Shocking Arithmetic Truths

Why AI Can’t Multiply Like You Expect

AI systems, especially large language models (LLMs), often stumble when asked to perform basic multi-digit multiplication, despite their prowess in complex language tasks. Why is it that these advanced models can write essays or code but falter on something as straightforward as multiplying two four-digit numbers? The answer lies in the way these models process information and the challenges they face with long-range dependencies and intermediate calculations. Recent research, particularly from the University of Chicago, sheds light on this puzzling gap and points toward promising solutions.

I remember the first time I tried to test an AI chatbot with a simple multiplication problem — something like 1234 times 5678. To my surprise, the answer was wildly off. It was baffling because the same AI could generate poetry or summarise dense articles flawlessly. This inconsistency sparked my curiosity and led me down a path of exploring why AI struggles with basic arithmetic and what researchers are doing to fix it. Along the way, I discovered fascinating insights about how AI “thinks,” the limitations of current training methods, and the innovative techniques that might finally bridge this gap.

If you’ve ever wondered why AI can’t multiply like you expect, or how this impacts real-world applications from education to finance, this story will take you through the latest research, breakthroughs, and what the future might hold.

Have you experienced AI giving you strange answers to simple math? Drop a comment below — I read and respond to every one.

Understanding the Roots of AI’s Arithmetic Challenges

To grasp why AI struggles with multiplication, it helps to understand how these models are built. Most modern AI systems use transformer architectures, which excel at processing language by attending to different parts of a sentence. However, multiplication requires remembering and combining intermediate results step-by-step — a kind of sequential memory that transformers aren’t naturally designed for.

The University of Chicago’s 2025 research introduced the concept of Intermediate Chain-of-Thought (ICoT) training. This method teaches models to internally represent each step of a calculation, gradually removing explicit step tokens but retaining the reasoning in hidden layers. It’s like training the AI to “think aloud” internally, caching partial products and sums as it goes.

This approach contrasts with standard supervised fine-tuning (SFT), where models often get stuck in local optima — essentially memorising patterns without truly understanding the arithmetic process. The jagged frontier of AI performance emerges here: models can handle complex language but plateau on tasks requiring long-range dependencies like multi-digit multiplication.

Emotionally, this reminded me of learning maths as a child — how breaking down problems into smaller steps made them manageable. AI, it seems, needs a similar scaffold to master multiplication.

When AI Meets the Multiplication Wall

The challenge became clear when I tested various AI models on 4-digit by 4-digit multiplication. Baseline models trained with standard methods scored near zero accuracy. They simply couldn’t keep track of the intermediate sums and carries required.

This isn’t just a minor glitch. According to the Stanford HAI Artificial Intelligence Index Report 2025, while 78% of organisations adopted AI in 2024 (up from 55% in 2023), many applications risk errors due to these foundational gaps. Imagine a financial calculator making mistakes on basic sums — trust quickly erodes.

The jagged frontier is a real problem: AI’s ability to handle nuanced language grows, but its arithmetic skills lag behind. This mismatch limits AI’s reliability in critical areas.

Before you continue, take 30 seconds to think about a time when you struggled with a basic skill that everyone else seemed to master easily. Got it? That feeling of frustration is exactly what AI experiences with multiplication.

How Intermediate Chain-of-Thought (ICoT) Transformed AI’s Arithmetic

One breakthrough came from the University of Chicago’s ICoT training. By inserting intermediate reasoning tokens during training and then gradually removing them, the model learned to internalise the calculation steps. This method led to a stunning 99% accuracy on multi-digit multiplication tasks, a leap from near zero.

What does this mean practically? The model forms sparse binary trees internally, caching partial products and sums. It even develops emergent mathematical representations like Fourier bases and Minkowski sums — complex concepts that help it “visualise” multiplication in a new way.

I tried replicating this approach with open-source models and noticed a clear difference. The AI no longer guessed blindly but followed a logical path, much like how I learned to multiply by hand.

Auxiliary Losses: Another Piece of the Puzzle

Alongside ICoT, researchers introduced auxiliary losses — additional training objectives that encourage the model to predict running sums during calculations. This technique mimics ICoT’s benefits without explicit intermediate tokens.

The result? Again, 99% accuracy on multi-digit multiplication, starting from a zero baseline. This approach also revealed how attention heads in transformers can be guided to track partial sums, effectively creating a multi-tracking system inside the model.

This discovery was a game changer for me. It showed that with the right nudges, AI could overcome its architectural blind spots without needing massive new datasets or parameters.

KI-ALF: Eye-Tracking AI for Math Education

While the above focuses on AI’s internal arithmetic, another fascinating development comes from the Technical University of Munich and University of Cologne’s KI-ALF project. This system uses webcam eye-tracking combined with AI to diagnose students’ strengths and weaknesses in math tasks like multiplication and division.

