{"id":"ef727e37-b403-459d-87bd-c94d2ba1aa1d","arxiv_id":"2607.04045","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"SOGRAND specialized to SPC codes gives two hardware-friendly LDPC check-node updates that match or beat SPA and NMS on 5G codes with list size ~10.","lead":"This paper shows that SOGRAND, specialized to single-parity-check codes, yields two new check-node update rules for LDPC decoding. Simulations on 5G NR codes match or beat sum-product and normalized min-sum while using small fixed pattern tables that map cleanly to hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper's strongest claim is modest and fully supported by the material presented: two new, fully specified CN update rules obtained by specializing an existing SOGRAND formula to SPC codes, together with simulation evidence that they perform comparably to SPA and NMS on two standard 5G codes. The reader's weakest assumption correctly notes the empirical character of the pattern table and damping factor, but that character is ordinary for practical LDPC CN approximations (NMS itself relies on a hand-tuned alpha) and does not create a correctness risk for the claim as stated. No hidden assumption in the derivation, no circularity, and no contradictory evidence appear. Consequently the CONDITIONAL verdict already assigned by the reader remains appropriate; no adjustment is warranted.","tokens_in":10016,"tokens_out":561,"duration_ms":5225,"concrete_test":"Re-implement both CN rules (eqs. 5 and 6) with the exact ORBGRAND table of Fig. 1 (L=10) and alpha=0.9 inside a standard 5G NR LDPC decoder; recompute BLER/BER for the (256,128) and (1024,676) codes over the same SNR range as Figs. 4-5. If the curves remain within 0.05 dB of SPA, the empirical claim is confirmed; any larger gap would indicate an implementation detail that the paper omitted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption correctly flags that the fixed ORBGRAND pattern table (Fig. 1) and global damping alpha=0.9 are empirical rather than first-principles, yet this does not undermine the central claim. The paper never asserts that the table is optimal or universal; it only claims that specializing SOGRAND's SISO formulas (eqs. 5 and 6) to SPC codes yields two concrete CN updates that, with L approximately 8-10 and alpha=0.9, match or slightly exceed SPA/NMS BLER/BER on the two 5G NR LDPC codes examined (Figs. 2-6). The derivation of the APP expressions from the SOGRAND block-wise soft-output formula is short and algebraically correct, the even/non-even variants are shown to be equivalent, and the free parameters are ordinary engineering knobs whose modest sensitivity is reported. Hardware-friendliness is argued by reference to already-fabricated GRAND circuits that compute the required likelihoods and pattern lookups in a few clock cycles; that argument is plausible but secondary. No internal inconsistency, missing derivation step, or simulation artifact that would reverse the performance claim is present.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper specializes Soft-Output GRAND (SOGRAND) to single-parity-check (SPC) component codes and derives two explicit check-node (CN) update rules for iterative LDPC decoding (Eqs. 5 and 6). One rule exploits the even-code property of SPCs (requiring the parity likelihood Ψ of Eq. 3) while the other does not; both produce extrinsic LLRs from a short list of ORBGRAND noise patterns. With list size L ≈ 8–10 and a global damping factor α = 0.9, simulations on two 5G NR LDPC codes (n = 256 rate-1/2 and n = 1024 rate-2/3) over BPSK-AWGN show block- and bit-error rates that match or slightly improve upon Gallager SPA and normalized min-sum (Imax = 50). The authors argue that the updates are hardware-friendly because existing GRAND circuits already evaluate the needed likelihoods and fixed pattern tables in a few clock cycles.","tokens_in":10248,"tokens_out":1216,"duration_ms":19124,"significance":"If the reported performance holds more broadly, the work supplies a genuinely alternative CN philosophy: block-wise SOGRAND posteriors are marginalized rather than computing leave-one-out tanh products or min-sums. The algebraic specialization from the general SOGRAND formula (Eqs. 1/4) to the SPC APP expressions is short and correct, the even/non-even variants are shown to be equivalent, and the empirical match to SPA/NMS with only L ≈ 10 patterns is practically interesting. Hardware friendliness is made plausible by reference to already-fabricated GRAND chips. The free parameters (α, L, fixed pattern table) are ordinary engineering knobs whose modest sensitivity is acknowledged. The contribution is therefore a clean, usable algorithmic alternative rather than a fundamental capacity or complexity breakthrough.","major_comments":[{"comment":"Section V and Figs. 2–6 evaluate only two 5G NR LDPC codes. While the BLER/BER curves support the claim for these instances, the abstract and introduction present the CN updates as a general alternative to SPA/NMS. At least one additional irregular or higher-rate 5G base-graph code (or a classical regular ensemble) is needed to substantiate that the same small L and fixed ORBGRAND table remain competitive across the practical design space.","section":null},{"comment":"The fixed ORBGRAND pattern table (Fig. 1) and the single global damping α = 0.9 