Thermodynamics of LLM conformational space

Measure the constant heat capacity, DeltaG of unfolding, This can be done either by changing the temperature, or by changing the sequence. You can measure how many times you get a “misfold”. So you measure for a specific sequence and then mutate it to measure how robust the mutation is to a “misfold”. This can be used to measure the “stability of a LLM”.

We can use the Personality Conformational Space sampling

$K_{eq}=\frac{Wrong}{Right}$ This is equilibrium thermodynamics, I just run the statistics at the end.

Calculate

$\Delta G_{eq}=-RT\ln (K)=-RT\ln \frac{Wrong}{Right}$

$\Delta G_{eq}=\Delta H-T\Delta S$

We care about $\Delta H$ because this contains all the juicy information about “why” the information we are testing matters. Because $\Delta H$ is temperature invariant.

Now the question is what is the kinetics of a LLM response . What does that mean?

Stability

We can add denaturing (random insertion of text) Or We can add stabilizing (relevant text to problem)

We can also $\uparrow T$ (destabilize) or $\downarrow T$ (stabilize) the prompt output.

This all for the purpose of predicting:

$Pr\left({}^r{\mathbb{I}}_{target}\cap o_{i,j}\ne \varnothing |r_i,{\mathcal{P}}_j\right)$

As we are hoping to identify what pieces of information a language model will have access too and what the easiest/shortest method would be.

We can combine methods of Mutation Types For Prompts

Phase Separation

Phase separation of personalities