This article introduces Atom2.7m, a 2.74-million-parameter language model that reaches 69.24% accuracy on arithmetic benchmarks through an arithmetic-aware numeric representation rather than scale. The approach treats digits as atomic units, reverses integer spans to prioritize least-significant digits, and adds explicit place-value and operand-role features alongside standard BPE tokenization. The authors argue this shows arithmetic performance in language models depends heavily on how numbers are represented, letting a compact model outperform much larger baselines on the ArithMark2.0 benchmark.
