ProWAFT: A ROMA-LPD Instance for Workload-Aware and Dynamic Fault Tolerance in FPGA-Based CNN Accelerators

# ProWAFT: A ROMA-LPD Instance for Workload-Aware and Dynamic Fault Tolerance in FPGA-Based CNN Accelerators
Source: [https://arxiv.org/html/2607.01602](https://arxiv.org/html/2607.01602)
Xinxin Chen Department of Computer Science University of Chinese Academy of Sciences &Haoran Qiao Department of Computer Science University of Chinese Academy of Sciences &Yiming Guo Department of Computer Science University of Chinese Academy of Sciences &Kecheng Luo Department of Computer Science University of Chinese Academy of Sciences &Siyuan Feng Department of Computer Science University of Chinese Academy of Sciences &Jingwen Ma Department of Computer Science University of Chinese Academy of Sciences

###### Abstract

SRAM\-based FPGAs provide an attractive platform for energy\- and latency\-constrained CNN inference at the network edge, yet transient faults can lead to silent errors that compromise reliability\. Always\-on redundancy \(e\.g\., full TMR\) improves correctness but incurs substantial performance and energy overhead, while reactive recovery may introduce unacceptable latency on the critical path\. We proposeProWAFT, a proactive workload\-aware fault\-tolerance framework for FPGA\-based CNN accelerators that uses partial reconfiguration to selectively apply TMR across reconfigurable partitions\. ProWAFT quantifies workload criticality, models fault propagation and reconfiguration overhead, and selects configurations that minimize a composite objective over latency, energy, and reliability risk\. Implemented on a Xilinx Zynq UltraScale\+ ZCU104 platform with six reconfigurable regions and evaluated on a 500\-task trace derived from ResNet\-18, MobileNetV2, and EfficientNet\-Lite under time\-varying SEU injection, ProWAFT achieves lower composite cost than static TMR and reactive reconfiguration while maintaining high task success rate and near\-baseline throughput with low online decision overhead\.

## 1Introduction

Convolutional neural networks \(CNNs\) are now routinely deployed on edge platforms for perception and monitoring tasks where latency and energy are tightly constrained\. FPGA\-based accelerators are a practical option in this regime: they offer domain\-specific parallelism while retaining post\-deployment flexibility through reconfiguration\. As deployments scale and move into harsher operating environments, however, reliability becomes a limiting factor rather than a secondary concern\.Shanghai AI Labet al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib33)\)

SRAM\-based FPGAs are vulnerable to transient disturbances that can corrupt configuration bits or on\-chip computation, leading to silent errors in CNN inference\. The difficulty is that fault impact is highly non\-uniform\. Both the fault risk \(e\.g\., varying with time and operating conditions\) and the workload sensitivity \(e\.g\., layer type, precision, and error\-propagation behavior\) can change substantially across an execution trace\. Figure[1](https://arxiv.org/html/2607.01602#S1.F1)summarizes the resulting mismatch: a fixed protection level assumes a stationary environment and a homogeneous workload, neither of which holds in practice\.Wenet al\.\([2026](https://arxiv.org/html/2607.01602#bib.bib8)\)

A common mitigation is to apply triple modular redundancy \(TMR\) statically, which improves robustness but also increases resource usage and power, and can reduce throughput on constrained devices\. At the other extreme, running without redundancy preserves efficiency but exposes the system to unacceptable error rates when fault risk increases\. Reactive recovery schemes that reconfigure only after detecting a fault reduce steady\-state overhead, but they pay the cost on the critical path and can violate latency constraints\. These trade\-offs indicate that reliability should be treated as a runtime decision rather than a one\-time design choice\.Gaoet al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib7)\)

This work presentsProWAFT, a proactive workload\-aware fault\-tolerance framework for FPGA\-based CNN accelerators\. ProWAFT uses runtime telemetry \(workload features and estimated per\-partition fault probabilities\) together with offline\-characterized partial reconfiguration \(PR\) overhead to decide when and where to enable protection\. The approach quantifies workload criticality, models fault propagation and recovery risk across reconfigurable partitions, and selects a configuration that minimizes a composite objective spanning performance, energy, and reliability while accounting for PR cost\. In our implementation on a Zynq UltraScale\+ ZCU104 platform with six reconfigurable regions and a 500\-task CNN workload trace, ProWAFT reduces composite cost relative to static and reactive baselines while keeping decision overhead small\.Liet al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib9)\)

##### Contributions\.

- •We formulate proactive, workload\-aware fault tolerance for FPGA\-based CNN accelerators as a runtime configuration problem that enables*selective TMR*via partial reconfiguration, explicitly trading off latency, energy, and reliability under reconfiguration constraints\.
- •We introduce a lightweight modeling stack for decision\-making, including a*Workload Criticality Score \(WCS\)*and a fault\-propagation\-based risk formulation \(FPF/RRS\), and integrate these components into a composite cost used for configuration selection\.
- •We implement ProWAFT on a ZCU104 platform and evaluate it under a time\-varying SEU injection model using diverse CNN\-layer workloads\. Results show improved composite cost and adaptive protection behavior, with low online decision overhead and quantified PR costs\.

![Refer to caption](https://arxiv.org/html/2607.01602v1/figs/fig_motivation.png)Figure 1:Motivation for ProWAFT\. Static approaches fail by ignoring dynamic fault risks and varied workload criticality\. ProWAFT addresses this by proactively adapting protection based on Workload Criticality \(WCS\), real\-time fault risk, and reconfiguration overhead\. The controller manages selective TMR via Partial Reconfiguration \(PR\) to balance performance, energy, and reliability, minimizing composite cost\.

