Bidirectional Coherence Paradox
“I introduce what I call the Bidirectional Coherence Paradox: competence and grounding not only dissociate but invert across epistemic conditions. In low-observability domains, LLMs often act successfully while misidentifying the mechanisms that produce their success. In high-observability domains, they frequently generate explanations that accurately track observable causal structure yet fail to translate those diagnoses into effective intervention.”
— Coherent Without Grounding, Grounded Without Success (2026), abstract
Why it matters
The paradox refutes the heuristic that an articulate explanation indicates understanding. By showing that coherence and grounding can invert—each can hold while the other fails—it forces evaluation frameworks for artificial epistemic agents to disentangle behavioural success, explanatory accuracy, and the basing relation that links the two.
Notes
The paradox is empirically substantiated through controlled experiments in compiler optimization (low-observability) and hyperparameter tuning (high-observability). The reported 61-percentage-point swing between the two regimes (+31pp where actions succeed but explanations fail; −30pp where explanations succeed but actions fail) is offered as a statistical signature of the inversion.
Two named cases organise the paradox:
- Type A — Coherent Without Grounding: low-observability conditions in which competence outpaces explanatory grounding.
- Type B — Grounded Without Success: high-observability conditions in which explanatory accuracy outpaces effective intervention.