Block Diffusion Language Models (BDLMs): Overview

• Block Diffusion Language Models (BDLMs) are discrete diffusion language models that factorize sequences into blocks with clean, completed previous blocks and iterative denoising within each block.
• They offer a balance between AR and diffusion methods and utilize block-wise attention, which allows for efficient caching and permits flexible-length continuation.
• BDLMs have shown to improve generation quality and inference time, reaching state-of-the-art language modeling performances.

Block Diffusion LLMs (BDLMs) are discrete diffusion LLMs that factorize sequences autoregressively across contiguous generation blocks while modeling tokens within each block through parallel denoising. For a sequence $\mathbf{x}=\mathbf{x}^{1:B}$ with block size $L'$, their defining factorization is

$$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$

Earlier blocks are clean and fixed, whereas the current block is partially or fully masked and denoised jointly. BDLMs therefore interpolate between autoregressive (AR) LLMs and full-sequence diffusion: $L'=1$ recovers token-level AR modeling, $L'=L$ yields a single full-sequence diffusion block, and intermediate values produce semi-autoregressive generation. The architecture preserves causal blockwise context, KV-cache reuse, and flexible-length continuation while introducing bidirectional parallel prediction within the active block. The formulation was introduced as Block Diffusion or BD3-LM (Arriola et al., 12 Mar 2025).

1. Definition and architectural position

1.1 Blockwise factorization

Let $\mathbf{x}=[x^1,\ldots,x^L]$ be a token sequence and partition it into $B=L/L'$ consecutive blocks,

$\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$

The BDLM likelihood is decomposed as

$$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$

where $\mathbf{x}^{<b}$ denotes all completed preceding blocks. Unlike ordinary AR modeling, the conditional distribution of block $L'$0 is not represented by left-to-right next-token prediction. It is represented by a discrete diffusion model that denoises all $L'$1 positions jointly while conditioning on clean preceding blocks.

This produces an asymmetric dependency structure. Across blocks, generation is autoregressive; within a block, denoising is bidirectional and approximately parallel. The current block attends to noisy tokens in the same block and to clean tokens in earlier blocks, but not to later blocks. Once completed, a block becomes immutable context for subsequent generation and can be represented by cached attention states.

1.2 Interpolation between AR and diffusion

The block size controls the principal quality–parallelism trade-off:

• $L'$2: each block contains one token, recovering the AR factorization in expectation under the masked-diffusion objective.
• $L'$3: the entire sequence is one block, recovering a conventional fixed-length discrete diffusion model.
• Intermediate $L'$4: generation is semi-autoregressive, with sequential block dependence and intra-block parallelism.

Smaller blocks provide stronger causal conditioning and generally better likelihood, but approach AR decoding latency. Larger blocks reduce the number of sequential phases and increase parallelism, but make the diffusion likelihood bound looser and can reduce quality. For $L'$5, the diffusion objective is a bound rather than an exact token-level likelihood; the bound becomes progressively looser as blocks grow.

BDLMs thus differ from fully non-autoregressive diffusion models in their use of clean, immutable prefixes. They also differ from AR models because tokens within a block can be generated in parallel and refined before commitment.

1.3 Attention and cache structure

A BDLM is commonly implemented as a Transformer with a block-causal attention mask. For block $L'$6:

• tokens in block $L'$7 attend bidirectionally to the noisy current block;
• they attend to all clean tokens in blocks $L'$8;
• they cannot attend to later blocks.

The resulting cache contains keys and values for finalized blocks. During inference, the model processes the current noisy block against the prefix cache and appends its finalized representations once denoising is complete. This avoids recomputing completed blocks during later diffusion steps and permits generation beyond a fixed training context by appending additional blocks.

Fast-dLLM v2 extends this systems perspective by adapting pretrained Qwen2.5-Instruct models with blockwise masked supervision, hierarchical block- and sub-block caches, complementary masking, and a parallel decoding pipeline (Wu et al., 30 Sep 2025). Its reported settings include 1.5B- and 7B-parameter models, a block size of $L'$9, and up to $$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$0 speedup over standard AR decoding under particular hardware and decoding configurations.

2. Discrete diffusion within blocks

2.1 Absorbing-mask corruption

BDLMs generally use categorical masked diffusion. Each token is represented as a one-hot vector over a vocabulary augmented with a special mask token. Under an absorbing-mask process,

$$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$1

where $$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$2, $$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$3, and $$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$4 is the mask probability. Once a token becomes masked in the forward process, it remains masked. For a block, token corruption is independent:

$$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$5

More general categorical diffusion can use transition matrices $$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$6,

$$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$7

but absorbing masking is the principal formulation in the BDLM results supplied here.

