Enhancing Continual Learning for Software Vulnerability Prediction: Addressing Catastrophic Forgetting via Hybrid-Confidence-Aware Selective Replay for Temporal LLM Fine-Tuning

Xuhui Dou; Hayretdin Bahsi; Alejandro Guerra-Manzanares

cs.CRFeb 27, 2026

Enhancing Continual Learning for Software Vulnerability Prediction: Addressing Catastrophic Forgetting via Hybrid-Confidence-Aware Selective Replay for Temporal LLM Fine-Tuning

Authors: Xuhui Dou, Hayretdin Bahsi, Alejandro Guerra-Manzanares

Organizations: School of Computer Science, University of Nottingham, Nottingham, United Kingdom · School of Informatics and Computing, Northern Arizona University, Flagstaff, United States of America · School of Computer Science, University of Nottingham, Ningbo, China

Abstract

Recent work applies Large Language Models (LLMs) to source-code vulnerability detection, but most evaluations still rely on random train-test splits that ignore time and overestimate real-world performance. In practice, detectors are deployed on evolving code bases and must recognise future vulnerabilities under temporal distribution shift. This paper investigates continual fine-tuning of a decoder-style language model (microsoft/phi-2 with LoRA) on a CVE-linked dataset spanning 2018-2024, organised into bi-monthly windows. We evaluate eight continual learning strategies, including window-only and cumulative training, replay-based baselines and regularisation-based variants. We propose Hybrid Class-Aware Selective Replay (Hybrid-CASR), a confidence-aware replay method for binary vulnerability classification that prioritises uncertain samples while maintaining a balanced ratio of VULNERABLE and FIXED functions in the replay buffer. On bi-monthly forward evaluation Hybrid-CASR achieves a Macro-F1 of 0.667, improving on the window-only baseline (0.651) by 0.016 with statistically significant gains ( $p = 0.026$ ) and stronger backward retention (IBR@1 of 0.741). Hybrid-CASR also reduces training time per window by about 17 percent compared to the baseline, whereas cumulative training delivers only a minor F1 increase (0.661) at a 15.9-fold computational cost. Overall, the results show that selective replay with class balancing offers a practical accuracy-efficiency trade-off for LLM-based temporal vulnerability detection under continuous temporal drift.

Figures & tables

Component	Setting	Component	Setting
Base model	microsoft/phi-2	Learning rate	$2\times 10^{-4}$
Classification head	Binary (2 logits)	Weight decay	0
LoRA rank $r$	16	Training epochs	3
LoRA $\alpha$	32	Batch size (default)	32
LoRA dropout	0.05	Batch size (granularity)	4
Optimizer	AdamW	Evaluation frequency	Per epoch

Table 1: Training configuration maintained across all experimental conditions.

Figure 1: Distribution of vulnerability disclosures across bi-monthly windows (2018–2024), illustrating volume growth and temporal variability.

Strategy	Mechanism	Memory requirements
Replay-1P	Uniform sampling from previous window (Rebuffi et al., 2017).	Memory buffer storing samples from the immediately previous window.
Replay-3P	Uniform sampling from 3 previous windows (Rebuffi et al., 2017).	Memory buffer storing samples from the last 3 windows.
CASR	Uncertainty-based selective replay (Aljundi et al., 2019).	Memory buffer with prioritised vulnerable samples.
Hybrid-CASR	CASR + class balancing (this work).	Memory buffer with class-balanced sampling across VULNERABLE / FIXED.
LB-CL	Class-weighted loss (Wu et al., 2024).	No extra memory beyond model parameters.
OLoRA	Orthogonality constraints on LoRA updates (Lopez-Paz and Ranzato, 2017; Chaudhry et al., 2019).	LoRA adapter parameters (and small statistics for orthogonal subspace).

Table 2: Continual learning strategies with mechanisms and memory requirements.

Granularity	Windows	Mean F1	Std Dev	IQR	Min F1	Max F1
1-month	82	0.654	0.054	0.075	0.506	0.778
2-month	41	0.651	0.044	0.058	0.539	0.714
3-month	27	0.667	0.033	0.041	0.572	0.714
6-month	13	0.669	0.029	0.032	0.611	0.702
12-month	6	0.669	0.030	0.035	0.604	0.695

Table 3: Temporal granularity impact on vulnerability detection performance. All metrics are computed from forward evaluation on held-out windows.

Figure 2: Forward one-step F1 scores across different temporal granularities (1, 2, 3, 6, 12 months), showing variability and trends over chronological evaluation windows.

Method	Mean F1	Std Dev	Min F1	Max F1	Win Rate
Hybrid-CASR	0.667	0.037	0.566	0.728	46.3%
Cumulative	0.661	0.039	0.571	0.749	41.5%
Replay-1P	0.659	0.050	0.510	0.743	39.0%
CASR	0.659	0.038	0.580	0.737	36.6%
LB-CL	0.651	0.044	0.534	0.728	31.7%
Window-only	0.651	0.044	0.539	0.714	29.3%

Table 4: Forward evaluation performance of continual learning methods (37 bi-monthly windows).

Figure 3: Forward F1 comparison across all methods using bi-monthly windows, highlighting differences in performance level and stability over time.