are purely empirical. The paper reports that α ∈ [0.8,1] changes performance by only ~0.05 dB (data not shown) and that L ≥ 8 saturates, yet no analysis is given of why the particular 7–8 least-reliable-bit patterns suffice across SNR. A short ablation (different pattern generators, SNR-dependent L, or per-check α) would clarify whether the observed match to SPA is robust or fortuitous for the two tested codes.","section":null},{"comment":"Hardware claims (Section IV, final paragraph) rest entirely on citations to prior GRAND ASICs that compute likelihoods and pattern lookups in 1–3 cycles. No gate-count, critical-path, or area comparison against a standard SPA or NMS CN is supplied for the proposed updates themselves. A brief complexity table (operations per CN, memory for the pattern table, estimated latency) would make the “hardware-friendly” assertion quantitative rather than referential.","section":null}],"minor_comments":[{"comment":"Abstract and Introduction: the phrase “alternative existing Check Node (CN) update” is awkward; “alternative to existing CN updates” is clearer.","section":null},{"comment":"Eq. (5) multiplies the residual probability mass by a factor of 2; a one-sentence justification (or a pointer to the non-even SOGRAND derivation) would help readers who have not memorized the earlier paper.","section":null},{"comment":"Section V: “data not shown” for the α-sensitivity study should be replaced by a short table or inset, given that α is a free parameter of the algorithm.","section":null},{"comment":"Fig. 1 caption and surrounding text: clarify whether the same fixed table is used for every check degree or whether degree-dependent tables are required for irregular 5G graphs.","section":null},{"comment":"References: several recent hardware-oriented LDPC CN approximations (e.g., adjusted min-sum variants) are omitted; a brief comparison or citation would situate the work more completely.","section":null},{"comment":"Notation: λ^n and N^n are used for both vectors and their components; a consistent bold/vector convention would improve readability.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is essentially a short, well-executed specialization paper. Its novelty is real but incremental; it would fit a solid IEEE Transactions letter or a conference proceedings better than a full-length journal article unless the experimental base is broadened. Citation density toward the authors’ own GRAND line is high but not improper given the technical lineage. No integrity concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: they take the already-published SOGRAND soft-output formula, specialize it cleanly to single-parity-check nodes, and get two practical CN update rules (with and without the even-code correction) that, with L around 8–10 and a global alpha of 0.9, match or slightly beat SPA and normalized min-sum on two 5G NR LDPC codes. That is new and useful.\n\nWhat they do well is keep the derivation short and correct. Equations 5 and 6 follow directly from the block-wise SOGRAND posterior once you restrict to SPC; the even/non-even variants are shown to be equivalent in performance; and the fixed ORBGRAND pattern tables (Fig. 1) make the hardware story concrete by reference to already-fabricated GRAND circuits that can evaluate the needed likelihoods and lookups in a few cycles. Simulations are standard BPSK-AWGN, Imax=50, MATLAB toolbox baselines, and the free parameters are ordinary engineering knobs whose modest sensitivity is reported. Circularity is low—the SOGRAND formula is an independent approximation that has been checked elsewhere.\n\nSoft spots are real but proportionate. The pattern table and the single global alpha are empirical, not derived; they work on the two codes shown, but there is no first-principles guarantee they remain optimal across rates, lengths, or channels. No silicon measurements of the new CN itself, and no code release, so reproducibility is limited to re-implementing the fully specified algorithms. None of that overturns the central claim.\n\nThis is for people who design LDPC decoders or care about low-complexity CN kernels for 5G/6G hardware. It will not change asymptotic coding theory, but it is a clean, honest specialization that deserves a serious referee. I would engage with it if I were working on iterative soft decoding or GRAND hardware.","headline":"Solid, modest engineering paper: two concrete SPC CN updates from SOGRAND that match SPA/NMS on 5G LDPC with small fixed tables; useful for VLSI people, not a foundational shift.","tokens_in":10905,"tokens_out":485,"would_cite":true,"duration_ms":4727,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Specializing SOGRAND to single-parity-check codes yields two hardware-friendly check-node updates that match or beat sum-product and normalized min-sum on 5G LDPC codes.","keywords":["LDPC codes","Soft Input","Iterative Decoding","GRAND","SOGRAND","Check Node update","Single Parity Check"],"falsifier":"Re-run the same 5G NR LDPC simulations (n=256 rate-1/2 and n=1024 rate-2/3) with L=10 and α=0.9; if either SOGRAND rule produces a clear BLER or BER gap of more than a few tenths of a decibel relative to SPA or normalized min-sum at any operating point of interest, the central performance claim is refuted.","tokens_in":10856,"feed_emoji":"⚡","tokens_out":1081,"duration_ms":11714,"temperature":0.7,"pith_summary":"Long