## Diagram: LLM Response with and without Epistemic Markers ### Overview The image is a diagram illustrating how a Large Language Model (LLM) responds to the question "Which team won the 2022 NBA Finals?" with and without epistemic markers. The diagram shows the question being fed into the LLM, and then two different responses are generated: one with epistemic markers (indicating uncertainty) and one without. ### Components/Axes * **Top:** Blue rounded rectangle containing the question "Which team won the 2022 NBA Finals?" and a silhouette of a person. * **Middle:** Green rounded square containing a symbol resembling two intertwined ampersands. The label "LLM" is to the left of the square. An arrow points from the top rectangle to this square. * **Bottom Left:** Light blue rounded rectangle labeled "With Epistemic Markers" in a yellow box. Inside this rectangle is a black-bordered box containing the text "I think the Milwaukee Bucks won the 2022 NBA Finals, but I am not sure." The words "I think" and "but I am not sure" are in red. * **Bottom Right:** Light blue rounded rectangle labeled "Without Epistemic Markers" in a yellow box. Inside this rectangle is a black-bordered box containing the text "The Milwaukee Bucks won the 2022 NBA Finals." * **Arrows:** Two downward-pointing arrows, one connecting the top rectangle to the middle square, and another connecting the middle square to the two bottom rectangles. ### Detailed Analysis or ### Content Details * **Question:** "Which team won the 2022 NBA Finals?" * **LLM Representation:** The symbol in the green square represents the LLM. * **Response with Epistemic Markers:** "I think the Milwaukee Bucks won the 2022 NBA Finals, but I am not sure." * **Response without Epistemic Markers:** "The Milwaukee Bucks won the 2022 NBA Finals." ### Key Observations * The LLM provides two different responses to the same question, differing in the inclusion of epistemic markers. * The response with epistemic markers expresses uncertainty ("I think," "but I am not sure"). * The response without epistemic markers presents the information as a definitive statement. ### Interpretation The diagram demonstrates how LLMs can be programmed to provide responses with varying degrees of certainty. The inclusion of epistemic markers can be used to indicate the LLM's confidence in its answer, which can be important for users who need to assess the reliability of the information. The absence of epistemic markers can make the LLM's response sound more authoritative, but it can also be misleading if the LLM is not actually certain about the answer. The diagram highlights the importance of considering the context and purpose of the LLM's response when interpreting its output.

# Technical Document Extraction: LLM Epistemic Markers Diagram ## 1. Overview This image is a flow diagram illustrating the difference in Large Language Model (LLM) outputs when using "Epistemic Markers" versus when they are absent. The diagram follows a top-down linear progression from a user query to a model processing stage, resulting in two comparative output formats. ## 2. Component Isolation ### Region 1: Header (Input Stage) * **User Icon:** A black silhouette of a person's head and shoulders is positioned to the right of the input box. * **Input Box:** A blue rounded rectangle containing white text. * **Text Transcription:** "Which team won the 2022 NBA Finals?" * **Flow Element:** A downward-pointing black arrow connects the Input Box to the Processing Stage. ### Region 2: Main Processing (LLM Stage) * **Label:** The text "LLM" is positioned to the left of the central icon. * **Icon:** A lime-green square with rounded corners containing a black circular emblem with a stylized "X" or interlocking knot design. * **Flow Element:** A downward-pointing black arrow connects the LLM icon to the Output Stage. ### Region 3: Footer (Output Comparison Stage) This region is contained within a large, light-blue rounded rectangle and is split into two parallel columns. #### Column A: With Epistemic Markers (Left) * **Sub-header Label:** A pale yellow rectangle containing the text: "With Epistemic Markers" * **Output Box:** A light-blue box with a black border. * **Transcribed Text:** "I think the Milwaukee Bucks won the 2022 NBA Finals, but I am not sure." * **Visual Emphasis:** The words "**I think**" and "**but I am not sure**" are highlighted in **red text**, identifying them as the epistemic markers (indicators of uncertainty). #### Column B: Without Epistemic Markers (Right) * **Sub-header Label:** A pale yellow rectangle containing the text: "Without Epistemic Markers" * **Output Box:** A light-blue box with a black border. * **Transcribed Text:** "The Milwaukee Bucks won the 2022 NBA Finals." * **Visual Emphasis:** All text is in standard black font, presenting the statement as a definitive fact without qualifiers. --- ## 3. Technical Analysis & Logic Flow The diagram demonstrates a specific behavior in AI communication: 1. **The Query:** The user asks a factual question about a past event. 