AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

Beyond Reactivity: Proactive Adaptive Conformal Inference for Online LLM Factuality

Xinyu Liu, Jun Wu

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a proactive adaptive conformal inference method for online LLM factuality that handles non-exchangeable query streams, providing guarantees on output correctness.

Problem

LLM hallucinations limit reliability; existing conformal prediction approaches for LLM factuality assume exchangeability, which fails in online settings where user queries and interests shift over time.

Approach

Shifts from reactive adaptive conformal inference to a proactive approach for online LLM factuality under non-exchangeable query streams, providing correctness guarantees on LLM outputs. Exact mechanism not stated in the available abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace reactive DtACI with PACE (proactive shift signal plus reactive error)DtACI · Synthetic abrupt shiftrecovery time (steps): 120.66 -> 46.06Improved
replace DtACI with PACEDtACI · Synthetic abrupt shifttracking RMSE: 0.0461 -> 0.0349Improved
replace DtACI with PACEDtACI · Synthetic smooth shifttracking RMSE: 0.0325 -> 0.0238Improved
replace DtACI with PACEDtACI · MMLU streamtracking RMSE: 0.0175 -> 0.0068Improved
replace DtACI with PACEDtACI · MMLU streamrecovery time (steps): 24.6 -> 3.0Improved
replace fixed-step ACI (gamma=0.2) with PACEACI gamma=0.2 · WikiData streamclaims retained (%): 72.17 -> 72.27Improved
replace DtACI with PACEDtACI · WikiData streamtracking RMSE: 0.1656 -> 0.1670No clear effect
replace static SCP with PACESCP · MMLU streamfactuality (%): 88.07 -> 89.85Improved
use sliding window n=25 instead of n=5 (ablation)PACE · Synthetic abrupt shifttracking RMSE: 0.0467 -> 0.0401Improved
use sliding window n=25 instead of n=100 (ablation)PACE · Synthetic abrupt shifttracking RMSE: 0.0496 -> 0.0401Improved
use sliding window n=5 instead of n=100 (ablation)PACE · Synthetic abrupt shiftrecovery time (steps): 80.27 -> 38.20Improved
increase weight of proactive signal (WikiData) (ablation)PACE · WikiData streamclaims retained (%): 71.4 -> 72.3Improved

Experimental setup

Builds on: conformal prediction adaptive conformal inference

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy