{"id":"14dc41e5-27ff-4f5d-b777-1bd6e3943b9b","arxiv_id":"2509.04801","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"KGRAG-SC transmits only knowledge-graph entity IDs and reconstructs text with an LLM, improving low-SNR semantic fidelity and cutting transmitted bits versus text-based baselines.","lead":"This paper combines knowledge graphs, retrieval-augmented generation, and large language models into a semantic communication system that sends only short entity IDs instead of full text. A smart generalist might read it to see how structured knowledge can make wireless communication more robust under noise and cut transmission costs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The low-SNR advantage is confounded by an unequal coding baseline: the 'traditional' scheme uses no channel coding while KGRAG-SC uses rate-1/2 convolutional codes, so the gain may reflect FEC rather than KG-based semantics.","rationale":"The reader's verdict is conditional and my concern is consistent with a conditional outcome rather than rejection: the underlying architecture may still be sound, but the reported experiments do not isolate the claimed mechanism. I agree with the reader's note that a channel-coding baseline is missing; however, the reader's identified weakest assumption (KG coverage) is about out-of-distribution generalization. My concern targets the in-distribution headline result itself: even within WebNLG, the 0.780 vs 0.285 comparison is not interpretable as a semantic-coding gain because the two systems use different channel coding. This is a more direct threat to the central claim. The KG coverage issue is real but secondary; if the FEC confound is resolved and the semantic advantage persists, the scheme would still need out-of-distribution tests to support deployment claims. I am not accusing the authors of inconsistency; the omission is in the experimental design, not in the argument structure. The paper does provide a clear system description and a reproducible-looking pipeline, but the missing baseline makes the headline quantitative claim currently underdetermined.","tokens_in":8994,"tokens_out":4611,"duration_ms":41148,"concrete_test":"Re-run the Section IV-B experiment with a baseline that applies the same rate-1/2 convolutional code (constraint length 7) to Huffman-coded text before 16QAM, and also run KGRAG-SC without any channel coding. Compare semantic similarity versus SNR and count total transmitted bits including coding overhead. If the coded text baseline closes most of the 4-dB gap, or if uncoded KGRAG-SC is no better than the uncoded text baseline, then the reported low-SNR advantage is explained by FEC rather than by KG-RAG semantic processing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline quantitative evidence is in Fig. 2: at 4 dB, KGRAG-SC achieves 0.780 semantic similarity versus 0.285 for the baseline. Section IV-A specifies that the baseline transmits Huffman-coded text over 16QAM 'without any channel coding,' whereas KGRAG-SC transmits node IDs where high-importance nodes are protected by a rate-1/2 convolutional code (constraint length 7). The comparison therefore varies two things at once: the semantic representation and the channel code. A conventional FEC-coded text transmission would also be expected to improve dramatically at low SNR, so the reported gap is not evidence that the KG-RAG semantic pipeline is responsible for the robustness. This is load-bearing because the central claim is 'superior semantic fidelity in low SNR conditions' specifically attributed to the scheme's semantic design. Similarly, the transmission-efficiency claim (Figs. 3 and 4) omits the overhead of the convolutional code; if the reported bits are only the payload node IDs, KGRAG-SC's true spectral cost is understated. The paper provides no matched baseline that isolates the contribution of semantic compression and LLM reconstruction from the contribution of channel coding, so the main empirical support for the central claim is currently confounded.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes KGRAG-SC, a knowledge-graph-assisted semantic communication framework. The transmitter performs GraphRAG-style entity extraction over a shared KG constructed offline from WebNLG, builds a one-hop subgraph around selected entities, and transmits only the corresponding node IDs, applying importance-aware unequal error protection based on centrality metrics. The receiver recovers the subgraph from the KG and uses an LLM with a structured prompt to reconstruct the original text. Experiments report semantic-similarity gains at low SNR relative to a Huffman-coded text baseline with no channel coding, and reduced transmission bits per sentence versus ASCII and Huffman text.","tokens_in":9297,"tokens_out":3599,"duration_ms":34006,"significance":"The paper