- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper presents SWIFT, a framework that co-serves text generation and watermarking using a shared LLM and vLLM engine. It combines LLM-generated synonym candidates, key-conditioned substitution, and adaptive request admission. Experiments evaluate output utility, watermark detectability, robustness, and serving efficiency across multiple tasks and models.
Critical strengths
-
Practical system integration. The framework addresses the deployment overhead of watermarking by sharing the generation model and inference engine. Its sentence-level pipeline provides a concrete approach to coordinating text generation and watermark construction within an existing serving stack.
-
Preservation of downstream task performance. The reported results preserve the no-watermark baseline accuracy on GSM8K and MedMCQA while maintaining detectable watermarks. This supports the practical feasibility of applying the method to tasks where textual changes could affect answer correctness.
Critical weaknesses
- Novelty of the proposed method is limited. The core watermarking algorithm is based on existing work: SafeSeal. SWIFT mainly use the same watermarking LLM as raw text generation, and deploy it on vLLM inference engines using an adaptive scheduling strategy. does not clearly establish a substantially new watermarking algorithm, and the technical contribution of its adaptive serving framework design beyond existing vLLM capabilities and standard request orchestration remains insufficiently demonstrated.
- Insufficient illustration of system efficiency gap compared to baselines (especially SafeSeal). Although Appendix B specifies matched hardware and uses vLLM for SafeSeal’s text generation, it provides limited details on the optimization of its subsequent watermarking pipeline. In particular, it is unclear whether the caching, batching, and GPU parallelization described in the original SafeSeal paper were retained. The existing ablations evaluate SWIFT’s own components but do not fully resolve this issue.
- Limited evaluation under dynamic workloads. Although Fig. 5(a) compares adaptive scheduling with fixed admission ratios, the benefits under changing request arrival rates or bursty traffic remain insufficiently demonstrated. Evaluations across varying load levels, reporting queueing delays and user-facing tail latency, would better support the claimed advantages of adaptive co-serving.
Questions for authors
- Could the authors clarify whether SafeSeal’s original caching, batching, and GPU parallelization optimizations were retained, and report a comparison with these optimizations enabled where applicable under the same hardware and workloads? Does the reported speedup persist?
- Could the authors provide a runtime breakdown for both SWIFT and SafeSeal, covering text generation, candidate generation, candidate filtering, and other post-processing? Which components account for most of the reported 5.9× speedup?
- Could the authors evaluate changing request arrival rates, transitions between low and high load, and bursty traffic? Please compare adaptive scheduling with tuned fixed admission ratios while keeping the model, watermarking pipeline, and vLLM configuration unchanged.
3 comments
Thank you for your experiment and clarification. Your rebuttal resolved most of my technical concerns, but I still have concerns about the paper's novelty and contribution to this domain. So I'll keep my score at 2(Weak Reject).
Clarification of Remaining Novelty Concerns
We thank the reviewer for considering our rebuttal and acknowledging that most technical concerns have been resolved.
We would like to reiterate that SWIFT's primary novelty lies in the systems-level co-design of (i) efficient watermark construction within a shared LLM, (ii) asynchronous co-serving of interdependent generation and watermarking workloads, and (iii) adaptive watermark admission beyond native vLLM scheduling. Our additional experiments demonstrate 1.05×–1.48× speedups over the best fixed admission ratios and a 38.0% reduction in mean user E2E latency compared with native vLLM scheduling.
We respect the reviewer's assessment but would appreciate clarification on which specific aspects of these systems contributions remain insufficiently novel, or whether the concern primarily relates to the absence of a new watermarking primitive.
Thank you again for your constructive feedback.
Response to Reviewer: Systems Novelty, SafeSeal Optimization, and Adaptive Scheduling
We thank the reviewer for their constructive feedback. We revised the manuscript (changes in blue) to clarify SWIFT's systems novelty (Sec. 4.2, Table 3), expand SafeSeal optimizations and runtime analysis (Tables 6, 18), and strengthen scheduling evaluations and limitations (Sec. 4.2, Appx. B).
