Parallel Box Decoding (PBD)

• Parallel Box Decoding (PBD) is a decoding strategy that replaces sequential token decisions with block-structured, parallel units, reducing latency and enhancing throughput.
• It employs domain-specific block designs—using atomic geometric units in vision-language tasks and candidate clusters in MIMO detection—to ensure structural coherence and hardware efficiency.
• Empirical evaluations reveal that PBD methods achieve significant speed improvements and competitive accuracy, effectively balancing throughput and performance across various applications.

Parallel Box Decoding (PBD) is a domain-specific term used in several strands of recent arXiv literature for decoding schemes that replace fine-grained sequential decisions with larger, box-structured or block-structured units that can be processed in parallel. In vision-language grounding, PBD denotes decoding bounding boxes and points as atomic geometric units rather than as serialized coordinate tokens, as introduced by "LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding" (Wang et al., 26 May 2026). In MIMO detection, the same acronym denotes a hardware-friendly realization of Box Decoding that searches fixed candidate boxes around a zero-forcing estimate and applies deterministic, sort-free pruning (Yang et al., 29 Nov 2025). Related formulations appear in "Blockwise Parallel Decoding for Deep Autoregressive Models" (Stern et al., 2018) and in "$\ell_p$-Box Decoding of LDPC Codes" (Wu et al., 2017). A plausible unifying interpretation is that these methods all seek to shorten the critical sequential path by aligning the decoding unit with the underlying structure of the problem.

1. Terminological scope and major usages

In the cited literature, PBD is not a single canonical algorithm. It denotes distinct methods in vision-language modeling, wireless detection, sequence generation, and coding theory.

Domain	Meaning of PBD	Representative paper
Vision-language grounding and detection	Parallel Box Decoding of atomic geometric blocks	(Wang et al., 26 May 2026)
MIMO detection	Parallel Box Decoding as a parallel realization of Box Decoding with sort-free pruning	(Yang et al., 29 Nov 2025)
Autoregressive generation	Blockwise parallel decoding of future tokens with longest-prefix validation	(Stern et al., 2018)
LDPC decoding	$\ell_p$-box decoding with distributed ADMM updates	(Wu et al., 2017)

The diversity of usage matters because the term can otherwise be mistaken for a single technique transferable without modification across domains. In practice, each formulation defines its own decoding unit, factorization, and hardware or algorithmic objective. In vision-language grounding, the atomic unit is a box or point. In MIMO detection, it is a small QAM candidate box centered at a ZF-derived reference. In autoregressive generation, it is a block of future tokens. In LDPC decoding, the “box” is part of an exact continuous reformulation of binary constraints.

2. Geometric PBD in vision-language grounding

In LocateAnything, PBD is introduced to address a specific mismatch in vision-LLMs that formulate grounding and detection as a coordinate-token generation problem (Wang et al., 26 May 2026). Standard systems serialize $2\text{D}$ geometry into $1\text{D}$ text tokens, such as textual digits like “1024” emitted as “1”, “0”, “2”, “4”, or coordinate tokens emitted as $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$. The reported limitations are geometric incoherence, because the four coordinates of a box are decoded largely independently; a sequential bottleneck, because token-by-token decoding increases latency and reduces throughput; and the inadequacy of structure-agnostic multi-token prediction, because arbitrary token grouping is misaligned with box geometry and can induce spurious dependencies.

PBD replaces this token-level view with block-aligned decoding. Geometric elements—bounding boxes and points—are treated as atomic units, and the full set of coordinates is predicted in a single parallel step. A block $b$ is a fixed-length unit with $L=6$ that includes structural tokens such as $\langle box\rangle$ and $\langle/box\rangle$ together with the four quantized coordinates. Unoccupied positions are padded with $\langle null\rangle$ for uniform tensor shapes. The paper attributes intra-box geometric coherence to bidirectional attention across tokens within the same block and to box-aligned supervision, in which blocks are defined per box rather than as arbitrary token chunks.

The formulation distinguishes several functional block types: semantic block, box block, negative block, and end block. The semantic block carries identity or text semantics and may spill across multiple blocks. The box block contains the four discretized coordinates. The negative block provides an explicit “no object” signal to reduce hallucinations. The end block marks generation termination.

LocateAnything uses corner coordinates for boxes. Let $\ell_p$0. Coordinates are normalized to $\ell_p$1 and discretized into tokens, with the illustrative discretization

$\ell_p$2

where $\ell_p$3 are image width and height. Standard token-by-token decoding factorizes as

$\ell_p$4

where $\ell_p$5 is the image and $\ell_p$6 is the text query. PBD instead factorizes at block level:

$\ell_p$7

where $\ell_p$8 are visual tokens and $\ell_p$9 is the text query. Within an active block, masked positions are predicted simultaneously, and the intra-block attention is bidirectional. If the block size equals $2\text{D}$0, the PBD formulation reduces to standard NTP.