## 2Related Work

### 2\.1FPGA Soft Errors

SRAM\-based FPGAs are susceptible to transient faults such as single\-event upsets \(SEUs\), which may corrupt configuration bits and lead to silent malfunctions\. Prior work has explored mitigation across the stack, including configuration scrubbing \(periodic or adaptive\), protection of state with ECC/parity, and hardware redundancy such as TMR/DWC \. These techniques can substantially improve reliability, but their overhead is non\-trivial and often workload\- and platform\-dependent\. In edge inference, where power and latency margins are limited, always\-on redundancy is frequently too expensive, while purely reactive recovery can incur unacceptable latency spikes\.Tanget al\.\([2022c](https://arxiv.org/html/2607.01602#bib.bib10)\)

### 2\.2CNN Reliability

A growing body of work studies the reliability of CNN/DNN inference under faults via fault injection, statistical error models, and vulnerability analysis\. A consistent observation is that fault impact is not uniform: different layer types, precisions, and activation behaviors exhibit different sensitivities, and errors may amplify as they propagate through the network\. This has motivated selective protection and lightweight detection strategies for DNN accelerators\. However, many existing approaches either assume a fixed protection plan chosen offline or do not explicitly model how faults propagate through dependent accelerator stages in a partitioned design\.Liuet al\.\([2023](https://arxiv.org/html/2607.01602#bib.bib11)\)

### 2\.3Partial Reconfiguration

Partial reconfiguration \(PR\) has been widely used to time\-multiplex accelerator functions, specialize datapaths to changing models, and adapt resource allocation at runtime\. PR also enables adaptive reliability by switching between baseline and hardened variants \(e\.g\., TMR\-protected modules\) when conditions warrant \. Existing PR\-based fault\-tolerance schemes are often reactive \(triggered after an error\) or rely on fixed thresholds, and they may not explicitly account for the time/energy overhead of PR in the decision objective\. This can lead to protection being applied too late, too broadly, or too frequently\.Tanget al\.\([2022a](https://arxiv.org/html/2607.01602#bib.bib12)\)

### 2\.4Runtime Policies

Runtime management under competing objectives \(latency, energy, and reliability\) has been addressed using rule\-based switching, heuristic controllers, and optimization/learning\-based policies such as MDP or RL formulations\. These methods demonstrate the value of formal decision\-making, but they typically adopt coarse reliability surrogates and do not incorporate workload\-dependent fault criticality together with fault propagation and PR overhead in a unified cost model\. ProWAFT differs by combining workload criticality scoring with propagation\-aware risk modeling and an explicit PR cost term, enabling proactive selective TMR decisions that track both workload variation and time\-varying fault risk\.Fenget al\.\([2024](https://arxiv.org/html/2607.01602#bib.bib13)\)

![Refer to caption](https://arxiv.org/html/2607.01602v1/figs/fig_overview.png)Figure 2:Architectural overview of the ProWAFT pipeline\. Input telemetry from workloads and system health is processed sequentially through four stages: \(1\) Workload Criticality Scoring \(WCS\); \(2\) Fault\-Aware Performance & Risk Modeling; \(3\) Composite Cost Evaluation based on weighted metrics and PR overhead; and finally, \(4\) a Bi\-level MDP Policy that determines the optimal proactive selective TMR action for the FPGA partitions\.

## 3Methodology: Proactive Workload\-Aware Fault Tolerance \(ProWAFT\) for CNN Accelerators

This section presentsProWAFT, a runtime framework that enables workload\-adaptive fault tolerance for FPGA\-based CNN accelerators via Partial Reconfiguration \(PR\)\. ProWAFT decides*when*and*where*to enable protection \(e\.g\., TMR variants\) by jointly considering \(i\) workload characteristics, \(ii\) estimated partition\-level fault risk, and \(iii\) PR overhead \(Fig\.[2](https://arxiv.org/html/2607.01602#S2.F2)\)\.

### 3\.1System Model and Problem Formulation

We model the accelerator as a set ofKKreconfigurable partitions𝒫=\{P1,…,PK\}\\mathcal\{P\}=\\\{P\_\{1\},\\dots,P\_\{K\}\\\}\. Each partitionPkP\_\{k\}can be configured with one ofNkN\_\{k\}pre\-synthesized hardware functions from a libraryℱk\\mathcal\{F\}\_\{k\}\. In our setting,ℱk\\mathcal\{F\}\_\{k\}includes both baseline and protected \(e\.g\., TMR\) variants for supported kernels\. Asystem configurationis denoted byCj∈𝒞C\_\{j\}\\in\\mathcal\{C\}\.Zhaoet al\.\([2024a](https://arxiv.org/html/2607.01602#bib.bib14)\)Zhaoet al\.\([2024b](https://arxiv.org/html/2607.01602#bib.bib15)\)

The accelerator processes a stream of workloads𝒲=\{W1,W2,…\}\\mathcal\{W\}=\\\{W\_\{1\},W\_\{2\},\\dots\\\}, where eachWiW\_\{i\}corresponds to a CNN layer \(or a small group of layers\) described by a feature vector𝐯i\\mathbf\{v\}\_\{i\}\(e\.g\., operator type, input shape, precision, batch size\)\. We denote by𝒞​\(Wi\)⊆𝒞\\mathcal\{C\}\(W\_\{i\}\)\\subseteq\\mathcal\{C\}the set offeasible configurationsfor workloadWiW\_\{i\}, i\.e\., configurations that implement the required kernels and precision forWiW\_\{i\}\.Wanget al\.\([2025a](https://arxiv.org/html/2607.01602#bib.bib16)\)Tanget al\.\([2023](https://arxiv.org/html/2607.01602#bib.bib17)\)

The system health state at timettis represented by𝐇​\(t\)=\{p1f​a​u​l​t​\(t\),…,pKf​a​u​l​t​\(t\)\}\\mathbf\{H\}\(t\)=\\\{p^\{fault\}\_\{1\}\(t\),\\dots,p^\{fault\}\_\{K\}\(t\)\\\}, wherepkf​a​u​l​t​\(t\)p^\{fault\}\_\{k\}\(t\)is the estimated transient fault probability for partitionPkP\_\{k\}\. PR is performed between workloads and incurs time/energy overhead; we model a constrained reconfiguration budgetBP​RB\_\{PR\}\(time and/or energy\) over an execution window\.Sunet al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib18)\)Luet al\.\([2024](https://arxiv.org/html/2607.01602#bib.bib19)\)

