TernaryLLM preprint reports accuracy gains with lower-bit model weights

Researchers report in a preprint that TernaryLLM improves language-model benchmark results while representing weights with three values. On LLaMA-3, its W1.58A16 configuration outperforms a previous W2A16 method by 5.8 in C4 perplexity and by 8.2% in average accuracy on zero-shot tasks. The approach learns both scales and shifts when converting weights, addressing asymmetric outliers and non-zero means. It also transfers information from the floating-point model by matching internal features through cosine similarity. The reported gains concern text-generation and zero-shot benchmarks across different model families; the conversion still requires recovering information lost when weights are reduced to extremely few values.