2.2 Reverse denoising

The ideal reverse transition depends on the posterior

$$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$8

Because the clean token $$p_\theta(\mathbf{x})=\prod_{b=1}^{B}p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}).$$9 is unknown at inference, a neural denoiser predicts

$L'=1$0

and the reverse transition averages the exact posterior over the predicted clean-token distribution:

$L'=1$1

In the SUBS parameterization, an already unmasked token is copied forward and is never remasked, while masked positions are assigned clean-token identities. This makes the denoising process absorbing and allows efficient first-hitting samplers.

2.3 Block diffusion objective

Applying a diffusion ELBO independently to each block yields

$L'=1$2

For continuous-time masked diffusion, the objective is a weighted masked-token cross-entropy:

$L'=1$3

with the negative sign understood when the expression is interpreted as a loss. Only masked positions contribute effective supervision.

The $L'=1$4 case is particularly important. Under masking diffusion, the objective reduces in expectation to the ordinary AR negative log likelihood. However, the variance of the stochastic diffusion estimator can be substantially worse than that of conventional AR training because only a subset of positions is masked at each update. BD3-LM therefore introduces data-driven masking schedules and empirical variance estimation. On LM1B, a tuned full-mask schedule reduced measured objective variance from $L'=1$5 to $L'=1$6 and restored the reported AR perplexity of $L'=1$7 in the $L'=1$8 experiment (Arriola et al., 12 Mar 2025).

3. Training, optimization, and alignment

3.1 Efficient BDLM training

A naïve implementation evaluates the denoiser separately for every block. The BD3-LM training procedure instead precomputes clean keys and values for the full sequence and evaluates all noisy blocks using the relevant prefix caches:

1. sample a diffusion time independently for each block;
2. corrupt each block;
3. run the clean sequence to obtain prefix key-value states;
4. denoise all blocks in parallel using the cache of preceding blocks;
5. concatenate logits and compute the block ELBO;
6. backpropagate.

A concatenated noisy-and-clean representation with a specialized block mask permits all block losses to be evaluated in one attention operation. The reported implementation achieved a measured $L'=1$9–$L'=L$0 training speedup over two separate passes. FlexAttention with sparse block masks produced an approximately $L'=L$1 speedup over naïve PyTorch scaled-dot-product attention in the reported A5000 experiment and an approximately $L'=L$2 end-to-end forward-pass speedup. Training nevertheless remained more expensive than ordinary diffusion training, at less than approximately twice diffusion-training speed in the supplied report (Arriola et al., 12 Mar 2025).

3.2 Mask-schedule and gradient-variance effects

The masking schedule is not merely a sampling detail. Very light masking makes reconstruction trivial, whereas masking every token can reduce the learning signal to token marginals. The optimal schedule depends on block size. Reported LM1B results indicate that smaller blocks benefit from heavier masking, while larger blocks prefer lighter or intermediate masking.

Clipped schedules sample mask rates from

$L'=L$3

with $L'=L$4 selected by grid search to minimize empirical NELBO variance. Low-discrepancy time sampling assigns different subintervals of $L'=L$5 to sequence/block pairs to reduce Monte Carlo error.

3.3 Supervised fine-tuning

Standard diffusion SFT randomly masks response tokens across the entire sequence. This can corrupt tokens that should constitute a clean prefix, expose future suffix tokens that will be unavailable during inference, and distribute loss across blocks that are not currently being generated.

Blockwise SFT addresses this mismatch by selecting one active response block per update. Preceding tokens remain clean, the active block is partially masked, future blocks are fully hidden, and loss is computed only on the active block. The resulting supervision matches the BDLM decoding state:

$L'=L$6

On GSM8K, MATH, and MetaMathQA using LLaDA-8B-Instruct with LoRA, Blockwise SFT achieved reported Pass@1 scores of $L'=L$7 on GSM8K and $L'=L$8 on MATH, compared with $L'=L$9 and $\mathbf{x}=[x^1,\ldots,x^L]$0 for classical SFT. Performance was highest when training and inference block sizes matched, supporting training–inference structural alignment (Sun et al., 27 Aug 2025).

