{"id":"e5d90078-4241-4cb2-8335-4374fdef8d65","arxiv_id":"2607.18640","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using an LLM to correct misdecoded text and reuse those corrections as virtual pilots improves MIMO channel estimation, approaching oracle accuracy.","lead":"This paper uses a large language model to correct text errors in a wireless signal, then feeds the corrected words back as extra reference signals to improve channel estimation in MIMO systems. The idea is that fixing a wrong symbol helps more than reusing a right one, which simulations confirm.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof assumes the noise on a corrected pilot is independent of the selection outcome, but Layer 2 selects corrections using that same received signal; the claimed strict expected NMSE reduction may not follow.","rationale":"The reader's weakest_assumption identifies the independence of n from the detection outcome, which I agree is the single most load-bearing flaw in the theoretical core. The proof of Theorem 1 is the advertised contribution, and it is materially weakened by conditioning on selection events that depend on n. The paper is otherwise a reasonable engineering proposal with consistent simulations, though without error bars and with a tuned threshold. Since the reader already recommended CONDITIONAL, my stress-test does not move the verdict. I do not elevate to REJECT because the empirical results may well be correct despite the flawed proof; the concern is about the theoretical guarantee, not about the plausibility of the framework.","tokens_in":18194,"tokens_out":4158,"duration_ms":48434,"concrete_test":"Monte Carlo at SNR=5 dB in the 2x2 QPSK setup: for every Layer-2 accepted corrected symbol, record the true transmitted symbol x, the received y, and the pilot-only error E; compute the actual per-symbol NMSE reduction Delta_emp = ||E||_F^2 - ||H_new - H||_F^2 when that symbol (with its true x) is appended as an additional pilot, and average over accepted symbols. Compare to the Lemma 1 prediction a*||E x||^2 - b using the same E. Compute the same average for MAP-correctly-decoded symbols selected by Layer 1. If the empirical ordering E[Delta_C] > E[Delta_D] fails, or if the Lemma 1 formula is systematically biased for Layer-2-selected symbols, the independence assumption is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim, Theorem 1, depends on Lemma 1, whose proof in Appendix A states: 'We assume that n is independent of the detection outcome that determines the reference type of x.' In the actual pipeline, a Layer-2 corrected symbol is selected precisely because its received signal y[n] caused a MAP misdetection and the LLM correction passes the d_ratio cross-validation (22). Both events are functions of the same n that appears in the pilot update (36)-(37). Therefore, for actually selected corrected pilots, E[n | selected] is not zero, and the term nx^H/D in (37) does not vanish in expectation. The Lemma 1 formula (43), which subtracts aE[R(x;E)|x] - b with b computed under the nominal zero-mean noise, is not the true expected NMSE reduction for selected Layer-2 pilots.\n\nA second, related gap is that Theorem 1 assumes both x_D and x_C equal the true transmitted symbol, while the Layer-2 accepted set has precision around 70-79% at 5 dB (Table IV, Fig. 10). A large fraction of accepted corrections are wrong, and the theorem does not analyze the effect of those erroneous pilots. The nearest-neighbor approximation in Appendix B, Step 2, also reduces the K=16 MAP decision to a single pairwise comparison without justification.\n\nThe empirical results may still support the engineering claim, but the advertised proof of a strict advantage for corrected symbols is not established for the actual selection mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a semantic-aware data-aided channel estimation framework for MIMO systems. An LLM is fine-tuned to correct typographical errors in the decoded text; a Needleman–Wunsch alignment module maps corrected characters to symbol slots; and a two-layer selection mechanism admits verified and corrected symbols as additional pilots for LMMSE channel refinement. The central theoretical claim (Theorem 1) is that a corrected symbol—one initially misdecoded but then corrected—yields a strictly larger expected reduction in channel estimation error than an initially correctly decoded symbol. Simulations on a 2x2 QPSK system compare the proposed method against pilot-only, reinforcement-learning-based data-aided, LLM-all-corrected, Layer-1-only, MAP-oracle, and perfect-CSI baselines in terms of NMSE, BER, and pilot-selection quality.","tokens_in":18602,"tokens_out":4984,"duration_ms":53256,"significance":"The idea of using LLM-based semantic correction as a source of side information for channel estimation is novel and timely. If the theoretical and empirical claims hold, the paper would demonstrate a genuinely new mechanism for data-aided estimation that goes beyond physical-layer statistics. The manuscript is clearly structured, describes the full pipeline in detail, and includes appropriate oracle baselines (MAP Oracle, Perfect CSI) that help quantify the achievable gain. The