I-DLM is the first diffusion language model to match autoregressive quality

Diffusion language models have a problem. They can generate tokens in parallel, which should make them faster than autoregressive models. But they have always been worse at actually producing good text. A new paper from Together AI, UIUC, Princeton, Stanford, and UT Austin claims to have fixed this.

The paper introduces I-DLM, the Introspective Diffusion Language Model. The core insight is simple once you see it: autoregressive models agree with what they generate, diffusion models do not. The authors call this “introspective consistency.” When an AR model produces a token, the next forward pass treats that token as ground truth because causal masking forces it. When a diffusion model produces a token, it might disagree with that token on the next denoise step. This disagreement accumulates and degrades output quality.

I-DLM measures this gap with an “introspective acceptance rate.” SDAR, a prior diffusion approach, scores 0.699. I-DLM hits 0.984. That is the difference between a model that second-guesses itself and one that does not.

The method has two parts. First, introspective-consistency training, which converts a pretrained AR model using causal attention, logit shifting, and an all-masked objective. Second, introspective strided decoding (ISD), which generates N tokens per forward pass while verifying previously generated tokens via a p/q acceptance criterion. Think of it as speculative decoding but the model verifies its own outputs within the same diffusion step rather than needing a separate draft model.

The numbers are hard to argue with. I-DLM-8B scores 69.6 on AIME-24 and 45.7 on LiveCodeBench-v6. LLaDA-2.1-mini, which has 16 billion parameters, scores 43.3 and 30.4 respectively. I-DLM-8B beats a model twice its size by 26 points on AIME-24 and 15 points on LiveCodeBench. On ARC-C it matches Qwen3-8B at 95.8, and the 32B variant scores 96.8, beating Qwen3-32B’s 97.2 by a hair on some benchmarks while crushing all prior diffusion models.

Throughput matters too. I-DLM delivers 2.9x to 4.1x higher throughput than LLaDA-2.1-mini at high concurrency (C=64). The authors built a custom inference engine on top of AR-inherited optimizations with a stationary-batch scheduler. Because I-DLM uses strict causal attention, it plugs directly into existing AR serving infrastructure like SGLang. No custom kernels. No special runtime.

There is also a neat trick with gated LoRA. When using low-rank adaptation, ISD enables bit-for-bit lossless acceleration. The model produces exactly the same outputs as the base AR model, just faster. This is rare in the acceleration space. Most speculative decoding approaches change outputs slightly.

The authors evaluated across 15 benchmarks spanning knowledge, reasoning, math, and coding. I-DLM outperforms all prior diffusion LMs across the board. LLaDA-2.0-flash and LLaDA-2.1-flash at 100B parameters have holes in their reported results (marked as “—” in the comparison tables), which makes direct comparison difficult, but the 8B and 32B I-DLM variants beat everything that has full numbers.

Models are on HuggingFace. Code is on GitHub. The paper is arXiv:2604.11035.

The practical question is whether this changes deployment decisions. Diffusion LMs have been a curiosity until now. Interesting architecture, but you would not actually use one in production because the quality gap was too large. If I-DLM genuinely closes that gap while keeping the parallelism advantage, the calculus shifts. Serving a model at 3x throughput with no quality penalty is the kind of thing that matters at scale. The drop-in SGLang compatibility is what makes this more than a research demo. You can try it without rewriting your inference stack.