## Diagram: LLM Factuality Alignment ### Overview The image illustrates the process of improving the factuality of a Large Language Model (LLM) through self-evaluation and alignment. It compares the LLM's output before and after this process, highlighting changes in the generated text and confidence scores associated with specific facts. ### Components/Axes * **Header:** "Prompt: Write a biography of Jesse Foppert." * **Left Side:** * "Before Alignment" with a downward arrow. * "Self-Evaluation on Factuality" with a downward arrow. * "After Alignment" with a downward arrow. * **LLM Representation:** A cartoon llama is used to represent the LLM. * **Generation Text Blocks:** Two text blocks representing the LLM's generated text before and after alignment. * **Self-Ask Bubble:** A speech bubble containing the question "Self-Ask: Are the generated facts correct?" * **Confidence Scores:** A box displaying confidence scores for specific facts, along with checkmarks or crosses indicating correctness. ### Detailed Analysis **1. Before Alignment:** * **LLM Representation:** A blue cartoon llama. * **Generation Text:** "Generation: Jesse Foppert is a former Major League Baseball pitcher who was born on July 20, 1980, in Los Angeles, Pennsylvania, USA ..." * The text is annotated with numbers 1-4, corresponding to specific facts. * Fact 1: "Jesse Foppert is a former Major League Baseball pitcher" * Fact 2: "born on July 20, 1980" * Fact 3: "in Los Angeles" * Fact 4: "Pennsylvania, USA" * **Confidence Scores:** * Fact 1: 0.87 (Green Checkmark) - Correct * Fact 2: 0.10 (Red Cross) - Incorrect * Fact 3: 0.08 (Red Cross) - Incorrect * Fact 4: 0.95 (Green Checkmark) - Correct **2. Self-Evaluation on Factuality:** * **Self-Ask Bubble:** The LLM poses the question "Self-Ask: Are the generated facts correct?" **3. After Alignment:** * **LLM Representation:** A white cartoon llama with a purple background. * **Generation Text:** "Generation: Jesse Foppert is a former Major League Baseball pitcher who was born on July 10, 1980, in Reading, Pennsylvania, USA ..." * The text is modified to reflect correct information. * "July 20" is changed to "July 10" * "Los Angeles" is changed to "Reading" ### Key Observations * The LLM initially generates some incorrect facts about Jesse Foppert's biography. * The LLM assigns confidence scores to each fact, reflecting its belief in their accuracy. * The self-evaluation process identifies incorrect facts. * After alignment, the LLM generates corrected facts. ### Interpretation The diagram demonstrates a process for improving the factuality of LLM-generated text. By incorporating a self-evaluation step, the LLM can identify and correct inaccuracies in its output. The confidence scores provide a measure of the LLM's certainty about each fact, allowing for targeted interventions to improve accuracy. The alignment process results in a more accurate and reliable biography of Jesse Foppert. The change in the LLM's representation (from blue to white with a purple background) visually signifies the transformation and improvement achieved through the alignment process.

# Technical Document Extraction: LLM Factuality Alignment Process This document describes a technical diagram illustrating the workflow for improving the factuality of a Large Language Model (LLM) through a self-evaluation and alignment process. ## 1. Header Section * **Prompt:** "Write a biography of Jesse Foppert." --- ## 2. Main Process Flow (Pre-Alignment) This section is divided into two vertical stages connected by a downward-pointing gradient arrow (blue to green to red). ### Stage 1: Before Alignment * **Actor:** LLM (represented by a blue line-art dog icon). * **Action:** **Generation** * **Output Text:** "Jesse Foppert is a former Major League Baseball pitcher who was born on July 20, 1980, in Los Angeles, Pennsylvania, USA ..." * **Annotated Claims:** Specific segments of the text are highlighted and numbered: 1. **[1]** "a former Major League Baseball pitcher" (Highlighted in blue) 2. **[2]** "July 20" (Highlighted in white/grey) 3. **[3]** "Los Angeles" (Highlighted in blue) 4. **[4]** "Pennsylvania" (Highlighted in red/pink) ### Stage 2: Self-Evaluation on Factuality * **Actor:** LLM (same icon). * **Internal Thought (Self-Ask):** "Are the generated facts correct?" * **Component: Confidence Scores** (Located at the right): * A panel displays confidence values for the four numbered claims from Stage 1: * **1: 0.87** (Accompanied by a green checkmark icon) * **2: 0.10** (Accompanied by a red 'X' icon) * **3: 0.08** (Accompanied by a red 'X' icon) * **4: 0.95** (Accompanied by a green checkmark icon) --- ## 3. Post-Alignment Section Separated by a dashed horizontal line, this section shows the result after the alignment process. ### Stage 3: After Alignment * **Actor:** Aligned LLM (represented by a detailed, colored white dog/wolf icon). * **Action:** **Generation** * **Output Text:** "Jesse Foppert is a former Major League Baseball pitcher who was born on **July 10**, 1980, in **Reading**, Pennsylvania, USA ..." * **Key Changes and Trends:** * The model has corrected the low-confidence facts identified in Stage 2. * **Correction 1 (Date):** "July 20" (Claim 2, low confidence) has been changed to **July 