DualGuard: Dual-Mode Quality Control for Logic-Preserving Data Augmentation

Shenghao Li; Lin Zhao

cs.CLSep 26, 2026

DualGuard: Dual-Mode Quality Control for Logic-Preserving Data Augmentation

Authors: Shenghao Li, Lin Zhao

Organizations: Guangzhou University, Guangzhou, China

Abstract

Large language models provide a practical way to generate augmented data for logical reasoning at scale, but a larger generation volume does not guarantee semantic, label, or logical reliability. Existing work has improved generation quality through generation constraints, candidate validation, filtering, and feedback-based revision; however, once a quality judgment is available, deciding whether a candidate should be retained, filtered, or repaired remains an important control problem. We propose DualGuard, a dual-mode quality-control framework for logic-preserving data augmentation. The first mode uses the current instance and candidate batch for selective retention, filtering, attribution, and targeted feedback. The second mode accumulates cross-instance execution records of augmentation actions on top of per-sample diagnosis and attribution, compares new executions against each action's own historical behavior, and supports retrospective anomaly inspection, targeted rollback, and bounded repair. Both modes share semantic verification and additionally use symbolic verification when a reliable logical form is available. Across seven downstream tasks in the Two-Stage Transfer setting, DualGuard achieves the highest Accuracy on five tasks and outperforms the no-augmentation BERT baseline on all seven. Controlled ablations further show complementary roles for Memory, Z3, and history-aware anomaly control.

Figures & tables

Figure 1: Overall framework of DualGuard . The shared generation and validation workflow feeds one task-preconfigured quality-control mode: current-batch control in Difficult Mode or history-aware control in Variance Mode.

Figure 2: Detailed dual-mode quality-control flow. (a) Difficult Mode uses the current instance and current candidate batch for selective retention, filtering, attribution, replenishment, and final review. (b) Variance Mode performs per-sample diagnosis first and then uses action-level historical behavior to detect low historical anomalies and support targeted repair.

Method	FOILO	ProofWriter	LogicQA	ReClor	RTE	CB	LogDed
BERT	51.89	52.87	29.70	46.20	65.22	85.12	17.33
AugGPT	50.57	59.63	30.41	47.13	69.80	83.93	24.00
AttrPrompt	56.27	56.94	30.62	48.13	70.04	85.71	21.33
CoTAM	56.65	54.81	29.95	47.73	71.24	87.50	20.89
FlipDA	53.04	56.57	29.54	49.40	70.16	90.48	24.89
DualGuard	57.47	61.02	31.49	49.67	69.92	86.31	25.67

Table 1: Two-Stage Transfer Accuracy (%), averaged over three predefined learning rates using the same fixed random seed across methods. Best result per task is bold.

Figure 3: BERT versus DualGuard in Two-Stage Transfer. Labels show the absolute Accuracy gain of DualGuard over BERT for each task.

Method	FOILO	ProofWriter	RTE	CB
BERT	42.67	45.33	50.00	70.83
AugGPT	43.33	48.67	53.00	75.00
AttrPrompt	44.67	53.33	54.99	72.92
CoTAM	43.33	38.00	66.01	75.00
FlipDA	44.67	41.33	64.05	77.08
DualGuard	44.67	54.67	57.04	79.17

Table 2: Direct Augmentation Accuracy (%), averaged over three fixed validation folds. Best result per task is bold; FOILO has a three-way tie.

Variant	FOILO	ProofWriter	LogicQA	ReClor	RTE	CB	LogDed
Full	59.28	61.02	31.49	50.20	71.08	85.12	25.78
w/o Memory	57.97	57.78	29.24	45.73	70.95	85.12	22.67
w/o Z3	57.31	58.98	28.42	46.13	69.61	85.12	21.78
w/o Variance	59.11	56.85	29.60	47.07	66.93	80.36	24.89

Table 3: Controlled ablation Accuracy (%).

Figure 4: Accuracy drop relative to the Full controlled ablation system after removing Memory, Z3, or Variance.

Dataset	Sample validity	Logic/label	Transformation	Language	Difference
CB	25/25	25/25	25/25	4.96	3.45
FOILO	25/25	25/25	25/25	4.92	4.05
ProofWriter	25/25	25/25	25/25	4.95	3.75
RTE	25/25	25/25	25/25	4.96	3.35
Simple Logic	25/25	25/25	25/25	4.88	4.10

Table 4: Quality of selected DualGuard augmentations (25 per dataset). The first three metrics are pass counts; the last two are mean ratings on a five-point scale.

