CS Major to Present AI Research at NeurIPS 2026

When asked to multiply two numbers, a language model may predict the answer based on patterns in its training data rather than perform the calculation directly. University of Maryland undergraduate Eric Gilerson is investigating a way to change that.

Gilerson, a junior computer science major, has had a single-author paper accepted for a poster presentation at the 2026 Conference on Neural Information Processing Systems, known as NeurIPS. He will present his findings in Sydney, Australia, in December.

The conference brings together researchers in artificial intelligence and machine learning. For Gilerson, the acceptance marks his first experience submitting research to an academic conference.

“Having my first-ever research submission accepted to NeurIPS 2026 as an undergraduate is an incredible milestone that feels surreal and deeply rewarding,” Gilerson said.

His paper, “Don’t Learn What You Can Compute: Arithmetic Residual Blocks for Exact Arithmetic in Transformers,” examines how language models can perform calculations directly rather than rely on predictions that may produce incorrect results.

The approach adds an arithmetic component to the model’s architecture, letting it calculate answers without relying entirely on learned patterns.

“To solve this, I engineered an arithmetic residual block that calculates exact arithmetic results directly inside the transformer, injecting the answer into the token stream rather than forcing the model to approximate it,” Gilerson said.

In one experiment, the modified model achieved 99.95% accuracy across 36,000 short arithmetic expressions, compared with 4.56% for the unmodified model under the same testing conditions. Further research is needed to determine how the approach performs on more complex calculations and larger models.

Beyond raw calculation accuracy, Gilerson uncovered a broader benefit: parameter reallocation. By relieving the neural network from memorizing or approximating arithmetic, the model preserved parameter capacity, allowing it to learn unrelated tasks significantly faster and achieve higher benchmark performance.

“When you stop forcing a model to use representational capacity on deterministic rules such as arithmetic that computation can execute perfectly, you can free up its parameters for more complex, high-level reasoning,” Gilerson said. “My experiments showed that offloading arithmetic didn’t just eliminate computational errors; it freed the network to learn unrelated tasks faster and more effectively.”

Developing the system also required Gilerson to demonstrate its performance through experiments and respond to questions raised during the conference’s peer-review process.

“The review and rebuttal cycle was definitely intense, but I really appreciated how the reviewers challenged my initial assertions because it pushed me to design deeper experiments and empirically validate claims that had previously only seemed sound conceptually,” he said.

Gilerson describes himself primarily as a software engineer interested in building practical systems. That perspective also shapes how he thinks language models should handle tasks requiring precise calculations.

“Much like how humans naturally reach for a calculator or spreadsheet instead of wasting brain power on manual long division, language models should delegate mechanical steps,” Gilerson said.

Although the current implementation focuses on arithmetic, Gilerson sees potential applications in other rule-based operations, including unit conversions, calendar calculations and symbolic transformations.

His next steps include testing the architecture under different conditions and investigating whether language models can learn to formulate mathematical expressions that the component can solve directly.

As he prepares for NeurIPS, Gilerson hopes to use the poster presentation to discuss his findings and receive feedback from researchers working on related problems.

“I am now incredibly excited to head to Sydney to present the poster and discuss these findings directly with other researchers in the area,” Gilerson said.

—Story by Samuel Malede Zewdu, CS Communications