3.4 Curriculum and multi-scale training

Large blocks are attractive at inference but can be harmful during training, particularly for long-chain reasoning. DreamReasoner-8B reports that direct large-block training produces poor reasoning performance, whereas training initially with block size $\mathbf{x}=[x^1,\ldots,x^L]$1 and then introducing larger or mixed block sizes improves robustness across inference block sizes. Its curriculum uses three epochs at $\mathbf{x}=[x^1,\ldots,x^L]$2 followed by mixed block sizes $\mathbf{x}=[x^1,\ldots,x^L]$3, and achieves competitive results with Qwen3-8B-Thinking on mathematical and code reasoning benchmarks (Wu et al., 17 Jun 2026).

Adaptive Block Diffusion generalizes this principle by sampling prefix/window configurations during training rather than training a specialist at one fixed block size. Its objective is

$\mathbf{x}=[x^1,\ldots,x^L]$4

where $\mathbf{x}=[x^1,\ldots,x^L]$5 is the clean-prefix length and $\mathbf{x}=[x^1,\ldots,x^L]$6 is the active denoising-window length. The method is intended to prevent off-grid collapse when inference uses different block sizes, sliding windows, or adaptive schedules. The supplied theoretical claims require qualifications: support coverage removes an obvious configuration mismatch, but shared neural parameters do not generally guarantee pointwise optimality at every configuration (Jain, 28 Jun 2026).

4. Inference and generation strategies

4.1 Standard blockwise generation

A conventional BDLM sampler proceeds sequentially over blocks:

1. initialize the next block as fully masked;
2. predict clean-token distributions;
3. unmask or refine multiple positions in parallel;
4. leave uncertain positions masked;
5. finalize the block;
6. append its key-value states to the prefix cache;
7. continue with the next block.

With absorbing masking, first-hitting sampling can sample the next diffusion time analytically. If $\mathbf{x}=[x^1,\ldots,x^L]$7 tokens remain masked,

$\mathbf{x}=[x^1,\ldots,x^L]$8

This reduces random-variable generation from approximately $\mathbf{x}=[x^1,\ldots,x^L]$9 to $B=L/L'$0, where $B=L/L'$1 is the nominal diffusion horizon and $B=L/L'$2 is vocabulary size. In the reported experiments, a nominal horizon of $B=L/L'$3K steps did not imply $B=L/L'$4K effective model evaluations because tokens were never remasked.

4.2 Confidence-based commitment

Many BDLM samplers use token confidence,

$B=L/L'$5

to determine which masked positions can be committed. A fixed threshold provides a direct quality–parallelism control: higher thresholds defer more tokens and increase computation, while lower thresholds accept more tokens simultaneously but can increase errors.

Bounded Adaptive Confidence Decoding clips a running mean of historical confidence scores between $B=L/L'$6 and $B=L/L'$7:

$B=L/L'$8

Tokens with confidence above $B=L/L'$9 are unmasked, and if none qualifies, the highest-confidence masked token is forcibly unmasked. The reported principal settings were $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$0 and $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$1. On the cited reasoning benchmarks, removing the lower bound increased effective parallelism but substantially reduced accuracy, supporting the role of $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$2 as a quality guard (Lu et al., 10 Feb 2026).

4.3 Adaptive boundaries and deferred commitment

Fixed positional boundaries can separate semantically related tokens and force uncertain boundary tokens to commit without useful future context. Swordsman uses entropy shifts between adjacent positions to select boundaries and dynamically adjusts confidence thresholds within each block. Its reported settings include a minimum entropy-shift threshold $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$3 and an initial confidence threshold $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$4. On LLaDA-8B-Instruct GSM8K with dual caching, Swordsman reached $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$5 accuracy compared with $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$6 for Fast-dLLM under the same cache setting (Zhang et al., 4 Feb 2026).

Deferred Commitment Decoding instead maintains a confidence-aware sliding window over masked positions. Low-uncertainty tokens are committed early, while high-uncertainty tokens remain masked until the window expands and additional context becomes available. Across the reported models and benchmarks, DCD improved metrics by an average of $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$7 percentage points and reduced decoding time by $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$8 on average relative to fixed block or sub-block baselines (Shu et al., 5 Jan 2026).

4.4 Revision after commitment

Ordinary BDLM decoding treats finalized blocks as immutable. Diffusion in Diffusion restores partial reversibility by generating a small-block draft, recording snapshot confidence at token-generation time, remasking low-confidence tokens, and applying larger-block or full-sequence refinement. With a $\mathbf{x}^b=\mathbf{x}^{(b-1)L'+1:bL'}.$9-token drafting stage and a $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$0-token global stage, the reported generative perplexity on OpenWebText at length $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$1 improved from $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$2 for BD3-LM to $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$3, using $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$4 of the baseline fine-tuning budget (Ma et al., 20 Jan 2026).

