IFS

Concepts

The Internal Family Systems is a psychotherapy proposed by Richard Schwartz in the 1980s. In this model, everybody is a combination of different parts like those characters in the film, Inside Out. But the difference is that only emotions are anthropomorphized in the animation while in IFS, a part can also be a piece of thought or mindset. For example, if you have a social appearance anxiety, there would always be a voice in your mind criticizing your appearance. In IFS we consider the source this voice as a part of your mind. It is kind of like the dramaturgical theory with IFS focusing more on internal feelings and dramaturgical theory emphesizing more on social characters.

(Source: Pixar/Disney-Pixar, via Associated Press)

When some part of your brain takes control of your central console, you behave like that part. So it is reasonable of one person saying some disgruntled words in a tavern to be like a gentleman in front of his child. Both of them are real. Both are an aspect of that person. Occasionally, these parts have conflict opinions and fight for the control of body. When such turbulent scuffle occurs, it’s like there are two debating voices in your mind. And that’s the source of annoyment and trouble.

No specific part in our mind on the console means we are in a sedate status and IFS calls it self. In the theory, the self has characteristics with the mnemonic “8 Cs” which represents Calmness, Compassion, Curiosity, Courage, Connectedness, Confidence, Creativity and Clarity. In concise words, reaching the self state is like in the zen mode.

Then let’s delve into details of the parts in our body. Richard Schwartz thinks there are two kinds of parts(I think what he means is passive parts) – exiles and protectors.

When we are severely hurt, a part would undertake this pain. Our body does not want to feel this pain, so this part is exiles to a hidden liminal space. This is what exiles mean. The psychological age of exiles will be fixed at the moment of exile. For example, you were scolded and beaten at the age of 8 because those stuff in your room was jumbled up, which left you a mysophobic exile. After growing up, whenever you endure some plight and cannot control yourself, the exile would escape from the prison and you have a fanatic impulse to arrange all the things around you in order.

A protector tries to reduce the perception of pain by dissociating, numbing emotions, diverting your attention, etc. The bahavior pattern of defenders will also be frozen in the past, using the anachronistic policy to protect themselves. Protectors are catogorized into two types:

  • Manager: Manage behavior plans. Negative ones can lead to obsessive-compulsive disorder, workaholism, and perfectionism.

  • Firefighter: Put out the fire when the pain recurs. Negative ones involve overeating, criticizing others, and addictive behaviors.

Goal

The goal of IFS is not to expel those passive parts but to learn the correct interactive relation in your internal family bacause those passive parts also want to help you but they fail to find the right way – for instance, pessimism protects you from being hurt and self-criticizing is eager to see a better yourself.

+-----------------------------------+---------------------------+
|              Concepts             ‖          Features         |
+===================================+===========================+
|                Self               ‖        love & peace       |
+-------+---------------------------+---------------------------+
| Parts |           Exiles          ‖      pain, depression     |
|       +------------+--------------+---------------------------+
|       | Protectors |   Managers   ‖      OCD, mysophobia      |
|       |            +--------------+---------------------------+
|       |            | Firefighters ‖ substance abuse, gluttony |
+-------+------------+--------------+---------------------------+

From the Lens of AI

A message contains both semantics and syntax. We can express the same semantics in different ways. For example, in ancient China, there is a book The Mathematical Classic of Sun Zi recording a number theory problem:

今有物不知其数，三三数之剩二，五五数之剩三，七七数之剩二，问物几何？

whose English meaning is

There are a number of things. But we do not know the number. When divided by 3, it leaves a remainder of 2. When divided by 5, it leaves a remainder of 3. When divided by 7, it leaves a remainder of 2. How many things are there?

We may ask the same problem in variable ways with the same meaning. And the answer to the question can be “二十三” in Chinese or “twenty three” in English based on which language is used in the question.

In the following post, we can also call semantics as meaning and syntax as “carrier”.

There is a world model in everybody’s mind predicting future. Sometimes this model would sink into some special state in which its prediction to the future is faster. That state is the mechanics of emotion or in a more general word parts. For example, a self-contemptuous person is likely to have a protector whose mindset is merely criticizing themselves instead of analyzing what’s the most important part of this problem and what you can do to tackle the predicament objectively. Such an energy-saving but emotional mindset that is easy to trigger produces a glooming future dominantly whenever they come across any obstacles.

Especially, let us regard this world modal as a general LLM. Different internal characters are actually those distinct hidden state activated by distinct context, controlling the style of the output sentences. However, semantics and syntax are not totally orthogonal to this model, which means the inputs with the same semantics and different syntax could result in semantically varying outputs. It is like the model is controlled by part of its network that was formed based on its past experience or rote memory, instead of meditating on the problem objectively.

We try our best to get rid of the influence that those carriers leave us to keep our predicition from different parts coherent. And the mindset generated from a distribution not influenced by posteriori, a more rational mindset , is what makes self. The exiles are painful and unforgetable memory. That’s why we are going to expel them away. Protectors are those neural circuits that were formed in pains, mostly in childhood when the brain was growing and evolving fast. These circuits are stubborn and conservative because they were developed in early stages of the neural network’s training process which made them components of the fundamental pilars of the whole complex system. It is like a low-level function that is frequently invoked in billions lines of codes. Even if the function proves fallible, we just cannot cast away this snippet – you never know whether there is any subtile code lurking in the shadows, depending on this function’s fallacy to work. If we view the brain development in childhood as pretraining and learning new knowledge in adulthood as finetuning, as shown in this research, finetuning would not create new circuite but reuse previous ones, revealing the developmental conservatism of protectors.

IFS posits that the source of pain lies in the conflicts between internal family members. This can be explained as incongruity in the world model’s predictions about the future. Sometimes, the part of the world model accustomed to negative outcomes gets activated, leading to lower reward predictions, while at other times, another part may generate higher rewards. In the context of LLMs (Large Language Models), this can be seen as the model’s output incongruity in semantics when the semantics of the input remains the same but the syntax varies. Some researchers define the degree of such incongruity as “semantic entropy”, which is studied in relation to hallucinations.

(Source: the “semantic entropy” paper)

In practice, loving every internal part involves letting each part express its true feelings. This corresponds to rational attribution, understanding the reasons behind certain thought patterns. From a neural perspective, we have triggers in our minds, and when these triggers are activated by certain patterns in the input, some parts are brought to the forefront. Adopting an antagonistic attitude towards these parts only increases the weight on the trigger, making it easier for the part to be activated. In a short-term therapy, we can’t immediately eliminate a part due to its conservative nature; instead, we can only perform some fine-tuning work. We add a meta-trigger to the initial trigger, which activates when the original trigger is activated. Being aware of the existence of our parts is our first step to reduce their influence. This helps us identify when a part has come to the forefront and choose to ignore the quick and shortcut-like conclusions given by the parts in favor of rational thinking, thus immersed into a state of self.

Digression

It is important to note that IFS does not guarantee a cheerful, confident, and optimistic psychological state; it only ensures a congruent personality. A bad person or a psychopath can also be congruent. For instance, in Raging Loop, Fusaishi Haruaki’s theory about villains is that they are bad, by nature, innatedly, liberally, unrestricted, without concern for social morality – badness is a villain’s purest state. Therefore, a bad person can achieve IFS congruence with all their parts entirely evil as well. Psychotherapy usually presupposes a cheerful, confident, and optimistic psychological state, which may be true for most people. However, from an evolutionary perspective, a person who derives pleasure from the suffering of others is also naturally possible, much like a person with six fingers due to a genetic mutation. A congruent bad person is entirely possible within the IFS framework.

(Source: Raging Loop CG)