{"id":"897e266e-b966-477c-b255-ceb895923b07","arxiv_id":"2501.12167","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A soft belief-propagation decoder for LDPC-based quantitative group testing beats the hard-decision peeling decoder in simulations of misdetection rate.","lead":"Researchers propose a belief-propagation decoder that uses soft information to find defective items in quantitative group testing, where each test reports how many defective items are in a pool. In simulations, the new decoder correctly identifies defective items at higher prevalence levels than the previous hard-decision peeling decoder.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed gains over peeling decoder may hinge on exact knowledge of prevalence δ; no mismatched-prior experiments are reported, so the advantage could be an artifact of feeding the true prior.","rationale":"The reader identified the exact same load-bearing concern: the soft decoder requires the true prevalence δ as a prior, while the peeling baseline does not, and the paper never tests mismatched δ. This is a genuine threat to the central claim because the reported gains could be entirely due to the unfair advantage of knowing the ground-truth prevalence. The proposed concrete test directly addresses this: it varies δ_est while holding the true δ and graph fixed, and compares the resulting PMD to the peeling baseline. If the gain survives moderate mismatches, the claim is robust; if not, the conclusion must be qualified. The reader's CONDITIONAL verdict is appropriate: the paper is technically sound but needs this robustness check and perhaps error bars or code before full acceptance. I found no other issue that is more load-bearing; the underspecified graph construction is secondary because it affects both decoders equally and can be clarified. My recommendation is therefore to keep the reader's verdict unchanged, conditional on the sensitivity analysis being performed.","tokens_in":8655,"tokens_out":4345,"duration_ms":44610,"concrete_test":"Re-run the experiments of Section V for the (3,6) regular graph with n = 1024 and n = 8192. For each true δ in {0.1, 0.2, 0.3, 0.4}, run Algorithm 1 with the decoder prior δ_est set to values in a grid around the true δ, e.g., δ_est ∈ {0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50}. Plot PMD versus δ_est after L = 100 iterations and compare against the peeling decoder PMD at the same true δ (dashed curves in Figs. 3 and 4). If PMD at δ_est = true δ ± 0.05 rises to or above the peeling PMD, the claimed gain is not robust to prevalence mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed BP decoder outperforms the peeling decoder of [15] in misdetection rate for noiseless non-adaptive quantitative group testing. Algorithm 1 takes δ as an explicit input, and the variable node update (3) uses the true prior Pr_Di(b) = (1−δ, δ). The peeling decoder baseline in [15] does not use δ. In all simulations (Section V), the soft decoder is fed the exact true δ, while the baseline is not given this information. The paper never tests sensitivity to a mismatched δ. If δ is unknown or estimated from data, the reported gains (0.02 to 0.17 in δ at PMD ≈ 10^−3) may shrink or disappear. This is the weakest point in the argument because the informational advantage of the soft decoder is confounded with its algorithmic advantage. The claim 'outperforms the hard-decision decoder' is therefore only established under an idealized assumption that is not stated as a limitation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a belief-propagation (BP) decoder for LDPC-code-based noiseless non-adaptive quantitative group testing. The decoder exchanges soft messages between variable nodes (items) and constraint nodes (tests), with the constraint-node update performing exact marginalization over the test constraint. The authors derive the variable-node and constraint-node update rules in Section IV, provide a computational complexity analysis, and present Monte Carlo simulations in Section V showing that the proposed soft decoder achieves lower misdetection rate than the hard-decision peeling decoder of Mashauri et al. [15] for regular graphs with (dv,dc) = (3,6) and (3,9), for population sizes ranging from n=128 to n=16384. The reported gain in prevalence δ at PMD ≈ 10^-3 ranges from about 0.02 (short lengths) to 0.17 (moderate lengths).","tokens_in":8820,"tokens_out":5030,"duration_ms":53340,"significance":"If the claimed improvement is robust, the proposed decoder is a useful contribution to quantitative group testing: it is a natural, correctly derived belief-propagation adaptation of the LDPC framework, and the exact constraint-node update (5) is a sound marginalization for the noiseless adder-channel constraint. The paper is clearly written and the algorithm is straightforward to implement. However, the central claim is currently supported only by simulations that are confounded by an asymmetry in side information and that lack statistical error characterization. The authors do not