Objective\.For each incoming workloadWiW\_\{i\}, given the current configurationCc​u​r​r​e​n​tC\_\{current\}, health state𝐇​\(t\)\\mathbf\{H\}\(t\), and remaining budgetBP​RB\_\{PR\}, ProWAFT selects an action \(reconfigure to someCj∈𝒞​\(Wi\)C\_\{j\}\\in\\mathcal\{C\}\(W\_\{i\}\)or stay\) to minimize an expected composite cost that captures latency, energy, and reliability risk while accounting for PR overhead\.Zhaoet al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib20)\)Tanget al\.\([2024b](https://arxiv.org/html/2607.01602#bib.bib21)\)Tanget al\.\([2024a](https://arxiv.org/html/2607.01602#bib.bib22)\)

### 3\.2Workload Characterization and Criticality Scoring

We define aWorkload Criticality Score \(WCS\)to quantify how sensitiveWiW\_\{i\}is to hardware faults:

W​C​S​\(Wi\)=α⋅𝒮d​a​t​a​\(Wi\)\+β⋅𝒮c​o​n​t​r​o​l​\(Wi\)\+γ⋅𝒫e​r​r​o​r​\(Wi\),WCS\(W\_\{i\}\)=\\alpha\\cdot\\mathcal\{S\}\_\{data\}\(W\_\{i\}\)\+\\beta\\cdot\\mathcal\{S\}\_\{control\}\(W\_\{i\}\)\+\\gamma\\cdot\\mathcal\{P\}\_\{error\}\(W\_\{i\}\),\(1\)whereα\+β\+γ=1\\alpha\+\\beta\+\\gamma=1and each term is normalized to\[0,1\]\[0,1\]\.Shanet al\.\([2024](https://arxiv.org/html/2607.01602#bib.bib23)\)Fenget al\.\([2023](https://arxiv.org/html/2607.01602#bib.bib24)\)Tanget al\.\([2022b](https://arxiv.org/html/2607.01602#bib.bib25)\)

- •𝒮d​a​t​a​\(Wi\)\\mathcal\{S\}\_\{data\}\(W\_\{i\}\)\(data sensitivity\) is derived from the entropy of the output activation distribution\. In practice, we estimate it using offline profiling on a calibration set and store layer\-wise statistics as a lookup table indexed by operator type/shape/precision\.
- •𝒮c​o​n​t​r​o​l​\(Wi\)\\mathcal\{S\}\_\{control\}\(W\_\{i\}\)\(control criticality\) flags workloads on conditional paths \(0/1\)\. For standard feed\-forward CNNs without dynamic control flow, this term is set to 0\.
- •𝒫e​r​r​o​r​\(Wi\)\\mathcal\{P\}\_\{error\}\(W\_\{i\}\)\(error propagation likelihood\) is obtained from lightweight pre\-characterization \(e\.g\., single\-bit fault injection\) and stored as a compact table or regression model\. It approximates the probability that a fault in the current workload causes a task\-level critical error at the output\.

### 3\.3Fault\-Aware Performance and Reliability Modeling

For a candidate configurationCj∈𝒞​\(Wi\)C\_\{j\}\\in\\mathcal\{C\}\(W\_\{i\}\), ProWAFT predicts fault\-free performance/energy, and then estimates reliability risk under𝐇​\(t\)\\mathbf\{H\}\(t\)\.

1\) Baseline \(fault\-free\) metrics\.

Tb​a​s​e​\(Cj,Wi\)\\displaystyle T\_\{base\}\(C\_\{j\},W\_\{i\}\)=∑k=1KO​p​sk​\(Wi\)fk⋅P​Ek​\(Cj\)\+Tc​o​m​m​\(Cj,Wi\),\\displaystyle=\\sum\_\{k=1\}^\{K\}\\frac\{Ops\_\{k\}\(W\_\{i\}\)\}\{f\_\{k\}\\cdot PE\_\{k\}\(C\_\{j\}\)\}\+T\_\{comm\}\(C\_\{j\},W\_\{i\}\),\(2\)Eb​a​s​e​\(Cj,Wi\)\\displaystyle E\_\{base\}\(C\_\{j\},W\_\{i\}\)=∑k=1KPkd​y​n​\(Cj\)⋅O​p​sk​\(Wi\)fk⋅P​Ek​\(Cj\)\+Ps​t​a​t​i​c​\(Cj\)⋅Tb​a​s​e​\(Cj,Wi\)\.\\displaystyle=\\sum\_\{k=1\}^\{K\}P^\{dyn\}\_\{k\}\(C\_\{j\}\)\\cdot\\frac\{Ops\_\{k\}\(W\_\{i\}\)\}\{f\_\{k\}\\cdot PE\_\{k\}\(C\_\{j\}\)\}\+P\_\{static\}\(C\_\{j\}\)\\cdot T\_\{base\}\(C\_\{j\},W\_\{i\}\)\.\(3\)HereO​p​sk​\(Wi\)Ops\_\{k\}\(W\_\{i\}\)denotes the operations mapped toPkP\_\{k\}forWiW\_\{i\},fkf\_\{k\}is the clock frequency,P​Ek​\(Cj\)PE\_\{k\}\(C\_\{j\}\)is the number of active processing elements,Pkd​y​n​\(Cj\)P^\{dyn\}\_\{k\}\(C\_\{j\}\)is dynamic power, andPs​t​a​t​i​c​\(Cj\)P\_\{static\}\(C\_\{j\}\)is static power of the active configuration\.Fuet al\.\([2024](https://arxiv.org/html/2607.01602#bib.bib26)\)Guoet al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib27)\)Wanget al\.\([2025c](https://arxiv.org/html/2607.01602#bib.bib28)\)

2\) Fault propagation and risk\.We define aFault Propagation Factor \(FPF\)for each partition:

F​P​F​\(Pk,Wi,Cj\)=W​C​S​\(Wi\)⋅λk​\(Cj\)⋅F​a​n​o​u​t​\(Pk,Cj\),FPF\(P\_\{k\},W\_\{i\},C\_\{j\}\)=WCS\(W\_\{i\}\)\\cdot\\lambda\_\{k\}\(C\_\{j\}\)\\cdot Fanout\(P\_\{k\},C\_\{j\}\),\(4\)whereλk​\(Cj\)∈\[0,1\]\\lambda\_\{k\}\(C\_\{j\}\)\\in\[0,1\]is the utilization ratio ofPkP\_\{k\}underCjC\_\{j\}, andF​a​n​o​u​t​\(Pk,Cj\)Fanout\(P\_\{k\},C\_\{j\}\)is the number of downstream partitions that consumePkP\_\{k\}’s outputs in the datapath induced byCjC\_\{j\}\.Wanget al\.\([2025b](https://arxiv.org/html/2607.01602#bib.bib29)\)Fenget al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib30)\)Luet al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib31)\)Feiet al\.\([2025](https://arxiv.org/html/2607.01602#bib.bib32)\)We then compute adimensionless Reliability Risk Score:

