Paper Overview
- Field: Machine Learning
- Author: Edward Y. Chang
- Published: 2025-06-01
- arXiv: 2606.04421
- Temporal regret captures *when* failure persists: how long a miscalibrated causal model is tolerated before correction.
- Epistemic regret captures *why* failure persists: residual uncertainty or error in the working causal model.
- Outcome regret captures the *what*: divergence between predicted and actual results.
Summary
Many current agentic systems and LLM pipelines correct mistakes by optimizing outcome reward. This addresses only the *what* of failure: when an outcome diverges from prediction, the *why* and *when* of the mismatch are not systematically logged, reviewed, or corrected, so the same error can recur episode after episode. The paper argues this is a structural problem, not merely a model-capacity one.
Three Regrets as First-Class Objectives
Theoretical Results
Modeling an agent as a stream of E episodes, under explicit assumptions on causal probing, persistence, and detectability, the paper proves three conditional results:
1. Under observationally equivalent confounding, outcome-only learning cannot distinguish causal structure from spurious structure without an intervention channel — temporal miscalibration can persist linearly even when outcome regret is driven to zero. 2. With persistent causal logs and budgeted probing, total probing complexity is logarithmic in episodes, yielding O(log E) temporal regret. 3. With K detectable change points, the bound extends to O(K log E).
Experiments
Trivium is instantiated with five pre-registered falsifiable predictions. On CausalBench-Seq, Trivium follows the predicted logarithmic envelope, while outcome-only baselines grow linearly. A pilot study on real LLM streams provides preliminary external validity evidence, spanning one full E=500 run and three E=100 frontier-model pilots.
Notably, self-learning in this framework means revising an external causal model, not retraining LLM weights.