LLM-Driven Offloading Decisions for Edge Object Detection in Smart City Deployments
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors- GPT-4o-mini is queried K×I times (K candidates, I iterations). Costs, API latency, and token-rate limits are not reported; nor is the variance of generated rewards
- The DQN observes per-node RAM, CPU load, queue length, request size, but not fine-grained GPU memory, power cap, or wireless channel quality—factors that dominate real city deployments.
- Decisions are made slot-by-slot with no look-ahead; no comparison with Model Predictive Control or multi-step roll-outs.
- ILP is solved per time slot without future knowledge, so it is really a “genie” upper bound; comparing with modern heuristics (e.g., HEFT, GA, or A*-search) would be fairer. Some related works can be included for comaprison, such as: "Energy-Efficient Task Scheduling Based on Traffic Mapping in Heterogeneous Mobile-Edge Computing: A Green IoT Perspective," in IEEE Transactions on Green Communications and Networking and "Behavior Cloning With Fuzzy Logic for Accelerating Early-Stage DRL in Edge Offloading," in IEEE Transactions on Consumer Electronics.
- Fine-tune a 7 B-parameter code-model locally so that reward generation runs on-edge, eliminating cloud LLM dependency.
- Extend to multi-objective RL (Pareto) so the operator can slide along the latency-vs-energy frontier without re-prompting.
The presentation is fairly OK.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors claim that they introduce a new paradigm for enabling DRL schedulers in edge-based object detection to adapt seamlessly to evolving optimization objectives. Their framework directly incorporates the transformed optimization objectives into natural language prompts, enabling the LLM to automatically redesign the reward function. They evaluate the proposed LLM-guided DRL scheduling strategy against the RL-based approach, as well as two additional baselines: ILP and a greedy scheduling algorithm. The experimental results show that the proposed approach achieves the highest performance.
The readers would like to know the correctness rate of reward function generation, the reason of choosing a proprietary model alone, the latency of generating a correct reward function from the model, and the reason of choosing the specific datasets for evaluation.
The authors mentioned "The emergence of ChatGPT has sparked a surge of research into LLMs, which are now being applied across diverse domains due to their remarkable generalization, contextual reasoning, and code generation capabilities." Can ChatGPT generate reward function without any errors with the engineered prompts ? There are many open-source LLMs nowadays. Can open-source LLMs generate reward functions successfully ? It will be interesting to know whether the LLM-based approach is effective using open-source LLMs. Are the datasets used in the manuscript representative of other workloads in other big and smart cities ? What exactly is the energy constraint in the experiment ?
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis article presents a very specific and original use of large language models. The application is innovative and highlights a unique perspective on how LLMs can be applied beyond their conventional domains. The work places strong emphasis on algorithm description and rigorous mathematical formalization, which demonstrates the technical depth and precision of the authors’ approach.
However, the paper could be made more accessible to a broader research audience. The introduction is dense and provides limited context for readers who may not have a strong background in the field. Additional illustrations, explanatory figures, or intuitive examples would greatly improve clarity, especially for researchers outside the immediate area of expertise.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for Authors- As the author pointed out in the paper, the potential volatility of LLM outputs appears to be central to system performance, yet this aspect seems to have been significantly underestimated. The computational overhead associated with reward verification also appears likely to constitute a major component of system performance, yet this aspect too seems to have been overlooked. Consideration of these aspects will be necessary in the experiments.
- Chapter 4's experimental scenarios are overly simplistic. To adequately demonstrate the paper's excellence, more diverse experiments should be included. Crucially, each point of excellence asserted in the paper should be demonstrated experimentally; however, these aspects are either missing or difficult to verify properly. The difference between Experiments 2-4 and 5-6 lies solely in the latency difference, yet it is challenging to discern the significance of this latency variation.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsNo more comments.
Author Response
We sincerely thank Reviewer 1 for the time and effort dedicated to reviewing our manuscript. We appreciate the confirmation that no further comments or concerns remain.
Reviewer 2 Report
Comments and Suggestions for AuthorsThank you for updating the manuscript. It will be helpful for readers to understand why the Llama3 8B model requires more tokens, longer latency, and has less consistent refinement quality.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe revised paper appears to have addressed all the review comments raised. Thank you for your hard work.
Author Response
We sincerely thank Reviewer 4 for the time and effort dedicated to reviewing our manuscript. We sincerely appreciate your recognition of our revisions.