1. Systems Novelty (Weakness 1). We acknowledge that SWIFT builds upon SafeSeal's lexical watermarking, rather than introducing a new watermark primitive. Its novelty is the systems co-design of watermark construction, asynchronous co-serving, and adaptive scheduling. Related extensions include SWEET (ACL 2024), SIR (ICLR 2024), and Unigram Watermark (ICLR 2024), building on KGW. SWIFT advances beyond SafeSeal and native vLLM in three ways:
(a) Efficient construction. SafeSeal uses spaCy, NLTK, RoBERTa, and BLEURT. SWIFT instead uses the same generation LLM for entity protection, watermarkable-word identification, and candidate generation, processing multiple words in a single sentence-level request instead of per-word RoBERTa inference. It uses structured decoding, batching, and prefix caching.
(b) Asynchronous co-serving. SWIFT shares one vLLM engine between two competing, interdependent workloads. It begins watermarking each completed sentence while subsequent text generation proceeds concurrently, coordinating sentence readiness, generation progress, watermark construction, and ordered streaming output - not simply issuing independent vLLM requests.
(c) Adaptive admission control. Unlike native vLLM scheduling, SWIFT accounts for watermark-specific dependencies by adjusting admission using queue length and waiting-time pressure. Experiments show 1.05×–1.48× speedups over the best fixed ratios across four tasks; versus a native-vLLM No-Controller baseline, mean user E2E latency falls 38.0%, P95 tail latency 19.0%, and workload duration 17.7%. See Sec. 4.2, Table 3, and the ablation.
2. SafeSeal Optimizations and Runtime Breakdown (Weakness 2; Questions 1–2). We conducted matched-hardware, matched-workload comparisons with Gemma-3-27B to distinguish inference-engine acceleration from watermark construction and concurrent serving.
(a) Baseline fairness. We tested four sequential configurations:
SafeSeal (vLLM) is the original reported baseline: only text generation moves to vLLM, preserving RoBERTa and BLEURT. The scoring variant replaces BLEURT with BERTScore-style scoring via vLLM, modifying the similarity method. SafeSeal performs RoBERTa inference per eligible word, with candidates scored in batches per word; SWIFT processes multiple eligible words jointly within sentence-level requests. We clarified caching and GPU parallelization in Tables 6 and 18.
(b) Runtime breakdown. The corresponding component times are:
Moving SafeSeal generation to vLLM alone leaves watermarking at 5.056s, about 95% of total runtime. Optimizing similarity reduces the total to 2.776s; sequential SWIFT reaches 2.713s without auxiliary models or similarity scoring. SWIFT further overlaps generation and watermarking with adaptive scheduling, achieving 0.905s latency (about 3.0× faster than sequential SWIFT and 5.9× than SafeSeal vLLM under reported measurements). These measurements identify watermark construction as the primary bottleneck. Finer-grained profiling of candidate generation, filtering, and post-processing would provide further runtime attribution.
3. Adaptive Scheduling and Dynamic Workloads (Weakness 3; Question 3). We extended fixed-ratio comparisons (0.1–0.9) beyond C4 to summarization, GSM8K, and MedMCQA. The best fixed ratio differs by task (0.7–0.9), while adaptive SWIFT achieves 1.05×–1.48× speedups without task-specific tuning. We also added a No-Controller baseline retaining the same model, watermark pipeline, asynchronous streams, and vLLM configuration, but removing watermark-aware admission. Although the latency comparison slightly favors No-Controller (0.899 s vs. 0.905 s), a separate user-level experiment finds SWIFT reduces mean E2E latency by 38.0%, P95 by 19.0%, P99 by 17.7%, and workload duration by 17.7% versus native scheduling. These findings support adaptation across heterogeneous workloads; explicit bursty traffic, arrival-rate transitions, and queue-delay profiling remain future work. We updated Sec. 4.2 and Appx. B.