Cross-Stage Conditional Processing

• Cross-stage conditional processing is a design pattern where state or decisions from one stage explicitly guide subsequent computations, optimizing performance.
• In tasks like iterative transformer detection and 3D Gaussian splatting, the method refines predictions or skips redundant processing based on prior outputs.
• This pattern spans diverse applications—from enhancing speech extraction and segmentation supervision to optimizing accelerator scheduling and quantum operations.

Cross-stage conditional processing is a design pattern in which state, observations, or decisions produced at one stage explicitly determine how later stages execute, rather than treating stages as isolated modules. In iterative transformer detection, the issue appears when decoder stage $t+1$ must refine the box predicted at stage $t$ using the previous stage’s localization state (Liu et al., 2023). In 3D Gaussian Splatting acceleration, the term is defined at the dataflow level: preprocessing and rendering are interleaved so rendering state can suppress future Gaussian preprocessing (Pei et al., 21 Jul 2025). Across the literature, the same structural idea recurs in staged pseudo-supervision, continuation-passing request routing, adaptive experimental promotion, conditional estimation, and dynamic temporal control.

1. Core idea and recurring structure

Representative instances span iterative detection, 3D Gaussian Splatting, semi-supervised segmentation, and adaptive trial design (Liu et al., 2023, Pei et al., 21 Jul 2025, Liu et al., 2023, Broberg et al., 2016).

Domain	Stage-conditioned signal	Later-stage effect
DETR-style detection	Previous box $(c_x,c_y,w,h)$	Head-specific reference points for next-stage cross-attention
3DGS acceleration	Accumulated transmittance and early termination	Skip deeper preprocessing and rendering
Semi-supervised segmentation	Conditional class label and hard-organ subset	Organ-specific pseudo-supervision in later training
Two-stage adaptive design	Interim mean $Y_1$	Choose $R$ and Stage 2 sample size

A common pattern is visible. One stage computes a structured state that is richer than a scalar score: a box extent, a transmittance field, a class-conditioned pseudo-label, or an interim sufficient statistic. A later stage then uses that state either to change its search origin, suppress work, alter supervision, or switch inference regime. This suggests that cross-stage conditional processing is less a single algorithm than a family of mechanisms for preserving task-relevant state across stage boundaries.

The literature also shows that the conditional signal may be geometric, probabilistic, semantic, or operational. In Box-DETR it is geometric state derived from a previous box (Liu et al., 2023). In GCC it is runtime rendering state, especially accumulated transmittance and early-termination outcomes (Pei et al., 21 Jul 2025). In adaptive clinical estimation it is the interim-driven decision variable $R$, which changes the conditional law of the final estimator (Broberg et al., 2016). The unifying feature is that later computation is not merely sequenced after earlier computation; it is explicitly conditioned by it.

2. Iterative inference and representation transfer in neural models

In "Box-DETR" (Liu et al., 2023), cross-stage conditional processing is identified as the core issue in iterative box refinement. Conditional DETR decomposes cross-attention weights into a content term and a spatial term,

$$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$

with the spatial query generated conditionally from decoder embedding $\mathbf{f}$ and reference point $\mathbf{s}$ through

$$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$

DAB-DETR iteratively refines boxes, but the reference point used in cross-attention at stage $t$0 is only the center of the box predicted by stage $t$1. Box-DETR argues that this leaves width and height unavailable as explicit reference geometry in dense cross-attention. Its Box Agent converts the previous box into head-specific agent points,

$t$2

with $t$3, and then forms per-head conditional spatial queries

$t$4

The transferred cross-stage state therefore changes from a head-shared center to head-specific points informed by the full previous box. Empirically, the paper reports that with ResNet-50 at 50 epochs, AP improves from $t$5 to $t$6, the starred version reaches $t$7 AP, and fixed agent points reduce AP from $t$8 to $t$9 (Liu et al., 2023).

