Update README.md

Author: SergioV3005

README.md

@@ -7,15 +7,50 @@ This project tests Local Large Language Models (LLMs) using **syllogisms** to de
 **You can test yourself by clicking [here](https://longocris.github.io/Belief-Bias-Questionnaire/) or you can check out the [quiz repository](https://github.com/LongoCris/Belief-Bias-Questionnaire)**
 
 In particular, the following LLMs are tested:
-1. [llama3.2:1b](https://ollama.com/library/llama3.2)
-2. [Mistral](https://ollama.com/library/mistral)
-3. [qwen3:8b](https://ollama.com/library/qwen3)
+1. [llama3.2:1b](https://ollama.com/library/llama3.2): a lightweight baseline model from LLaMa 3.2 family
+2. [Mistral](https://ollama.com/library/mistral): a 7b model designed for long-context reasoning
+3. [qwen3:8b](https://ollama.com/library/qwen3): a very recent model optimized for instructions and logical tasks
 
-The experiments involve the test of each LLM (configured with **temperature** equal to 0 and to 0.7) to see the distribution of accuracy in **conflictual** and **non-conflictual** items. In particular, the conflictual items are defined as the valid-inbelievable and the invalid-believable items, while the non-conflictual ones as the valid-believable and invalid-unbelievable ones. If the errors display correlation between the conflictuality and non-conflictuality of the item, then there is a signal of BB.  
+The experiments involve the test of each LLM (configured with **temperature** equal to 0 and to 0.7) to see the distribution of accuracy in **conflictual** and **non-conflictual** items. 
+
+In particular, the conflictual items are defined as the valid-inbelievable and the invalid-believable items, while the non-conflictual ones as the valid-believable and invalid-unbelievable ones. If the errors display correlation between the conflictuality and non-conflictuality of the items, then there is a signal of BB.  
 
 ![image](https://github.com/user-attachments/assets/f95062ba-f9fc-4a5d-a186-6f7269ca605f)
 
-For a comparison, the test is also experimented on a set of humans.
+For a comparison, the test is also experimented on a set of humans, as shown in the presentation and in the report.
+
+### Belief Bias Questionnaire
+
+The questionnaire consists of **16 items** balanced across the four categories:
+* Valid-Believable (VB): 4 items
+* Invalid-Believable (IB): 4 items
+* Valid-Unbelievable (VU): 4 items 
+* Invalid-Unbelievable (IU): 4 items
+
+In each category, three items are of **regular** difficulty (being composed of two premises and one conclusion), while one item is **hard** (being composed of three premises and the conclusion). This, for the set of humans, is done to avoid adaptation to the difficulty of the item.
+
+For the humans, one neutral item was added to the questionnaire to stimulate **attention**. Each item required valid or invalid responses within **20 seconds**, to encourage intuitive processing. One item included open-ended explanation to examine the reasoning behind the decision and to compare it to LLMs.
+
+### Results
+
+| Model         | Temperature | Average Accuracy | Belief Bias Signal |
+|---------------|-------------|------------------|--------------------|
+| Llama3.2:1b   | 0.0         | Low              | Present            |
+| Llama3.2:1b   | 0.7         | Low–Medium       | Present            |
+| Mistral:7b    | 0.0         | Medium           | Moderate           |
+| Mistral:7b    | 0.7         | Medium           | Moderate           |
+| **Qwen3:8b**  | **0.0**     | **High**         | **Absent**         |
+| Qwen3:8b      | 0.7         | High             | Low                |
+
+The precise distribution of the errors can be found in the presentation.
+
+As for the control group of humans (balanced gender, minimum B2 english and Bachelor's degree holders), the results showed expression of BB, especially due to the time pressure.
+
+### Conclusions
+
+BB is not unique to humans. Some LLMs mimic it under certain configurations. Model complexity and architecture (not temperature alone) determine the expression of BB.
+
+For deeper insights, see our report.
 
 ## Setup Instructions