3. Architecture, supervision, and large-scale data in LocateAnything

LocateAnything is implemented as a native-resolution VLM with a Moon-ViT vision encoder, a Qwen2.5 LLM language decoder, and an MLP projector bridging visual tokens $2\text{D}$1 to the language decoder (Wang et al., 26 May 2026). Coordinates are discretized to $2\text{D}$2 tokens and organized into fixed-length blocks with structural tokens. Boxes and points are therefore represented as atomic units via box or point blocks. The decoding pattern is non-autoregressive within a block and semi-autoregressive across blocks: block-causal attention preserves stable inter-block dependencies, reduces duplicates and misses, and maintains KV cache compatibility.

The training objective is a dual formulation over the same ground truth:

$2\text{D}$3

Both terms are cross-entropy losses over discrete token predictions in the NTP and block-wise MTP streams. The paper also emphasizes design-level constraints rather than regression losses: structural tokens constrain syntax and box structure; the attention mask enforces causal inter-block dependence and bidirectional intra-box coherence; and negative block supervision teaches abstention when a query has no target. It explicitly does not report using box regression losses such as $2\text{D}$4, Smooth $2\text{D}$5, IoU, GIoU, or DIoU, and it does not use Hungarian or bipartite matching. Nor does it use DETR-style set prediction. Instead, a deterministic box ordering policy is used, with X–Y corner order reported as best in ablation.

A second pillar of the method is LocateAnything-Data, curated by a scalable data engine. The dataset statistics are $2\text{D}$6M unique images, $2\text{D}$7M natural-language queries, and $2\text{D}$8M annotated bounding boxes. The task mixture by queries is reported as general object detection $2\text{D}$9, GUI grounding $1\text{D}$0, referring comprehension $1\text{D}$1, text localization $1\text{D}$2, document/scene layout $1\text{D}$3, and point-based localization $1\text{D}$4. For labeled detection datasets such as OpenImages and Objects365, category labels are used to prompt Qwen3-VL and synthesize rich queries, while Molmo predicts candidate points and points inside corresponding ground-truth boxes are retained as reliable supervision. For unlabeled images such as Unsplash and SA-1B, Qwen3-VL generates diverse queries directly; either Molmo is prompted for points followed by SAM 3 to produce boxes, or Rex-Omni predicts boxes directly; final post-verification by Qwen3-VL filters inconsistent boxes. Negative samples are explicitly constructed across domains, with the example of $1\text{D}$5M negatives in Detection.

4. Inference modes, complexity, and empirical frontier in LocateAnything

LocateAnything reports three inference modes: Slow Mode, which uses NTP token-by-token for maximum stability; Fast Mode, which uses MTP PBD to predict full boxes or points per block in one parallel step; and Hybrid Mode, which defaults to PBD and falls back to NTP if format irregularity or spatial ambiguity is detected (Wang et al., 26 May 2026). The ambiguity trigger is defined by two simultaneous conditions: top-1 coordinate token probability $1\text{D}$6, and the max–min range among top-5 coordinate tokens $1\text{D}$7 in $1\text{D}$8 space. During an MTP step, the current block uses bidirectional intra-block attention; after prediction, only committed tokens are retained in the KV cache, while mask tokens and duplicated anchors are evicted. The reported decoding setup uses BF16 precision, block size $1\text{D}$9, nucleus sampling with temperature $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$0, top-$$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$1 $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$2, and repetition penalty $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$3.

The complexity claim is explicit. If each box requires $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$4 tokens in NTP, autoregressive decoding takes $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$5 sequential steps, and throughput degrades linearly with the number of boxes. PBD reduces per-box steps to $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$6 by predicting all masked positions in a single block step. As target boxes grow from $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$7 to $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$8, NTP latency increases sharply, whereas PBD generation time increases slightly, and throughput rises from $$x_1 \rightarrow y_1 \rightarrow x_2 \rightarrow y_2$$9 BPS to $b$0 BPS.