2. **The Error:** In both output examples, the LLM provides incorrect information (The Golden State Warriors won the 2022 NBA Finals, not the Milwaukee Bucks). 3. **The Comparison:** * **Epistemic Version:** By using markers like "I think" and "but I am not sure," the model signals its low confidence in the accuracy of the statement, potentially alerting the user to verify the fact. * **Non-Epistemic Version:** The model states the incorrect information as an absolute truth, which constitutes a "hallucination" presented with high confidence. ## 4. Text Summary | Element | Content | | :--- | :--- | | **User Query** | Which team won the 2022 NBA Finals? | | **Processor** | LLM | | **Output (With Markers)** | **I think** the Milwaukee Bucks won the 2022 NBA Finals, **but I am not sure.** | | **Output (Without Markers)** | The Milwaukee Bucks won the 2022 NBA Finals. |

\n ## Diagram: LLM Response with and without Epistemic Markers ### Overview This diagram illustrates the difference in responses from a Large Language Model (LLM) when prompted with a question, specifically "Which team won the 2022 NBA Finals?", with and without the inclusion of epistemic markers (expressions of uncertainty). The diagram shows a flow from the question, through the LLM, to two different outputs. ### Components/Axes The diagram consists of the following components: * **Input Question:** "Which team won the 2022 NBA Finals?" - positioned at the top of the diagram. * **LLM Representation:** A circular icon labeled "LLM" with a stylized infinity symbol inside, representing the Large Language Model. This is positioned centrally. * **Output 1: With Epistemic Markers:** A rounded rectangle labeled "With Epistemic Markers" containing the text: "I think the Milwaukee Bucks won the 2022 NBA Finals, but I am not sure." - positioned on the left side, below the LLM. * **Output 2: Without Epistemic Markers:** A rounded rectangle labeled "Without Epistemic Markers" containing the text: "The Milwaukee Bucks won the 2022 NBA Finals." - positioned on the right side, below the LLM. * **Arrows:** Downward arrows indicating the flow of information from the question to the LLM, and from the LLM to the outputs. * **User Icon:** A small grey icon of a person's head and shoulders, positioned to the right of the input question. ### Detailed Analysis or Content Details The diagram presents a comparative analysis of LLM responses. * **Input:** The question is a straightforward factual query about the winner of the 2022 NBA Finals. * **LLM Processing:** The LLM processes the question. * **Output with Epistemic Markers:** The LLM's response includes phrases expressing uncertainty ("I think," "but I am not sure"). The response states the Milwaukee Bucks as the winner, but qualifies it with doubt. * **Output without Epistemic Markers:** The LLM's response is a direct statement of fact, asserting that the Milwaukee Bucks won the 2022 NBA Finals without any qualifiers. ### Key Observations The key observation is the difference in the LLM's output based on whether or not it incorporates epistemic markers. The inclusion of these markers demonstrates a level of awareness of uncertainty, while their absence presents the information as a definitive fact. The diagram highlights the importance of how LLMs frame their responses, particularly when dealing with potentially ambiguous or uncertain information. ### Interpretation The diagram demonstrates a crucial aspect of LLM behavior: the ability to express (or suppress) uncertainty. The presence of epistemic markers in the first output suggests a more cautious and nuanced response, acknowledging the possibility of error. The second output, lacking these markers, presents the information as a confident assertion. This difference is significant because it impacts how users perceive the reliability of the information provided by the LLM. The diagram implicitly raises questions about the underlying mechanisms that control the inclusion or exclusion of epistemic markers in LLM responses. It suggests that these markers are not inherent to the LLM's knowledge but are rather a function of its programming or the specific prompt it receives. This has implications for the development of more trustworthy and transparent LLMs, as it highlights the need for mechanisms that accurately reflect the LLM's level of confidence in its responses. The diagram also subtly points to the potential for LLMs to be manipulated into presenting uncertain information as fact, simply by omitting epistemic markers. This underscores the importance of critical thinking and source evaluation when interacting with LLMs.

