Researchers independently reproduce Orthrus, a hybrid autoregressive-diffusion architecture that claims lossless speculative decoding via an intra-model consensus mechanism, and test its exact-match claim under different numerical precisions. Under BF16, they find exact trajectory matching in only 45% of cases for the original checkpoint and 43% for their own retrained model across 1,190 prompts spanning 12 domains, while FP32 inference achieves exact matching on all evaluated prompts. Despite this trajectory divergence, downstream lm-eval-harness benchmark scores show no systematic degradation. The authors conclude that Orthrus’s practical losslessness depends on numerical precision and that exact trajectory equivalence should be evaluated separately from downstream task accuracy.