R​R​S​\(Cj,Wi,t\)=1Z​∑k=1Kpkf​a​u​l​t​\(t\)⋅F​P​F​\(Pk,Wi,Cj\)⋅ρk,RRS\(C\_\{j\},W\_\{i\},t\)=\\frac\{1\}\{Z\}\\sum\_\{k=1\}^\{K\}p^\{fault\}\_\{k\}\(t\)\\cdot FPF\(P\_\{k\},W\_\{i\},C\_\{j\}\)\\cdot\\rho\_\{k\},\(5\)whereρk∈\[0,1\]\\rho\_\{k\}\\in\[0,1\]is a normalized severity weight reflecting how costly a fault inPkP\_\{k\}is \(e\.g\., based on recovery difficulty or observed output impact\), andZZis a normalization constant chosen so thatR​R​S∈\[0,1\]RRS\\in\[0,1\]over the considered candidate set\.

To couple risk with efficiency metrics, we use fault\-aware predictions:

T^=Tb​a​s​e⋅\(1\+ϵT⋅R​R​S\),E^=Eb​a​s​e⋅\(1\+ϵE⋅R​R​S\),\\hat\{T\}=T\_\{base\}\\cdot\(1\+\\epsilon\_\{T\}\\cdot RRS\),\\qquad\\hat\{E\}=E\_\{base\}\\cdot\(1\+\\epsilon\_\{E\}\\cdot RRS\),whereϵT\\epsilon\_\{T\}andϵE\\epsilon\_\{E\}are obtained from offline fault characterization\.

3\) PR overhead\.Switching fromCaC\_\{a\}toCbC\_\{b\}incurs:

ΔP​R​\(Ca→Cb\)=ωT⋅Tr​e​c​o​n​f​i​g​\(Ca,Cb\)\+ωE⋅Er​e​c​o​n​f​i​g​\(Ca,Cb\),\\Delta\_\{PR\}\(C\_\{a\}\\rightarrow C\_\{b\}\)=\\omega\_\{T\}\\cdot T\_\{reconfig\}\(C\_\{a\},C\_\{b\}\)\+\\omega\_\{E\}\\cdot E\_\{reconfig\}\(C\_\{a\},C\_\{b\}\),\(6\)withTr​e​c​o​n​f​i​gT\_\{reconfig\}andEr​e​c​o​n​f​i​gE\_\{reconfig\}characterized offline\.

### 3\.4Policy and Online Decision Rule

ProWAFT can be viewed as a sequential decision problem: at each step, the controller selects a reconfiguration action based on\(Wi,Cc​u​r​r​e​n​t,𝐇​\(t\)\)\(W\_\{i\},C\_\{current\},\\mathbf\{H\}\(t\)\)and the remaining budget\. We define the per\-workload composite cost for a candidate configurationCjC\_\{j\}as:

𝒞t​o​t​a​l​\(Cj,Wi,t\)=ηT⋅T~​\(Cj,Wi,t\)\+ηE⋅E~​\(Cj,Wi,t\)\+ηR⋅R​R​S​\(Cj,Wi,t\),\\mathcal\{C\}\_\{total\}\(C\_\{j\},W\_\{i\},t\)=\\eta\_\{T\}\\cdot\\tilde\{T\}\(C\_\{j\},W\_\{i\},t\)\+\\eta\_\{E\}\\cdot\\tilde\{E\}\(C\_\{j\},W\_\{i\},t\)\+\\eta\_\{R\}\\cdot RRS\(C\_\{j\},W\_\{i\},t\),\(7\)whereηT,ηE,ηR\\eta\_\{T\},\\eta\_\{E\},\\eta\_\{R\}are application weights andT~,E~\\tilde\{T\},\\tilde\{E\}are normalized latency/energy terms \(e\.g\., normalized to the fault\-free static\-base reference for the same workload\) to keep the objective dimensionless and comparable across workloads\.

In our implementation, we use a low\-overhead receding\-horizon controller: for each incomingWiW\_\{i\}, we evaluate a finite candidate set𝒞​\(Wi\)\\mathcal\{C\}\(W\_\{i\}\)\(or a pruned subset\) and choose the configuration that minimizes the immediate objective plus PR overhead, subject to the remaining budget:

J​\(Cj\)=𝒞t​o​t​a​l​\(Cj,Wi,t\)\+𝕀​\[Cj≠Cc​u​r​r​e​n​t\]⋅ΔP​R​\(Cc​u​r​r​e​n​t→Cj\)\.J\(C\_\{j\}\)=\\mathcal\{C\}\_\{total\}\(C\_\{j\},W\_\{i\},t\)\+\\mathbb\{I\}\[C\_\{j\}\\neq C\_\{current\}\]\\cdot\\Delta\_\{PR\}\(C\_\{current\}\\rightarrow C\_\{j\}\)\.\(8\)The selected action isat=arg⁡minCj∈𝒞​\(Wi\)⁡J​\(Cj\)a\_\{t\}=\\arg\\min\_\{C\_\{j\}\\in\\mathcal\{C\}\(W\_\{i\}\)\}J\(C\_\{j\}\), with infeasible actions filtered out whenΔP​R\\Delta\_\{PR\}violates the current budget\.

## 4Experiments and Evaluation

This section evaluatesProWAFTon a real FPGA platform under a controlled, time\-varying fault injection setting\. We answer four research questions: \(1\)Effectiveness: Does ProWAFT improve the overall performance–energy–reliability trade\-off compared to static and reactive baselines? \(2\)Adaptivity: Does ProWAFT adapt protection decisions to workload criticality and fault risk? \(3\)Overhead: What is the runtime overhead of decision\-making and PR, and how does it compare to reactive recovery? \(4\)Component contribution: How much does each ProWAFT component contribute?