authors also make a serious attempt at a formal proof of the proposed advantage, which is valuable even if incomplete. The main weaknesses are the selection-bias gap in the proof, the idealization that corrected symbols equal the true transmitted symbols, and the absence of statistical confidence measures in the simulations.","major_comments":[{"comment":"The proof explicitly assumes that n is independent of the detection outcome that determines the reference type of x. This is not true in the actual pipeline: a Layer-2 corrected pilot is accepted only if (a) the MAP detector made an error on y[n]=Hx+n (so n contributed to the error event) and (b) the LLM-corrected vector passes the d_ratio test (22), which is computed from the same y[n]. Both events depend on n, so E[n | x_C selected] ≠ 0 and E[n n^H | selected] ≠ σ² I. Consequently the cancellation of the nx^H/D term in (37) and the noise power term in (38) are not valid for actually selected symbols, and (43) is not the true expected reduction. Please re-derive the conditional moments under the selection rule, or clearly separate an idealized theorem (reference provided independently of n) from the engineered selection mechanism.","section":"Appendix A, Lemma 1 (Eqs. (37)–(43))"},{"comment":"The theorem assumes both reference types equal the true transmitted symbol. However, Table IV and Fig. 10 report a Layer-2 precision of only 70.7% at 5 dB on the tuning set (78.7% on the test set), so a substantial fraction of accepted corrections are wrong. The proof does not analyze the contribution of erroneous pilots to the refined estimate; the strict inequality (33) therefore does not follow for the actual X_C. Please extend the analysis to a mixed model that includes incorrect corrections (e.g., parametrized by precision), or explicitly present Theorem 1 as an ideal-case result and explain why the empirical gain persists despite imperfect precision.","section":"Section IV-A and Theorem 1"},{"comment":"The proof reduces the K=16 MAP decision to a single pairwise comparison with the minimum-margin competitor, calling it a standard nearest-neighbor approximation, but gives no justification or validation for this reduction in the considered QPSK 2x2 setting. The strict monotonicity p'(r)<0 in (55) is derived under this approximation. If the approximation is not accurate, Proposition 1 is not established. Please justify the approximation (e.g., via a union bound over the 15 competitors) or provide numerical verification that it is tight in the SNR range used.","section":"Appendix B, Step 2 (Eq. (51))"},{"comment":"All simulation results are point estimates; the number of Monte Carlo runs is not reported, and no confidence intervals or error bars are given. Many reported differences are small (e.g., BER at 5 dB: Proposed .0505 vs. MAP Oracle .0504; at 8 dB .0190 vs .0190), so it is difficult to assess whether the claimed advantages are statistically significant. Please report the number of trials and standard errors or confidence intervals for the key NMSE/BER comparisons, especially where the proposed method is said to outperform the MAP Oracle.","section":"Section VI (Figs. 7–11, Table III)"}],"minor_comments":[{"comment":"The threshold τ_d is selected on a tuning set at 5 dB and then applied across all SNRs. Please quantify the sensitivity of the results to τ_d across the full SNR range, since Fig. 10 is only at 5 dB.","section":"Section VI-B, Table IV"},{"comment":"The notation \\bar t(i)_LLM is used in the text before the alignment output is formally defined in (9); consider reordering the definitions to avoid confusion.","section":"Section II-C"},{"comment":"The alignment example shows 'awound' versus 'abou-t', but the traceback path in the score matrix is not shown; a small arrow or highlighted path would clarify the gap placement.","section":"Fig. 4"},{"comment":"The legend 'Layer 1Layer 2' appears without colors in the caption; please add a legend with clearly labeled colors or patterns.","section":"Fig. 9"},{"comment":"The statement 'closely approaching the performance of an oracle estimator' should be qualified by the SNR range and by the fact that at some SNRs the BER is identical to MAP Oracle (e.g., 8 dB and 10 dB in Table III).","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a topic of potential interest to the signal processing community. The main risk is the theoretical overclaim: the proof of Theorem 1 relies on an independence assumption that is violated by the actual selection mechanism, and on the idealization that all accepted corrections are correct. These issues are fixable by either re-deriving conditional bounds or downgrading the theorem to an idealized statement and supporting the empirical gain with proper statistical evidence. If the authors address these points, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this paper on using an LLM to correct decoded text and then feed corrected symbols back as pilots. The engineering story is better than the theory. The system—LLM correction, Needleman-Wunsch alignment, two-layer selection—is new, and the simulations are consistent with the main claim. But the proof of Theorem 1, the strict expected NMSE reduction for corrected symbols, has a real gap.