The Value of Mechanistic Priors in Sequential Decision Making

Itai ShufaroTechnion, Gal BenorTechnionShie MannorTechnion, NVIDIA Research · 2026

Read The Value of Mechanistic Priors in Sequential Decision Making by Itai ShufaroTechnion, Gal BenorTechnionShie MannorTechnion, NVIDIA Research (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-10018.txt.

Description

Shufaro and two colleagues studied mechanistic priors in step-by-step choice making. Hybrid models with physics-based priors and learned extras promise to cut the data needed for good choices, but have no usable test. The paper defines mechanistic info as shared info between a model's picked policy and the best policy, scored via a use-weighted bias.

In the large-N late stage, bounds show Bayesian regret scales with residual entropy, giving an in-theory hardness cut versus a blank start. The paper also gives a model check for seen yield. In the small-N burn-in regime, which matters in clinics, the paper sets a floor under sure-prior penalties. It tests the ideas on 5-fluorouracil dosing runs from FOLFOX data, where a hybrid prior gains in burn-in. LLMs as priors suffer severe mechanistic-information losses, pointing to physics-based priors for safety-first use.

Catalog record

CREATORItai ShufaroTechnion, Gal BenorTechnionShie MannorTechnion, NVIDIA Research
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE84 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.10018

Browse the full library · About the Tower of Babel