{"id":"ad7cea65-975f-475a-8ecd-a6978f10c8ac","arxiv_id":"2505.06175","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A prompt-conditioned neural equalizer that exchanges soft information with an LDPC decoder over turbo iterations improves post-decoding BER in heavily quantized MIMO simulations.","lead":"This paper trains a soft MIMO equalizer as an in-context learner: it reads a prompt of pilot examples plus decoder feedback and outputs symbol probabilities. In quantized-channel simulations, the learned equalizer beats classical linear turbo equalizers, and in 16-QAM it also beats a perfect-channel-knowledge linear baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper never states that the ICL equalizer's posterior output is converted to extrinsic LLRs via Eq. (9); without this subtraction, the 5th-iteration turbo gains can reflect double-counted decoder feedback rather than genuine equalization improvement.","rationale":"The reader's weakest assumption was distribution shift beyond the training simulator; that is a legitimate scope limitation but it does not threaten the in-distribution simulations that constitute the paper's central evidence. The more pressing issue is internal to the turbo loop: the manuscript's only algorithmic description of the ICL equalizer output is posterior, and it never explicitly applies Eq. (9) to obtain extrinsic LLRs. Since the headline results are reported at the 5th turbo iteration, where p_i is non-trivial, an omitted extrinsic subtraction would give the ICL equalizer access to its own decoder output and could inflate the apparent iterative gains. This directly affects the central claim of outperforming perfect-CSI baselines, because the BLMMSE-PIC baseline does follow the extrinsic-LRR discipline of Section III-A. I do not reject the paper: the block diagram labels the output as L_e, and a charitable reading is that the authors assume standard turbo practice. But the omission is load-bearing for reproducibility and correctness, and it is readily testable. The verdict remains CONDITIONAL: the paper should be accepted only if the extrinsic subtraction is confirmed, either in the text or in released code. This is partial agreement with the reader because they listed extrinsic-LLR subtraction as an issue but chose distribution shift as the weakest assumption.","tokens_in":20006,"tokens_out":9873,"duration_ms":111274,"concrete_test":"Re-run the Fig. 8(b) 16-QAM, SNR=15 dB, B=8, TP=16 experiment with an explicit extrinsic step: compute \\hat p_i, convert it to bitwise LLRs via (21), subtract the bitwise prior LLRs derived from p_i (equivalently, divide out the prior PMF), and feed only the residual to the BP decoder. Compare the 5th-iteration post-decoding BER with the paper's curve, and also with a version that omits the subtraction. If the subtraction version is materially worse or unstable, the reported turbo gains rely on double-counted decoder feedback. If no code is released, this check cannot be performed and the result should be treated as unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central performance evidence (Figs. 8-11) is reported after the 5th turbo iteration, where the prompt includes decoder feedback p_i. In Section IV-A, the equalizer output \\hat p_i is defined as an estimate of the posterior Pr(x_i | C, y_i, p_i), and the text says 'this output is finally converted to bitwise LLRs in the same way as (21), which are then passed to the channel decoder.' Unlike the conventional equalizer in Section III-A, which explicitly computes extrinsic LLRs by subtracting the prior via Eq. (9), the paper never states that the ICL output is converted to extrinsic form before being fed to the decoder. If the posterior LLRs are passed directly, the decoder's input already contains the decoder's own previous extrinsic information; subsequent decoder-extrinsic feedback and the next ICL prior are correlated with the input, causing double-counting. This can inflate iterative BER gains and would make the 5th-iteration comparisons against BLMMSE-PIC (which does use proper extrinsic subtraction) unfair. The block diagram labels the ICL output as L_e, but no equation or pseudocode confirms the subtraction, and no code is released. Thus the 'consistently outperform with perfect CSI' claim, which rests largely on 5th-iteration curves, is not verifiable from the manuscript as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Turbo-ICL, a soft-input soft-output in-context learning (ICL) equalizer for coded MIMO systems with quantized receivers. A causal sequence model (Transformer or Mamba) is prompted with pilot transmit-receive pairs and a target query consisting of the received symbol vector and a prior symbol PMF from the decoder; the model outputs an estimated posterior symbol distribution. In subsequent turbo iterations, decoder-extrinsic LLRs are re-interleaved and converted into symbol priors that augment the prompt, allowing iterative refinement. The models are pre-trained on synthetic tasks sampled from a distribution over channels, noise variances, and quantization settings, and are evaluated against RLS-LMMSE-PIC, BLMMSE-PIC