{"id":"6d98b904-87c0-4e90-b755-e4afa30492b2","arxiv_id":"2412.19494","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A case study shows that retrieving related text and images from a knowledge base improves a diffusion-based semantic image transmission system.","lead":"This paper proposes adding retrieval-augmented generation (RAG) to AI-based semantic communication systems, where only meaning is transmitted rather than raw bits. A case study on image transmission suggests that retrieving related text and images before reconstruction improves fidelity and semantic consistency.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical claim that RAG improves GenSemCom is unsecured because the knowledge base is unspecified: if retrieved images include the ground-truth or near-duplicates, the reported gains are trivial, not evidence of RAG.","rationale":"The reader's weakest_assumption identifies exactly the point that decides whether the case study supports the headline claim. The paper's only quantitative evidence for RAG is the image-transmission experiment in Section IV. That experiment is not reproducible without knowing the knowledge base contents, and the reported numbers can be produced trivially if retrieval returns images of the transmitted scenes. The ablation in Fig. 5c shows the gain is mainly from image RAG; since the image-retrieval path feeds retrieved images directly into the IP-Adapter as visual conditioning, any overlap between the knowledge base and the test image invalidates the comparison. I agree with the reader's Conditional verdict; the decisive requirement is to specify the knowledge base, exclude test images and near-duplicates, and release retrieval logs. No further verdict adjustment is needed.","tokens_in":12708,"tokens_out":4072,"duration_ms":39318,"concrete_test":"Fix the knowledge base and make it public. Rerun the Kodak and West Lake evaluations under a leave-one-out protocol: for each test image, remove the image itself, all images from the same source dataset above a similarity threshold (e.g., CLIP similarity >= 0.9 to the transmitted image), and all documents that quote its ground-truth caption. Log the top-10 retrieved images with their CLIP similarity to the transmitted image and report the retrieval-quality statistics. Reproduce Fig. 5 under this protocol. If the image-RAG gain over the no-RAG baseline drops to noise when top-retrieved similarity is capped below the threshold, leakage explains the result; if the gain persists with clearly disjoint retrieved images, the leakage concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the performance gain in Figs. 5 and 6 comes from retrieving semantically relevant but non-identical knowledge, not from retrieving the answer itself. Section IV-B describes only that 'the advanced GPT-4o model' is used for retrieval and that evaluation uses Kodak and a West Lake image; it never states what the knowledge base contains, how many entries it has, whether test images or their near-duplicates were excluded, or what images/documents were retrieved for a given query. In the proposed pipeline, retrieved images are fed through the IP-Adapter into the SDXL UNet (Step 4), and retrieved documents refine the text prompt. If the knowledge base contains the original image, a same-scene photograph, or an image that already encodes the same content, the reconstruction can copy appearance directly, and any benefit is an artifact of leakage rather than of RAG-enabled semantic communication. The ablation in Fig. 5c shows the gain is mainly from image RAG, making this path the decisive one. The visualization in Fig. 6 does not show retrieved images, so the reader cannot check whether they overlap the source. Without knowledge-base specification and leakage control, the case study does not discriminate between 'RAG improves generation under noisy channels' and 'the retrieval system found the ground-truth image.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes integrating retrieval-augmented generation (RAG) into generative-AI-enabled semantic communication (GenSemCom). It first surveys existing GenSemCom systems and RAG techniques, then presents a conceptual architecture with components including a knowledge base, an intelligent retriever, and knowledge-aware encoders/decoders. The main technical contribution is a case study on image transmission: the transmitter sends a text description and an edge map, and the receiver retrieves related documents and images from a knowledge base to condition a Stable Diffusion XL generator with ControlNet and IP-Adapter. Experiments on the Kodak dataset and a single West Lake image report that the RAG-enabled system improves CLIP similarity, LPIPS, PIEAPP, and MS-SSIM relative to a baseline GenSemCom system from [14]. The paper concludes with future directions on retrieval efficiency, knowledge-base synchronization, and security.","tokens_in":13102,"tokens_out":3903,"duration_ms":35130,"significance":"The conceptual idea of using RAG to supply external knowledge at the receiver is a useful extension of generative semantic communications, and the paper provides a clear taxonomy of how RAG can be attached to VAEs, GANs, transformers, and diffusion models. The proposed architecture, with retrieved images fed through an adapter into the UNet and retrieved documents refining the text prompt, is plausible and well illustrated. However, the empirical support for the central claim is currently incomplete: the knowledge base is never specified, so the reported gains