provide code, number of Monte Carlo trials, or sensitivity analysis, so the magnitude of the reported gains is not yet established. The work has potential, but the evidence as presented is insufficient to fully validate the central outperformance claim.","major_comments":[{"comment":"The comparison between the proposed decoder and the peeling decoder of [15] is confounded: the BP decoder is given the true prevalence δ as the prior in the VN update (3) and as an input to Algorithm 1, while the peeling decoder baseline does not use δ. The paper never tests sensitivity to a mismatched prior, so the reported gains in δ (0.02–0.17) may partly or entirely stem from this extra information rather than from the soft-information update rule. To substantiate the claim that the proposed decoder outperforms the hard-decision decoder, the authors should add experiments with a misspecified prior, provide a baseline decoder that also knows δ, or explicitly state and justify the assumption that δ is known exactly.","section":"Section V, Algorithm 1, Eq. (3)"},{"comment":"The numerical results are presented without error bars, confidence intervals, or any statement of the number of Monte Carlo trials. Since the target PMD values are as low as 10^-3, the plotted points are estimates of rare-event probabilities and their reliability is unclear. Without a statistical characterization, the claimed prevalence gains (e.g., 0.14–0.18 for moderate lengths) cannot be assessed for significance. Please specify the number of simulation runs, the simulation procedure, and add error bars or confidence bands to the figures.","section":"Section V, Figs. 3–5"}],"minor_comments":[{"comment":"The notation [x] is defined as the set {1,...,x}, but square brackets are also used for citations throughout the text; this dual use is confusing and should be disambiguated (e.g., use a different symbol for the set).","section":"Section II"},{"comment":"The abstract mentions 'Mashauri et al.' without a citation number, and the first author's affiliation contains a typo ('T echnical'). Please fix these presentation issues.","section":"Abstract and author affiliation"},{"comment":"The quantity r_i^(ℓ) in the second bullet point is used but never defined; please define it explicitly or rephrase the sentence.","section":"Section IV-B, second bullet"},{"comment":"The footnote says that slightly fewer iterations yield similar performance, but the paper does not report how the performance degrades with L. Since L is a free parameter, a small experiment varying L would help the reader understand the sensitivity of the results to this choice.","section":"Section V, first paragraph"},{"comment":"The legends in the figures repeat the same n values for both the solid and dashed curves, which makes the figures hard to read. Please use a combined legend entry (e.g., 'BP, n=128' and 'Peeling, n=128') or otherwise clarify the correspondence.","section":"Figures 3–5"}],"recommendation":"major_revision","confidential_remarks":"The central concern is the confounded comparison: the BP decoder uses the true prevalence δ while the peeling decoder does not. If the authors can demonstrate that the gain persists with a mismatched prior or that the baseline is also given the same information, the paper may be acceptable. The lack of error bars is also a substantive issue for the quantitative claims. I recommend major revision and would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent, well-scoped paper. The new piece is a belief-propagation decoder for the LDPC-based quantitative group testing scheme of [15], with a constraint-node update that exactly marginalizes the adder-channel constraint. The derivation is correct, the pseudocode is clear, and the authors are honest about complexity being exponential in check-node degree. The comparison to peeling is the right benchmark, and the reported gains are consistent with what you would expect from soft information. I believe the central claim is plausible.\n\nWhat it does well: the CN update (5) is the natural exact marginalization for the sum constraint; it generalizes the hard-decision update of [15], and the authors explicitly note the relation to counter braids [18]. That is honest. The simulation section covers short and moderate blocklengths, two regular graph degrees, and reports meaningful gains in tolerable prevalence at PMD around 1e-3. The writing is concise and mostly precise. I checked the citation pattern: it points to the relevant sparse-graph group testing and compressed-sensing literature, and the self-citations are topical, not padding.