📊 statistics

Active Inference with People: a general approach to real-time adaptive experiments

This paper introduces a unified, real-time adaptive experimental framework that combines active inference with the PsyNet platform to optimize diverse behavioral studies—such as adaptive testing and treatment assignment—across multiple modalities, significantly improving efficiency and accuracy compared to conventional fixed designs.

Original authors: Lucas Gautheron, Nori Jacoby, Peter Harrison

Published 2026-04-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lucas Gautheron, Nori Jacoby, Peter Harrison

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Smart Shopper" of Experiments

Imagine you are a shop owner trying to figure out exactly how much money a customer has in their pocket.

The Old Way (Static Experiment):
You have a standard set of 15 questions to ask every single customer: "Do you have a penny?" "Do you have a dollar?" "Do you have a million dollars?"

If the customer is a billionaire, asking about pennies is a waste of time.
If the customer is broke, asking about millions is confusing and annoying.
You ask everyone the same 15 questions, regardless of what they've already told you. This is slow, boring, and inefficient.

The New Way (Adaptive Experiment):
You hire a "Smart Shopper" (the computer algorithm). This person asks one question, listens to the answer, and immediately decides the perfect next question based on that answer.

If you say "I have a dollar," the Smart Shopper skips the penny question and asks, "Do you have a five?"
If you say "No," they stop asking about money and move on.
Result: They figure out your exact net worth in half the time, with half the questions, and you (the customer) are less bored.

This paper introduces a new, unified way to build these "Smart Shoppers" for all kinds of scientific experiments, not just money questions.

The Two Main Ingredients

The authors combined two powerful tools to make this happen:

1. Active Inference: The "Curious Detective"

Think of Active Inference as a detective who has two goals:

To learn the truth (Epistemic): "I need to know if this suspect is guilty."
To get a specific result (Pragmatic): "I need to catch the bad guy before they escape."

Most old methods were like detectives who only wanted to learn the truth, even if it took forever. Others were like detectives who only wanted to catch the bad guy, even if they didn't understand the crime.

Active Inference is special because it balances both. It asks: "Which question will teach me the most about the truth AND help me solve the case right now?"

Analogy: Imagine you are playing a video game. You want to find the treasure (the goal), but you also want to learn the map (the knowledge). Active Inference is the strategy that tells you exactly which path to take to get the treasure fastest while still learning the map along the way.

2. PsyNet: The "Universal Remote Control"

Even with a brilliant detective, you need a way to talk to the people in the experiment. That's where PsyNet comes in.

Analogy: Think of PsyNet as a universal remote control for the internet. Whether you are showing people a picture, playing a sound, asking a text question, or showing a video game, PsyNet handles the technical "plumbing."
It allows researchers to plug in their "Smart Detective" (the math) and instantly start running experiments with thousands of people online, without needing to be a coding wizard.

The Two Experiments in the Paper

The authors tested their idea with two different "games" to prove it works.

Experiment 1: The "Quiz Show" (Adaptive Testing)

The Goal: Figure out how smart a person is at trivia (like Solar System facts) as quickly as possible.
The Problem: If you give a genius a question about "What is 2+2?", it's useless. If you give a child a question about "Quantum Physics," it's impossible.
The Solution: The system asks a question. If you get it right, it immediately picks a harder one. If you get it wrong, it picks an easier one.
The Result: They found they could determine a person's knowledge level with 30–40% fewer questions than the old way, without losing any accuracy. It's like finishing a marathon in record time without getting tired.

Experiment 2: The "Magic Trick" (Adaptive Treatment)

The Goal: Find the one specific trivia question that perfectly separates people with a college degree from those without one.
The Problem: In a normal experiment, you might ask 100 people 15 different questions. You waste time asking questions that everyone gets right or everyone gets wrong.
The Solution: The system learns as it goes. If it notices that "Who was the first President?" is the perfect question to tell the difference, it starts asking that question to almost everyone, while ignoring the useless ones.
The Result: The system found the "magic question" 3 times more accurately than the old method. It was like a chef tasting a soup and immediately adding the exact right amount of salt, rather than guessing.

Why Does This Matter?

It Saves Time and Money: Researchers can run experiments faster and cheaper because they don't need as many participants or as many questions.
It's Better for People: Participants don't get bored answering questions that are too easy or too hard. They stay engaged.
It Unifies Science: Before this, "testing ability" and "testing treatments" were done with totally different math and software. This paper says, "Hey, they are actually the same problem!" It gives scientists one toolkit to solve many different puzzles.

The Bottom Line

This paper is like handing scientists a smart, self-driving car for their research. Instead of manually steering through every bump and turn (asking every question to everyone), they can set a destination (the research goal), and the car (Active Inference + PsyNet) automatically figures out the fastest, most efficient route to get there, adjusting to traffic (the participants' answers) in real-time.

1. Problem Statement

Adaptive experiments optimize their design iteratively based on collected data, offering significant advantages over static designs (e.g., reduced trial counts, higher precision, better resource allocation). However, the field is currently fragmented:

Conceptual Silos: Scenarios like Computerized Adaptive Testing (CAT), Adaptive Treatment Assignment (e.g., clinical trials), and Active Learning are often treated as distinct problems with separate algorithms and software frameworks.
Implementation Barriers: Real-time adaptive experiments face substantial computational challenges. Calculating optimal designs (e.g., Expected Information Gain) often requires intractable integrals or heavy Monte Carlo sampling, which is difficult to execute within the tight latency constraints of online human-computer interaction.
Lack of Unification: There is no unified framework that can handle diverse modalities (text, audio, video) and different objectives (pure knowledge acquisition vs. maximizing specific outcomes) within a single software architecture.

