## Diagram: An HMM on Belief Tree
### Overview
The image presents a Hidden Markov Model (HMM) on a belief tree, illustrating the relationships between ground-truth factuality (hidden states) and LLM's prior beliefs (observed states). The diagram shows the flow of information and probabilities within the model, including transition probabilities, emission probabilities, and belief propagation.
### Components/Axes
* **Title:** An HMM on Belief Tree
* **Nodes:**
* `Z_u` (Ground-truth factuality, hidden): Represented by dashed circles with gray fill.
* `S_u` (LLM's prior belief, observed): Represented by solid circles with white fill.
* **Edges:**
* Solid black arrows: Represent transition probabilities `p(Z_{c(u)}|Z_u)`.
* Curved lines with arrows pointing upwards: Represent emission probabilities `p(S_u|Z_u)`.
* Dotted red arrow: Represents belief propagation.
* **Posterior Belief Box:** A rounded red rectangle containing `p(Z_0 = T|{s_u})`.
* **Legend (Right Side):**
* Dashed circle with gray fill: Ground-truth factuality (hidden)
* Solid circle with white fill: LLM's prior belief (observed)
* Solid black arrow: Transition probability `p(Z_{c(u)}|Z_u)`
* Curved line with arrow pointing upwards: Emission probability `p(S_u|Z_u)`
* Dotted red arrow: Belief propagation
* Rounded red rectangle: Posterior belief
### Detailed Analysis or Content Details
The diagram depicts a tree-like structure with nodes representing states and edges representing probabilities.
* **Node Labels:** The nodes are labeled as follows: `S0`, `Z0`, `S1`, `Z1`, `S2`, `Z2`, `S3`, `Z3`, `S4`, `Z4`.
* **Prior Belief:** `S0` is labeled as "Prior belief" with a blue arrow pointing to it.
* **Posterior Belief:** A red box at the bottom-left contains the expression `p(Z_0 = T|{s_u})` and is labeled "Posterior belief" with a red arrow pointing to it.
* **Connections:**
* `Z0` connects to `S0` (emission probability), `Z1` (transition probability), and `Z2` (transition probability).
* `Z1` connects to `S1` (emission probability).
* `Z2` connects to `S2` (emission probability), `Z3` (transition probability), and `Z4` (transition probability).
* `Z3` connects to `S3` (emission probability).
* `Z4` connects to `S4` (emission probability).
* **Belief Propagation:** A dotted red arrow labeled "Belief prop." points from the posterior belief box to the right.
### Key Observations
* The diagram illustrates a hierarchical structure where hidden states (`Z_u`) influence observed states (`S_u`).
* The transition probabilities define how the hidden states evolve, while the emission probabilities define how the observed states are generated from the hidden states.
* Belief propagation is used to update the posterior belief based on the observed data.
### Interpretation
The diagram represents a Hidden Markov Model applied to a belief tree. This model is likely used to infer the ground-truth factuality (`Z_u`) based on the observed beliefs of a Large Language Model (LLM) (`S_u`). The model captures the dependencies between the hidden states and the observed states, allowing for probabilistic reasoning about the underlying facts. The posterior belief `p(Z_0 = T|{s_u})` represents the updated belief about the initial state `Z_0` given the observed data `{s_u}`. The diagram provides a visual representation of the model's structure and the flow of information within it.
```
