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

Linqian Wu, Diaoxuan HearXiv 2026 · arXiv

TL;DR

This paper rigorously examines whether mechanism-level properties—like the integrated information ($\varphi$) used to assess machine consciousness—can actually be measured from finite, noisy experiments. The authors identify three fundamental barriers (behavioral blindness, hidden units, and mathematical discontinuities) but demonstrate that while precise point estimates are often impossible, one-sided "certificates" can reliably prove a system meets a minimum threshold of integration.

What Problem Does This Paper Address?

As AI systems become more complex, we increasingly judge them not just by what they do (behavior), but by how they do it (mechanism). This is critical for AI auditing, mechanistic interpretability, and controversial assessments of machine consciousness. Integrated Information Theory (IIT) is a leading, albeit debated, framework that quantifies this "how" via a metric called $\varphi$.

However, there is a massive gap between the mathematical definition of $\varphi$—which requires perfect knowledge of a system's causal transition law—and the reality of scientific experiments, which only have access to finite, partial, and noisy data. The authors ask: Is $\varphi$ actually identifiable? Is it stable enough to be measured? And can we ever be certain of its value given finite data?

Methods

The authors move away from large-scale approximations and instead focus on small, stochastic networks of four and five binary units. This allows them to compute the "ground truth" $\varphi$ exhaustively and check every statistical inference against it.

They employ a "design-first" approach, simulating experiments that:

  1. Intervene: Prepare the system in a known state.
  2. Record: Observe a subset of units after a specific delay.
  3. Infer: Build models that fit the observed data.
  4. Certify: Use statistical tools (like Clopper–Pearson and Hoeffding intervals) to make one-sided claims about $\varphi$.
The experimental workflow and the three barriers identified by the authors.

Datasets

Rather than using standard machine learning benchmarks, the study utilizes synthetic stochastic systems where causal laws are known:

  • Parity Networks: Systems where units fire based on the parity of their neighbors, allowing for closed-form solutions of $\varphi$.
  • Random Logistic Networks: 400 random networks with excitatory or mixed-sign weights to test the prevalence of mathematical "cliffs" in the $\varphi$ landscape.
  • Same-Function Pairs: Specifically constructed pairs of recurrent and feed-forward networks that produce identical input-output behavior but have different internal mechanisms.

Key Insights and Ideas

  • Behavior is Blind: The authors prove that for any recurrent network, one can construct a feed-forward "unfolding" that behaves identically but has zero integrated information ($\varphi=0$). This means behavioral testing alone can never certify integration.
  • The Hidden Unit Floor: If even one unit is unobserved, there is a "precision floor." No matter how many trials you run, you cannot resolve $\varphi$ beyond a certain width (e.g., 0.30–1.02 bits in their examples). More data doesn't help; only better access (recording more units) does.
  • The "Cliff" Problem: The way $\varphi$ is calculated involves selecting a "minimum information partition." This selection process makes $\varphi$ mathematically discontinuous. An infinitesimal change in a system's parameters can cause the "best" partition to flip, resulting in a jump of 25–100% in the reported $\varphi$ value.
  • Certificates over Estimates: Because of these cliffs, trying to estimate the exact value of $\varphi$ is often a fool's errand. However, the authors show you can still certify that $\varphi$ is at least some value $c$ by proving the entire region of mechanisms compatible with your data stays above that threshold.
Illustration of how $\varphi$ contains "cliffs" where small parameter changes lead to massive jumps in the metric.

Key Results

  • Recurrent vs. Feed-forward: A recurrent network with $\varphi = 0.28$ bits and a feed-forward chain with $\varphi = 0$ were shown to have identical input-output laws. While behavior saw no difference, internal recordings could separate them in as few as 4 to 53 trials.
  • Quantized Jumps: In 400 random networks, the authors found 849 "edges" or cliffs. 99% of these edges changed $\varphi$ by at least 25% of its value.
  • Robust Certification: The authors successfully proved $\varphi > 1.004$ bits throughout a 64-dimensional region of mechanism space. Certifying a system at the center of this region required roughly 7.5 million controlled transitions to reach a 95% confidence level.

Limitations and Future Work

  • Scalability: The exact computation of $\varphi$ is computationally expensive, growing super-exponentially with the number of units. Applying these rigorous certification methods to large-scale AI models remains a major challenge.
  • Structural Knowledge: The "certificates" rely on having a good model class (e.g., knowing the wiring). If the underlying model is wrong, the certificate may be invalid.
  • Simplified Sensors: The study uses a relatively simple sensor noise model. Real-world neural recordings or AI logging may involve more complex distortions.

The authors suggest that future work should focus on applying these tools to coarse-grained models or tractable proxies of $\varphi$ to see if the "barriers" remain as formidable at larger scales.

Summary updated

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. We ask when such a quantity is identifiable from an experiment's access, continuous in the mechanism and certifiable from finite data, and answer exactly for a deliberately demanding case - the integrated information $\varphi$ of integrated information theory - in stochastic networks of four and five units with known ground truth. 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, yet internal recordings separate them within 4-53 trials. Partial access can be just as blind: in our examples, units left out of a one-step recording impose precision floors of 0.30-1.02 bits that no sample size lowers, and in all-but-self parity networks, recordings of any length see several hidden units only through the product of their fidelities - a symmetry that one intervention broke in exact checks. The measure has sharp edges: because $\varphi$ selects a partition, and the cause and effect it specifies, by one criterion and reports another, infinitesimal changes of mechanism that cross an edge change it by 25-100% at 99% of the edges we found in random networks, and about 1% of surrogate models that fit interventional data almost perfectly misreport it by more than 20%. One-sided claims survive: the same data certify a lower bound on both sides of an edge, and we prove $\varphi>1.004$ bits throughout a 64-dimensional region of mechanisms and certify it from about $10^7$ controlled transitions with false-certificate probability at most $α$. Measurability is thus a design question with computable answers.
BibTeX
@misc{wu2026same,
  author = {Linqian Wu and Diaoxuan He},
  title = {{Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information}},
  year = {2026},
  eprint = {2611.08469},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.08469},
}
Figure