Candidate set	$n$	Mean reward	Logic pass	High risk
Initial candidates	30	6.86	86.67%	13.33%
Low-reward outliers	8	1.59	50.00%	50.00%
Attribution-guided refills	35	8.35	91.43%	8.57%
Final accepted	32	9.09	100.00%	0.00%

Table 5: Difficult Mode on eight naturally triggered FOILO instances (from 24 traceable runs). Rows describe different candidate sets, not paired before–after repairs.

Outcome	Count
Completed controlled trials	30
Historical low-reward outliers	6
Strict repair successes	3
Valid repairs without reward gain	1

Table 6: Variance Mode under a fixed action-history snapshot. Counts describe 30 independent controlled trials, not the anomaly rate of normal operation.

Data Augmentation

Data augmentation expands or transforms training examples to improve learning, data efficiency, and robustness to distribution shifts.

See all 242 papers

$\mathrm{R}{\scriptstyle\mathrm{EAL}}\mathrm{G}{\scriptstyle\mathrm{UARD}}$ : Reliable, Efficient, and Auditable LLM Safeguards via Latent Reasoning
Aug 4, 2026Zhinan Liu, Jie Li, Mingyu Kang +1LLM Auditing LLM Guardrails
Xiamen University · University of Science and Technology of China
Guardian-as-an-Advisor: Advancing Next-Generation Guardian Models for Trustworthy LLMs
Apr 8, 2026Yue Huang, Haomin Zhuang, Jiayi Ye +6LLM Guardrails LLM Safety Alignment
University of Notre Dame · University of California, Los Angeles · MIT-IBM Watson AI Lab +1
Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails
Sep 27, 2026Gert Lek, Abele Malan, Chaoyi Zhu +3LLM Safety Benchmarks LLM Guardrails
University of Neuchâtel · Delft University of Technology · IBM Research +1

Aug 4, 2026J/K move · Enter open · S save

$\mathrm{R}{\scriptstyle\mathrm{EAL}}\mathrm{G}{\scriptstyle\mathrm{UARD}}$ : Reliable, Efficient, and Auditable LLM Safeguards via Latent Reasoning

Zhinan Liu, Jie Li, Mingyu Kang, Jiayi Ji

Xiamen University · University of Science and Technology of China

LLM Auditing LLM Guardrails Latent Reasoning

Open paper AskPDF arXiv

Reasoning-based large language model (LLM) safeguards improve safety classification accuracy by generating explicit rationales before making decisions. However, generating explicit rationales adds substantial decoding overhead to every safety decision, while readable explanations are needed only for on-demand audits. This mismatch motivates

\mathrm{R}{\scriptstyle\mathrm{EAL}}\mathrm{G}{\scriptstyle\mathrm{UARD}}

, a reliable, efficient, and auditable framework for LLM safeguards that decouples safety reasoning from rationale generation via latent reasoning. During training, a task-aligned curriculum progressively replaces explicit rationales with continuous latent reasoning under supervision from the remaining rationale text and final safety labels. In parallel, a lightweight audit decoder is trained to generate explanations conditioned on interaction text and latent states under supervision from condensed task-level rationales. At inference, the guard model makes safety decisions through continuous latent reasoning with adaptive stopping, while the audit decoder generates concise explanations when needed. Evaluations of 1B, 3B, and 8B models show that

\mathrm{R}{\scriptstyle\mathrm{EAL}}\mathrm{G}{\scriptstyle\mathrm{UARD}}

improves mean F1 score by up to 1.76 points over GuardReasoner, with up to 15.7

\times

speedups and up to 166.7

\times

reductions in reasoning token usage during routine safety classification, while its audit decoder achieves up to 87.73% explanation acceptability under LLM-based evaluation.

Abstract

Data Augmentation

Explore similar work

REALGUARD\mathrm{R}{\scriptstyle\mathrm{EAL}}\mathrm{G}{\scriptstyle\mathrm{UARD}}REALGUARD: Reliable, Efficient, and Auditable LLM Safeguards via Latent Reasoning

Guardian-as-an-Advisor: Advancing Next-Generation Guardian Models for Trustworthy LLMs

Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails

$\mathrm{R}{\scriptstyle\mathrm{EAL}}\mathrm{G}{\scriptstyle\mathrm{UARD}}$ : Reliable, Efficient, and Auditable LLM Safeguards via Latent Reasoning