The single-H100, batch-size-$b$1 throughput numbers place LocateAnything on a distinct speed–accuracy frontier. LocateAnything-3B Hybrid is reported at $b$2 BPS, Rex-Omni-3B at $b$3 BPS, and Qwen3-VL-8B at $b$4–$b$5 BPS. In ablation, Fast Mode reaches up to $b$6 BPS, while Hybrid keeps most of the speed gain and preserves accuracy. On detection, the reported LVIS F1 mean is $b$7 for LocateAnything-3B versus $b$8 for Rex-Omni-3B, and LVIS F1@IoU$b$9 is $L=6$0 versus $L=6$1. On COCO, F1 mean is $L=6$2 versus $L=6$3, and F1@IoU$L=6$4 is $L=6$5 versus $L=6$6. On dense detection, Dense200 F1 mean is $L=6$7 versus $L=6$8, and VisDrone F1 mean is $L=6$9 versus $\langle box\rangle$0. On document layout and OCR, DocLayNet F1 mean is $\langle box\rangle$1 versus $\langle box\rangle$2, M6Doc F1 mean is $\langle box\rangle$3 versus $\langle box\rangle$4, and TotalText F1 mean is $\langle box\rangle$5 versus $\langle box\rangle$6. On GUI grounding, ScreenSpot-Pro average is $\langle box\rangle$7. On referring expression comprehension, HumanRef F1 mean is $\langle box\rangle$8, and RefCOCOg test F1 mean is $\langle box\rangle$9. On pointing tasks, COCO F1@Point is $\langle/box\rangle$0, LVIS $\langle/box\rangle$1, Dense200 $\langle/box\rangle$2, and RefCOCOg test $\langle/box\rangle$3.

The ablations clarify what is specific to PBD rather than to coordinate quantization alone. In coordinate representation, NTP Textual gives F1 $\langle/box\rangle$4 at $\langle/box\rangle$5 BPS, NTP Quantized gives F1 $\langle/box\rangle$6 at $\langle/box\rangle$7 BPS, PBD Slow gives F1 $\langle/box\rangle$8 at $\langle/box\rangle$9 BPS, PBD Fast gives F1 $\langle null\rangle$0 at $\langle null\rangle$1 BPS, and PBD Hybrid gives F1 $\langle null\rangle$2 at $\langle null\rangle$3 BPS. Structure-agnostic MTP methods identified as SDLM/Block Diffusion show lower F1, reported as $\langle null\rangle$4, with throughput $\langle null\rangle$5 BPS. Joint training improves Slow Mode F1 from $\langle null\rangle$6 to $\langle null\rangle$7, and on Qwen3-VL-4B the Hybrid re-instantiation improves COCO F1 from $\langle null\rangle$8 to $\langle null\rangle$9 while throughput increases from $\ell_p$00 to $\ell_p$01 BPS.

5. PBD in MIMO detection

In "Box Decoding with Low-Complexity Sort-free Candidate Pruning for MIMO Detection", PBD refers to a hardware-friendly realization of Box Decoding that exploits fixed-size, ZF-centered candidate boxes and deterministic, sort-free pruning rules to process candidate clusters concurrently (Yang et al., 29 Nov 2025). The system model is

$\ell_p$02

with $\ell_p$03, $\ell_p$04, and $\ell_p$05. After QR decomposition $\ell_p$06, the transformed observation is

$\ell_p$07

The partial Euclidean distance recursion is

$\ell_p$08

The defining idea is that Box Decoding forms a fixed-size “box” of constellation points around a ZF-centered reference $\ell_p$09, with grid spacing $\ell_p$10, and that pruning inside each box can be done without global sorting. For a candidate $\ell_p$11 in the layer-$\ell_p$12 box cluster, the PED increment simplifies to

$\ell_p$13

Because $\ell_p$14, $\ell_p$15, and $\ell_p$16 are shared within a cluster, local displacement determines ordering. The paper then introduces three deterministic pruning devices: a QAM-grid symmetry rule, a relative displacement rule around the ZF-centered reference, and sort-free local ordering with early rejection bounds. Metric 1 chooses the local minimum using a few sign checks, while Metric 2 fully orders the four nearest neighbors without computing full squared distances.

These local rules define three pruning strategies. Single-Step Candidate Pruning (SCP) selects exactly one child per parent box cluster and realizes maximal parallelism. Iterative Candidate Pruning (ICP) treats each box cluster as a pre-ordered source list and uses a small multiway merge to pick the $\ell_p$17 globally best children across clusters. Hybrid SICP$\ell_p$18 applies ICP only in the first $\ell_p$19 layers and then reverts to SCP. The hardware mapping is explicitly pipeline-oriented: QR preprocessing and observation transform, per-layer concurrent processing of surviving parents, local pruning units, PED updates only for survivors, and lightweight buffering of $\ell_p$20 and small queues.

The paper’s complexity results are stated in terms of visited nodes per layer. With fixed $\ell_p$21, K-best has $\ell_p$22, DKB has $\ell_p$23, ICP has $\ell_p$24, and SCP has $\ell_p$25. For $\ell_p$26, the totals reported are: for $\ell_p$27 MIMO, SCP $\ell_p$28, ICP $\ell_p$29, DKB $\ell_p$30; for $\ell_p$31, SCP $\ell_p$32, ICP $\ell_p$33, DKB $\ell_p$34. The paper summarizes these as about $\ell_p$35 reduction for Box-SCP relative to DKB and about $\ell_p$36 reduction for Box-ICP, with savings independent of $\ell_p$37. The BER trade-off is also explicit: for $\ell_p$38 MIMO at BER $\ell_p$39, Box-SCP incurs approximately $\ell_p$40 dB, $\ell_p$41 dB, and $\ell_p$42 dB SNR losses at 4-QAM, 16-QAM, and 64-QAM; Box-SICP$\ell_p$43 closes most of the gap and approaches K-best. For $\ell_p$44 MIMO, the corresponding Box-SCP losses are approximately $\ell_p$45 dB, $\ell_p$46 dB, and $\ell_p$47 dB, while one ICP layer improves by approximately $\ell_p$48 dB, $\ell_p$49 dB, and $\ell_p$50 dB.

