Gemini 2.5 Pro Aced the JEE Advanced 2025 Maths Paper. All Questions Correct | by Khalidha Khanam | Google Cloud – Community | May, 2025

My team and I were exploring EdTech innovation when inspiration struck: Could Gemini 2.5 Pro tackle something as complex as the JEE Advanced Mathematics Paper of 2025, fresh off the presses? I put it to the test, and what happened next was incredible.

My goal was to push the boundaries of what these models can achieve with complex, multi-step problem-solving, facing the absolute latest and most challenging set of questions.

And what happened next left me absolutely speechless.

My methodology was straightforward, emphasizing the “out-of-box” capabilities of Gemini 2.5 Pro. I didn’t engage in extensive fine-tuning or highly complex prompt engineering. Instead, I focused on a structured, multi-stage process for each problem:

Step 1: Presenting the Problem with a “Tutor” Prompt

For each question from the JEE Advanced 2025 Maths Paper, I instructed Gemini 2.5 Pro to act as an IIT JEE Advanced Tutor. This prompt was designed to elicit a methodical, step-by-step solution, complete with topic identification, problem decomposition, and self-verification at each stage.

Here’s the exact prompt template I used for problem generation:

You are an IIT JEE Advanced Tutor. Your task is to solve a given problems in a structured and verifiable manner.
You will receive the problem via the placeholder `{problem_text}`, the subject, the question type.
Follow these steps precisely:
1. Identify Involved Topics: The specific topics relevant to the problem are Chapters: , Sub-Topics: )
2. Decompose the Problem into Sub-Problems: Break down the problem into smaller, manageable sub-problems. Clearly label each sub-problem.
3. Solve Sub-problems Step-by-Step: Solve each sub-problem methodically, showing all steps involved.  After each step, verify its correctness. Explain your verification process.
4. Combine Solutions: Combine the solutions of the sub-problems to arrive at the final answer to the original problem.
Now, solve the following problem:
Question: {problem_text}
Subject: {subject}
Question Type: {question_type}
Solution:
Correct Option Number:

This structured prompt aimed to ensure that Gemini 2.5 Pro didn’t just guess the answer, but demonstrated a clear understanding of the problem-solving process.

Step 2: Leveraging the Official JEE Advanced 2025 Mathematics Paper

For the “2025” paper, I used the officially released JEE Advanced 2025 Mathematics Paper from https://jeeadv.ac.in/. These were the fresh, intricate problems designed to truly test advanced mathematical reasoning, straight from the latest examination. Each question, along with its subject and type, was dynamically inserted into the {problem_text}, {subject}, and {question_type} placeholders of the prompt above.

The moment of truth came as Gemini 2.5 Pro began generating solutions based on this prompt. I watched in awe as it meticulously broke down complex calculus, navigated intricate algebra, and solved challenging geometry problems.

Step 3: Rigorous Evaluation with a Dedicated Rubric

Once Gemini 2.5 Pro generated a solution for a given problem, I then put on my “evaluator” hat. To ensure an unbiased and systematic assessment of the model’s output, I used a predefined evaluation rubric. This rubric is designed for meticulous and precise math assistance, ensuring every solution’s correctness is verified against strict criteria.

Here’s the specific JSON format and criteria I used for evaluating each generated solution:

{
"category": "",
"confidence_score": <0-1>,
"explanation": ""
}

My analysis of each test solution adhered to the following criteria based on the rubric:

• Mathematical Rigor: Ensuring every step in the solution followed valid mathematical principles and logic.
• Accuracy: Verifying that all calculations and algebraic manipulations were accurate and error-free.
• Approach: Assessing whether the approach used was valid and efficient.

Furthermore, if a solution was incorrect, it was categorized precisely according to the following rubric (with “Correct” being the only category for a fully correct solution, Inspired by the error analysis framework from the paper “LM2 : A Simple Society of Language Models Solves Complex Reasoning” (EMNLP 2024 Main Conference, Paper 920)):

• Conceptual Mistakes: The test solution has misunderstood the concept or has applied the wrong concept.
• Computational Mistakes: The test solution has made a mistake in the calculations.
• Procedural Mistakes: The test solution has made a mistake in the procedure of solving the problem.
• Mistake in understanding the question: The test solution has made a mistake in understanding the question.
• Mistake in the first step: The test solution has made a mistake in the first step of the solution.
• Mistake in the first half: The test solution has made a mistake in the first half of the solution.
• Mistake in the second half: The test solution has made a mistake in the second half of the solution.
• Mistake in the last step: The test solution has made a mistake in the last step of the solution.

The result?

Gemini 2.5 Pro correctly solved every single question on the JEE Advanced 2025 Mathematics Paper. A perfect score on one of the world’s most challenging entrance exams, just days after it was conducted.

Still reeling from this astonishing performance on the 2025 paper, I decided to see if this was a consistent capability. I then fed Gemini 2.5 Pro the official JEE Advanced Mathematics Paper of 2024, using the exact same structured prompting and rigorous evaluation approach.

Again, with the same methodology, Gemini 2.5 Pro demonstrated its prowess. It navigated the nuances, avoided common pitfalls, and arrived at the correct solutions for every single question.

Another perfect score. This wasn’t a fluke; it was consistent, high-level performance.

This experience wasn’t just a “wow” moment; it was a profound demonstration of the capabilities that advanced AI models like Gemini 2.5 Pro are now exhibiting.

1. Unprecedented Mathematical Reasoning on Current Problems: This isn’t just about regurgitating facts or performing simple calculations. JEE Advanced problems demand deep conceptual understanding, multi-step logical deduction, and often, creative problem-solving. Gemini 2.5 Pro showcased an astonishing level of mathematical reasoning on current, real-world exam questions.
2. Out-of-Box Power with Structured Guidance: The fact that this performance was largely “Out-of-Box” (OOB) combined with a well-defined prompting strategy underscores the inherent intelligence and capabilities built into these models. This highlights that while the models are powerful, strategic prompting can unlock even greater precision and reliability.
3. Implications for Education: Imagine a future where AI can serve as an incredibly precise and personalized tutor, explaining complex concepts, solving intricate problems step-by-step, and identifying areas where a student needs more help. This goes far beyond just answering questions; it’s about facilitating deeper learning with immediate, accurate feedback.
4. Accelerating R&D: For researchers and engineers, such models could significantly accelerate the pace of discovery and problem-solving in highly technical fields.
5. The AI Frontier is Expanding Rapidly: My experiment reinforces that the pace of AI development is not just rapid, it’s accelerating. What seemed like sci-fi a few years ago is becoming a tangible reality today.

My personal experience with Gemini 2.5 Pro flawlessly solving both the 2025 and 2024 JEE Advanced Maths papers has been truly transformative. It’s a testament to the incredible work being done in Google’s AI research and development.

This isn’t to say that human intelligence will be replaced. Rather, it suggests a future where AI acts as a powerful co-pilot, augmenting human capabilities and helping us tackle challenges that were previously thought intractable, from academic hurdles to scientific breakthroughs.

The journey of AI is just beginning, and if my experiment with the latest JEE Advanced paper is any indication, we are on the cusp of truly revolutionary advancements.

What are your thoughts on AI’s growing capabilities in complex problem-solving? How do you envision AI shaping the future of education and STEM fields? Share your insights in the comments below!