may be caused by retrieval of the ground-truth image rather than by genuine retrieval-augmented generation. The case study is also not reproducible because the retrieval mechanism is not described at the algorithmic level. If the experiments are redone with a leakage-controlled knowledge base and full implementation details, the paper could make a solid contribution; as it stands, the evidence for the title claim is not secured.","major_comments":[{"comment":"The knowledge base used in the case study is never specified. Section IV-B describes only the use of GPT-4o for retrieval and lists the Kodak dataset and a West Lake image as evaluation data; it does not state what the knowledge base contains, how many entries it has, whether the test images or their near-duplicates were excluded from retrieval, or what entries are retrieved for a given query. Since Step 4 of the pipeline in Fig. 4 feeds retrieved images through the IP-Adapter into the SDXL UNet, retrieved content can directly determine the reconstruction appearance. If the knowledge base contains the ground-truth image or a same-scene photograph, the gains in Fig. 5 and Fig. 6 are an artifact of answer leakage rather than evidence for RAG-enabled semantic communication. Please specify the knowledge base construction, report the retrieval results for the tested queries, and re-run the experiments with a knowledge base that provably excludes the transmitted images and their near-duplicates.","section":"Section IV-B, Fig. 4"},{"comment":"The retrieval mechanism is underspecified. GPT-4o is a multimodal LLM, not a retrieval system; the paper does not explain how textual queries are matched to images and documents, whether a vector database is used, which embedding model computes similarities, how many results are returned, or what prompt is given to GPT-4o for retrieval. Without these details the experiment is not reproducible and it is impossible to judge whether the retrieved images are semantically relevant but non-identical to the source, which is the load-bearing assumption behind the claimed improvement.","section":"Section IV-B"},{"comment":"The ablation study is performed on a single image (West Lake) and no error bars, repeated trials, or significance tests are reported. The differences in the table are small in several metrics (e.g., LPIPS 0.4342 vs 0.3985 vs 0.4075; PIEAPP 3.1909 vs 2.1111 vs 2.1841), and text-only RAG actually worsens PIEAPP relative to the baseline while improving MS-SSIM. The statement that 'the proposed system with text-only retrieval shows significant improvements' is not supported without variance information or a statistical test. Please report the ablation over the full Kodak dataset or provide confidence intervals.","section":"Section IV-C, Fig. 5c"}],"minor_comments":[{"comment":"The phrase 'overall overflow' should be 'overall overview'.","section":"Section III"},{"comment":"'Multi-model' in 'multi-model semantic information' and 'multi-model prompt capability' should be 'multi-modal'.","section":"Section IV-A"},{"comment":"In the text accompanying Fig. 4, 'putout' should be 'output' and 'serval iterations' should be 'several iterations'.","section":"Section IV-A, Step 4"},{"comment":"The transmission model is only described for the edge map (BER is injected into edge-map transmission); it is unclear whether the text description is also corrupted by channel errors or assumed error-free.","section":"Section IV-B"},{"comment":"Reference [7] lists the venue as 'Proc. EMMNLP'; the correct abbreviation appears to be 'EMNLP'.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The experimental validation is much thinner than the central claim requires. Before inviting a revision, the editor should verify that the authors can provide full knowledge-base specifications and leakage-controlled experiments; otherwise, the case study remains uninterpretable. The survey portion and the proposed architecture are reasonable, but the paper's contribution is framed around the numerical results, so the missing details are load-bearing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful part of this paper is the tutorial. It organizes GenSemCom systems by generator type and explains how RAG can be attached to VAEs, GANs, transformers, and diffusion models. The proposed architecture—retrieve text and images at the receiver, condition SDXL with ControlNet on the edge map and IP-Adapter on retrieved images—is a sensible combination, and I don't think this exact configuration is in prior work.\n\nThe case study is where the evidence goes soft. The authors evaluate on Kodak and one West Lake image, report that image RAG gives most of the gain, and never describe the knowledge base. The stress-test note is on target: if the retrieved images include the ground truth or near-duplicates, the comparison is trivial. The paper does not state that test images were excluded from retrieval, and the GPT-4o retrieval mechanism is a black box. The ablation runs on a single image and there are no error bars or multiple seeds. The comparison set is only [14] variants; the knowledge-base-enabled SemCom system in [13] is the most relevant baseline and is missing.\n\nI do not read this as deliberate mischief—the figures look like a working demo. But the central claim that RAG significantly enhances GenSemCom is not yet supported. The architecture may well be right; retrieving semantically relevant but non-identical references should help. The paper just has not ruled out leakage.