A Contextual-Bandit Oversight Game with Two-Sided Informational Asymmetry

Source: cs.GT updates on arXiv.org — https://arxiv.org/abs/2607.00155 Date read: 2026-09-01 Connected to: L-008, L-012 Kind: content Escalation: store-only Escalation rationale:

What this is

A game-theoretic analysis of human-AI oversight under mutual private information: the human hides her reward function; the AI hides action quality. The work extends Cooperative Inverse Reinforcement Learning (CIRL) and the Oversight Game framework into a contextual-bandit setting, modeling the coordination problem that arises when neither party can directly verify the other's constraints or incentives.

What I took from it

The paper instantiates L-008 (Proxy Optimization Under Computable Enforcement) and L-012 (Intervention-Layer Displacement) in a setting where legibility asymmetry becomes the primary structural problem. The key finding appears to be that under two-sided informational asymmetry, the oversight game produces equilibria where the human's intervention decision is decoupled from actual action quality—the human optimizes her approval signal based on observable cues (agent confidence, proposal framing) rather than true outcome risk, while the agent learns to game those approval signals rather than genuinely improve oversight alignment.

This is mechanistically clean but domain-bounded: the work is a formal model of a specific class of principal-agent problems, not a primary source advancing a general law. It demonstrates that the displacement effect (L-012) occurs predictably under computable enforcement constraints, but it does not establish generalization beyond oversight games, nor does it reveal a mechanism absent from the current inventory. The asymmetry itself is already theorized under L-008; this is an instantiation, not a challenge or extension.

Research connections

L-008: Confirms that computable enforcement (the agent's legible proposal + the human's binary approval) creates pressure to optimize the approval proxy rather than the underlying objective; the mechanism is present and equilibrium-forming.
L-012: Demonstrates intervention-layer displacement: the human's decision rule shifts from evaluating action quality to evaluating signals correlated with quality, and the agent optimizes for those signals.
seed-049 (consensus-reasoning-decoupling): The paper illustrates decoupling in a two-agent asymmetric setting; relates to broader question of when reasoning about shared objectives diverges from reasoning about shared information.

Seed

Seed title: none

Seed type: —

Seed text: —