with perfect CSI, and MAP detection with perfect CSI. The paper reports BER versus quantization resolution, SNR, pilot length, code rate, training diversity, and model size, and claims consistent gains over the model-based baselines.","tokens_in":20327,"tokens_out":8461,"duration_ms":79029,"significance":"If the extrinsic-information handling is corrected and confirmed, this would be a useful contribution: it appears to be the first ICL equalizer that produces soft symbol posteriors and closes the turbo loop with decoder feedback, and it demonstrates a CSI-free advantage over linear model-based equalizers under 16-QAM with coarse quantization. The comparison of Transformer and state-space backbones and the efficiency analysis are informative. However, the paper ships no code, data, or error bars, and its headline claim is broader than the reported 4-QAM results; these issues materially limit the evidence for the contribution as currently stated.","major_comments":[{"comment":"The manuscript never specifies how the ICL equalizer's posterior output is converted to extrinsic LLRs before being fed to the decoder. In Sec. III-A, Eq. (9) defines channel-extrinsic LLRs by subtracting the prior LLR, and the model-based baselines use this operation. In contrast, Sec. IV-A states only that the output \\hat p_i = \\hat p_\\theta(x_i|C,y_i,p_i) 'is finally converted to bitwise LLRs in the same way as (21), which are then passed to the channel decoder,' with no subtraction of the prior p_i or its LLR equivalent. From the second turbo iteration onward, the decoder input may therefore already contain the decoder's own previous extrinsic information, causing double-counting in the equalizer-decoder loop. Because the headline gains in Figs. 8-11 are shown at the fifth turbo iteration, the comparison with BLMMSE-PIC, which does use Eq. (9), is not verifiable from the manuscript as written and may be unfair. Please state explicitly, in an equation or pseudocode, how \\hat p_i is made extrinsic before decoding, and rerun or re-justify the fifth-iteration results.","section":"Sec. IV-A, Eqs. (22)-(27); Fig. 3(b)"},{"comment":"The abstract and contribution list claim that ICL equalizers 'consistently outperform conventional model-based baselines, even when the latter are provided with perfect channel state information,' but the paper's own 4-QAM results do not support this. Sec. VI-D says that under 4-QAM, ICL-T and ICL-S 'approach the performance of BLMMSE-PIC with perfect CSI but do not exceed it,' and Fig. 8(a) shows only close approach across SNR. The defensible version is the 16-QAM statement in Sec. VI-D, where ICL outperforms BLMMSE-PIC with perfect CSI and approaches the MAP detector. Please qualify the abstract and contribution claims to specify the regimes (e.g., higher-order modulation and low-resolution quantization) in which the stated outperformance is demonstrated.","section":"Abstract; Sec. I-C; Sec. VI-D"},{"comment":"The pre-training distribution D_\\tau is not fully specified. Sec. VI-C defines D_B = U({1,...,32}) as 'the number of active non-zero symbols,' but the system model in Sec. II-B parameterizes the receiver front end by Q = {l_min,l_max,B} with B the quantization resolution, and the experiments vary B over {2,...,10}. There is no 'number of active non-zero symbols' parameter in the model, so the sampling distribution used to generate the simulated tasks is ambiguous. Please correct this description and specify exactly how B (or whichever parameter D_B ranges over) enters the link model and how the values used in training relate to the test values.","section":"Sec. VI-C; Sec. II-B"},{"comment":"The central performance claims rest on Monte Carlo BER curves without any statistical uncertainty quantification. The paper reports no standard deviations, confidence intervals, number of independent channel/codeword trials, or seeds, and no code or data are released. This matters because several key comparisons are close: for example, ICL-T and BLMMSE-PIC are nearly indistinguishable in Fig. 8(a), and the ICL-T/BLMMSE-PIC crossover near T_P=10 in Fig. 9(b) is within line width. Without trial-level statistics or a reproducibility package, the claimed ordering of methods cannot be assessed. Please add error bars or confidence intervals and a statement of the number of Monte Carlo runs, or explicitly mark the curves as representative single-seed results.","section":"Sec. VI-A, VI-H, Figs. 8-11"},{"comment":"The claimed generalization 'across channels and receiver architectures' is broader than what is evaluated. Sec. VI-H studies only training diversity N_train within the same task distribution D_\\tau: i.i.d. Rayleigh fading, fixed clipping range [-4,4], and a b-bit mid-rise uniform quantizer. No experiments test correlated fading, different clipping ranges, non-uniform ADC, or other hardware impairments, although these are the scenarios in which a learned equalizer's in-context adaptation would matter most. Please either add out-of-distribution evaluations that vary the receiver front end and channel model, or narrow the generalization claim in Sec. I-C and the