2. Methodology

The authors propose a unified framework combining Active Inference (a Bayesian framework from cognitive neuroscience) with PsyNet (a modular Python platform for online behavioral experiments).

A. Theoretical Framework: Active Inference

The paper utilizes Active Inference as a generalization of Bayesian Adaptive Design (BAD). It unifies two distinct goals under a single objective function: Expected Free Energy (G).

Objective Function: $G(d) \approx -[EIG(d) + U(d)]$ $G (d) \approx - [E I G (d) + U (d)]$
- Epistemic Drive (EIG): Expected Information Gain. Maximizes knowledge about latent parameters (e.g., participant ability).
- Pragmatic Drive (U): Expected Utility. Maximizes specific outcomes or minimizes costs (e.g., finding the best treatment, avoiding long tasks).
Mechanism: The system selects the design $d$ $d$ (e.g., the next question or treatment) that minimizes $G$ $G$ .
- In Experiment 1 (Adaptive Testing), the utility term is zero ( $U=0$ ), reducing the objective to pure EIG maximization (standard BAD).
- In Experiment 2 (Treatment Assignment), the utility term is non-zero, allowing the system to balance exploration (learning about treatments) with exploitation (administering the best-known treatment).
Inference Engine: To handle real-time constraints, the authors employ Variational Inference (VI) rather than Markov Chain Monte Carlo (MCMC). VI approximates the posterior distribution using a tractable family of distributions (e.g., independent normals), optimizing parameters via stochastic gradient descent. This allows for rapid updates of the posterior after every participant response.

B. Implementation Platform: PsyNet

PsyNet serves as the execution layer, bridging the gap between complex Bayesian computation and real-time web-based experiments.

Architecture: A modular, object-oriented Python framework built on Dallinger. It supports deployment on cloud servers (AWS, Heroku) and integration with recruitment platforms (Prolific, MTurk).
Modularity: The system separates the Timeline (experiment flow), Trial Makers (logic for selecting tasks), Trials (stimuli/response), and an Optimization Module.
Adaptive Logic: The TrialMaker class is overloaded to call an external optimizer. This optimizer computes the Expected Free Energy for all candidate stimuli and selects the optimal one in real-time, passing the result back to the trial maker.
Scalability: The architecture supports concurrent users and real-time data propagation, essential for updating priors based on the collective data of all participants.

3. Key Contributions

Unified Mathematical Framework: The paper demonstrates that CAT and Adaptive Treatment Assignment can be solved using the same Active Inference machinery, differing only in the specification of the utility function ( $U$ ).
Real-Time Feasibility: By combining Variational Inference with PsyNet, the authors prove that complex Bayesian adaptive designs can be executed in real-time with human participants, overcoming previous computational bottlenecks.
Practical Tooling: The authors provide a fully open-source implementation (Python/Pyro) that allows researchers to switch between static and adaptive designs, or between different optimization strategies, with minimal code changes.
Handling Un-calibrated Data: The approach works effectively even when item difficulties or treatment effects are unknown a priori, learning them dynamically during the experiment.

4. Experimental Results

The authors validated the approach through two experiments involving 200 human participants (plus simulations using an "oracle" dataset where all answers were known).

Experiment 1: Adaptive Testing (Knowledge Assessment)

Goal: Estimate participant ability ( $\theta$ ) using trivia questions with unknown difficulty ( $\delta$ ).
Strategy: Maximize EIG to select the most informative question.
Results:
- Efficiency: Reduced the number of trials required by 30–40% compared to a static design (average 9.6 trials vs. 15).
- Accuracy: The ability estimates from the adaptive design correlated highly with the "oracle" estimates ( $R^2 = 0.96$ ).
- Information Retention: The adaptive design retained 96% of the total information gain achievable by the full oracle dataset.

Experiment 2: Adaptive Treatment Assignment (Discrimination)

Goal: Identify trivia questions that best discriminate between college-educated and non-college-educated participants.
Strategy: Minimize Expected Free Energy, balancing EIG (learning question difficulty) and Utility (maximizing the probability of a correct answer for college-educated participants and incorrect for others).
Results:
- Accuracy: The active inference approach identified the optimal treatment (best discriminating question) up to 3 times more accurately than a fixed design (even sampling).
- Dynamic Balance: The system successfully shifted from exploration (testing various questions to learn their properties) to exploitation (focusing on the best questions) as data accumulated.
- Comparison: In simulations, Active Inference outperformed standard bandit algorithms (Thompson Sampling, Exploration Sampling) when leveraging contextual information (participant education level) was crucial.

5. Significance and Future Directions

Bridging Disciplines: This work breaks down barriers between psychometrics, clinical trials, and machine learning, offering a single "toolbox" for adaptive experimentation.
Scalability: The use of PsyNet and Variational Inference makes these advanced methods accessible to behavioral scientists who may not have deep expertise in computational statistics.
Real-Time Model Evaluation: The paper introduces methods for real-time model checking (reliability diagrams) to detect misspecification during data collection.
Future Work: The authors suggest extending the framework to Deep Adaptive Design (using neural networks to learn policies) and Online Learning (updating posteriors incrementally without re-computing from scratch) to further improve scalability for massive datasets.

In conclusion, the paper provides a robust, unified, and practically implementable solution for real-time adaptive experiments, demonstrating that Active Inference can significantly improve efficiency and accuracy across diverse behavioral science applications.