trigger families — event-triggered, optimal stopping, SPRT, CUSUM, Bayesian — all fall out as special cases. Accepted at ECML PKDD 2026 (Research Track).

VIEW_SOURCE() DOCS()
status COMPLETED
type Research
started 2025-10-01
completed 2026-05-28
stack Python PyTorch vLLM CMAPSS GRU TCN Transformers

// DESCRIPTION

Accepted at ECML PKDD 2026 (Research Track). arXiv 2607.13048 · archived code release 10.5281/zenodo.20783549.

The question

Streaming inference pipelines increasingly pair a lightweight fast model with an LLM that supplies semantic understanding at substantial cost. When to invoke the LLM has had very little formal treatment — it is usually a hand-tuned threshold. This work casts it as a risk-based sequential stopping problem: a trigger policy fires when a risk functional R(H_t) crosses a threshold θ.

Six results

  1. A minimum inter-event time bound that excludes trigger chattering
  2. Optimality of threshold policies, via smooth pasting
  3. Approximate SPRT guarantees under estimated parameters
  4. O(√(T log T)) regret on stationary streams, degrading to O(√((C_T+1)·T log T)) under C_T changepoints
  5. O(1/√T) convergence of online gradient descent for adaptive thresholds
  6. A calibration-to-miss-rate transfer inequality

Event-triggered control, optimal stopping, SPRT, CUSUM and Bayesian triggers are recovered as special cases of the same framework rather than treated as competing families.

Evaluation

Turbofan degradation data (CMAPSS) with real LLM calls, not simulated costs: the assumptions are verified rather than assumed, the risk-function design is ablated, six baselines are compared (including a RouteLLM-style router and contextual bandits), and both cost sensitivity and LLM failure modes are analysed.

// HIGHLIGHTS

  • ECML PKDD 2026 Research Track — accepted
  • Six formal guarantees, including O(√(T log T)) regret and a chattering bound
  • Classical trigger families unified as special cases of one risk functional
  • Evaluated on CMAPSS with real LLM calls against six baselines
  • Code archived at Zenodo with a DOI; reproduction map in REPRODUCE.md
< BACK_TO_PROJECTS()