6. Blockwise parallel decoding as a precursor in autoregressive generation

"Blockwise Parallel Decoding for Deep Autoregressive Models" studies a different setting but provides a closely related template: propose multiple future outputs in parallel, then validate the longest consistent prefix (Stern et al., 2018). For an input $\ell_p$51 and output $\ell_p$52, the standard autoregressive model is

$\ell_p$53

Greedy decoding commits one token per iteration. The proposed scheme chooses a block size $\ell_p$54, predicts $\ell_p$55 in parallel, and then verifies the largest $\ell_p$56 such that each proposed token matches the greedy choice of the base scoring model conditioned on the already accepted prefix.

Under exact-match validation, the output is guaranteed to be identical to greedy decoding. The expected accepted prefix length $\ell_p$57 determines the iteration reduction, with

$\ell_p$58

and the expected number of decoding iterations approximately $\ell_p$59. With a combined scoring and proposal head, the model-call speedup is approximately $\ell_p$60. The measured results show up to approximately $\ell_p$61 wall-clock speedup in machine translation and approximately $\ell_p$62 in image super-resolution. On WMT14 En–De, the fine-tuned and distilled setting with $\ell_p$63 reports BLEU $\ell_p$64 with wall-clock speedup approximately $\ell_p$65, versus distilled greedy BLEU $\ell_p$66. On CelebA image super-resolution, the best setting reports wall-clock speedup approximately $\ell_p$67 at $\ell_p$68, with human evaluation indicating comparable or slightly preferred output quality.

The relation to later PBD terminology is conceptual rather than literal. The parallel unit in this work is a future token block, not a geometric box or a QAM candidate box. Nevertheless, it formalizes a recurring pattern: exploit hardware-parallel scoring inside a block, retain a causal interface between committed blocks, and use a verification or backoff rule to preserve stability.

7. $\ell_p$69-box LDPC decoding and broader distinctions

In "$\ell_p$70-Box Decoding of LDPC Codes", the term “box decoding” refers to an exact continuous surrogate for binary constraints, combined with a distributed ADMM realization that updates variable nodes and check nodes in parallel (Wu et al., 2017). For an LDPC code with parity-check matrix $\ell_p$71, the decoder minimizes $\ell_p$72 subject to code constraints. The key reformulation replaces $\ell_p$73 with

$\ell_p$74

For $\ell_p$75, this intersection equals $\ell_p$76. The practical method uses $\ell_p$77, because projection onto the $\ell_p$78 sphere is closed-form.

The ADMM splitting introduces an auxiliary $\ell_p$79 for the sphere constraint and $\ell_p$80 copies for local parity-polytope constraints. The updates then decouple naturally: variable-node updates are componentwise and depend only on local check neighborhoods; check-node updates project independently onto the parity polytope; and the $\ell_p$81-update is a single Euclidean projection onto the sphere

$\ell_p$82

with $\ell_p$83 and $\ell_p$84. The paper describes this as parameter-free at the modeling level because it eliminates the penalty weight $\ell_p$85 used in penalized LP decoding. Reported runtime on CPU is $\ell_p$86 s at SNR $\ell_p$87 dB and $\ell_p$88 s at SNR $\ell_p$89 dB for $\ell_p$90-box ADMM, versus BP at $\ell_p$91 s and $\ell_p$92 s and penalized $\ell_p$93-ADMM at $\ell_p$94 s and $\ell_p$95 s. The reported WER behavior shows lower WER than BP and penalized ADMM, especially as SNR increases and error floors become prominent.

Across these literatures, one recurring misconception is to equate PBD with generic parallel decoding. The papers do not support that simplification. In LocateAnything, structure-agnostic MTP is explicitly reported as insufficient because arbitrary token grouping is misaligned with box geometry. In MIMO detection, the defining property is not merely parallelism but sort-free, comparator-based local pruning around ZF-centered candidate boxes. In blockwise autoregressive decoding, correctness under exact validation comes from longest-prefix verification against the base model. In LDPC decoding, the “box” is a constraint set rather than a generated object. PBD is therefore better understood as a family of structurally aligned parallel decoding strategies rather than as a single algorithmic primitive.