Many recent and legacy advances in AI often split the responsibility of a learning system into a primary agent and a "secondary" agent (or dual) agent to maximize the primary objective. Lets call this the dual-agent optimization paradigm, where two processes interact to improve learning. Think Actor-Critic, Discriminator-Generator, GAN, Convex optimization which has a primal and to solve we form the dual which relaxes the primal, quasi-Newton L-BFGS, Adaptive Stress Testing, MARL (Multiagent Reinforcement Learning), and more recently learning to learn by gradient descent by graident descent.. In all cases, one “primary” agent works on a main task (e.g., classifying images or predicting text), while a “secondary” agent optimizes a secondary "dual" objective (such as guiding learning or adversarially testing the primary agent). This interplay is very powerful. Some have speculated that such dual-agent dynamics are analogous to how human consciousness arises, with a core learner guided by a meta-level process. But I have yet to see anything mapping all of these together to the same paradigm?
Many machine learning algorithms can be viewed as having two agents or levels optimizing different objectives:
The concept of dual and primal problems is fundamental in convex optimization. The primal problem represents the original optimization objective, while the dual problem is derived by forming constraints and Lagrange multipliers to reformulate the problem. The optimal solutions of both are deeply related through strong duality in convex settings (Boyd & Vandenberghe, 2004). This provides a mathematical grounding for dual-agent learning frameworks where the primary agent optimizes a base task, while the secondary agent optimizes constraints or an alternate objective.
Why do these methods work so much better? Is this what is going on in our brain? Could consciousness arise from a similar split-architecture; a primary cognitive process coupled with a meta-level process? Is this supported in consciousness literature? I wanted to know! Here is what I found.
Global Workspace Theory (GWT) posits that many unconscious processes run in parallel, and consciousness arises when a subset is broadcast globally (Baars, 1988). This resembles a multi-agent system, where secondary processes direct attention and guide the primary agent’s focus.
Integrated Information Theory (IIT) suggests that consciousness corresponds to the integration of information within a system (Tononi, 2004). The dual-agent paradigm increases this integration, making the whole system less decomposable.
As per Douglas Hofstadter in GEB, there are two agents in a strange loop where one agent recognizes the other, and visa versa.
Julian Jaynes discusses the right lobe going "dormant" to allow the left lobe (in wernicke's area specifically) to program the right and "quell the voices of the god's" which was the right lobe. This submission of the right lobe allowed one to be programmed and to allow consciousness and not a deferral to "god" (the right lobe/instincts).
Before we assert that this is a "necessary" condition of consciousness, we have to prove that LLMs that do not have a dual agent structure do not have consciousness. Well we need a test. Too much to discuss on the topic and everytime I talk to philosophy phds on this topic they say 'dont say the c-word'.. lol. So not much help there. But every test that has been suggested (the mirror test, turing, etc) has been debunked. Perhaps its simple. Do you know you exist? Consciousness and self awareness are the same thing. Thats it. Consciousness is simply the following property: are you modeling yourself (your own actions) in your model of the world. Do you know what impact your actions will have on your environment? Do LLMs have that property? Do they know whats going to happen if they tell you how to tie a nousse or something worse? I don't think so.. but what do others say on this topic?
No one really knows and I do not find these answers compelling in either direction TBH.
A dual agent paradigm is any training setup where one agent is trying to learn the primary objective such as classifying images or next token prediction and the other is trying to learn a dual objective which is how good it is at training the primary agent on the primary objective. In the case of LBFGS we are fitting the local loss function of the primary objective with a paraboloid and then jumping to the minimum of that paraboloid. That fitting is a secondary model/agent (dual agent) that is aiding the learning of the primary agent.. In the case of GANs, adaptive stress testing, etc the dual model or agent is more explicit but has a secondary objective which is to help the primary agent achieve the primary objective as much as possible. Perhaps this is much like what the conscious brain is doing to the subconscious brain. The conscious brain (the dual agent) emits probe directions into the subconscious brain (the primary agent). A probe direction (like in MeZO or Stochastic Gradient-Free Descent) is an educated guess of how to to change the primary agent to get better at the primary objective. The goal of the dual agent is to get better at that educated guess measured as the improvement. It is able to achieve in the primary agent. If we were to apply a mezo (memory efficient zero order) style approach to optimization where we pick a random probe direction, temporarily step the primary agent into that probe direction both plus and minus or just one way from where it is now to a random direction, then check some attributes about the local landscape in that perturbed (or dirty) step direction, then make an educated guess based on that information of the gradient estimate, and then using that educated guess, then we step the primary agent. Now if we make this more complex and add more information into the educated guess. Instead of random perhaps we use a fitting function to fit a hyperplane of 100 perturbations And step based on the hyperplane.. let’s make this more complex so we sample random directions, get loss on those perturbed directions, and have it fit a paraboloid (like BFGS or other quasi-newton methods) and then our dual agent fits it, and steps or jumps to the global minimum of the paraboloid. let’s make this even more complex and now use a neural network to learn a much lower dimensional agent to simulate the primary loss landscape and we train it to learn this surrogate loss function which approximates the real loss landscape, mapping from the design space to a scalar space, taking in probe directions and outputting an approx loss difference, and then we can search with the dual agent quickly but randomly perhaps and then select the most promising directions based on our predicted estimated loss trained on our dual objective, and then use that as our sampling procedure for the next 100 agents NOT subject to any input data just that’s it’s biased sampling directions. We could also have the dual agent follow a diffusion process where it inputs a current point in the design space and outputs probe direction gradient guesses directly. And the dual loss function is how good those probe directions are. The dual agent can use the full Stochastic gradient decent but the primary agent is far too big for SGD. so in this way, the dual agent is the conscious. It is thinking how to change the primary agent a.k.a. the policy so that next time we do better on the primary objective, it is like program induction. The dual agent is the programmer which humans have in their prefrontal cortex. It tries to reason about the samples generated by testing the local loss landscape and how to improve based on the geometry of the local loss landscape.
TLDR, I have been pondering.. is my conscousness just my dual agent which is trying to train/optimize the primal agent? Is this the limiting step inside our architectures? Is this what Julian Jaynes was talking about in the Bicameral Mind? Is this what GEB means when they talk about a "strange-loop"? Separating the networks allows for a clear distinction between "the programmer" and "the program". Moving from transduction to a more meta ability like program induction.
TLDR; I really like the dual-agent paradigm although it is not used in LLMs today. It is a compelling lens for thinking about consciousness, perhaps this is the missing link. Future AI systems may incorporate such 2 agent architectures resembling our own 2 lobes, but whether that leads to consciousness or just better simulations remains an open question. I want to train a "sovler agent" thats objective is to train a "primary agent" like L2L w/ GD w/ GD but not HPPs.. actually outputting synthetic gradients! Need to test!