combines knowledge graphs, GraphRAG-style retrieval, and LLM-based generation for semantic communication, which is a timely and interesting direction. The proposed pipeline is transparent and interpretable compared with end-to-end deep SC systems. If the reported gains are accurately attributed, the scheme would be a meaningful step toward structure-aware semantic compression with robust low-SNR behavior. The authors also honestly acknowledge the high-SNR crossover where the traditional baseline outperforms KGRAG-SC. However, the current empirical evaluation does not isolate the contribution of the semantic design from that of channel coding, and the in-distribution evaluation limits the generality of the claims.","major_comments":[{"comment":"The low-SNR comparison is confounded: the baseline transmits Huffman-coded text with 16QAM and no channel coding, while KGRAG-SC protects high-importance nodes with a rate-1/2 convolutional code. This simultaneously varies the semantic representation and the channel code, so the reported 0.780 vs. 0.285 gap at 4 dB cannot be attributed to the KG-RAG pipeline. A matched text baseline with the same convolutional code, or a KG-based variant without UEP, is needed to isolate the semantic contribution.","section":"Section IV-A, Fig. 2"},{"comment":"The transmission-overhead comparison appears to exclude the overhead of the rate-1/2 convolutional code used for high-importance nodes in KGRAG-SC. If the reported bits are payload bits before channel coding, the true spectral cost of KGRAG-SC is understated. The paper should report end-to-end transmitted bits after channel coding (including rate overhead) for all schemes.","section":"Section IV-C, Figs. 3 and 4"},{"comment":"The shared KG is constructed from the same WebNLG dataset used for evaluation, so the entity linking and reconstruction operate strictly in-distribution. For any sentence whose entities or relations are absent from the KG, Algorithm 1 cannot link them, the MCSG may be empty or incomplete, and the LLM has no grounding for reconstruction. The paper provides no out-of-distribution evaluation or coverage analysis, making the central coverage assumption untested.","section":"Section III-B1 and Section IV-A"},{"comment":"Several load-bearing parameters are unspecified: the importance weight alpha in Eq. (7), the SNR-dependent importance thresholds, the Top-3 expansion in Eq. (3), and the one-hop radius of the MCSG. No sensitivity analysis, error bars, or sample sizes are provided. This makes the reported quantitative results non-reproducible and does not establish that the claimed robustness is statistically significant.","section":"Section III-C and Section IV-A"}],"minor_comments":[{"comment":"The object defined by Eq. (5) is called a \"Minimum Connected Subgraph,\" but the manuscript explicitly states that it \"may consist of multiple disconnected components.\" The terminology should be corrected or the definition should be amended to require connectivity.","section":"Section III-B3, Eq. (5)"},{"comment":"No error bars or number of test sentences are reported for the semantic-similarity results. Please include the test-set size and repeated-run variability.","section":"Section IV-B, Fig. 2"},{"comment":"The exact prompts used for LLM-based entity selection and text reconstruction are not given, and the generation of entity descriptions and community summaries is only described qualitatively. Including the prompts would substantially improve reproducibility.","section":"Section IV-A"},{"comment":"The caption and text should explicitly state whether the bit counts in Figs. 3 and 4 include the convolutional-code overhead, since this directly affects the efficiency comparison.","section":"Section IV-C, Fig. 3"},{"comment":"The sentence \"The current schemes relies on a static, pre-shared knowledge graph\" contains a subject-verb agreement error; it should read \"The current scheme relies.\"","section":"Section V"},{"comment":"The function ConstructPrompt is not defined; the manuscript should describe how the sentence and candidate entities are formatted into the LLM prompt.","section":"Algorithm 1, line 9"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the proposed architecture is appealing, but the eval-uation confounds the semantic-pipeline contribution with the channel-coding contribution. The in-distribution KG construction is also a significant generality concern. These issues are fixable in a revision, so I recommend major revision rather than rejection. The authors should also clarify whether the bit-count comparisons include coding overhead."