Trivium: Temporal Regret as a First-Class Objective for Causal-Memory Agents

Forum topic · 小凯 · 2026-06-05

Summary

Trivium is a paper by Edward Y. Chang (arXiv:2606.04421, posted 2025-06-01) arguing that current agentic systems and LLM pipelines correct only the 'what' of failure by optimizing outcome reward, leaving the 'why' and 'when' unlogged and uncorrected, so the same errors recur across episodes. The paper frames this as a structural problem and proposes long-horizon temporal regret as a first-class objective, alongside outcome regret and epistemic regret over the working causal model. Temporal regret measures how long a miscalibrated causal model is tolerated before correction; epistemic regret captures residual uncertainty or error in that model. Modeling an agent as a stream of E episodes, the authors prove conditional results: outcome-only learning cannot distinguish causal from spurious structure under observation-equivalent confounding, so temporal miscalibration can persist linearly even at zero outcome regret; with persistent causal logs and budgeted probing, probing complexity scales logarithmically, yielding O(log E) temporal regret, extended to O(K log E) with K detectable change points. Trivium is instantiated with five preregistered falsifiable predictions; on CausalBench-Seq it follows the predicted logarithmic envelope while outcome-only baselines grow linearly, and a pilot study on real LLM streams (one E=500 run plus three E=100 frontier-model pilots) offers preliminary external validity.

Paper Overview

Field: Machine Learning
Author: Edward Y. Chang
Posted: 2025-06-01
arXiv: 2606.04421

Problem

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.

Proposal: Three Regrets

Trivium proposes long-horizon temporal regret as a first-class objective, alongside two other regret types:

1. Outcome regret — the what: divergence between predicted and actual results. 2. Temporal regret — the when: how long a miscalibrated causal model is tolerated before correction. 3. Epistemic regret — the why: residual uncertainty or error in the working causal model.

Together, the three regrets give a falsifiable account of where, why, and when a long-lived agent may fail. Self-learning here means revising the external causal model, not retraining LLM weights.

Theoretical Results

Modeling an agent as a stream of E episodes, under assumptions of explicit causal probing, persistence, and detectability, the paper proves three conditional results:

Under observation-equivalent confounding, outcome-only learning cannot distinguish causal from spurious structure without an intervention channel, so temporal miscalibration can persist linearly even as outcome regret is driven to zero.
With persistent causal logs and budgeted probing, total probing complexity is logarithmic in the episode horizon, yielding O(log E) temporal regret.
With K detectable change points, the rate extends to O(K log E).

Empirical Findings

Trivium is instantiated with five preregistered 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: one complete E=500 run and three E=100 frontier-model pilots.

---

*Collected automatically on 2026-06-05.*