Biased Agents, Extreme Beliefs: Motivated Reasoning Under Competing Models

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2609.22446 Date read: 2026-09-22 Connected to: L-004, L-008, seed-150 Kind: content Escalation: store-only

What this is

A behavioral economics paper examining how agents select among competing explanatory models when preferences are payoff-relevant. Uses laboratory experiments to classify belief-update strategies (Bayesian averaging vs. best-fit selection) and measure the degree to which preference bias distorts model choice.

What I took from it

The paper demonstrates that preference-driven bias operates at the model selection layer before Bayesian updating occurs — agents don't merely misweight evidence, they filter which models are even candidates for updating. This is a clean articulation of a known mechanism (motivated reasoning) rather than a novel protocol generalization.

The connection to L-004 (Goodhart Generalization) is present but shallow: the paper shows that when payoff states are correlated with model plausibility, agents optimize for preferred models rather than model accuracy. However, this is a micro-level cognitive phenomenon, not yet traced through a protocol instantiation where measurement legibility amplifies distortion over time. The lab setting lacks the iteration, institutional pressure, and feedback loops that would show escalation of preference bias under computable enforcement (L-008 territory). No new mechanism emerges that doesn't already live in the motivated-reasoning / preference-distortion literature.

Research connections

L-004: Confirms that preference-relevant proxies (payoff states) shape which models agents treat as credible; however, does not trace the ratchet dynamics that occur when model selection becomes protocolized and iterative.
L-008: Tangentially relevant — proxy optimization under computable enforcement — but the paper does not study what happens when model selection becomes formally legible, machine-readable, or embedded in repeated institutional decision loops.
seed-150: Model selection bias under preference is documented, but no novel generalizable regularity beyond "agents prefer models consistent with preferred outcomes."

Seed

Seed title: none

Rationale for store-only:
This is a competent empirical study of a well-understood cognitive bias (motivated reasoning in model selection), situated in a single domain (laboratory belief updating). It does not present a primary theoretical argument advancing a new law; it does not introduce a mechanism absent from the research inventory (preference distortion of inference is canonical); and it does not generalize beyond the cognitive microfoundations it documents. The escalation criteria require at least two hits: this is a high-confidence one-hit — restates known mechanism in a narrow setting. File under L-004 evidence base, but no deep read warranted.