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely new pipeline — GraphRAG-style entity linking, one-hop MCSG ID transmission, and LLM-based reconstruction — applied to KG-aided semantic communication. The authors build it cleanly and the compression story is coherent. But the headline low-SNR advantage (0.78 vs 0.28 at 4 dB) is not a clean comparison: KGRAG-SC gives high-importance nodes a rate-1/2 convolutional code while the \"traditional\" baseline sends Huffman-coded text with no channel coding at all. Any reasonable FEC on the baseline would close much of that gap. The authors also report transmission bits without counting the convolutional coding overhead, so the efficiency claim is overstated. The KG is built from WebNLG and tested on WebNLG, making the entity linking and reconstruction in-distribution; there's no out-of-distribution evaluation, no error bars or sample sizes, and alpha and the importance thresholds are left unspecified. These are fixable, but they currently leave the central claim under-supported.\n\nWhat's genuinely good: the idea of transmitting node IDs rather than triples is a nice step beyond prior KG-aided SC, and the use of community summaries to guide entity linking is a sensible adaptation of GraphRAG. The high-SNR crossover (baseline 0.997 vs KGRAG 0.882 at 12 dB) is honest about the lossy compression ceiling. The paper is clearly written and the system is described in enough detail to rebuild.\n\nMy overall read: the novelty is real but incremental; the evaluation is not yet convincing. A serious referee should see this, but the revision needs matched baselines (e.g., same channel coding for both), overhead-inclusive bit counts, and an OOD or held-out KG test. I would not cite it as evidence for anything yet, but I'd keep it in mind as related work.","headline":"A promising but under-validated GraphRAG+KG semantic communication scheme whose low-SNR win is confounded by an unequal channel-coding baseline.","tokens_in":9780,"tokens_out":2269,"would_cite":false,"duration_ms":20168,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"KGRAG-SC proposes that transmitting compact knowledge-graph node IDs instead of raw text preserves semantic fidelity at low SNR while sharply cutting transmission overhead.","keywords":["semantic communication","knowledge graph","retrieval-augmented generation","minimum connected subgraph","unequal error protection","LLM text reconstruction","low-SNR communication","compression"],"falsifier":"Take a set of sentences whose entities are absent from the knowledge graph, run the transmitter pipeline, and measure the fraction that produce a non-empty MCSG and the resulting semantic similarity; if most produce empty payloads or similarity near zero, the coverage assumption fails.","tokens_in":8824,"feed_emoji":"📡","tokens_out":4437,"duration_ms":36850,"temperature":0.7,"pith_summary":"KGRAG-SC proposes that a transmitter can compress a sentence into the node IDs of a small subgraph taken from a shared knowledge graph, and a receiver with a large language model can turn those IDs back into fluent, meaning-preserving text. The paper argues that this beats transmitting the text itself in two ways: it needs far fewer bits per sentence, and it keeps semantic similarity high when the channel is noisy. At 4 dB signal-to-noise ratio the scheme reaches a semantic similarity of 0.780 against 0.285 for a traditional text-with-compression baseline, while the advantage reverses above about 8 dB. The motivating claim is that meaning, not bits, should be the unit of transmission.","feed_headline":"Sending graph node IDs beats sending text at low SNR","feed_subtitle":"At 4 dB, semantic similarity reaches 0.78 versus 0.29 for text, with far fewer bits per sentence.","key_machinery":"The Minimum Connected Subgraph (MCSG) is the central object: the set of selected entities plus their one-hop neighbors in the shared knowledge graph, transmitted as an ordered node-ID sequence. Its compression power comes from the fact that both sides share the graph, so structure and entity descriptions never need to be sent. The importance-aware unequal error protection, weighting normalized degree and betweenness centrality, is what keeps core nodes intact when bits flip, and the large-language-model reconstruction step converts the surviving subgraph into text.","core_discovery":"Under the proposed scheme, the unit of transmission is a list of knowledge-graph entity IDs forming a Minimum Connected Subgraph (MCSG), not the source text and not semantic triples. The transmitter extracts entities from the sentence, expands each to a one-hop neighborhood in a shared knowledge graph, and sends only the unique node IDs. An importance score based on degree and betweenness centrality decides which IDs get convolutional error protection. The receiver looks up the IDs in its local graph copy, discards invalid IDs, and prompts a large language model with the surviving subgraph and entity descriptions to regenerate the sentence. The experimental claim is that this pipeline preserves