\n\nSoft spots: the main one is the prior. Algorithm 1 takes delta as an input, the VN update uses the true Bernoulli prior, and every simulation feeds the exact true delta to the soft decoder while the peeling baseline gets no prevalence information. That is an informational asymmetry, not just an algorithmic one. The paper never tests a mismatched delta. If delta is estimated or wrong, the 0.02-0.17 prevalence gain could shrink, though for a Bayesian decoder with exact constraint nodes I would expect graceful degradation rather than collapse. Since the authors frame the decoder for practical applications like federated learning, they should include a sensitivity plot. Second, there are no error bars, no confidence intervals, and no code release. The curves are smooth and the gaps are large, so I do not doubt the qualitative result, but reproducibility is weaker than it should be. Third, the LDPC graph construction is not fully specified (random regular? progressive edge growth? girth?), which makes exact replication harder. These are fixable issues, not a load-bearing flaw.\n\nBottom line: worth a serious referee and likely publishable after a mismatched-prior experiment and some reproducibility details. It is an incremental but useful contribution for people working on code-based group testing, not a fundamental advance. If I were reviewing, I would ask for the delta-robustness experiment and a code/data statement, then accept.","headline":"A clean, modest extension of belief propagation to quantitative group testing, with a real but untested caveat: all reported gains come from feeding the true prevalence only to the proposed decoder.","tokens_in":9344,"tokens_out":2309,"would_cite":true,"duration_ms":25685,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B35","94B70"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces a belief-propagation decoder with purely soft messages for LDPC-code-based quantitative group testing, and shows by simulation that it lowers the misdetection rate relative to the hard-decision peeling decoder, with…","keywords":["quantitative group testing","belief propagation","soft-decision decoding","LDPC codes","misdetection rate","prevalence","non-adaptive group testing","iterative decoding"],"falsifier":"Run the soft decoder on a $(3,6)$ regular graph with $n=1024$ and true prevalence $\\delta=0.3$, but with the decoder using a fixed prior $\\delta=0.1$, and compare its misdetection rate with the peeling decoder; if the soft decoder no longer reaches the $10^{-3}$ target while the peeling decoder does, the claimed advantage depends on perfect prior knowledge.","tokens_in":8457,"feed_emoji":"🧪","tokens_out":8211,"duration_ms":73006,"temperature":0.7,"pith_summary":"This paper proposes a belief-propagation decoder for noiseless, non-adaptive quantitative group testing, a setting in which each test reports the exact number of defective items in a pool rather than just a positive or negative result. The decoder works on the bipartite graph defined by an LDPC-code assignment matrix and passes soft probability messages between item nodes and test nodes. By simulation, the authors show that this soft-information decoder achieves a lower misdetection rate than the hard-decision peeling decoder of [15] for every population size and prevalence tested. At a target misdetection probability of $10^{-3}$, the gain in tolerated prevalence is about 0.02 for short populations and up to 0.17 for moderate populations, meaning the same test budget can identify defective items when a larger fraction of the population is defective.","feed_headline":"Soft decoding beats hard-decision group testing","feed_subtitle":"At a 1-in-1000 misdetection target, it handles defect prevalence up to 0.17 higher than the hard-decision decoder.","key_machinery":"The load-bearing object is the bipartite factor graph of the assignment matrix, with $n$ variable nodes for items and $r$ constraint nodes for tests, together with the two message-update rules. The constraint-node update (equation 5) computes the belief that an item is defective by summing over all assignments of the other neighbors whose total equals the observed test outcome minus the item's candidate state, which is the step that extracts soft information from the exact count. The variable-node update (equation 3) is the product of the prevalence prior and all incoming constraint messages except the one on the edge being updated. Because the constraint sum ranges over all configurations with a given count, the update complexity is $O(2^{d_c-1})$, exponential in the check-node degree.","core_discovery":"The central claim is that a belief-propagation decoder with purely soft messages strictly improves the misdetection rate of LDPC-code-based quantitative group testing compared with the hard-decision peeling decoder of [15]. The decoder models each test outcome as a constraint that the sum of the states of the items in the pool equals the observed count, and it computes each constraint-to-variable message by summing over all configurations of the other pool members consistent with that count. The variable-node update multiplies the incoming constraint beliefs by a Bernoulli prior set by the prevalence $\\delta$. Simulations on regular $(3,6)$ and $(3,9)$ assignment graphs with population sizes from 128 to 16,384 show consistent gains that grow with population size, and the soft update rules reduce to the peeling decoder's declarations as extreme