A research paper out of arXiv proposes a way to have parallel text generation and fast cached inference at the same time — something existing diffusion language models haven't managed.
Discrete diffusion language models decode tokens in parallel rather than one at a time, which sounds fast until you hit the architecture wall: bidirectional attention gives you quality but kills KV caching, while causal attention keeps caching but throws away right-side context. The new paper introduces Bifocal dLLMs and a concrete implementation called R2LM (Right-to-Left Mamba). R2LM runs standard causal attention for left-side context — preserving full KV cache compatibility — while a lightweight reverse Mamba state-space model feeds in compressed right-side context without breaking cacheability. The researchers continued pretraining Qwen3-1.7B on 60 billion tokens to test it.
The throughput numbers are the headline: 2.4x to 12.9x faster than bidirectional diffusion models in batch serving, and 1.9x to 2.9x faster than standard autoregressive generation. On quality, R2LM exceeds the causal-only baseline on most benchmarks and surpasses the bidirectional model on average — a meaningful result, though not a clean sweep across every task.
The broader significance is that the KV cache problem has been the quiet ceiling on diffusion language model deployment; every lab experimenting with parallel decoding hits it. R2LM's Mamba sidecar approach is a pragmatic engineering answer rather than a theoretical one, which makes it easier to adopt — assuming the gains hold outside a single continued-pretraining setup.