semantics under low SNR while using markedly fewer bits per sentence than raw text transmission.","pith_inferences":["The paper does not test sentences whose entities are absent from the shared knowledge graph, so the scheme's real-world coverage is an open question; extending it to open-domain text would require a much larger or dynamically updated graph.","The importance threshold and the balance between degree and betweenness centrality are chosen heuristically; an end-to-end trained version could adapt them per sentence and channel state.","The approach points toward a notion of semantic entropy measured by graph coverage rather than character entropy, which could connect to rate-distortion theory for meaning."],"forward_implications":["At low SNR, the scheme preserves meaning where bit-oriented text transmission collapses; the 4 dB gap (0.780 versus 0.285 semantic similarity) is the headline result.","Transmission cost tracks semantic complexity rather than sentence length, so verbose sentences cost about the same as terse ones and cumulative bandwidth savings grow over long sessions.","The performance crossover above about 8 dB means the scheme is suited to noisy, bandwidth-constrained links, not clean channels where raw text is already near-perfect.","Because corrupted IDs are dropped and the large language model fills gaps from graph context, partial information loss degrades gracefully instead of producing garbled text."],"supporting_citations":[{"why":"Establishes the paradigm of knowledge-graph-enabled semantic communication with unequal error protection, which this scheme extends.","marker":"[13]"},{"why":"Supplies the retrieval-augmented generation concept that motivates retrieving compact context instead of transmitting raw text.","marker":"[10]"},{"why":"Introduces the graph-based retrieval approach that the MCSG construction adapts.","marker":"[11]"},{"why":"Provides the sentence-embedding model used for both retrieval and the semantic similarity evaluation metric.","marker":"[16]"},{"why":"Supplies the fast similarity-search index used for community-guided entity linking.","marker":"[18]"},{"why":"Defines the betweenness centrality metric used in the importance-aware error protection scheme.","marker":"[19]"},{"why":"Supplies the dataset from which the shared knowledge graph is built and on which the experiments are evaluated.","marker":"[20]"},{"why":"Provides the large language model used at the receiver for knowledge-driven text reconstruction.","marker":"[21]"}],"fun_headline_variants":["Knowledge graph IDs survive low SNR better than text","Graph entity IDs beat raw text in noisy channels","Semantic comms via graph node IDs: less bits, more meaning","KGRAG-SC: sending graph node IDs, not text, at low SNR","GraphRAG-inspired SC: transmit entity IDs, not sentences"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire scheme assumes the shared knowledge graph already contains the entities and relations expressed in the source sentence; the evaluation uses sentences drawn from the same dataset the graph was built from, so coverage of unseen sentences is never tested.","fun_headline_variants_meta":{"raw":{"variants":["Knowledge graph IDs survive low SNR better than text","Graph entity IDs beat raw text in noisy channels","Semantic comms via graph node IDs: less bits, more meaning","KGRAG-SC: sending graph node IDs, not text, at low SNR","GraphRAG-inspired SC: transmit entity IDs, not sentences"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":2072,"prompt_tokens":891,"completion_tokens":1181,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":1094}},"tokens_in":507,"tokens_out":1181,"duration_ms":7844,"temperature":1.0,"reasoning_tokens":1094,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:26:15.040500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of sentences whose entities are absent from the knowledge graph, run the transmitter pipeline, and measure the fraction that produce a non-empty MCSG and the resulting semantic similarity; if most produce empty payloads or similarity near zero, the coverage assumption fails.","supporting_citations":[{"cited_title":"Reliable semantic communication system enabled by knowledge graph,","cited_arxiv_id":null,"evidence_quote":"Establishes the paradigm of knowledge-graph-enabled semantic communication with unequal error protection, which this scheme extends."},{"cited_title":"Betweenness centrality in large complex networks,","cited_arxiv_id":null,"evidence_quote":"Defines the betweenness centrality metric used in the importance-aware error protection scheme."},{"cited_title":"Creat- ing training corpora for nlg micro-planning,","cited_arxiv_id":null,"evidence_quote":"Supplies the dataset from which the shared knowledge graph is built and on which the experiments are evaluated."}],"review_version":2}