Diffusion language models (dLLMs) have long been trapped by a binary choice between "KV cache" and "bidirectional context": bidirectional attention is high quality but cannot use KV cache; pure causal can use cache but loses all right-side context. arXiv 2606.27732 introduces the Bifocal dLLM paradigm and instantiates it as R2LM (Right-to-Left Mamba): the main path keeps standard causal attention compatible with KV caching to handle exact left-side context, while a lightweight reverse Mamba SSM on a side-path compresses the right-side context, and the two "focal" sides are concatenated for bidirectional information.
The paper continues pretraining on Qwen3-1.7B with 60B tokens: R2LM achieves 2.4×–12.9× throughput over bidirectional dLLMs, 1.9×–2.9× speedup over AR baselines in batched serving, and surpasses pure causal baselines on most benchmarks while beating bidirectional dLLM on the average. The main decoding path still uses cacheable attention; the side-path is just an SSM state, so it stacks directly with existing vLLM and KV-compression solutions.