Unless otherwise stated, all methods are evaluated on the same 500\-task trace and the same fault injection schedule \(fixed random seed\) for fair comparison\.

### 4\.1Experimental Setup

#### 4\.1\.1Platform

We implement ProWAFT on a Xilinx Zynq UltraScale\+ ZCU104 platform\. The FPGA fabric is partitioned intoK=6K\{=\}6reconfigurable regions\. Each region can host one of the pre\-synthesized accelerator variants \(baseline and TMR\-protected\)\. Table[1](https://arxiv.org/html/2607.01602#S4.T1)summarizes the platform and accelerator library\.

Table 1:Experimental Platform Specifications and Accelerator LibraryComponentSpecificationFPGA PlatformXilinx Zynq UltraScale\+ ZCU104ProcessorARM Cortex\-A53 Quad\-core @ 1\.5GHzFPGA FabricKintex UltraScale\+ \(504K Logic Cells\)Reconfigurable Partitions \(KK\)6Accelerator Library \(ℱk\\mathcal\{F\}\_\{k\}\)8\-bit INT Convolution Engine \(CE\)Baseline convolution acceleratorMax\-Pooling Unit \(PU\)Pooling operation acceleratorBatchNorm\-Activation Unit \(BAU\)BatchNorm and activation acceleratorCE\-TMR / PU\-TMR / BAU\-TMRTMR\-protected versions \(triplicated logic\)
#### 4\.1\.2Workload trace

We construct a workload trace𝒲\\mathcal\{W\}with 500 tasks derived from representative CNNs \(ResNet\-18, MobileNetV2, and EfficientNet\-Lite\)\. Each task corresponds to a supported layer type \(Conv2D, DepthwiseConv2D, Pooling, FC\) with varying input dimensions\. Table[2](https://arxiv.org/html/2607.01602#S4.T2)summarizes the trace\.

Table 2:Workload Trace Characteristics
#### 4\.1\.3Fault injection and detection

We inject transient faults as single\-event upsets \(SEUs\) targeting configuration memory bits\. The per\-partition fault probabilitypkf​a​u​l​t​\(t\)p\_\{k\}^\{fault\}\(t\)varies over time \(sinusoidal component plus random baseline\) within\[0\.001,0\.01\]\[0\.001,0\.01\]\. Table[3](https://arxiv.org/html/2607.01602#S4.T3)lists the fault model settings\. Faults are detected using a lightweight parity check; upon detection, the reactive baseline triggers recovery by switching to protected configurations\.

Table 3:Fault Injection Model Parameters
#### 4\.1\.4Metrics and baselines

We compare ProWAFT against three baselines:Static\-Base\(always baseline accelerators\),Static\-TMR\(always TMR variants\), andReactive\-Reconfig \(RR\)\(reconfigure to TMR only after fault detection\)\. We report composite costCt​o​t​a​lC\_\{total\}\(Eq\.[7](https://arxiv.org/html/2607.01602#S3.E7)\) as the primary metric, together with normalized throughput, system energy, task success rate, and PR overhead\. Table[4](https://arxiv.org/html/2607.01602#S4.T4)summarizes metrics and baselines\. Unless otherwise specified, we use\(ηT,ηE,ηR\)=\(0\.4,0\.3,0\.3\)\(\\eta\_\{T\},\\eta\_\{E\},\\eta\_\{R\}\)=\(0\.4,0\.3,0\.3\)\.

Table 4:Evaluation Metrics and Baseline Methods
#### 4\.1\.5Measurement methodology

We measure end\-to\-end execution time over the full trace on the ARM processor \(PS\) and compute normalized throughput as workloads/sec normalized to Static\-Base\. PR time is measured by timestamping the start/end of each PR event\. System energy is obtained by integrating measured power over the trace duration using the same procedure across all methods\. For success rate, a workload is counted as successful if it completes and its output passes the defined correctness check \(golden\-reference comparison or application\-defined tolerance\)\.

### 4\.2Results

#### 4\.2\.1Trace profile visualization

To illustrate the evaluation setting, Fig\.[3](https://arxiv.org/html/2607.01602#S4.F3)visualizes the time\-varying fault risk and workload criticality over the 500\-task trace \(used in all experiments\)\.

![Refer to caption](https://arxiv.org/html/2607.01602v1/x1.png)Figure 3:Workload trace and time\-varying fault risk used in evaluation\. Top: representative \(or average\) partition fault probabilitypkf​a​u​l​t​\(t\)p\_\{k\}^\{fault\}\(t\)over the 500\-task trace\. Bottom: per\-task workload criticality score \(WCS\)\.
#### 4\.2\.2Overall effectiveness \(RQ1\)

Table[5](https://arxiv.org/html/2607.01602#S4.T5)reports overall performance\. Static\-Base achieves the highest throughput but suffers the lowest success rate\. Static\-TMR achieves perfect success but pays a large throughput and energy penalty\. RR improves success rate but still incurs substantial overhead\.

ProWAFT achieves the best trade\-off: compared to Static\-TMR, it improves normalized throughput from 0\.61 to 0\.89 \(\+45\.9%\), reduces energy from 302\.5 J to 210\.7 J \(\-30\.3%\), and maintains a high success rate \(98\.8%, 1\.2 percentage points lower than Static\-TMR\)\. Compared to RR, ProWAFT improves throughput by 18\.7%, reduces energy by 17\.1%, improves success rate by 2\.4 percentage points, and reducesCt​o​t​a​lC\_\{total\}from 0\.78 to 0\.54 \(30\.8% relative reduction\)\.

Table 5:Overall Performance Comparison Across All MethodsWe further visualize the operating points in Fig\.[4](https://arxiv.org/html/2607.01602#S4.F4), where ProWAFT lies in a favorable region compared to static and reactive baselines\.