"Separate in the Speech Chain: Cross-Modal Conditional Audio-Visual Target Speech Extraction" (Mu et al., 2024) uses a different but closely related staged design. AVSepChain splits audio-visual target speech extraction into a speech perception stage and a speech production stage. In the first stage, audio is dominant and vision is conditional; in the second, vision is dominant and the mel-spectrogram $(c_x,c_y,w,h)$0 of the preliminary separated speech is the conditional input. The final waveform is

$(c_x,c_y,w,h)$1

The second stage is therefore not a generic post-filter: it is a stage-conditioned residual producer whose input is an intermediate representation from stage 1. The paper further adds a contrastive semantic matching loss between AV-HuBERT visual embeddings and HuBERT audio embeddings. On LRS2-2Mix, AVSepChain reports SI-SNRi $(c_x,c_y,w,h)$2, SDRi $(c_x,c_y,w,h)$3, PESQ $(c_x,c_y,w,h)$4, and WER $(c_x,c_y,w,h)$5, while removing the AV-Synthesizer reduces SI-SNRi to $(c_x,c_y,w,h)$6 and raises WER to $(c_x,c_y,w,h)$7 (Mu et al., 2024).

A related, less explicit formulation appears in "Vision Backbone Enhancement via Multi-Stage Cross-Scale Attention" (Shang et al., 2023). MSCSA pools feature maps from several backbone stages to a common spatial size, concatenates them into a multi-stage tensor $(c_x,c_y,w,h)$8, and applies attention with

$(c_x,c_y,w,h)$9

while keys and values are built from $Y_1$0, $Y_1$1, and $Y_1$2. The paper frames this as multi-stage and cross-scale interaction rather than explicit stage-conditioned control, but the effect is that later feature updates depend on features originating from all selected backbone stages. This suggests a nearby concept: cross-stage conditioning can also appear as shared latent interaction, not only as hard gating (Shang et al., 2023).

3. Dataflow-level conditioning in accelerators and sparse inference

"GCC: A 3DGS Inference Architecture with Gaussian-Wise and Cross-Stage Conditional Processing" (Pei et al., 21 Jul 2025) gives the most explicit hardware definition of the term. Prior accelerators such as GSCore use a decoupled preprocess-then-render pipeline plus tile-wise rendering. GCC identifies two inefficiencies: in GSCore, roughly $Y_1$3 of runtime is spent in preprocessing while over $Y_1$4 of preprocessed Gaussians remain unused during rendering, and the same Gaussian is loaded $Y_1$5 to $Y_1$6 times on average depending on the scene. GCC therefore interleaves preprocessing and rendering so that rendering state gates future preprocessing. With standard alpha compositing,

$Y_1$7

once transmittance falls below a threshold, deeper Gaussians can be skipped. GCC realizes this through depth grouping, a Gaussian-wise four-stage pipeline, a transmittance mask $Y_1$8, and controller logic that can bypass deeper groups entirely. Across six benchmark scenes, GCC achieves $Y_1$9 to $R$0 area-normalized speedup over GSCore, with geometric mean $R$1, and $R$2 area-normalized energy-efficiency improvement on average (Pei et al., 21 Jul 2025).

"Designing Spatial Architectures for Sparse Attention: STAR Accelerator via Cross-Stage Tiling" (Wang et al., 23 Dec 2025) applies the same principle to sparse Transformer inference under large-scale token parallelism. The paper characterizes prior dynamic sparsity accelerators as stage-isolated: predict, sort, attend, and write back are locally optimized but insufficiently coordinated. STAR instead forwards compact metadata across stages. Differential leading zero scheme (DLZS) produces approximate sparsity information, SADS performs segmented top-$R$3 selection, and sorted indices are then consumed by sorted-updating FlashAttention. The predicate is represented as index lists and binary masks, and coordinated tiling ensures that these metadata remain tile-local across prediction, sorting, and formal compute. The paper reports that DLZS reduces complexity by an average of $R$4 relative to a baseline using 4-bit multiplication in prediction, SADS and SU-FA together contribute another $R$5, and STAR’s software-level optimizations cut computational complexity by $R$6 for the same token sparsity. Architecturally, memory-access reduction rises to $R$7 after incorporating SU-FA and tiled dataflow, and Spatial-STAR achieves a $R$8 throughput improvement over the baseline spatial design (Wang et al., 23 Dec 2025).

Taken together, these systems papers show a particularly strong form of cross-stage conditional processing: the output of an early approximate or rendering stage is not only consumed by the next stage, but used to restructure later arithmetic, memory movement, and scheduling. This suggests that the principle is especially valuable when redundant preprocessing and off-chip traffic dominate total cost.