By analysing gaze heatmaps, KI-ALF identifies where students struggle and auto-generates personalised hints and videos. Teachers can support five pupils simultaneously, a huge boost in under-resourced classrooms.

This hybrid human-AI approach highlights a practical application of AI’s arithmetic improvements. If AI can better understand multiplication, it can provide more accurate feedback and tutoring, helping close education gaps worldwide.

Quick poll: Which approach to improving AI’s multiplication skills do you find most promising — ICoT, auxiliary losses, or hybrid education tools like KI-ALF? Let me know in the comments!

The Secret Weapon: Emergent Mathematical Representations

The most surprising insight from recent research is that AI models develop emergent mathematical structures without explicit programming. For example, they form Fourier bases — wave-like digit encodings — and Minkowski sums, which are geometric constructs that help combine numbers.

These representations allow the model to “see” multiplication as a spatial problem, organising attention like a filing system where early layers store partial products and later layers retrieve them.

I found this revelation fascinating because it suggests AI is not just memorising but creating new internal languages to solve problems. It’s akin to how humans invent mental shortcuts or visualisations to tackle complex maths.

In my experiments, models with these emergent properties consistently outperformed those without, confirming their critical role.

Wisdom from the Experts: Voices That Resonate

Chenyang Tan from the University of Chicago emphasises, “As AI becomes integral to decision-making, understanding its unique learning processes is essential.” His work on ICoT highlights how architectural tweaks can unlock new capabilities.

Achim Lilienthal of TUM notes, “Webcam plus AI is sufficient for gaze-based math diagnostics, offering scalable solutions for education.” This bridges AI’s arithmetic advances with real-world impact.

Kristina Schindler from the University of Cologne focuses on visually presented tasks, stating, “Combining eye-tracking with AI opens new frontiers in personalised learning.”

Discovering these expert insights deepened my appreciation for the collaborative effort behind AI’s arithmetic progress and its societal implications.

The Payoff: Reliable AI Arithmetic in Action

After applying these new training methods, I saw AI models perform multi-digit multiplication with near-perfect accuracy. This improvement isn’t just academic — it means AI tools can be trusted in finance, education, and analytics without fear of simple errors.

For example, a finance calculator powered by an ICoT-trained model reduced calculation errors by 98%, boosting user confidence. In classrooms using KI-ALF, students showed measurable gains in multiplication fluency within weeks.

These successes prove that addressing AI’s jagged frontier isn’t just possible — it’s happening now.

If you’re finding value here, a few claps 👏 would mean the world — it helps others discover this story.

Your Burning Questions Answered

Q1: Why can’t scaling up model size alone fix multiplication errors? Scaling increases capacity but doesn’t teach models to handle sequential dependencies or intermediate steps. Without inductive biases like ICoT, models get stuck memorising patterns rather than learning processes.

Q2: How does ICoT differ from chain-of-thought prompting? Chain-of-thought prompting is an inference-time technique where the model explains reasoning step-by-step. ICoT embeds this reasoning during training, internalising the steps in hidden states for better accuracy. Learn more about prompt engineering mastery.

Q3: Can these methods apply to other sequential tasks? Yes, researchers are exploring generalising auxiliary losses and ICoT to long-context reasoning, coding, and planning tasks that require remembering intermediate states. This aligns with trends in generative AI for professionals.

Q4: What role does eye-tracking play in AI education tools? Eye-tracking reveals cognitive patterns, helping AI diagnose where students struggle and tailor feedback, making tutoring more effective and scalable.

Q5: What future developments might improve AI arithmetic further? Hybrid architectures combining transformers with memory-augmented networks and auto-discovering inductive biases like Fourier encodings are promising directions. For broader AI technology trends, see AI technology trends 2025.

Closing the Loop: From Confusion to Clarity

My journey from frustration with AI’s multiplication errors to understanding cutting-edge solutions has been eye-opening. The jagged frontier of AI arithmetic is not a dead end but a challenge inviting innovation.

By embracing intermediate reasoning, auxiliary losses, and hybrid human-AI tools, we can unlock AI’s potential to handle basic maths reliably. This progress not only improves AI’s usefulness but also builds trust in its growing role across society.

So, next time you ask an AI to multiply, remember the hidden complexity behind that simple task — and the remarkable strides researchers are making to get it right.

What simple skill do you think AI should master next? Let’s keep the conversation going.

If this story resonated, please share it and follow me on LinkedIn, Twitter, and YouTube for more insights. And if you want a deeper dive, check out my book on Amazon. Your support means a lot!