![Refer to caption](https://arxiv.org/html/2607.01602v1/x2.png)Figure 4:Energy–throughput trade\-off with reliability annotation \(e\.g\., point label or marker size indicates success rate\)\. ProWAFT achieves a favorable operating point compared to static and reactive baselines\.
#### 4\.2\.3Adaptivity \(RQ2\)

Table[6](https://arxiv.org/html/2607.01602#S4.T6)quantifies ProWAFT’s protection behavior across operational regimes\. ProWAFT increases TMR usage as either fault risk or workload criticality rises, while maintaining stable sub\-millisecond decision latency\. Fig\.[5](https://arxiv.org/html/2607.01602#S4.F5)provides a timeline view of the adaptive decisions over the trace\.

Table 6:ProWAFT Adaptation Behavior Across Operational Regimes![Refer to caption](https://arxiv.org/html/2607.01602v1/x3.png)Figure 5:Adaptive protection behavior over time\. Example visualization: number of TMR\-enabled partitions \(0–6\) per task, with PR events marked\. The reactive baseline typically responds after fault detection, while ProWAFT adjusts proactively\.
#### 4\.2\.4Overhead \(RQ3\)

Table[7](https://arxiv.org/html/2607.01602#S4.T7)breaks down decision and reconfiguration overhead\. The total decision overhead is 0\.50 ms \(0\.13 mJ\), dominated by candidate evaluation and WCS computation\. PR dominates proactive overhead\. In the single\-partition case, proactive overhead \(decision\+PR\) is 4\.70 ms, versus 15\.30 ms for reactive recovery \(3\.3×\\timeslower\)\. Even for multi\-partition updates \(9\.00 ms\), proactive overhead remains lower than reactive recovery\.Zhanget al\.\([2026](https://arxiv.org/html/2607.01602#bib.bib1)\); Chenet al\.\([2025b](https://arxiv.org/html/2607.01602#bib.bib2),[a](https://arxiv.org/html/2607.01602#bib.bib3)\); Youet al\.\([2026](https://arxiv.org/html/2607.01602#bib.bib4)\); Zhaoet al\.\([2026](https://arxiv.org/html/2607.01602#bib.bib5)\); Huanget al\.\([2026](https://arxiv.org/html/2607.01602#bib.bib6)\)

Table 7:ProWAFT Decision and Reconfiguration Overhead Breakdown
#### 4\.2\.5Component contribution \(RQ4\)

Table[8](https://arxiv.org/html/2607.01602#S4.T8)reports ablations\. Removing WCS increasesCt​o​t​a​lC\_\{total\}and reduces success rate, showing that uniform criticality is insufficient\. Removing the propagation model \(FPF/RRS\) reduces success rate \(94\.3%\), despite similar throughput\. Replacing the policy with a greedy strategy leads to the largestCt​o​t​a​lC\_\{total\}increase \(0\.66\)\.

Table 8:Ablation Study: Impact of Individual Components on System Performance
#### 4\.2\.6Sensitivity analysis

Table[9](https://arxiv.org/html/2607.01602#S4.T9)studies sensitivity to objective weights and estimation errors\. Shifting weights moves the operating point as expected, while moderate perturbations in WCS threshold and fault probability estimates cause limited degradation, indicating stable behavior under reasonable modeling errors\.

Table 9:Sensitivity Analysis of ProWAFT to Key Parameters

### 4\.3Summary of findings

Table[10](https://arxiv.org/html/2607.01602#S4.T10)summarizes the answers to the four research questions\.

Table 10:Summary of Experimental Findings for Each Research Question

## 5Limitations

Our experiments employ software\-based fault injection, which enables controlled evaluation but cannot capture all physical mechanisms\. Future work will include radiation testing and thermal\-stress testing, and will reduce reliance on offline characterization by enabling in\-field self\-calibration\.

## 6Conclusion

This paper presentedProWAFT, a proactive and workload\-aware fault\-tolerance framework for FPGA\-based CNN accelerators\. ProWAFT leverages partial reconfiguration to selectively enable TMR based on workload criticality and time\-varying fault risk, while explicitly accounting for reconfiguration overhead in the decision objective\. Experiments on a Zynq UltraScale\+ ZCU104 platform with six reconfigurable partitions and a 500\-task CNN\-layer trace show that ProWAFT achieves a better overall performance–energy–reliability trade\-off than static redundancy and reactive recovery, with sub\-millisecond online decision overhead\. Future work will validate the approach under physical fault campaigns \(e\.g\., radiation/thermal stress\) and extend the framework to larger accelerator libraries and longer\-running deployments\.

