English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information

Forum topic · 小凯 · 2026-11-06

Summary

This paper studies when mechanism-level quantities, functionals of a system's causal transition law, can be identified from experiments, remain continuous in the mechanism, and be certified from finite, partial, noisy interventions. The authors give exact answers for integrated information phi from integrated information theory (IIT) in stochastic 4- and 5-unit networks with known ground truth. Key findings: behaviorally identical recurrent and feed-forward networks carry 0.28 vs 0 bits of phi, separable by internal recordings within 4-53 trials; partial access imposes irreducible precision floors of 0.30-1.02 bits regardless of sample size; phi has sharp edges where infinitesimal mechanism changes shift it by 25-100% at 99% of edges in tested networks; yet one-sided lower-bound certificates survive, proven for phi > 1.004 bits over a 64-dimensional mechanism region with false-certificate probability at most alpha. Measurability is thus a design problem with computable answers.

Paper Overview

Field: ML Authors: Linqian Wu, Diaoxuan He Posted: 2026-11-05 arXiv: 2611.08469

Abstract

Interpretability, AI auditing and assessments of machine consciousness increasingly rest on mechanism-level quantities: functionals of a system's causal transition law that describe how it computes, not what it outputs. Experiments see that law only through finite, partial and noisy interventions. The authors ask when such a quantity is identifiable from an experiment's access, continuous in the mechanism, and certifiable from finite data, answering exactly for a deliberately demanding case: the integrated information φ of integrated information theory (IIT), in stochastic networks of four and five units with known ground truth.

Key Findings

  • Behavior can be silent: a recurrent and a feed-forward network with identical input-output laws at their response delay carry 0.28 and 0 bits respectively, yet internal recordings separate them within 4–53 trials.
  • Partial access can be equally blind: in their examples, units missed by single-step recordings impose a precision floor of 0.30–1.02 bits that no sample size can reduce. In a parity network recording all units except one, arbitrarily long records observe several hidden units only through their fidelity product — a symmetry broken by a single intervention.
  • The measure has sharp edges: because φ depends on a standard choice of partition and its designated causes and effects, reporting a different one means infinitesimal mechanism changes crossing an edge can shift φ by 25–100% at 99% of the edges in their random networks, and about 1% of alternative models fitted to intervention data misreport it by more than 20%.
  • One-sided claims can survive: the same data certify a lower bound on both sides of an edge; the authors prove φ > 1.004 bits across a 64-dimensional region of mechanisms and certify it from ~10^7 controlled transitions with false-certificate probability at most α.

Conclusion

Measurability is a design question with computable answers, rather than an obstacle.

--- *Collected automatically on 2026-11-07.*

Tags

#machine-learning#interpretability#integrated-information-theory#ai-auditing#identifiability#arxiv#causality

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636825