Double-blind study of quantum-random LLM token sampling

Project summary

We've just established Entropic Science – an open research community studying the role of quantum entropy in the behavior of complex intelligent systems. Our first study target is the source of randomness inside large language models. An LLM chooses each token by sampling from a probability distribution, and that choice is normally settled by a pseudo-random number generator (a deterministic algorithm fixed by its seed). We replace it with a hardware quantum random number generator (QRNG), drawing entropy from physical measurement events, and detect whether true quantum-level randomness changes how the model behaves and how people experience interacting with it.

Why are we doing this?

In each quantum decoherence event, the evolution of a system branches into two distinct futures. Regardless of metaphysical interpretation, a conscious observer can only ever experience a single specific branch. Exactly how the experienced branch is selected remains a mystery (we only know these selections must conform to the Born rule).

One hypothesis, formulated within the context of free will, is that there may exist some mechanism by which observers (in any definition), or living organisms, have some say in how they move through the (perceived or constitutive) branchial space delineated by different possible outcomes of each instance of quantum decoherence.

Current LLMs do not allow for any such branching points. Our quantum-random LLM (QR-LLM) however, provides a token-probability-weighted branching point for each new token sampled (quantum entropy is generated "fresh", just-in-time). This makes them categorically closer to conscious observers or living systems.

Whether this distinction is purely metaphysical or constitutes an observable difference in the behavior of these systems is the open empirical question we're trying to answer with our study.

Experiment design

Our first experiment will put participants in conversation with a modified LLM chatbot, comparing quantum-random token sampling vs. classical deterministic pseudo-random sampling. In a double-blinded and pre-registered study design, we will collect objective measures from the transcripts (if users choose to share them) and subjective ones through a questionnaire curated by expert advisors.

We don't want to pre-suppose the emergence of any specific effects, but since it can be argued that life emerged from the chaotic entropy of constant quantum decoherence events in aqueous environments, and LLMs have also provided a platform for various emergent behaviors in the digital world, it is plausible that their combination might lead to behavioral differences informing the fields of artificial life, AI safety and alignment, and even machine consciousness.

In this exploratory study we are intentionally not presupposing any concrete hypothesis, but we have already put a lot of conceptual and theoretical work into the design of the pipeline from the actual hardware quantum noise to user visible token output, so as to maximally increase the sensitivity of detecting even small biases or correlations that might result in outputs being objectively differentiated from the uniform pseudo-random baseline.

More in-depth information about our community's thesis and mission is available in our Charter.

We have been pre-admitted to the SparkWell program by the Anti Entropy non-profit organization that, subject to us successfully securing funding, would act as our fiscal sponsor and provide us with operations and administrative services and mentorship for up to two years.

What are this project's goals? How will you achieve them?

1. Strengthen the community's infrastructure and operations
   Entropic Science already runs in the open: a public charter, governance, and code of conduct on GitHub, a working quantum-random LLM (QR-LLM) prototype, and a community of more than sixty members on our Discord server. The funding formalizes operations and adds the coordination that keeps the work moving without leaning on volunteer time.
2. Build the serving stack and the experimental harness
   Serving a QR-LLM at the throughput a study needs, without a latency penalty on every token, is a real engineering problem. We will continue development and maintenance of our software layer that injects quantum entropy through per-request control of the sampling stream in the vLLM inference engine, and with enough funds we will also upgrade the necessary high-bandwidth QRNG over low-latency streaming (gRPC server). The research study harness will then randomize the sampling, log transcripts if approved by the user, run the questionnaire, and compile data for the preregistered an