4. Realizing cross-stage control flow across software, reconfigurable hardware, and quantum operations

"Actor Continuation Passing: Efficient and Extensible Request Routing for Event-Driven Architectures" (Plantikow, 2011) studies staged systems in which each stage runs in its own thread, stages do not share state, and communication occurs by message passing. Its response to cross-stage conditional routing is actor continuation passing. Request logic is written as a script, and the continuation of the remaining computation is sent between stages as a one-argument anonymous function. The Souffleuse DSL exposes remember, compute, goto, run, and asyncRun, and stage-local outcomes can determine later goto(stage) calls without forcing intermediate state into ad hoc event types. The evaluation reports that actor continuation passing performs about equally or better compared to coordinator stages for scripts of at least 3 sequential steps (Plantikow, 2011).

"Supporting Dynamic Control-Flow Execution for Runtime Reconfigurable Processors" (Nassar et al., 20 May 2026) addresses the same problem at the microcode level. Special instructions are executed as VLIW microcode on runtime-reconfigurable accelerator slots, and the paper extends the fabric execution controller with loops, conditional jumps, stalls, and exceptions. Four parameter sets are reserved for jump control, each with a 12-bit counter, and accelerator-output-based branches use 2-bit control signals. The instruction set includes NO_JMP, ALW_JMP, counter-based conditions such as JMP_IF_CNT_EQ, and accelerator-based conditions such as JMP_IF_ACC_EQ. This allows one microcode stage to redirect later stages based on counter values or accelerator outputs. On four benchmark applications, the reported speedup reaches up to $R$9, with average $R$0, at a relatively low overhead of $R$1 (Nassar et al., 20 May 2026).

"Realizing Scalable Conditional Operations through Auxiliary Energy Levels" (Zhang et al., 2024) transfers the idea to NISQ hardware. The paper proposes a transition composite gate scheme in which conditional logic is implemented by routing selected basis states through auxiliary energy levels. Its controlled-unitary family is built as

$R$2

so the conditional branch is realized by a fixed diabatic $R$3 bridge and a tunable single-qutrit $R$4 transition. In entangled state preparation, this reduces circuit depth for three-qubit GHZ and W states by approximately $R$5-$R$6 compared to circuits using only CZ gates, with fidelity improvements of $R$7 and $R$8, respectively. The abstract also reports a $R$9 reduction in circuit depth for a quantum comparator, and the sampled CU family reaches process fidelities from $$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$0 to $$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$1 (Zhang et al., 2024).

These works occupy very different substrates, but they instantiate the same architectural move: conditions are encoded as first-class execution state and propagated forward, rather than rederived or flattened into static sequencing.

5. Training-time and evaluation-time staged conditioning

"Context-aware Conditional Cross Pseudo Supervision for Semi-supervised Medical Image Segmentation" (Liu et al., 2023) introduces a training-time form of cross-stage conditional processing. The framework uses a regular multi-class segmenter

$$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$2

and a conditional binary segmenter

$$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$3

where $$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$4 is the selected class label. On unlabeled data, pseudo-supervision is conditioned on the chosen class $$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$5, and context-awareness is introduced by averaging predictions over the overlap region $$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$6 from two overlapping patches before generating pseudo-labels. A later training-stage mechanism, Hard Organ Learning, changes the sampling set $$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$7: it is the full label set in the first $$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$8 iterations and the hard-organ subset afterward, with $$\mathbf{c}_q^T \mathbf{c}_k + \mathbf{p}_q^T \mathbf{p}_k,$$9. On BCV, mean DSC improves from $\mathbf{f}$0 to $\mathbf{f}$1; on MMWHS, it improves from $\mathbf{f}$2 to $\mathbf{f}$3. The ablation reports $\mathbf{f}$4 DSC for CCPS plus context-awareness, and $\mathbf{f}$5 after adding HOL (Liu et al., 2023).

"Task Expansion and Cross Refinement for Open-World Conditional Modeling" (Brahmavar et al., 3 Mar 2026) organizes staged conditioning as a semi-supervised pipeline for open-world conditional modeling. TEXR first generates uninstantiated semantic contexts

$\mathbf{f}$6

then weakly instantiates them into

$\mathbf{f}$7

filters synthetic values by

$\mathbf{f}$8

trains split-specific models $\mathbf{f}$9 and $\mathbf{s}$0, cross-refines $\mathbf{s}$1 with $\mathbf{s}$2 and $\mathbf{s}$3 with $\mathbf{s}$4, and finally trains the full model on

$\mathbf{s}$5

The paper reports average F1 gains of $\mathbf{s}$6 in few-shot and $\mathbf{s}$7 in many-shot settings, while removing the $\mathbf{s}$8-score filter reduces performance from $\mathbf{s}$9 to $$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$0 (Brahmavar et al., 3 Mar 2026).

