Infosiz AI paper out in TACL, 8B model tops GPT-5 on table math
Infosiz said its TaNOS-trained 8B model reached 85.38% accuracy on financial table numerical reasoning and was presented orally at EMNLP in Budapest.
Infosiz announced on October 27 that its research paper on AI numerical reasoning, co-authored with researcher Jo Han-jun as first author, was published in TACL and accepted for presentation at an international conference. The company said the work was also presented orally at EMNLP in Budapest, Hungary, according to datanet.co.kr.
The paper is titled "Generalizing Numerical Reasoning in Table Data through Operation Sketches and Self-Supervised Learning." It describes a training framework the researchers call TaNOS, short for Table Numerical Reasoning with Operation Sketches.
An 8B-parameter AI model trained with TaNOS scored 85.38% accuracy on a financial table numerical reasoning evaluation. Under the researchers' own evaluation conditions, the model outperformed GPT-5, Claude Sonnet 4.5 and Gemini 2.5 Pro.
The team also reported that the model's accuracy fell by less than 2 percentage points when it was trained on financial data and then applied to table data from other fields, such as biotech and mechanical engineering. Trained on only 10% of the data, the model scored 80.13% on the same evaluation, above GPT-5.
Infosiz said it is applying the synthetic data generation methodology to its enterprise AI solution Axiom, which builds benchmarks for evaluating AI performance from internal corporate documents.
Quick answers
What accuracy did the TaNOS-trained model achieve?
The 8B-parameter model scored 85.38% on a financial table numerical reasoning evaluation, and 80.13% when trained on only 10% of the data.
Where was the paper published?
It was published in TACL and accepted for presentation at an international conference, with an oral presentation at EMNLP in Budapest, Hungary.
What is TaNOS?
TaNOS is the researchers' training framework for table numerical reasoning, short for Table Numerical Reasoning with Operation Sketches.