Method	IBR@1	IBR@3	IBR@5	IBR@6	Decay rate	AUC a
Replay-1P	0.791	0.747	0.734	0.729	7.8%	0.750
Hybrid-CASR	0.741	0.726	0.716	0.710	4.2%	0.723
CASR	0.734	0.719	0.707	0.706	3.8%	0.717
LB-CL	0.718	0.703	0.691	0.687	4.3%	0.700
Window-only	0.713	0.701	0.689	0.693	2.8%	0.699
Replay-3P	0.702	0.688	0.676	0.673	4.1%	0.685

Table 5: Knowledge retention measured by Immediate Backward Retention at multiple lags. Decay rate is defined as

(\mathrm{IBR@1}-\mathrm{IBR@6})/\mathrm{IBR@1}

Figure 4: Plasticity–stability trade-off, showing the relationship between forward learning capability and performance consistency. Methods in the upper-right region achieve the best balance.

Method	Time (s)	Time (min)	Speedup	GPU (MB)	F1/min	Efficiency
Hybrid-CASR	432	7.2	1.20 $\times$	19,847	0.093	124%
Window-only	520	8.7	1.00 $\times$	13,525	0.075	100%
Cumulative	8,291	138.2	0.06 $\times$	13,525	0.005	6%

Table 6: Resource consumption and efficiency. Metrics are averaged across 41 bi-monthly training windows.

Method	2019_M05–06 F1	$\Delta$ F1	2020_M05–06 F1	$\Delta$ F1
Window-only	0.539	—	0.548	—
Hybrid-CASR	0.598	+0.059	0.612	+0.064
CASR	0.606	+0.067	0.597	+0.049
Replay-1P	0.584	+0.045	0.588	+0.040
Cumulative	0.585	+0.046	0.574	+0.026
LB-CL	0.557	+0.018	0.561	+0.013

Table 7: Method performance on challenging temporal windows. Values show F1 and improvement over the window-only baseline.

Software Vulnerability Detection

Software vulnerability detection identifies insecure code, pinpoints weaknesses, and helps developers validate or prevent exploitable bugs.

See all 139 papers

Calibration Without Comprehension: Diagnosing the Limits of Fine-Tuning LLMs for Vulnerability Detection in Systems Software
Jun 18, 2026Arastoo Zibaeirad, Marco VieiraLLM Evaluation Software Vulnerability Detection
University of North Carolina at Charlotte, Charlotte, NC, USA
VulTriage: Triple-Path Context Augmentation for LLM-Based Vulnerability Detection
May 10, 2026Wenxin Tang, Xiang Zhang, Junliang Liu +11Software Vulnerability Detection Static Code Analysis
Tsinghua University · Henan University · Dalian Maritime University +8
PromptAudit: Auditing Prompt Sensitivity in LLM-Based Vulnerability Detection
May 22, 2026Steffen J. Camarato, Yahya Hmaiti, Mandana Ghadamian +1Prompt Sensitivity LLM Auditing
University of Central Florida Orlando, Florida, USA

Jun 18, 2026J/K move · Enter open · S save

Calibration Without Comprehension: Diagnosing the Limits of Fine-Tuning LLMs for Vulnerability Detection in Systems Software

Arastoo Zibaeirad, Marco Vieira

University of North Carolina at Charlotte, Charlotte, NC, USA

LLM Evaluation Software Vulnerability Detection Language Model Calibration LLMs for Cybersecurity Data Contamination in Language Models

Open paper AskPDF arXiv

Whether LLMs scoring well on vulnerability benchmarks genuinely reason about security or merely pattern-match on contaminated data remains unresolved. We present CWE-Trace, a framework for LLM vulnerability detection built from 834 manually curated Linux kernel samples spanning 74 CWEs. The framework enforces a strict temporal split (pre-2025 historical set / post-cutoff leakage-free set), preserves context-aware vulnerable--patched pairs, and introduces two diagnostic metrics: the Directional Failure Index (DFI) and Hierarchical Distance and Direction (HDD). We evaluate eight vanilla LLMs and 15 LoRA fine-tuned variants across non-targeted detection, targeted detection, and CWE classification. Our analysis yields two key results. First, data contamination provides no measurable advantage. Function-level analysis shows that 84% of nominally contaminated samples carry no usable memorization signal: vulnerable functions are absent or cross-mapped across datasets, and ~31% of contaminated samples carry CWE misclassification. Second, backbone directional priors dominate fine-tuning. Models exhibit stable, systematic failure modes (DFI ranging from -85.5 to +94.8 pp) that persist from historical to post-cutoff data and resist correction. Fine-tuning shifts the output threshold without changing the decision policy. This is calibration without comprehension: output distributions adapt to training data while the underlying security reasoning remains absent. The weakest backbone at binary detection (DeepSeek-R1) gains the most in coarse CWE classification, revealing that detection and understanding are decoupled capabilities. The best detection score reaches only 52.1% (+2.1 pp above chance); exact CWE ranking remains below 1.3% Top-1 accuracy, confirming that current LLMs lack reliable security reasoning for systems software, regardless of fine-tuning strategy.

Abstract

Software Vulnerability Detection

Explore similar work

Calibration Without Comprehension: Diagnosing the Limits of Fine-Tuning LLMs for Vulnerability Detection in Systems Software

VulTriage: Triple-Path Context Augmentation for LLM-Based Vulnerability Detection

PromptAudit: Auditing Prompt Sensitivity in LLM-Based Vulnerability Detection