"Small Experiments, Cheaper Decisions: A Case Study in Staged Promotion for Micro-Pretraining" (Polania, 9 Jun 2026) shows that cross-stage conditioning also applies to experimental resource allocation. The protocol fixes budgets of 2 minutes, 5 minutes, 10 minutes, 60 minutes, and 12 hours, with frozen promotion rules before expensive continuations. The promoted set is conditioned not only on rank but also on predeclared roles such as the bridge reference and the greedy comparator. The executed 12-hour branch spends $$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$1 GPU-hours, and the full staged protocol records $$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$2 training GPU-hours including screening stages. The corresponding accounting counterfactuals are $$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$3 GPU-hours for continuing all four 60-minute candidates and $$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$4 GPU-hours for continuing all nine replicated 10-minute candidates (Polania, 9 Jun 2026).

These examples suggest that training-time cross-stage conditioning serves two distinct roles. It can refine the supervision signal itself, as in class-conditioned pseudo-supervision and cross-refined synthetic data, or it can govern which candidates are allowed to consume later budget, as in staged promotion.

6. Formal statistical, optimization, and temporal-control formulations

In "Conditional Estimation in Two-stage Adaptive Designs" (Broberg et al., 2016), cross-stage conditional processing is formalized as inference after stage-1-driven adaptation. Stage 1 produces

$$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$5

and the interim decision

$$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$6

The final estimator must then be interpreted conditional on the realized regime $$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$7, because stage-1 adaptation changes the conditional law of the full-sample mean. The paper develops the unconditional MLE, the conditionally unbiased Rao-Blackwell estimator, the conditional median unbiased estimator, and conditional maximum-likelihood estimators with and without bias correction. The central point is that stage 1 does not merely precede stage 2; it changes the sample size and therefore the correct inferential model (Broberg et al., 2016).

"On conditional cuts for Stochastic Dual Dynamic Programming" (Van-Ackooij et al., 2017) gives an optimization analogue. For Markov dependence

$$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$8

the Bellman recursion contains conditional expectations $$\mathbf{p}_q = \mathbf{T}\mathbf{p}_s = \boldsymbol{\lambda}_q \odot \mathbf{p}_s.$$9. The paper replaces exact state augmentation by regression-based approximation of those conditional expectations,

$t$00

and defines conditional cuts

$t$01

The cut remains affine in the resource state, but its coefficients are functions of the currently observed exogenous information state. This preserves cross-stage dependence without enlarging the physical state, although the paper also notes that with too many meshes relative to backward trajectories, the backward value can fall below the forward one (Van-Ackooij et al., 2017).

Hunsberger, Posenato, and Combi’s "The Dynamic Controllability of Conditional STNs with Uncertainty" (Hunsberger et al., 2012) places the idea in temporal reasoning. A CSTNU combines labeled time-points and constraints with contingent links, and its execution uncertainty is described by a drama $t$02, where $t$03 is a scenario over propositions and $t$04 is a situation over contingent durations. Dynamic controllability requires an execution strategy that is both viable and dynamic$t$05: decisions at time $t$06 may depend only on observations and contingent durations revealed strictly before $t$07. The paper’s label modification rule shows the cross-stage logic sharply. If events $t$08 and $t$09 must occur before observation $t$10, then the timing of $t$11 relative to $t$12 cannot legitimately depend on proposition $t$13; the label on the corresponding constraint must be strengthened accordingly (Hunsberger et al., 2012).

Across these formal treatments, the same principle reappears in progressively different language: conditional estimation after adaptation, conditional expectations in Bellman recursion, and dynamic controllability under observations and uncertainty. This suggests that cross-stage conditional processing is not merely an engineering heuristic. It is also a statistical and control-theoretic requirement whenever earlier-stage information changes either the feasible set or the probability law faced by later stages.

The literature also makes the tradeoff clear. Cross-stage conditioning can transfer richer state, reduce redundant work, and preserve decision relevance, but it also introduces additional control logic, regression or calibration burden, and new failure modes when the transferred condition is too weak, too noisy, or improperly aligned with later-stage decisions.