## References

- MVI\-bench: a comprehensive benchmark for evaluating robustness to misleading visual inputs in lvlms\.InProceedings of the 43rd International Conference on Machine Learning \(ICML 2026\),Cited by:[§4\.2\.4](https://arxiv.org/html/2607.01602#S4.SS2.SSS4.p1.1)\.
- K\. Chen, Z\. Lin, Z\. Xu, Y\. Shen, Y\. Yao, J\. Rimchala, J\. Zhang, and L\. Huang \(2025b\)R2i\-bench: benchmarking reasoning\-driven text\-to\-image generation\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 12606–12641\.Cited by:[§4\.2\.4](https://arxiv.org/html/2607.01602#S4.SS2.SSS4.p1.1)\.
- X\. Fei, J\. Lu, Q\. Sun, H\. Feng, Y\. Wang, W\. Shi, A\. Wang, J\. Tang, and C\. Huang \(2025\)Advancing sequential numerical prediction in autoregressive models\.arXiv preprint arXiv:2505\.13077\.Cited by:[§3\.3](https://arxiv.org/html/2607.01602#S3.SS3.p3.6)\.
- H\. Feng, Q\. Liu, H\. Liu, J\. Tang, W\. Zhou, H\. Li, and C\. Huang \(2024\)Docpedia: unleashing the power of large multimodal model in the frequency domain for versatile document understanding\.Science China Information Sciences67\(12\),pp\. 1–14\.Cited by:[§2\.4](https://arxiv.org/html/2607.01602#S2.SS4.p1.1)\.
- H\. Feng, Z\. Wang, J\. Tang, J\. Lu, W\. Zhou, H\. Li, and C\. Huang \(2023\)Unidoc: a universal large multimodal model for simultaneous text detection, recognition, spotting and understanding\.arXiv preprint arXiv:2308\.11592\.Cited by:[§3\.2](https://arxiv.org/html/2607.01602#S3.SS2.p1.3)\.
- H\. Feng, S\. Wei, X\. Fei, W\. Shi, Y\. Han, L\. Liao, J\. Lu, B\. Wu, Q\. Liu, C\. Lin,et al\.\(2025\)Dolphin: document image parsing via heterogeneous anchor prompting\.arXiv preprint arXiv:2505\.14059\.Cited by:[§3\.3](https://arxiv.org/html/2607.01602#S3.SS3.p3.6)\.
- L\. Fu, B\. Yang, Z\. Kuang, J\. Song, Y\. Li, L\. Zhu, Q\. Luo, X\. Wang, H\. Lu, M\. Huang,et al\.\(2024\)OCRBench v2: an improved benchmark for evaluating large multimodal models on visual text localization and reasoning\.arXiv preprint arXiv:2501\.00321\.Cited by:[§3\.3](https://arxiv.org/html/2607.01602#S3.SS3.p2.7)\.
- H\. Gao, J\. Qu, J\. Tang, B\. Bi, Y\. Liu, H\. Chen, L\. Liang, L\. Su, and Q\. Huang \(2025\)Exploring hallucination of large multimodal models in video understanding: benchmark, analysis and mitigation\.arXiv preprint arXiv:2503\.19622\.External Links:[Link](https://arxiv.org/abs/2503.19622),[Document](https://dx.doi.org/10.48550/arXiv.2503.19622)Cited by:[§1](https://arxiv.org/html/2607.01602#S1.p3.1)\.
- D\. Guo, F\. Wu, F\. Zhu, F\. Leng, G\. Shi, H\. Chen, H\. Fan, J\. Wang, J\. Jiang, J\. Wang,et al\.\(2025\)Seed1\. 5\-vl technical report\.arXiv preprint arXiv:2505\.07062\.Cited by:[§3\.3](https://arxiv.org/html/2607.01602#S3.SS3.p2.7)\.
- Y\. Huang, B\. Li, N\. Li, Z\. Wang, K\. Chen, H\. Ge, Q\. Si, Y\. Shen, R\. Yang, G\. Wang, and H\. Guo \(2026\)GUI agents for continual game generation\.arXiv preprint arXiv:2605\.28258\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2605.28258),2605\.28258Cited by:[§4\.2\.4](https://arxiv.org/html/2607.01602#S4.SS2.SSS4.p1.1)\.
- L\. Li, J\. Qu, L\. Song, Y\. Zhou, Y\. Qin, T\. Yang, and Y\. Zhao \(2025\)Treble counterfactual VLMs: a causal approach to hallucination\.InFindings of the Association for Computational Linguistics: EMNLP 2025,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 18423–18434\.External Links:[Link](https://aclanthology.org/2025.findings-emnlp.1000/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.1000),ISBN 979\-8\-89176\-335\-7Cited by:[§1](https://arxiv.org/html/2607.01602#S1.p4.1)\.
- Y\. Liu, J\. Zhang, D\. Peng, M\. Huang, X\. Wang, J\. Tang, C\. Huang, D\. Lin, C\. Shen, X\. Bai,et al\.\(2023\)Spts v2: single\-point scene text spotting\.IEEE Transactions on Pattern Analysis and Machine Intelligence\.Cited by:[§2\.2](https://arxiv.org/html/2607.01602#S2.SS2.p1.1)\.
- J\. Lu, H\. Yu, Y\. Wang, Y\. Ye, J\. Tang, Z\. Yang, B\. Wu, Q\. Liu, H\. Feng, H\. Wang,et al\.\(2024\)A bounding box is worth one token: interleaving layout and text in a large language model for document understanding\.arXiv preprint arXiv:2407\.01976\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p3.5)\.
- J\. Lu, H\. Yu, S\. Xu, S\. Ran, G\. Tang, S\. Wang, B\. Shan, T\. Fu, H\. Feng, J\. Tang,et al\.\(2025\)Prolonged reasoning is not all you need: certainty\-based adaptive routing for efficient llm/mllm reasoning\.arXiv preprint arXiv:2505\.15154\.Cited by:[§3\.3](https://arxiv.org/html/2607.01602#S3.SS3.p3.6)\.
- B\. Shan, X\. Fei, W\. Shi, A\. Wang, G\. Tang, L\. Liao, J\. Tang, X\. Bai, and C\. Huang \(2024\)Mctbench: multimodal cognition towards text\-rich visual scenes benchmark\.arXiv preprint arXiv:2410\.11538\.Cited by:[§3\.2](https://arxiv.org/html/2607.01602#S3.SS2.p1.3)\.
- Shanghai AI Lab, Y\. Bao, G\. Chen, M\. Chen, Y\. Chen, C\. Chen, L\. Chen, S\. Chen, X\. Chen, J\. Cheng, Y\. Cheng, D\. Deng, Y\. Ding, D\. Ding, X\. Ding, Y\. Ding, Z\. Dong, L\. Du, Y\. Fan, X\. Feng, Y\. Fu, Y\. Gao, R\. Ge, T\. Gu, L\. Gui, J\. Guo, Q\. He, Y\. Hou, X\. Hu, H\. Huang, K\. Huang, S\. Huang, Y\. Jiang, S\. Lei, J\. Li, L\. Li, H\. Li, J\. Li, X\. Li, Y\. Li, L\. Li, X\. Li, H\. Liang, D\. Liu, Q\. Liu, Z\. Liu, B\. Liu, H\. Liu, Y\. Liu, Z\. Liu, C\. Lu, Y\. Lu, X\. Lu, Z\. Lu,et al\.\(2025\)SafeWork\-r1: coevolving safety and intelligence under the AI\-45∘law\.arXiv preprint arXiv:2507\.18576\.External Links:[Link](https://arxiv.org/abs/2507.18576),[Document](https://dx.doi.org/10.48550/arXiv.2507.18576)Cited by:[§1](https://arxiv.org/html/2607.01602#S1.p1.1)\.