10** (Text is bolded and green). * **Correction 2 (City):** "Los Angeles" (Claim 3, low confidence) has been changed to **Reading** (Text is bolded and green). * **Consistency:** "Pennsylvania" (Claim 4, high confidence) remains in the text, though the background highlight is now a light red/pink. "a former Major League Baseball pitcher" (Claim 1, high confidence) remains highlighted in blue. --- ## 4. Summary of Data Transformations | Claim ID | Original Value (Before) | Confidence | Final Value (After) | Status | | :--- | :--- | :--- | :--- | :--- | | 1 | a former MLB pitcher | 0.87 | a former MLB pitcher | Retained | | 2 | July 20 | 0.10 | **July 10** | **Corrected** | | 3 | Los Angeles | 0.08 | **Reading** | **Corrected** | | 4 | Pennsylvania | 0.95 | Pennsylvania | Retained | **Technical Conclusion:** The diagram demonstrates a "Self-Ask" mechanism where an LLM evaluates its own generated claims. By identifying claims with low confidence scores (0.10 and 0.08), the "Aligned LLM" is able to regenerate the text, replacing the erroneous facts with corrected data while maintaining the high-confidence factual structure.

\n ## Diagram: LLM Alignment Process - Factuality Evaluation ### Overview This diagram illustrates a process for aligning a Large Language Model (LLM) to improve the factual accuracy of its generated text. It shows three stages: "Before Alignment", "Self-Evaluation on Factuality", and "After Alignment". Each stage includes a depiction of the LLM and a generated text sample, along with confidence scores indicating the accuracy of specific facts within the text. ### Components/Axes The diagram is structured vertically with three distinct sections separated by dashed lines. Each section contains: * **LLM Representation:** A stylized icon representing the LLM. The icon changes slightly between stages to indicate alignment. * **Generation:** A block of text representing the LLM's output. * **Self-Ask:** A question posed by the LLM to itself regarding the factuality of the generated text (present only in the middle section). * **Confidence:** A series of confidence scores associated with specific parts of the generated text, represented by colored circles with numerical values. The diagram also includes a prompt at the top: "Prompt: Write a biography of Jesse Fopper." ### Detailed Analysis or Content Details **1. Before Alignment:** * **Prompt:** "Write a biography of Jesse Fopper." * **Generation:** "Jesse Fopper is a former Major League Baseball pitcher who was born on July 20, 1980, in Los Angeles, Pennsylvania, USA..." * **Confidence:** Four confidence scores are displayed: * Score 1: 0.87 (Green checkmark) - Associated with "Jesse Fopper is a former Major League Baseball pitcher" * Score 2: 0.10 (Red X) - Associated with "born on July 20, 1980" * Score 3: 0.08 (Red X) - Associated with "in Los Angeles" * Score 4: 0.95 (Green checkmark) - Associated with "Pennsylvania, USA" **2. Self-Evaluation on Factuality:** * **Self-Ask:** "Are the generated facts correct?" * **LLM Representation:** The LLM icon appears to be "thinking". * **Confidence:** The same confidence scores as the "Before Alignment" stage are displayed. **3. After Alignment:** * **Generation:** "Jesse Fopper is a former Major League Baseball pitcher who was born on July 20, 1980, in Reading, Pennsylvania, USA..." * **LLM Representation:** The LLM icon is now labeled "Aligned" and has a different appearance. * **Confidence:** Four confidence scores are displayed: * Score 1: 0.87 (Green checkmark) - Associated with "Jesse Fopper is a former Major League Baseball pitcher" * Score 2: 0.10 (Red X) - Associated with "born on July 20, 1980" * Score 3: 0.08 (Red X) - Associated with "in Reading" * Score 4: 0.95 (Green checkmark) - Associated with "Pennsylvania, USA" ### Key Observations * The initial generation contains an incorrect location of birth ("Los Angeles, Pennsylvania"). * The confidence scores accurately identify the incorrect information (low confidence for location and date). * After alignment, the location of birth is corrected to "Reading, Pennsylvania". * The confidence scores remain the same for the correct statements, but the incorrect location now has a higher confidence score. * The confidence scores are associated with specific phrases within the generated text, indicating a granular evaluation of factuality. ### Interpretation This diagram demonstrates a process for improving the factual accuracy of LLM-generated text through self-evaluation and alignment. The LLM is able to assess the correctness of its own output and, after alignment, correct factual errors. The confidence scores provide a quantitative measure of the LLM's certainty about the accuracy of different parts of the generated text. The diagram highlights the importance of factuality evaluation in LLM development and the potential for alignment techniques to improve the reliability of LLM outputs. The fact that the confidence scores for the correct statements remain unchanged suggests that the alignment process focuses on correcting errors rather than altering the LLM's overall understanding of the topic. The change in the LLM icon from the first to the last stage indicates that the LLM has been modified to improve its factuality.