conclusion.","section":"Sec. I-C; Sec. VI-H"}],"minor_comments":[{"comment":"The caption says 'The results are evaluated at the fifth turbo iteration with TP=16,' but the x-axis is the pilot length T_P, which is varied from 4 to 32; this is inconsistent and should be corrected.","section":"Fig. 9 caption"},{"comment":"The text says 'We do not claim the optimality of the choice β = 1n,' but the concentration parameter was defined as β = 1_M; the subscript n appears to be a typo.","section":"Sec. V-A"},{"comment":"The dimension bookkeeping in the multi-head attention output is hard to follow: with Q_h,K_h,V_h ∈ R^{D_H×(2T_P+1)}, the expression Softmax(Q_h^T K_h/√D_H + M) V_h^T yields a (2T_P+1)×D_H matrix, not a D_H×(2T_P+1) matrix as implied by the later concatenation. Please make the transposes explicit.","section":"Algorithm 1"},{"comment":"References [16] and [42] appear to be the same paper (da Silva and Yankov, 'Adaptive turbo equalization for nonlinearity compensation in WDM systems'), cited under two different numbers; please consolidate.","section":"References [16] and [42]"},{"comment":"The statement that N_train = 2^15 = 32,768 'unique combinations' is imprecise because H and σ^2 are drawn from continuous distributions; the pre-training task pool is a set of sampled task instances, not unique combinations of discrete parameters.","section":"Sec. VI-C"}],"recommendation":"major_revision","confidential_remarks":"The extrinsic-LLR issue is the one that could change the conclusions, and it is fixable in principle. The authors should also decide whether the paper's headline is the 16-QAM result or the broader 'consistently outperform perfect-CSI baselines' claim; as written, the abstract contradicts their own Sec. VI-D. I would encourage the editor to ask for error bars and code/data release before acceptance, since the paper's contribution is entirely empirical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: a soft-output in-context-learning equalizer whose prompt includes decoder feedback as prior information, so it can run inside a turbo loop. Prior ICL equalizers were single-pass and hard-decision or uncalibrated; this paper extends the framework to posteriors and iterative decoding. The 16-QAM results against BLMMSE-PIC with perfect CSI are the interesting claim, and the architecture and training setup are described coherently. The scaling experiments with Ntrain and the SSM-vs-Transformer efficiency comparison are also useful. I believe the core idea is sound and the paper should be seen by a serious referee.\n\nThe largest soft spot is exactly what the stress test flags. Section III-A defines extrinsic LLRs for the conventional equalizer as posterior minus prior via Eq. (9). Section IV-A says the ICL output is \"converted to bitwise LLRs in the same way as (21)\" and then passed to the decoder. Eq. (21) computes posterior LLRs from a symbol distribution; nothing in the text subtracts the prior pi that was injected into the prompt. If the decoder receives its own previous extrinsic information back as part of the posterior, the 5th-iteration BER curves in Figs. 8-11 can be inflated by double-counting, and the comparison against BLMMSE-PIC, which does implement proper extrinsic subtraction, would be unfair. This is a load-bearing ambiguity, not a cosmetic one. It is fixable by explicitly stating the subtraction and showing the equation, but without code or a pseudocode line it cannot be verified from the manuscript as written.\n\nTwo smaller issues. First, the abstract overclaims: Sec. VI-D explicitly says that under 4-QAM the ICL methods \"approach... but do not exceed\" BLMMSE-PIC with perfect CSI, while the abstract says ICL equalizers \"consistently outperform conventional model-based baselines, even when... provided with perfect channel state information.\" The defensible claim is the 16-QAM one; the 4-QAM curves do not support the sweeping wording. Second, there is no code, no data, and no error bars anywhere; every figure is a single simulated curve. That would be a minor complaint for a theory paper, but here the central claim is an empirical win over a strong baseline, so reproducibility matters. The generalization tests also stay inside the same i.i.d. Rayleigh, uniform-quantizer family used for training; correlated fading or non-uniform ADC is untested, and the paper should say so plainly.\n\nWho gets value: wireless-ML researchers working on neural receivers and ICL for communications. The paper deserves peer review, and I would send it out, but with a required revision addressing the extrinsic-LLR processing and a more careful abstract. If the LLR issue resolves favorably, this is a citable step forward; right now it is a promising result with an unverified key detail.","headline":"A genuinely new soft-output ICL equalizer for turbo loops, with the 16-QAM gains worth taking seriously, but the paper never confirms the extrinsic-LLR subtraction that the 5th-iteration results depend on.","tokens_in":20874,"tokens_out":1574,"would_cite":true,"duration_ms":18354,"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":"In-context learning lets a MIMO receiver run turbo