Multi-Block Editing similarly reopens previously decoded blocks after later context is available. Training-Free MBE uses full attention over a short historical-plus-current window, edits only historical blocks, and recomputes standard-scope cache states for changed blocks. Full MBE with attention-mask SFT and progressively enlarged editing spans increased the reported 12-benchmark average from $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$5 to $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$6, including a $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$7-point gain on AIME 2025, while retaining $$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$8–$$\log p_\theta(\mathbf{x}) = \sum_{b=1}^{B} \log p_\theta(\mathbf{x}^b\mid\mathbf{x}^{<b}),$$9 of standard-decoding throughput (Mou et al., 29 Jun 2026).

5. Extensions of the BDLM paradigm

5.1 Dynamic and semantic grouping

Fixed contiguous blocks are not the only possible grouping mechanism. Dynamic Chunking Diffusion Model replaces positional blocks with learned, input-dependent semantic chunks. Each token receives a chunk assignment $\mathbf{x}^{<b}$0, and the chunk-causal mask is

$\mathbf{x}^{<b}$1

Tokens within a chunk attend bidirectionally, while chunks are ordered autoregressively. Chunking Attention uses learnable subspaces, soft aggregation, straight-through Gumbel-softmax load balancing, and running bias correction. At $\mathbf{x}^{<b}$2B parameters, reported average downstream performance was $\mathbf{x}^{<b}$3 for MDLM, $\mathbf{x}^{<b}$4 for positional BDLM, and $\mathbf{x}^{<b}$5 for DCDM; at $\mathbf{x}^{<b}$6B, the corresponding values were $\mathbf{x}^{<b}$7, $\mathbf{x}^{<b}$8, and $\mathbf{x}^{<b}$9 (Zhu et al., 15 May 2026).

5.2 Multi-block concurrent decoding

Multi-Block Diffusion LLMs extend single-block decoding to a bounded running set of consecutive blocks. Earlier blocks remain in the immutable prefix cache, while several subsequent blocks are simultaneously refined under heterogeneous mask ratios. Multi-block Teacher Forcing exposes the model to bounded noise-groups, randomized chain-uniform schedules, and group-aware dual-stream attention.

A fixed Block Buffer maintains static physical input shapes with dummy slots, enabling cache reuse and CUDA Graph execution while blocks enter, decode, and leave the running set. On LLaDA2-Mini, reported average Tokens Per Forward pass increased from $L'$00 for SingleBD to $L'$01 for the trained MultiBD model, while average accuracy increased from $L'$02 to $L'$03. With DMax, average TPF reached $L'$04 with a reported $L'$05-percentage-point accuracy reduction relative to SingleBD DMax (Jin et al., 28 Jun 2026).

5.3 Mamba–attention hybrids

BDLMs can also be combined with recurrent state-space layers. Full bidirectional Mamba scans are not exactly cacheable because the reverse scan over a prefix depends on the future noisy suffix. BDLM Mamba-H restricts the reverse Mamba scan to the active block while retaining a forward prefix scan. The prefix is cached as attention keys and values in attention layers and as recurrent boundary states in Mamba layers.

At 87M parameters, BDLM Mamba-H achieved a reported C4 validation perplexity of $L'$06, compared with $L'$07 for BDLM attention and $L'$08 for full-sequence DiffuMamba-H. At 65K tokens, it reached $L'$09 tokens per second compared with $L'$10 for full-sequence DiffuMamba-H; at 262K tokens, it reached $L'$11 tokens per second compared with $L'$12 for BDLM attention (Chaturvedi et al., 2 Jul 2026).

5.4 Security implications

Backdoor risks identified for masked diffusion models also have potential relevance to BDLMs. BadDLM increases the masking probability of target-relevant positions in poisoned examples, thereby concentrating denoising supervision on trigger-associated behavior. The paper directly evaluates LLaDA-8B-Instruct and Dream-Instruct-7B rather than an explicitly blockwise model; its transfer to BDLMs is therefore inferential.

