Model: Gemini 3.1 - Few-shot

Google DeepMind

Abstract

Gemini few-shot prompting

This solution that uses the gemini-3.1-pro-preview model. It is a few-shot prompting approach.

Key prompt used:

Classify the accuracy of the following student answer based on the question.

Definitions:
- Correct: the answer is correct and fully answers the question.
- Partially_correct_incomplete: everything stated in the answer is correct and the answer responds to part of the question, but does not fully answer the question.
- Contradictory: the answer contradicts scientifically correct information.
- Irrelevant: the answer contains domain content but does not respond to any part of the question.
- Non_domain:  if the student utterance does not include domain content, e.g., "I don't know", "what the book says", "you are stupid".

Decision rules:
- If the student answer is scientifically complete and correct, choose `correct`.
- If the answer has some relevant truth but is missing required information, choose `partially_correct_incomplete`.
- If the core claim conflicts with the scientifically correct information, choose `contradictory`.
- If the answer contains domain words but does not answer this question, choose `irrelevant`.
- If the answer is "I don't know", abusive, conversational, or otherwise lacks domain content, choose `non_domain`.

Here are some examples:
question text: "Look at the picture of a block of wood in a cup of water. What evidence indicates that the block of wood is less dense than the water?"

student answer: "Less dense because the block is floating on the water."
accuracy: correct

student answer: "The block is floating"
accuracy: partially_correct_incomplete

student answer: "Wood is a solid and water is a liquid."
accuracy: irrelevant

student answer: "It is denser than the water."
accuracy: contradictory

student answer: "I do not know."
accuracy: non_domain

Now classify this student answer based on the question:

Here is one example of a complete, correct answer to the question: "{reference_answer_text}"

question: "{question_text}", student answer: "{student_answer_text}".

Scores

Reference Scores
Name	Weighted F1	Classification rate
Gemini 3.1 - Few-shot	0.5545	0.5632

SAGE: Science Answer Grading & Evaluation

Navigation

Model: Gemini 3.1 - Few-shot

Abstract

Gemini few-shot prompting

Scores