error-correcting codes are built by stitching short component codes together and decoding them iteratively with soft-input soft-output updates. For the single-parity-check components that make up LDPC codes, the classical check-node update is Gallager’s sum-product formula (or its normalized min-sum approximation). This paper shows that Soft-Output GRAND, when specialized to those same single-parity-check codes, produces two alternative check-node rules. Both rules work by ranking a short fixed list of candidate noise patterns drawn from the least-reliable bits, computing block-wise posterior probabilities, and then marginalizing to obtain per-bit extrinsic LLRs. On the 5G NR LDPC codes of lengths 256 and 1024, a list of only eight to ten patterns already matches or slightly exceeds the block- and bit-error rates of sum-product and normalized min-sum, while the updates themselves reduce to table look-ups and a handful of parallel probability multiplications that fit comfortably in a few clock cycles of existing GRAND hardware.","feed_headline":"SOGRAND check-node rules match SPA on 5G LDPC codes","feed_subtitle":"A short fixed list of noise patterns replaces sum-product and min-sum with hardware-friendly arithmetic","key_machinery":"The two SOGRAND check-node formulae (equations 5 and 6): a short list L of ORBGRAND noise patterns of correct parity is scored by channel likelihood; residual probability mass is apportioned according to the per-bit hard decisions; the resulting block posteriors are marginalized to produce each bit’s APP LLR, from which the extrinsic message is obtained by subtraction and damping.","core_discovery":"When SOGRAND’s soft-output formula is restricted to single-parity-check codes, the resulting check-node update (either the even-parity version that needs the Gallager parity probability Ψ or the non-even version that simply doubles the list size) yields extrinsic LLRs whose iterative decoding performance on 5G LDPC codes equals or exceeds that of the classical sum-product algorithm and of normalized min-sum, provided the list contains roughly ten noise patterns and a mild global damping factor α ≈ 0.9 is applied.","pith_inferences":["If the residual-mass term in the SOGRAND formulae is replaced by a more accurate estimate of the unqueried probability, the required list size may drop below eight, further simplifying the hardware.","Because SOGRAND already produces block-wise soft output, the same check-node engine can be reused without redesign for hybrid decoding of product codes that mix LDPC and other component codes.","The method opens a path to joint multi-check-node updates that score a single noise list against several parity checks at once, potentially reducing message-passing iterations."],"forward_implications":["LDPC decoders can replace the classical check-node arithmetic with a short, parallel table-driven SOGRAND update without loss of error-rate performance.","Because the update re-uses existing ORBGRAND pattern generators already fabricated for soft GRAND chips, the same silicon can serve both universal soft decoding and LDPC check-node processing.","The even and non-even variants give designers a concrete trade-off: compute the Gallager parity probability once and use L patterns, or skip that probability and double the list size.","The same specialization technique can be applied to other short even component codes that appear in product, staircase or OFEC constructions."],"fun_headline_variants":["SOGRAND SPC check-node updates match SPA on 5G LDPC codes","Noise-list SOGRAND CNs rival sum-product and min-sum for LDPC","SOGRAND SPC formulas yield hardware-friendly LDPC check rules","Short noise patterns in SOGRAND equal SPA decoding of LDPCs","Even-parity SOGRAND CN update matches Gallager SPA performance"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That a fixed table of only the seven or eight least-reliable bit flips, damped by a single global factor, is enough to keep the quality of the extrinsic messages equal to exact sum-product across the practical SNR range.","fun_headline_variants_meta":{"raw":{"variants":["SOGRAND SPC check-node updates match SPA on 5G LDPC codes","Noise-list SOGRAND CNs rival sum-product and min-sum for LDPC","SOGRAND SPC formulas yield hardware-friendly LDPC check rules","Short noise patterns in SOGRAND equal SPA decoding of LDPCs","Even-parity SOGRAND CN update matches Gallager SPA performance"]},"model":"grok-4.5","effort":"low","cost_usd":0.005418,"raw_usage":{"total_tokens":1433,"prompt_tokens":699,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":54180000,"prompt_tokens_details":{"text_tokens":699,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":649,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":699,"tokens_out":85,"duration_ms":4979,"temperature":1.0,"reasoning_tokens":649,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T22:03:08.176392+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same 5G NR LDPC simulations (n=256 rate-1/2 and n=1024 rate-2/3) with L=10 and α=0.9; if either SOGRAND rule produces a clear BLER or BER gap of more than a few tenths of a decibel relative to SPA or normalized min-sum at any operating point of interest, the central performance claim is refuted.","supporting_citations":[],"review_version":1}