Potential BDLM-specific attack surfaces include target-bearing blocks, parallel reconstruction of malicious spans, early poisoned blocks that condition later blocks, and trigger-dependent denoising trajectories. The supplied results report attack success rates of $L'$13–$L'$14 for several target classes on LLaDA-8B-Instruct while largely preserving MMLU utility. Defenses therefore require auditing training-data provenance, mask and block schedules, intermediate denoising trajectories, adapters, and block-conditioned behavior rather than relying solely on AR trigger detection (Zhai et al., 10 May 2026).

6. Empirical properties, limitations, and research directions

6.1 Language modeling and generation quality

On LM1B after $L'$15B training tokens, BD3-LM achieved reported perplexities of at most $L'$16, $L'$17, and $L'$18 for block sizes $L'$19, $L'$20, and $L'$21, respectively, compared with $L'$22 for MDLM and $L'$23 for the AR Transformer. On OpenWebText after $L'$24B tokens, BD3-LM reached at most $L'$25, $L'$26, and $L'$27 for block sizes $L'$28, $L'$29, and $L'$30, compared with $L'$31 for MDLM and $L'$32 for AR. Thus BDLMs were reported as state of the art among the compared diffusion models while remaining behind the AR baseline (Arriola et al., 12 Mar 2025).

Fast-dLLM v2 reported benchmark averages of $L'$33 for its 1.5B model and $L'$34 for its 7B model, with stronger results on several coding and reasoning tasks than corresponding AR and diffusion baselines, although not uniformly across all tasks (Wu et al., 30 Sep 2025). DreamReasoner-8B achieved AIME 2024 scores of $L'$35, $L'$36, $L'$37, and $L'$38 for inference block sizes $L'$39, $L'$40, $L'$41, and $L'$42, respectively, demonstrating relatively stable reasoning quality under curriculum training (Wu et al., 17 Jun 2026).

6.2 Principal limitations

Several limitations follow from the blockwise construction:

1. Sequential block dependence: blocks remain sequential, so small blocks approach AR latency and large blocks are needed for substantial parallelism.
2. Likelihood looseness: for $L'$43, the diffusion objective is a variational bound rather than an exact likelihood.
3. Boundary sensitivity: tokens near the end of a block may lack future context, creating irreversible local errors.
4. Error propagation: an incorrect finalized block becomes conditioning context for all later blocks.
5. Granularity mismatch: training and inference block sizes, schedules, and active-window structures must be aligned or jointly covered.
6. Hardware irregularity: dynamic chunks, adaptive windows, editing spans, and multi-block buffers complicate efficient kernels and cache management.
7. Quality–speed trade-offs: aggressive parallel unmasking can harm code generation, reasoning, fluency, and factual consistency.
8. Evaluation scope: reported speedups depend on model size, hardware, batch size, cache implementation, threshold policy, and workload; TPF or FLOPs do not by themselves establish end-to-end latency.
9. Residual AR gap: BDLM perplexity and broad general-purpose performance remain worse than strong AR baselines in several reported comparisons.
10. Ordinary generative risks: hallucination, harmful or biased content, repetition, long-rollout degradation, and reproduction of copyrighted material remain possible.

6.3 Design principles

The collected results support several design principles:

• Match training and deployment granularity: Blockwise SFT, progressive block-size curricula, ABD, and MultiTF all treat training–inference alignment as central.
• Separate cache boundaries from commitment boundaries: DCD, MBE, and deferred refinement show that a block can remain cache-compatible while token commitment is delayed or selectively revised.
• Use adaptive granularity: CtrlDiff, Swordsman, TCCF, and semantic chunking vary block size or grouping according to semantic difficulty, entropy, or reasoning phase.
• Train for the intended noise geometry: MultiBD requires heterogeneous multi-block noise states rather than single-block teacher-forcing states.
• Preserve stable prefix representations: exact caching requires prefix states to depend only on clean finalized context and not on mutable suffixes or diffusion time.
• Treat confidence as a control signal, not a guarantee: confidence-aware decoding improves efficiency but remains sensitive to calibration and can fail when uncertainty is intrinsic or misleading.
• Combine drafting with revision: multi-stage refinement can recover some global planning and reversibility lost by strict block causality.

BDLMs are consequently best characterized as a family of semi-autoregressive diffusion models rather than a single fixed architecture. Their defining compromise remains

$L'$44

while current research extends this compromise through adaptive partitions, multi-scale training, concurrent running sets, confidence-aware commitment, post-commitment editing, semantic grouping, cache-compatible recurrent hybrids, and denoising objectives explicitly aligned with deployment-time states.