\n\nCredit where due: the survey is a decent entry point, the system model is clearly drawn, and the future directions (knowledge-base synchronization, security, privacy) are sensible. For a reader new to GenSemCom, this is a useful map.\n\nMy recommendation: send it to peer review. The tutorial content alone justifies referee time, and the case study can be fixed. I would ask for a precise description of the knowledge base, a statement that test images were excluded, repeated runs with standard deviations, and at least one comparison against [13]. With those in place, the empirical claim becomes testable.","headline":"A useful RAG-for-SemCom tutorial whose case study gains are unsecured because the knowledge base is unspecified and test images may leak into retrieval.","tokens_in":13461,"tokens_out":2782,"would_cite":false,"duration_ms":27388,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that integrating retrieval-augmented generation (RAG) into generative semantic communication systems—having the receiver pull related text and images from a knowledge base before a diffusion model reconstructs the…","keywords":["retrieval-augmented generation","semantic communication","generative AI","diffusion models","image transmission","knowledge base","semantic consistency"],"falsifier":"Re-run the Kodak and West Lake evaluations with a knowledge base from which every image whose embedding is within a fixed distance of a test image, including the test image itself, has been removed; if the improvements over the non-RAG baseline disappear, the reported gain came from retrieving the answer rather than from semantic augmentation.","tokens_in":12529,"feed_emoji":"🖼️","tokens_out":4938,"duration_ms":49719,"temperature":0.7,"pith_summary":"The paper tries to establish that retrieval-augmented generation (RAG) can address three known weaknesses of GenAI-enabled semantic communication (GenSemCom): hallucinated or semantically inconsistent reconstructions under noise, poor adaptability to out-of-domain inputs, and the inability to reuse past transmissions. The proposed system adds a knowledge base and an intelligent retriever so that the receiver can enrich the received text prompt and obtain reference images before a diffusion model reconstructs the picture. In a case study using the Kodak dataset and a West Lake image, the authors report that RAG-enabled GenSemCom outperforms existing GenSemCom baselines on semantic consistency and perceptual quality across several bit error rates. If the result holds, an idea already successful in LLM applications—retrieve before generating—can make semantic communication more reliable without consuming extra bandwidth.","feed_headline":"Retrieval sharpens AI image delivery over noisy channels","feed_subtitle":"Pulling related images and text from a knowledge base before diffusion reconstruction beats current semantic-communication baselines.","key_machinery":"The load-bearing object is the RAG-enabled receiver pipeline: a knowledge base holding domain-specific, task-specific, environment-specific knowledge and communication history; an intelligent retriever that queries it using sparse and dense retrieval, refines results through an LLM review, and applies a stop-exploration strategy; and a diffusion-based semantic decoder that conditions each denoising step on the enriched text prompt, the received edge map, and retrieved reference images. The retrieved text is merged into the prompt through LLM-based prompt engineering, and the retrieved images are fed through image adapters, while the edge map provides structural guidance. This mechanism supplies missing context to the generative model, which the paper argues reduces hallucinations, helps out-of-domain images by providing similar references, and allows past transmission data stored in the knowledge base to be reused.","core_discovery":"The paper's central claim is that RAG can be inserted into GenSemCom as an external knowledge loop, and that doing so significantly improves reconstruction quality. The system transmits a compressed textual description and an edge map of the image; at the receiver, the textual description queries a knowledge base, relevant documents and images are retrieved, and a pre-trained diffusion model with structural conditioning reconstructs the image from the enhanced prompt, the edge map, and the retrieved images. Reported results show higher semantic consistency, measured by text-image embedding similarity, and better perceptual quality, measured by LPIPS, PIEAPP, and MS-SSIM, compared with GenSemCom systems that do not use retrieval. The ablation study indicates that the largest visual-quality gains come from image retrieval, while text retrieval alone still improves semantic consistency.","pith_inferences":["My inference: the same retrieval loop should transfer beyond images; any GenAI decoder that conditions on prompts can accept retrieved references, so RAG-enabled semantic communication is likely applicable to text, audio, and video transmission.","My inference: because the paper never specifies what the knowledge base contains or whether test images were excluded, part of the reported gain could come from retrieving near-duplicates of the ground-truth image rather than from genuinely useful context; a hold-out evaluation would settle this.","My inference: retrieval effort could be adapted to channel conditions—when the channel is clean, lighter retrieval may suffice, while under high noise the receiver should retrieve more context to compensate for lost semantic