\n\nFirst, what is genuinely good. Bringing text semantics into pilot selection is a fresh angle. The baseline comparisons are fair: pilot-only, RL-based selection, oracle bounds. The ablation shows the fine-tuned LLM is necessary, Layer 2 contributes most of the gains, and using all corrections without filtering hurts. That part is convincing, and the details are enough to reproduce in principle, though no code or data is released.\n\nWhere it goes soft: Appendix A, Lemma 1, assumes the noise n is independent of the detection outcome that decides whether a symbol becomes a corrected pilot. In the actual pipeline, Layer 2 accepts a correction precisely when the received signal—including that same n—makes the d_ratio small. So for accepted corrected pilots, the conditional noise distribution is not the nominal CN(0, sigma^2 I). The proof's equality doesn't hold for the pilots the system actually uses. A second issue: Theorem 1 assumes both x_D and x_C equal the true transmitted symbol. But Layer 2's precision is 70–79% at 5 dB; a large share of accepted corrections are wrong. The theorem does not analyze that case, and it isn't obvious the negative effect is small. The nearest-neighbor reduction of a 16-vector MAP decision to a single pairwise comparison is also asserted without justification.\n\nEmpirically, the gains over the best practical baseline are real but modest—BER differences around 0.001 to 0.003 at most SNR points. The figures have no error bars or run counts, and tau_d is tuned on a validation set rather than fixed a priori. That is a standard weakness, not a fatal one.\n\nOverall: the central engineering claim likely survives, but the advertised theoretical guarantee does not. This paper should go to serious peer review, not be desk-rejected. The authors should either fix the proof by modeling selection-dependent noise properly, or downgrade Theorem 1 to a heuristic justification. They should also release code and error bars. Worth reading for anyone in semantic communication or data-aided estimation, and the proof gap is instructive.","headline":"Clever and plausible system; the proof of the headline theorem has a hole because selection uses the same noise the proof conditions on.","tokens_in":19029,"tokens_out":2268,"would_cite":true,"duration_ms":21541,"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 claims that LLM-corrected symbols, after physical-layer validation, reduce MIMO channel-estimation error more per symbol than already-correctly-decoded symbols, and that a two-layer semantic pilot built this way approaches oracle","keywords":["semantic communication","data-aided channel estimation","MIMO","large language model","semantic pilot","Needleman-Wunsch alignment","physical-layer cross-validation","channel estimation"],"falsifier":"Record the received signals of symbols admitted by Layer 2 and estimate the conditional distribution of the noise vector given 'corrected' status; if its mean or variance deviates materially from CN(0, σ²I), the Appendix A independence assumption fails and the per-symbol Δ(x_C) bound in Lemma 1 does not apply to the actually selected pilots. Alternatively, compare per-symbol NMSE reduction of corrected vs verified symbols on the test set; if corrected symbols do not show a strictly larger average reduction, Theorem 1's conclusion is not reproduced.","tokens_in":18084,"feed_emoji":"📡","tokens_out":7127,"duration_ms":62706,"temperature":0.7,"pith_summary":"The paper tries to establish that an LLM's ability to repair corrupted text can be turned into a channel-estimation advantage in MIMO wireless links. The idea is to treat the decoded payload as text, let a fine-tuned LLM fix the typos caused by channel noise, align the corrected text to the decoded text, and use the symbols the LLM fixed—after a physical-layer check—as extra pilots. The central formal claim is Theorem 1: a symbol that was initially misdetected and then corrected contributes a strictly larger expected reduction in estimation error than a symbol the detector got right in the first place. If true, this shifts data-aided estimation from selecting reliable detections to actively recovering unreliable ones, and simulations show the resulting estimate approaching an oracle that knows the true channel.","feed_headline":"LLM-corrected symbols shrink channel error more than clean ones","feed_subtitle":"A two-layer semantic pilot uses LLM text correction plus physical-layer checks, approaching oracle-level MIMO estimation.","key_machinery":"The load-bearing object is the realized error alignment R(x;E)=||Ex||², the squared channel-estimation error at the pilot-only estimate along the direction of a candidate reference symbol x. Lemma 1 shows the NMSE reduction from adding x as a pilot is a strictly increasing affine function of E[R(x;E)|x]; Proposition 1 shows this expected alignment is larger for corrected than for correctly decoded symbols because large alignment is what caused the initial misdetection; Theorem 1 combines them. The pipeline machinery is the two-layer gate: Layer 1 semantic verification (decoded and corrected characters agree) plus Layer 2 physical-layer cross-validation via the