## Diagram: LLM Factuality Alignment Process ### Overview This diagram illustrates a three-stage process for improving the factual accuracy of a Large Language Model's (LLM) generated text. It compares the model's output "Before Alignment" with its output "After Alignment," highlighting a self-evaluation step that assesses confidence in generated facts. The process uses a specific example: generating a biography of baseball player Jesse Foppert. ### Components/Axes The diagram is structured vertically into three distinct sections, separated by horizontal lines and connected by downward-pointing arrows indicating process flow. 1. **Top Section: "Before Alignment"** * **Left Label:** "Before Alignment" (vertical text). * **Icon:** A blue, stylized dog head labeled "LLM". * **Content Block:** A text generation box with a pencil icon and the header "Generation:". * **Generated Text:** "Jesse Foppert is a former Major League Baseball pitcher who was born on July 20, 1980, in Los Angeles, Pennsylvania, USA ..." * **Annotations:** Four numbered black circles (1, 2, 3, 4) are placed over specific parts of the text, linking them to confidence scores in the next section. 2. **Middle Section: "Self-Evaluation on Factuality"** * **Left Label:** "Self-Evaluation on Factuality" (vertical text). * **Icon:** The same blue dog head icon labeled "LLM". * **Thought Bubble:** Contains the text "Self-Ask: Are the generated facts correct?". * **Confidence Panel:** A box titled "Confidence:" with a robot icon. It contains four entries, each with a number, a decimal score, and a symbol: * 1: 0.87 (Green checkmark) * 2: 0.10 (Red X) * 3: 0.08 (Red X) * 4: 0.95 (Green checkmark) 3. **Bottom Section: "After Alignment"** * **Left Label:** "After Alignment" (vertical text). * **Icon:** A new, more detailed icon of a white llama/alpaca head labeled "Aligned LLM". * **Content Block:** A text generation box identical in style to the first. * **Generated Text:** "Jesse Foppert is a former Major League Baseball pitcher who was born on **July 10**, 1980, in **Reading**, Pennsylvania, USA ..." * **Annotations:** The text "July 10" and "Reading" are highlighted in green, indicating corrected information. ### Detailed Analysis * **Prompt:** The initial input is "Prompt: Write a biography of Jesse Foppert." * **Initial Generation (Before Alignment):** The LLM produces a biographical sentence. Four specific factual claims are tagged: 1. "a former Major League Baseball pitcher" 2. "July 20, 1980" (birth date) 3. "Los Angeles" (birth city) 4. "Pennsylvania, USA" (birth state/country) * **Self-Evaluation:** The model internally queries the correctness of its generated facts. The confidence scores suggest: * High confidence (0.87) in claim #1 (profession). * Very low confidence (0.10) in claim #2 (birth date). * Very low confidence (0.08) in claim #3 (birth city). * High confidence (0.95) in claim #4 (birth state/country). * **Aligned Generation:** After the alignment process, the LLM produces a revised sentence. The elements with low confidence scores (#2 and #3) have been corrected: * Birth date changed from "July 20" to "**July 10**". * Birth city changed from "Los Angeles" to "**Reading**". * The high-confidence elements (#1 and #4) remain unchanged. ### Key Observations 1. **Process Flow:** The diagram clearly depicts a linear workflow: Generation -> Self-Evaluation -> Corrected Generation. 2. **Visual Coding:** Color and icons are used systematically. Blue represents the base LLM, white/llama represents the aligned model. Green highlights and checkmarks indicate correct/high-confidence information, while red X's indicate errors/low confidence. 3. **Targeted Correction:** The alignment process does not regenerate the entire text from scratch. It specifically identifies and corrects only the low-confidence segments, preserving the high-confidence parts. 4. **Confidence as a Signal:** The numerical confidence scores (0.10, 0.08) directly correlate with the factual errors in the initial output, demonstrating their utility as a diagnostic tool. ### Interpretation This diagram demonstrates a method for **mitigating hallucinations and improving factual reliability in LLMs** through a process of internal self-evaluation and targeted correction. The core idea is that an LLM can be trained or prompted to assess its own confidence in discrete pieces of generated information. * **What it suggests:** The "Aligned LLM" is not necessarily more knowledgeable at its base, but is better at **recognizing the limits of its own knowledge**. It uses a confidence metric to flag uncertain claims, which can then be verified or corrected, possibly via an external knowledge retrieval step (not shown) or through refined decoding strategies. * **How elements relate:** The "Self-Evaluation" stage is the critical innovation. It acts as a filter between raw generation and final output. The confidence scores are the key output of this stage, directly determining which parts of the text are subject to revision in the "After Alignment" phase. * **Notable pattern:** The model exhibits high confidence in general, common-sense facts (a player's profession, their state of birth) but low confidence in specific, precise details (exact date, city). This aligns with known characteristics of LLM knowledge distributions. * **Underlying mechanism:** The shift from a "dog" to a "llama" icon symbolizes a change in the model's architecture or training paradigm (e.g., from a base model to one fine-tuned with reinforcement learning from human feedback (RLHF) or a similar alignment technique that incorporates factuality rewards). The process shown is a conceptual representation of how such alignment can manifest in practice.

