AI Insight
This paper introduces "temporal regret" as a new metric for AI systems that tracks how long an incorrect causal model persists before being corrected, complementing traditional outcome-based optimization. The authors prove that systems optimizing only for outcomes can achieve zero outcome regret while maintaining incorrect causal models indefinitely, but systems that log causal information and actively probe can identify errors with logarithmic complexity. In experiments on CausalBench-Seq, their Trivium system correctly identified and resolved causal misspecifications within an average of 7.8 episodes, while outcome-only controllers failed to detect structural errors.
Why it matters
This work addresses a critical blind spot in LLM-based agents and decision systems: the ability to learn why mistakes occur, not just that they occurred. By enabling systems to build and correct explicit causal models over time, this approach could reduce recurring errors in production AI systems and improve long-term reliability without requiring model retraining.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Abstract: Many agentic systems and LLM pipelines correct mistakes by optimizing outcome reward. This addresses only the what of failure; the why and when may go unlogged, allowing the same error to recur across episodes. We propose long-horizon temporal regret alongside outcome regret and epistemic regret. These are diagnostic quantities, not standard comparator-based online-learning regrets. Temporal regret captures how long an unresolved or incorrect causal model is tolerated; epistemic regret captures posterior error over that model. Over a stream of E episodes, we prove three conditional results under explicit probing, persistence, and detectability assumptions. First, under observationally equivalent confounding, outcome-only learning cannot separate causal from spurious structure, so miscalibration can persist after outcome regret reaches zero. Second, with a persistent causal log and budgeted probes, total probe complexity is logarithmic in E, inducing O(log E) delayed-identification temporal regret. The implemented clipped soft score adds a linear numerical floor, so this logarithmic claim applies only to identification delay. Third, under K detectable change-points, the rate extends to O(K log E). We instantiate Trivium and state five falsifiable predictions. On CausalBench-Seq, a hard structural readout records 7.8+-2.9 misidentified episodes per seed over 500 episodes and 20 seeds, with zero observed stationary errors, while outcome-only controllers remain misidentified throughout. Audit ablations show that continued posterior updating, not detector reopening or local repair, drives posterior recovery; local repair instead reduces committed-graph dispatch exposure. The prior logarithmic-envelope verdict based on a clipped soft score is withdrawn. A pilot real-LLM stream provides external evidence. Self-learning here means revising an external causal model, not retraining LLM weights.
Source: Trivium: Temporal Regret as a First-Class Objective for Causal-Memory Controllers