cases.","pith_inferences":["The reported gains assume the decoder knows the true prevalence $\\delta$; if $\\delta$ must be estimated from imperfect data, the advantage over hard-decision decoding may shrink, and the paper does not test that sensitivity.","The same message-passing structure could be augmented with an expectation-maximization step to estimate $\\delta$ on the fly, producing a self-calibrating decoder.","Since the decoder outputs an approximate posterior for each item, thresholds can be tuned to reflect asymmetric costs of false positives and false negatives, an option the paper does not explore.","The monotone improvement with $n$ hints at a finite-length threshold phenomenon similar to LDPC decoding, where performance sharpens once the graph becomes locally cycle-free."],"forward_implications":["At a fixed number of tests, the soft decoder can operate at a higher prevalence while meeting the same misdetection target, which lowers the cost of screening populations with more defective items.","For the regular graphs tested, the prevalence gain over the hard-decision decoder grows with the population size, from about 0.02 at $n=128$ to roughly 0.14-0.18 at $n=16{,}384$.","The decoder is compatible with irregular LDPC assignments as well as the regular graphs used in the simulations.","Because the messages are soft probabilities, the authors expect a natural extension to noisy test outcomes by replacing the exact constraint with the conditional test distribution, a route that hard-decision peeling cannot take directly.","The exponential check-node complexity bounds practical use to pools of small degree, so the reported gains apply in the low-degree regime."],"supporting_citations":[{"why":"Supplies the hard-decision peeling decoder and the LDPC-based quantitative group testing scheme that this paper builds on and benchmarks against.","marker":"[15]"},{"why":"Presents the optimal MAP decoder whose exponential complexity motivates the belief-propagation approximation.","marker":"[12]"},{"why":"Introduces generalized LDPC codes for quantitative group testing with BCH component codes, the prior scheme that [15] outperforms.","marker":"[13]"},{"why":"Provides the similar sum-constraint structure used for counter braids, informing the constraint-node update design.","marker":"[18]"}],"fun_headline_variants":["Belief-propagation decoder improves LDPC group testing","Soft-information decoding outperforms hard-decision group testing","Soft decision beats hard for LDPC group testing","Handles higher defect prevalence with soft decoding","Soft messages lift LDPC group testing accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The decoder assumes the prevalence $\\delta$ is known exactly and uses it as the prior in every variable-node update; all simulations feed the true $\\delta$, so a mismatched estimate could reduce or erase the reported gains.","fun_headline_variants_meta":{"raw":{"variants":["Belief-propagation decoder improves LDPC group testing","Soft-information decoding outperforms hard-decision group testing","Soft decision beats hard for LDPC group testing","Handles higher defect prevalence with soft decoding","Soft messages lift LDPC group testing accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000762,"raw_usage":{"total_tokens":3318,"prompt_tokens":820,"completion_tokens":2498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":2423}},"tokens_in":436,"tokens_out":2498,"duration_ms":19316,"temperature":1.0,"reasoning_tokens":2423,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:25:54.574871+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the soft decoder on a $(3,6)$ regular graph with $n=1024$ and true prevalence $\\delta=0.3$, but with the decoder using a fixed prior $\\delta=0.1$, and compare its misdetection rate with the peeling decoder; if the soft decoder no longer reaches the $10^{-3}$ target while the peeling decoder does, the claimed advantage depends on perfect prior knowledge.","supporting_citations":[{"cited_title":"Low- density parity-check codes and spatial coupling for quantitative g roup testing,","cited_arxiv_id":null,"evidence_quote":"Supplies the hard-decision peeling decoder and the LDPC-based quantitative group testing scheme that this paper builds on and benchmarks against."},{"cited_title":"Optimum detection o f defective elements in non-adaptive group testing,","cited_arxiv_id":null,"evidence_quote":"Presents the optimal MAP decoder whose exponential complexity motivates the belief-propagation approximation."},{"cited_title":"Sparse graph codes for non-adaptive quantit ative group testing,","cited_arxiv_id":null,"evidence_quote":"Introduces generalized LDPC codes for quantitative group testing with BCH component codes, the prior scheme that [15] outperforms."},{"cited_title":"Asymptotic analysis and spatial coupling of counter braids,","cited_arxiv_id":null,"evidence_quote":"Provides the similar sum-constraint structure used for counter braids, informing the constraint-node update design."}],"review_version":1}