residual ratio d_ratio = ||y −","core_discovery":"The paper's central claim is that payload semantics can be mined to produce better channel estimates than physical-layer statistics alone. At the receiver, the initially decoded text is fed to a fine-tuned LLM that corrects typographical errors; a Needleman-Wunsch alignment maps the LLM's token-level output back to characters. Layer 1 marks positions where decoded and corrected text agree as reliable; Layer 2 takes positions where the LLM changed a character and admits the corrected symbol only if a residual-ratio test on the received signal confirms it. The refined LMMSE estimator then uses these semantic pilots. The central theoretical result, Theorem 1, proves a corrected reference yields","pith_inferences":["The selection-bias mechanism behind Theorem 1 is not specific to LLMs: any reference acquisition rule that conditions on large estimation error should yield oversized per-symbol gains, so other side-information sources (structured code, checksums, generative priors) could be plugged into the same two-layer gate.","Because Layer 2's d_ratio test reuses the same received signal that caused the misdetection and an intermediate estimate built from Layer 1 symbols, the validation noise is not independent of the correction event; the paper's Appendix A independence assumption is therefore likely violated in the actual pipeline, and the theoretical guarantee should be re-derived under the true conditional distribu","The framework's advantage concentrates at low SNR (Layer 2 contributes up to 18.2% of pilots at 1 dB and shrinks at high SNR); a testable extension is SNR-adaptive τ_d scheduling, since the paper fixes τ_d=2 across all SNR.","One could test the same architecture with a smaller LLM or a masked language model to see how much of the gain comes from correction accuracy versus the selection gate; the paper's ablation with an unfine-tuned model suggests correction accuracy is the bottleneck."],"forward_implications":["Data-aided channel estimation need not be limited to picking reliable detections; actively recovering misdetected symbols via semantic correction gives a strictly larger per-symbol error reduction (Theorem 1).","The proposed semantic pilot, combining original pilots with Layer 1 verified and Layer 2 validated corrections, outperforms conventional data-aided schemes in NMSE and BER and approaches the MAP-oracle at high SNR.","A physical-layer validation gate is essential: using all LLM corrections without filtering performs worse than pilot-only estimation, confirming that LLM hallucinations must be screened.","The Needleman-Wunsch alignment is necessary for the two-layer selection to work when LLM tokenization changes text length; removing it degrades performance.","Fine-tuning the LLM on channel-corrupted text is what makes Layer 2 viable; without fine-tuning, corrections largely fail the d_ratio validation and the advantage disappears."],"fun_headline_variants":["LLM-corrected symbols yield larger channel error cut than clean ones","Semantic LLM pilots: text correction improves MIMO channel estimate","Theorem: LLM-corrected symbols beat clean ones for channel estimation","Two-layer LLM pilot selection approaches oracle MIMO estimation","LLM semantic checks pick reliable pilots, improving MIMO estimation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the noise on a symbol used as an additional pilot is independent of the detection outcome that decides whether it counts as 'corrected'—a premise stated in Appendix A that the actual pipeline likely violates, since a symbol becomes corrected precisely because its noise caused a misdetection and Layer 2 reuses that same received signal; the proof additionally relies on a standard nearest-neighbor approximation that collapses the 16-candidate MA","fun_headline_variants_meta":{"raw":{"variants":["LLM-corrected symbols yield larger channel error cut than clean ones","Semantic LLM pilots: text correction improves MIMO channel estimate","Theorem: LLM-corrected symbols beat clean ones for channel estimation","Two-layer LLM pilot selection approaches oracle MIMO estimation","LLM semantic checks pick reliable pilots, improving MIMO estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000562,"raw_usage":{"total_tokens":2491,"prompt_tokens":720,"completion_tokens":1771,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":1691}},"tokens_in":464,"tokens_out":1771,"duration_ms":14430,"temperature":1.0,"reasoning_tokens":1691,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T14:46:11.826744+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the received signals of symbols admitted by Layer 2 and estimate the conditional distribution of the noise vector given 'corrected' status; if its mean or variance deviates materially from CN(0, σ²I), the Appendix A independence assumption fails and the per-symbol Δ(x_C) bound in Lemma 1 does not apply to the actually selected pilots. Alternatively, compare per-symbol NMSE reduction of corrected vs verified symbols on the test set; if corrected symbols do not show a strictly larger average reduction, Theorem 1's conclusion is not reproduced.","supporting_citations":[],"review_version":1}