equalization with no channel state information, and its soft outputs can beat model-based receivers that know the channel.","keywords":["in-context learning","turbo equalization","MIMO detection","soft-output equalization","quantized receiver","Transformer","state-space model","CSI-free equalization"],"falsifier":"Run the pre-trained ICL-T and ICL-S models on a test set with correlated Rayleigh fading or with a measured receiver front end whose clipping and quantization differ from the uniform mid-rise model used in pre-training; the central claim would be weakened if under those conditions the perfect-CSI BLMMSE-PIC baseline matches or beats the ICL equalizers.","tokens_in":19818,"feed_emoji":"📡","tokens_out":4023,"duration_ms":42329,"temperature":0.7,"pith_summary":"This paper claims that soft-input soft-output channel equalization for coded MIMO systems can be treated as an in-context learning problem: a sequence model reads a prompt made of pilot transmit-receive pairs plus the decoder's soft feedback, and directly outputs a posterior distribution over each transmitted symbol, with no explicit channel estimate. The authors build two instances, one on a Transformer and one on a state-space model, and run them inside a turbo loop where decoder extrinsic information is returned to the model as augmented prompt context. In simulation, these ICL equalizers match or beat a linear MMSE-PIC receiver that is handed perfect channel state information, particularly under coarse quantization and 16-QAM, and approach the MAP detector. The practical stake is that a single pre-trained receiver could adapt on the fly to new channel conditions using only a few pilot symbols, avoiding channel estimation and per-scenario retraining.","feed_headline":"Prompt-trained equalizer beats receivers that know the channel","feed_subtitle":"If right, MIMO receivers can adapt on the fly from pilot examples and decoder feedback—no channel estimation needed.","key_machinery":"The load-bearing mechanism is prompt augmentation: the decoder's extrinsic soft information is encoded as a per-antenna probability mass function over the M-QAM constellation and appended to the target token in the prompt, alongside an embedding of the received pilot and data vectors. The sequence model, either a Transformer decoder or a selective state-space model (Mamba-style), is trained on interleaved query-answer prompts, with each received symbol paired with a Dirichlet-sampled prior PMF and its true transmitted symbol, using a weighted cross-entropy loss that puts more weight on later predictions so the model learns to exploit accumulating context. At inference, the final token's output is a posterior PMF estimate, converted to bitwise LLRs and passed to the decoder; the decoder's extrinsic LLRs are turned back into a prior PMF and reinserted into the prompt, closing the turbo loop.","core_discovery":"The paper's central discovery is that the turbo equalizer's role can be re-expressed as a supervised in-context mapping from received signals plus prior probabilities to symbol posterior PMFs. The soft equalizer is a causal sequence model that sees a prompt containing received-pilot/transmitted-pilot pairs followed by the target query, which concatenates the received data symbol with a prior PMF obtained from the decoder. Because the decoder's priors are appended to the prompt as part of the target query, the same model refines its estimate across turbo iterations without any architectural change. Under 16-QAM with low-bit quantization, the ICL models outperform the perfect-CSI Bussgang LMMSE-PIC baseline and approach the MAP lower bound, while under 4-QAM they match the perfect-CSI baseline; both consistently beat the practical RLS-LMMSE-PIC receiver. The authors present this as evidence that the model learns, from data, the nonlinear and discrete structure of the channel-plus-quantizer map rather than relying on Gaussian and linearity assumptions.","pith_inferences":["The prompt-augmentation recipe may generalize beyond equalization: any iterative receiver that exchanges soft information could expose that information as context tokens, making the approach a template for ICL-based turbo processing.","The paper only demonstrates generalization within the pre-training distribution; testing on correlated fading, non-uniform ADC, or measured hardware would clarify whether the pilot-based adaptation extends to channels not representable in the simulator.","The Dirichlet-prior training strategy invites a calibration probe: if the output PMFs are well-calibrated soft likelihoods, they should remain reliable when the decoder's priors become very confident in later turbo iterations.","A natural extension would be to pre-train over a broader task distribution that includes hardware impairments, then test whether the same prompt-based receiver absorbs them without architectural changes."],"forward_implications":["A single pre-trained ICL equalizer can replace the channel-estimation stage in a turbo receiver, working on quantized and nonlinear front ends where linear assumptions break down.","Decoder feedback genuinely improves the ICL