This article explores the differences and similarities between Active Inference from the Free Energy Principle and LLMs (Large Language Models), based on the paper "Predictive Minds: LLMs As Atypical Active Inference Agents."
- Introduction
- Summary in One Sentence
- Paper Notes: Predictive Minds: LLMs As Atypical Active Inference Agents
- Conclusion
- References
Introduction
As I continue using LLM like ChatGPT, I find these tools genuinely useful, yet I can't help but reflect on how they differ from living organisms and humans. I'm particularly interested in exploring their relationship with the Free Energy Principle/Active Inference, which is a hot topic in neuroscience.
When I searched for papers discussing these relationships, I found several, though they're still limited in number.
Today, we'll examine the differences and similarities between Active Inference and LLMs through one such paper: "Predictive Minds: LLMs As Atypical Active Inference Agents."
- Title: Predictive Minds: LLMs As Atypical Active Inference Agents
- Authors: Jan Kulveit, Clem von Stengel, Roman Leventov
- URL: https://arxiv.org/abs/2311.10215
Note: This article was translated from my original post.
Summary in One Sentence
The difference between LLMs and Active Inference is not as fundamental as previously argued, and can be viewed as quantitative rather than qualitative. In LLMs, the loop between action and resulting perception is barely closed (or the time delay is too long and the influence too weak), but by improving this aspect, we can develop Active LLMs that behave more like Active Inference agents.
Paper Notes: Predictive Minds: LLMs As Atypical Active Inference Agents
Below are my notes from the paper. Some sections include my own interpretation as supplementary explanation.
For complete details and accuracy, please refer to the original paper.
Background
How Have LLMs Been Conceptualized?
There has been ongoing debate about what happens inside LLMs and whether they truly "understand" the world.
Some argue that LLMs are merely passive predictors—statistical parrots—while others propose that since they learn from language describing the world, world models must emerge within LLMs. A diverse range of theories has been put forward.
Active Inference and Predictive Processing
Active Inference is a theory rooted in cognitive science and neuroscience.
Biological systems, including living organisms and human brains, constantly update their internal models while acting on the external environment. These two processes—action and internal model updating—can both be explained as processes that minimize the difference between predictions and actual sensory input (minimizing free energy).
Active Inference and LLMs/Generative AI are often considered fundamentally different, but the authors challenge this assumption.
Differences and Similarities Between Active Inference and LLMs
LLMs as a Special Type of Active Inference System
LLMs consume text from the internet and construct their internal models. In other words, the LLM learning process can be understood as a form of "perception" for the LLM.
LLM hallucinations can also be understood from an Active Inference perspective (*).
- I didn't fully grasp the hallucination-related discussion, so I'll omit it here.
What Constitutes Action in LLMs?
When discussing the differences between Active Inference and LLMs, the key distinction has traditionally been said to lie in the passivity of LLMs. In other words, the fundamental difference is that LLMs cannot act on the external environment.
However, the authors argue that this difference is quantitative rather than qualitative.
While it's true that LLMs cannot take physical actions like living organisms or robots, couldn't we say they are "acting" in the sense that LLM outputs ultimately affect the external environment?
The "external environment" here refers to the object of LLM perception—namely, the internet text that constitutes the LLM's training data. LLM outputs might be directly posted on the internet, or they might change the behavior of humans using the LLM, ultimately influencing internet data.
Output tokens from LLMs are, in essence, "micro-actions," and their accumulation ultimately affects the world, eventually influencing LLM perception—that is, training data.
Closing the Action-Perception Loop in Active Inference
Compared to living organisms that constantly run action-perception loops, the LLM action-perception loop cannot be said to be closed (or it takes far too long to close). LLMs typically undergo retraining—that is, perception—only once or a few times per year, and even during that training, they don't adequately receive feedback from their own outputs.
So how can we bridge this gap between action and perception? Here are some possible approaches:
- Use the model's output for training the next generation model, without adjusting or filtering the data.
- Use interaction data with the model for fine-tuning.
- Implement continuous online learning.
In any case, by closing this action-perception loop gap, model developers could realize LLMs as more autonomous, adaptive agents.
What Active LLMs Mean
As the action-perception loop in LLMs closes and they behave more like Active Inference agents, the model's self-awareness is expected to improve. The model will perceive more information about itself and strengthen its self-awareness by observing the results of its actions in the external environment.
Conclusion
We've examined the differences and similarities between LLMs and Active Inference based on the paper "Predictive Minds: LLMs As Atypical Active Inference Agents."
When I consider actually implementing the suggestions from this paper, I realize I still lack fundamental knowledge about both LLMs and Active Inference. I'll continue learning step by step.
[Related Articles]