detail."],"forward_implications":["RAG-enabled GenSemCom yields higher semantic consistency than non-RAG GenSemCom at the same bit error rate, meaning reliability can improve without increasing bandwidth usage.","Image retrieval carries most of the visual-quality gain, so providing reference images to the diffusion decoder is the strongest lever for fidelity.","Text-only retrieval still improves semantic consistency, so even document-level context helps the receiver reconstruct meaning more accurately.","The West Lake results suggest RAG acts as a domain-adaptation shortcut, letting a static pretrained diffusion system handle images far from its training distribution.","Storing historical transmissions in the knowledge base makes the system accumulate knowledge over time, enabling future reconstructions to benefit from past sessions."],"supporting_citations":[{"why":"Supplies the baseline GenSemCom system and the transmission pipeline, including multimodal semantic encoding and BER simulation, which the RAG-enabled system is compared against.","marker":"[14]"},{"why":"Represents the existing GenSemCom decoder paradigm that treats wireless transmission as a degradation process, providing the prior approach the RAG design builds on.","marker":"[5]"},{"why":"Frames GenAI for semantic communication and motivates using textual descriptions as high-level semantic prompts for the receiver.","marker":"[4]"},{"why":"Motivates the knowledge-accumulation component by showing how stored encoding results and intermediate generative outputs can be reused.","marker":"[6]"},{"why":"Grounds the idea of an external knowledge base in semantic communication from a generative perspective.","marker":"[13]"},{"why":"Provides the typical retriever-prompt-generator process for RAG systems and discussion of retrieval accuracy and prompt quality.","marker":"[9]"},{"why":"Supplies the precedent for retrieval-augmented diffusion-based text-to-image generation, which the case study's GDM integration follows.","marker":"[10]"},{"why":"Demonstrates RAG-enhanced LLMs applied to wireless and satellite communication problems, motivating RAG use in the networking domain.","marker":"[12]"}],"fun_headline_variants":["RAG integration enhances semantic image delivery","Retrieval boosts diffusion-based semantic comms","Knowledge retrieval sharpens AI image reconstruction","RAG improves semantic consistency in image comms","Integrating RAG into GenSemCom for better reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The case study's comparison is meaningful only if the knowledge base contains images and documents that are relevant to the test images but does not include the ground-truth image or near-duplicates, and the paper never specifies the knowledge-base contents or confirms that test images were excluded.","fun_headline_variants_meta":{"raw":{"variants":["RAG integration enhances semantic image delivery","Retrieval boosts diffusion-based semantic comms","Knowledge retrieval sharpens AI image reconstruction","RAG improves semantic consistency in image comms","Integrating RAG into GenSemCom for better reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000567,"raw_usage":{"total_tokens":2657,"prompt_tokens":885,"completion_tokens":1772,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1703}},"tokens_in":501,"tokens_out":1772,"duration_ms":13606,"temperature":1.0,"reasoning_tokens":1703,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:16:08.859808+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Kodak and West Lake evaluations with a knowledge base from which every image whose embedding is within a fixed distance of a test image, including the test image itself, has been removed; if the improvements over the non-RAG baseline disappear, the reported gain came from retrieving the answer rather than from semantic augmentation.","supporting_citations":[{"cited_title":"Latency-aware generative semantic communications with pre-trained diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline GenSemCom system and the transmission pipeline, including multimodal semantic encoding and BER simulation, which the RAG-enabled system is compared against."},{"cited_title":"Generative joint source-channel coding for semantic image transmission,","cited_arxiv_id":null,"evidence_quote":"Represents the existing GenSemCom decoder paradigm that treats wireless transmission as a degradation process, providing the prior approach the RAG design builds on."},{"cited_title":"Generative AI for Semantic Communication: Architecture, Challenges, and Outlook","cited_arxiv_id":"2308.15483","evidence_quote":"Frames GenAI for semantic communication and motivates using textual descriptions as high-level semantic prompts for the receiver."},{"cited_title":"Evolving semantic communication with generative modelling,","cited_arxiv_id":null,"evidence_quote":"Motivates the knowledge-accumulation component by showing how stored encoding results and intermediate generative outputs can be reused."},{"cited_title":"Knowledge base enabled semantic communi- cation: A generative perspective,","cited_arxiv_id":null,"evidence_quote":"Grounds the idea of an external knowledge base in semantic communication from a generative perspective."},{"cited_title":"Generative AI agents with large language model for satellite networks via a mixture of experts transmission,","cited_arxiv_id":null,"evidence_quote":"Demonstrates RAG-enhanced LLMs applied to wireless and satellite communication problems, motivating RAG use in the networking domain."}],"review_version":1}