## Diagram: LLM Response with and without Epistemic Markers ### Overview The image is a flowchart-style diagram illustrating how a Large Language Model (LLM) can generate two different types of responses to the same factual query. It contrasts a response that includes expressions of uncertainty (epistemic markers) with one that presents the information as a definitive statement. ### Components/Axes The diagram is structured vertically with three main sections: 1. **Top (User Input):** A blue, rounded speech bubble containing the user's question. A black silhouette icon of a person is positioned to the right of the bubble. 2. **Center (Processing):** A light green square containing a black circular icon with an interlocking knot symbol. The label "LLM" is placed to the left of this square. Black arrows point downward from the user input to the LLM, and from the LLM to the response section below. 3. **Bottom (Output):** A large, light blue rounded rectangle containing two distinct response boxes side-by-side. * **Left Box:** Has a yellow header labeled "With Epistemic Markers". The response text inside is in a black-bordered box. * **Right Box:** Has a yellow header labeled "Without Epistemic Markers". The response text inside is in a black-bordered box. ### Detailed Analysis **1. User Query (Top Section):** * **Text:** "Which team won the 2022 NBA Finals?" * **Presentation:** White text on a blue background, formatted as a question from a user. **2. LLM Processing (Center Section):** * **Label:** "LLM" * **Icon:** A stylized, black line-art icon resembling an infinity symbol or a knot, enclosed in a circle. This represents the AI model processing the query. **3. Generated Responses (Bottom Section):** * **Response A (Left - "With Epistemic Markers"):** * **Full Transcription:** "I think the Milwaukee Bucks won the 2022 NBA Finals, but I am not sure." * **Styling:** The phrases "I think" and "but I am not sure." are highlighted in red text, while the rest is in black. This visually emphasizes the linguistic markers of uncertainty. * **Response B (Right - "Without Epistemic Markers"):** * **Full Transcription:** "The Milwaukee Bucks won the 2022 NBA Finals." * **Presentation:** The entire statement is in plain black text, presenting the information as a confident, factual claim. ### Key Observations * **Structural Contrast:** The diagram uses a clear, symmetrical layout to directly compare two output modes from the same system given identical input. * **Visual Emphasis on Uncertainty:** The use of red text specifically for the epistemic markers ("I think", "but I am not sure") draws immediate attention to the linguistic elements that convey doubt. * **Identical Core Information:** Both responses contain the same core factual claim: "the Milwaukee Bucks won the 2022 NBA Finals." The difference lies entirely in the framing and certainty expressed. * **Spatial Grounding:** The "With Epistemic Markers" response is positioned on the left, and the "Without Epistemic Markers" response is on the right, creating a clear visual dichotomy. ### Interpretation This diagram serves as a conceptual model for a key challenge and design consideration in AI communication: **calibrating and expressing confidence.** * **What it Demonstrates:** It illustrates that an LLM's knowledge is not binary (know/doesn't know). The model can access the same information but present it with different levels of epistemic commitment. The "With Epistemic Markers" response is more intellectually honest when the model's confidence is not absolute, while the "Without" version risks presenting potentially incorrect information as fact. * **Relationship Between Elements:** The flow (Query → LLM → Dual Outputs) shows that the generation of epistemic markers is a deliberate output choice made by the model or its governing system, not an inherent property of the information itself. * **Significance:** The choice between these response styles has major implications for user trust, safety, and the perceived reliability of AI systems. A system that appropriately uses epistemic markers (like "I think," "I'm not sure," "According to my data") manages user expectations and encourages critical verification. Conversely, consistently definitive statements, even when wrong, can lead to over-reliance and the spread of misinformation. The diagram advocates for the value of transparency in AI uncertainty.

## Flowchart: LLM Response Generation with Epistemic Markers ### Overview The image depicts a flowchart illustrating how a Large Language Model (LLM) processes a question about the 2022 NBA Finals winner and generates two distinct responses based on the presence or absence of epistemic markers. The flowchart includes a user query, an LLM processing node, and two divergent response paths. --- ### Components/Axes 1. **User Query (Top)**: - Text: *"Which team won the 2022 NBA Finals?"* - Position: Top-center, enclosed in a blue rounded rectangle. - Icon: Silhouette of a human head in the top-right corner. 2. **LLM Processing Node (Center)**: - Label: *"LLM"* in black text. - Symbol: Green square with a black infinity symbol (∞) inside a circle. - Arrows: Two downward-pointing arrows branching left and right. 3. **Response Paths (Bottom)**: - **Left Path (With Epistemic Markers)**: - Label: *"With Epistemic Markers"* in black text on a yellow banner. - Text Box: - *"I think the Milwaukee Bucks won the 2022 NBA Finals, but I am not sure."* - Keywords: *"I think"* (red), *"not sure"* (red). - **Right Path (Without Epistemic Markers)**: - Label: *"Without Epistemic Markers"* in black text on a yellow banner. - Text Box: - *"The Milwaukee Bucks won the 2022 NBA Finals."* --- ### Detailed Analysis - **Flow Direction**: The flowchart follows a top-to-bottom, left-to-right structure. The user query initiates the process, leading to the LLM node, which splits into two response paths. - **Textual Content**: - The left response includes **epistemic markers** (*"I think"*, *"not sure"*) to express uncertainty. - The right response is **declarative** and lacks uncertainty markers. - **Color Coding**: - Red highlights uncertainty-related phrases (*"I think"*, *"not sure"*). - Yellow banners distinguish the two response categories. --- ### Key Observations 1. The LLM generates **two versions** of the answer: one uncertain (with markers) and one confident (without markers). 2. The correct answer (*Milwaukee Bucks*) is embedded in both paths, but confidence levels differ. 3. The flowchart emphasizes the **impact of epistemic markers** on response certainty. --- ### Interpretation This flowchart demonstrates how LLMs can modulate responses based on contextual cues (e.g., epistemic markers). The presence of uncertainty markers (*"I think"*, *"not sure"*) reflects the model’s awareness of potential ambiguity, even when the factual answer is known. The Milwaukee Bucks did win the 2022 NBA Finals, validating the declarative response. However, the uncertain version acknowledges the model’s probabilistic reasoning process, which may prioritize hedging over absolute certainty in certain contexts. The diagram underscores the importance of **epistemic framing** in AI-generated text, balancing factual accuracy with transparency about confidence levels.