Diffusion language models (dLLMs) are considered a possible alternative to the familiar autoregressive language models. Two new papers on arXiv show where this can pay off: in state management for AI agents and in speeding up text generation. For anyone building or choosing agents or language models, this is a sign that the new model family has strengths of its own.
What happened
The focus is on the paper "DiffuPlan: Editable Planning State for Diffusion Language Model Agents" from the Computation and Language category on arXiv. According to the authors, diffusion-based agents have so far mostly adopted an interaction pattern designed for autoregressive models. There, an intermediate state is written once. New observations can afterwards only be appended as additional text.
The authors consider this a poor fit. Diffusion models can reconstruct earlier text and take into account the context that comes before and after it. With mere appending, precisely this capability remains unused.
The DiffuPlan approach
DiffuPlan is a diffusion-based search agent. It maintains a single, editable planning state. If a clue found later resolves an entity or changes the search route, the agent reconstructs only the affected plan steps. The rest of the plan stays in place.
To do this, the authors added two components:
- Expansion-Partition Marginalization (EPM): A method for revisions of varying length. It marginalizes over equivalent expansion paths. That means it takes into account several routes that lead to the same result.
- Diffusion trajectory RL for multi-turn search: Reinforcement learning on diffusion paths is adapted so that it fits search dialogues with multiple rounds.
The results
On four multi-hop QA benchmarks, DiffuPlan achieves, according to the paper, an average F1 score of 50.24. This outperforms strong autoregressive search agents.
The controlled comparison is more interesting. With identical data and identical training, the editable planning state improves the diffusion agent by 2.44 F1. For the autoregressive baseline model, the improvement is only 0.47.
| Model type | Gain from editable planning state (F1) |
|---|---|
| Diffusion agent (DiffuPlan) | 2.44 |
| Autoregressive baseline model | 0.47 |
The authors conclude from this that diffusion models do not have to imitate autoregressive agents with "append-only" logic. Revising already committed states is, they say, an advantage in its own right.
A second paper: faster decoding
Another arXiv paper, "PV-Decoding: Progressive Verification Decoding with Decode Gating for Diffusion Language Models," deals with a different problem: efficiency during inference. Diffusion models update several tokens in parallel. The difficulty arises when tentative predictions are to be safely accepted.
In speculative decoding, speculative tokens within the same diffusion block can depend on one another. Accepting them too generously risks error propagation. Strict verification, in turn, is computationally expensive.
PV-Decoding divides masked positions into three groups using two confidence thresholds:
- directly decoded tokens,
- speculative tokens,
- tokens that are still unresolved.
The speculative tokens are then verified with progressively stronger strategies: single-batch, dual-batch, and multi-batch. Testing was carried out with three diffusion language models on GSM8K, MBPP, and HumanEval. According to the authors, the number of sequential model forward passes drops by 8 to 37 percent compared with confidence-based decoding. The largest observed accuracy loss is 1.3 percent.
Background
Most well-known language models are autoregressive: they generate text token by token from left to right. What has been written stays put. An agent built on this basis therefore keeps a growing log of thoughts, search queries, and observations. New material is added at the bottom, and old assumptions can at most be corrected later in the text.
Diffusion language models work differently. They start with masked positions and fill in the text over several passes. Multiple positions can be updated at the same time, and the model can re-evaluate text in light of what comes after it. DiffuPlan builds on this.
Why multi-hop search is a suitable test
With multi-hop questions, an agent has to link several pieces of information in sequence. The answer to the first step often determines what is searched for in the second. A plan can turn out to be incomplete or wrong midway through. If a later result is what clarifies which person or object is meant, it would make sense to adjust the affected plan steps directly rather than merely add to them.
Two sides of the same development
The two works address different weaknesses. DiffuPlan asks how an agent can make use of the special capabilities of diffusion models. PV-Decoding asks how inference can be sped up without greatly reducing accuracy. Both are part of the trend toward making diffusion models practical for everyday use, both in their behavior as an agent and in their computational cost.
The framing matters: these are research papers on a preprint server. The results come from the authors themselves and apply to the benchmarks named.
What this means
For developers
Those who build agents often adopt the "append the history" pattern without questioning it. DiffuPlan suggests that the type of model also influences which state management makes sense. With the autoregressive model, the editable planning state brought only a small gain; with the diffusion agent, a considerably larger one. It can therefore be worthwhile to consider agent architecture and model family together.
The idea of recomputing only the affected steps is also practical. Whether the effect holds beyond the tested benchmarks is not stated in the material and should be checked before deployment.
With PV-Decoding, it is worth looking at the trade-off between accuracy and efficiency. Weaker verification causes less overhead; stronger verification increases precision and robustness. Your own measurements on your own hardware remain advisable.
For small businesses
Small teams that want to use language models for research or information functions do not need to infer any urgency from this. The works are research results, not finished products. They do show, however, that alternatives to classic models are maturing. When choosing a model, one can in future ask not only about quality and computational cost but also about how well a model can retroactively correct an intermediate plan.
For everyone who uses AI tools
What is most interesting is the direction. A search assistant that cleanly adjusts its plan as soon as it finds something new could work more reliably on multi-step questions. A model that decodes faster could deliver answers more quickly. Both are currently prospects from experiments, not a guarantee for tools available today.
Open points
- The DiffuPlan results come from four multi-hop QA benchmarks. Transferability to other tasks remains to be verified.
- PV-Decoding was tested with three diffusion models on three benchmarks.
- Both works are preprints. Independent verification can bring additional clarity.
Frequently asked questions
What is DiffuPlan?
DiffuPlan is a search agent built on a diffusion language model. It maintains a single editable planning state and adjusts only the affected steps when new evidence arrives.
How does this approach differ from autoregressive agents?
Autoregressive agents can usually only incorporate new observations by appending more text. DiffuPlan can revise a planning state that has already been written, because diffusion models can reconstruct earlier text using the context before and after it.
What results are reported?
On four multi-hop QA benchmarks, DiffuPlan reaches an average F1 of 50.24 and outperforms strong autoregressive search agents.
How large is the benefit of the editable state?
With the same data and the same training, the editable planning state improves the diffusion agent by 2.44 F1, while the autoregressive baseline improves by only 0.47 F1.
What is Expansion-Partition Marginalization (EPM)?
EPM is a method that supports revisions of variable length. It marginalizes over equivalent expansion paths.