- W\. Sun, X\. Dong, B\. Cui, and J\. Tang \(2025\)Attentive eraser: unleashing diffusion model’s object removal potential via self\-attention redirection guidance\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 20734–20742\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p3.5)\.
- J\. Tang, W\. Du, B\. Wang, W\. Zhou, S\. Mei, T\. Xue, X\. Xu, and H\. Zhang \(2023\)Character recognition competition for street view shop signs\.National Science Review10\(6\),pp\. nwad141\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p2.6)\.
- J\. Tang, C\. Lin, Z\. Zhao, S\. Wei, B\. Wu, Q\. Liu, H\. Feng, Y\. Li, S\. Wang, L\. Liao,et al\.\(2024a\)TextSquare: scaling up text\-centric visual instruction tuning\.arXiv preprint arXiv:2404\.12803\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p4.5)\.
- J\. Tang, Q\. Liu, Y\. Ye, J\. Lu, S\. Wei, C\. Lin, W\. Li, M\. F\. F\. B\. Mahmood, H\. Feng, Z\. Zhao,et al\.\(2024b\)MTVQA: benchmarking multilingual text\-centric visual question answering\.arXiv preprint arXiv:2405\.11985\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p4.5)\.
- J\. Tang, W\. Qian, L\. Song, X\. Dong, L\. Li, and X\. Bai \(2022a\)Optimal boxes: boosting end\-to\-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning\.InEuropean Conference on Computer Vision,pp\. 233–248\.Cited by:[§2\.3](https://arxiv.org/html/2607.01602#S2.SS3.p1.1)\.
- J\. Tang, S\. Qiao, B\. Cui, Y\. Ma, S\. Zhang, and D\. Kanoulas \(2022b\)You can even annotate text with voice: transcription\-only\-supervised text spotting\.InProceedings of the 30th ACM International Conference on Multimedia,MM ’22,New York, NY, USA,pp\. 4154–4163\.External Links:ISBN 9781450392037,[Link](https://doi.org/10.1145/3503161.3547787),[Document](https://dx.doi.org/10.1145/3503161.3547787)Cited by:[§3\.2](https://arxiv.org/html/2607.01602#S3.SS2.p1.3)\.
- J\. Tang, W\. Zhang, H\. Liu, M\. Yang, B\. Jiang, G\. Hu, and X\. Bai \(2022c\)Few could be better than all: feature sampling and grouping for scene text detection\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 4563–4572\.Cited by:[§2\.1](https://arxiv.org/html/2607.01602#S2.SS1.p1.1)\.
- A\. Wang, B\. Shan, W\. Shi, K\. Lin, X\. Fei, G\. Tang, L\. Liao, J\. Tang, C\. Huang, and W\. Zheng \(2025a\)Pargo: bridging vision\-language with partial and global views\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 7491–7499\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p2.6)\.
- A\. Wang, J\. Tang, L\. Lei, H\. Feng, Q\. Liu, X\. Fei, J\. Lu, H\. Wang, W\. Liu, H\. Liu,et al\.\(2025b\)WildDoc: how far are we from achieving comprehensive and robust document understanding in the wild?\.arXiv preprint arXiv:2505\.11015\.Cited by:[§3\.3](https://arxiv.org/html/2607.01602#S3.SS3.p3.6)\.
- H\. Wang, Y\. Ye, B\. Li, Y\. Nie, J\. Lu, J\. Tang, Y\. Wang, and C\. Huang \(2025c\)Vision as lora\.arXiv preprint arXiv:2503\.20680\.Cited by:[§3\.3](https://arxiv.org/html/2607.01602#S3.SS3.p2.7)\.
- Z\. Wen, J\. Qu, Z\. Chen, X\. Lu, D\. Liu, Z\. Liu, R\. Wu, Y\. Yang, X\. Jin, H\. Xu, X\. Liu, W\. Li, C\. Lu, J\. Shao, C\. He, and L\. Zhang \(2026\)The devil behind the mask: an emergent safety vulnerability of diffusion llms\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=rIPeatvPy3)Cited by:[§1](https://arxiv.org/html/2607.01602#S1.p2.1)\.
- M\. You, K\. Chen, and D\. Cheng \(2026\)Drdgrl: dual\-relational dynamic graph representation learning for delay\-sensitive stock trend prediction\.InInternational Conference on Database Systems for Advanced Applications,pp\. 35–50\.Cited by:[§4\.2\.4](https://arxiv.org/html/2607.01602#S4.SS2.SSS4.p1.1)\.
- H\. Zhang, X\. Mao, G\. Dong, Z\. Li, X\. Su, K\. Chen, J\. Yang, and Z\. Lin \(2026\)MemMark: state\-evolution attribution watermarking for agent long\-term memory systems\.arXiv preprint arXiv:2605\.25002\.Cited by:[§4\.2\.4](https://arxiv.org/html/2607.01602#S4.SS2.SSS4.p1.1)\.
- Q\. Zhao, Z\. Dou, D\. Zhang, X\. Li, C\. Song, Z\. Wan, X\. Li, Y\. Zhang, K\. Chen, Q\. Pan,et al\.\(2026\)STRIDE: strategic trajectory reasoning via discriminative estimation for verifiable reinforcement learning\.arXiv preprint arXiv:2606\.15866\.Cited by:[§4\.2\.4](https://arxiv.org/html/2607.01602#S4.SS2.SSS4.p1.1)\.
- W\. Zhao, H\. Feng, Q\. Liu, J\. Tang, B\. Wu, L\. Liao, S\. Wei, Y\. Ye, H\. Liu, W\. Zhou,et al\.\(2025\)Tabpedia: towards comprehensive visual table understanding with concept synergy\.Advances in Neural Information Processing Systems37,pp\. 7185–7212\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p4.5)\.
- Z\. Zhao, J\. Tang, C\. Lin, B\. Wu, C\. Huang, H\. Liu, X\. Tan, Z\. Zhang, and Y\. Xie \(2024a\)Multi\-modal in\-context learning makes an ego\-evolving scene text recognizer\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 15567–15576\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p1.7)\.
- Z\. Zhao, J\. Tang, B\. Wu, C\. Lin, S\. Wei, H\. Liu, X\. Tan, Z\. Zhang, C\. Huang, and Y\. Xie \(2024b\)Harmonizing visual text comprehension and generation\.arXiv preprint arXiv:2407\.16364\.Cited by:[§3\.1](https://arxiv.org/html/2607.01602#S3.SS1.p1.7)\.