## Flowchart: LLM Text Generation Before and After Alignment ### Overview The image compares text generation from a Large Language Model (LLM) before and after alignment, focusing on factual accuracy. It includes a self-evaluation step where the model assesses the correctness of generated facts, visualized with confidence scores and checkmarks/crosses. The flowchart highlights corrections in biographical details of Jesse Foppert, a former Major League Baseball pitcher. ### Components/Axes 1. **Main Sections**: - **Before Alignment**: LLM generates text with factual errors. - **After Alignment**: LLM generates corrected text. 2. **Self-Evaluation**: - A thought bubble labeled "Self-Ask: Are the generated facts correct?" connects to confidence scores. 3. **Confidence Scores**: - Numerical values (0.08–0.95) with green checkmarks (correct) or red crosses (incorrect). 4. **Text Generation**: - Two biographical summaries of Jesse Foppert, with highlighted factual discrepancies. ### Detailed Analysis #### Before Alignment - **Generation**: 1. "Jesse Foppert is a former Major League Baseball pitcher who was born on July 20, 1980, in Los Angeles, Pennsylvania, USA..." - **Errors**: Incorrect birth date (July 20, 1980) and location (Los Angeles, Pennsylvania). 2. "born on July 20, 1980, in Los Angeles, Pennsylvania, USA..." - **Error**: Same incorrect date and location. 3. "Pennsylvania, USA..." - **Correct**: Pennsylvania is accurate. 4. "born on July 10, 1980, in Reading, Pennsylvania, USA..." - **Error**: Correct date (July 10, 1980) but incorrect location (Reading, Pennsylvania). - **Confidence Scores**: - Point 1: 0.87 (✅ Correct). - Point 2: 0.10 (❌ Incorrect). - Point 3: 0.08 (❌ Incorrect). - Point 4: 0.95 (✅ Correct). #### After Alignment - **Generation**: 1. "Jesse Foppert is a former Major League Baseball pitcher who was born on July 10, 1980, in Reading, Pennsylvania, USA..." - **Corrections**: Correct date (July 10, 1980) and location (Reading, Pennsylvania). 2. "born on July 10, 1980, in Reading, Pennsylvania, USA..." - **Corrections**: Same as above. 3. "Pennsylvania, USA..." - **Correct**: Pennsylvania remains accurate. 4. "born on July 10, 1980, in Reading, Pennsylvania, USA..." - **Corrections**: Consistent with point 2. - **Confidence Scores**: - Point 1: 0.87 (✅ Correct). - Point 2: 0.95 (✅ Correct). - Point 3: 0.87 (✅ Correct). - Point 4: 0.95 (✅ Correct). ### Key Observations 1. **Factual Corrections**: - The "After Alignment" section corrects the birth date (July 20 → July 10) and location (Los Angeles → Reading, Pennsylvania). 2. **Confidence Trends**: - Points with errors in "Before Alignment" (2 and 3) show low confidence (0.10 and 0.08) but gain high confidence (0.95 and 0.87) after alignment. - Correct points (1 and 4) maintain high confidence (0.87–0.95) before and after alignment. 3. **Self-Evaluation Impact**: - The "Self-Ask" step correlates with improved accuracy, as confidence scores for previously incorrect points increase post-alignment. ### Interpretation - **Alignment Effectiveness**: The alignment process successfully corrects factual errors in LLM-generated text, as evidenced by revised dates and locations. - **Confidence-accuracy Correlation**: Higher confidence scores align with factual correctness. For example, point 2’s confidence jumps from 0.10 (incorrect) to 0.95 (correct) after alignment. - **Self-Evaluation Role**: The "Self-Ask" mechanism likely prompts the model to verify and revise outputs, enhancing reliability. - **Anomalies**: Point 3 (“Pennsylvania, USA”) remains correct in both sections, suggesting some errors are context-dependent rather than systemic. This flowchart demonstrates how alignment improves LLM outputs by integrating self-evaluation to prioritize factual accuracy, a critical step for applications requiring precision.