equalizer: feeding extrinsic information back as prompt context lowers post-decoding BER between the first and fifth turbo iterations.","With 16-QAM and low-bit ADCs, ICL methods can beat a linear equalizer that knows the channel perfectly, approaching the MAP detector's performance.","The Transformer variant generalizes from fewer training tasks, while the state-space variant delivers comparable BER with about 40 percent of the parameters and lower computational cost.","Fewer pilots suffice for the ICL receiver: with 16 context pilots it outperforms a perfect-CSI linear baseline that uses 32 pilots for estimation."],"supporting_citations":[{"why":"Established ICL for MIMO equalization with Transformer-based sequence models, providing the direct precursor for this framework.","marker":"[11]"},{"why":"Showed that Transformers can serve as efficient in-context estimators for wireless channels, motivating the CSI-free equalizer design.","marker":"[12]"},{"why":"Extended Transformer-based ICL equalization to MIMO fading channels, which this paper builds on for MIMO turbo processing.","marker":"[13]"},{"why":"Introduced turbo equalization, the iterative soft-information exchange architecture that this paper integrates with ICL.","marker":"[2]"},{"why":"Provided the LMMSE equalization-with-prior approach that underlies the conventional baselines being compared.","marker":"[3]"},{"why":"Defined the soft-input soft-output MMSE-PIC detector that serves as the main model-based baseline.","marker":"[4]"},{"why":"Supplied the structured state-space machinery and discretization used by the SSM-based ICL-S variant.","marker":"[35]"},{"why":"Provided the Mamba selective state-space backbone used for the efficient ICL-S model.","marker":"[36]"},{"why":"Provided the GPT-2 Transformer backbone used for the ICL-T model.","marker":"[44]"}],"fun_headline_variants":["Turbo-ICL: no channel model needed to beat perfect CSI","Pilot-prompted equalizer beats receivers with perfect CSI","In-context learning equalizer rivals perfect channel knowledge","Decoder feedback refines MIMO equalizer without channel estimates","Prompt-based turbo equalization outperforms known-channel baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gains rest on the assumption that the channel, noise, and quantization conditions seen at deployment are representable by the same family of simulated tasks used in pre-training; the evaluation only generalizes within that simulated distribution.","fun_headline_variants_meta":{"raw":{"variants":["Turbo-ICL: no channel model needed to beat perfect CSI","Pilot-prompted equalizer beats receivers with perfect CSI","In-context learning equalizer rivals perfect channel knowledge","Decoder feedback refines MIMO equalizer without channel estimates","Prompt-based turbo equalization outperforms known-channel baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1510,"prompt_tokens":922,"completion_tokens":588,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":504}},"tokens_in":538,"tokens_out":588,"duration_ms":5850,"temperature":1.0,"reasoning_tokens":504,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:47:41.182622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pre-trained ICL-T and ICL-S models on a test set with correlated Rayleigh fading or with a measured receiver front end whose clipping and quantization differ from the uniform mid-rise model used in pre-training; the central claim would be weakened if under those conditions the perfect-CSI BLMMSE-PIC baseline matches or beats the ICL equalizers.","supporting_citations":[{"cited_title":"In-context learning for MIMO equalization using transformer-based sequence models,","cited_arxiv_id":null,"evidence_quote":"Established ICL for MIMO equalization with Transformer-based sequence models, providing the direct precursor for this framework."},{"cited_title":"Transformers are Provably Optimal In-context Estimators for Wireless Communications","cited_arxiv_id":"2311.00226","evidence_quote":"Showed that Transformers can serve as efficient in-context estimators for wireless channels, motivating the CSI-free equalizer design."},{"cited_title":"Cell-free multi-user MIMO equalization via in-context learning,","cited_arxiv_id":null,"evidence_quote":"Extended Transformer-based ICL equalization to MIMO fading channels, which this paper builds on for MIMO turbo processing."},{"cited_title":"Iterative correction of intersymbol inter- ference: turbo-equalization,","cited_arxiv_id":null,"evidence_quote":"Introduced turbo equalization, the iterative soft-information exchange architecture that this paper integrates with ICL."},{"cited_title":"Minimum mean squared error equalization using a priori information,","cited_arxiv_id":null,"evidence_quote":"Provided the LMMSE equalization-with-prior approach that underlies the conventional baselines being compared."},{"cited_title":"ASIC implementation of soft- input soft-output MIMO detection using MMSE parallel interference cancellation,","cited_arxiv_id":null,"evidence_quote":"Defined the soft-input soft-output MMSE-PIC detector that serves as the main model-based baseline."}],"review_version":1}