{"id":"ab46d799-bb86-4066-becb-2f57ee8d2e7d","arxiv_id":"2603.11021","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Leech-lattice vector quantization with codebook-free indexing and shape-gain search reaches state-of-the-art ~2-bit post-training LLM compression and reduces reliance on Hadamard rotations.","lead":"This paper builds a practical 24-dimensional vector quantizer from the Leech lattice so LLM weights can be stored near 2 bits each without a huge codebook table. It reports better perplexity and task scores than Quip#, QTIP, and related methods on several open models.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"SOTA claim rests on pipeline-matched recovery that may not isolate the Leech codebook from residual VQ / global FT differences.","rationale":"The reader's weakest_assumption correctly identifies the soft spot: comparison fairness under local Hessian + light scale FT, not an internal contradiction in the Golay-based search/indexing or the geometric claims. The paper is a solid systems+geometry contribution (indexing, multi-shell angular search, parallel dequantizer, shape–gain + Spherical GPTQ reinterpretation) and own-pipeline tables consistently favor LLVQ, including a notable Hadamard-free regime. That is enough for CONDITIONAL rather than REJECT. The concern does not invent fraud or demand consensus; it asks for the one control that would make the strongest claim (codebook-driven SOTA) land cleanly. No formal verification or released code is present, so reproducibility remains a secondary risk, but the primary load-bearing issue is attribution of the LLM gaps. Verdict stays CONDITIONAL: accept-shaped if the matched residual-free re-run preserves the gaps, otherwise the claim should be narrowed to \"strong Leech PTQ under this pipeline.\" Agreement with the reader is full on the weakest assumption.","tokens_in":20971,"tokens_out":669,"duration_ms":6900,"concrete_test":"Re-run the exact own-pipeline of Table 3 / Table 6 on Llama-2 7B and Llama-3 8B for Quip#/E8P and QTIP (or a residual-free E8P product) under identical DCLM-edu Hessians, no residual VQ, and the same shared-scale-only FT budget as LLVQ; if Wiki/MMLU/CSR gaps shrink below the reported margins (or reverse under matched residual), the SOTA attribution to Leech weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (SOTA ~2 BPW PTQ, including strong Hadamard-free results) attributes gains primarily to multi-shell Leech angular codes plus indexing. The load-bearing condition is that the authors' layer-local GPTQ-style Hessian corrections on a fixed ~6,100-sequence DCLM-edu set, plus optional shared input-scale fine-tuning only (§5.1–5.4, Alg. 1, App. D.2), constitute a fair and sufficient recovery procedure so that gaps vs Quip#/E8P, QTIP, AQLM, and PVQ can be credited to the codebook rather than to residual VQ, inter-layer / end-to-end fine-tuning, or unstated pipeline details. Table 3 is own-pipeline and favors LLVQ, but Table 5 mixes literature numbers with explicit caveats; QTIP/AQLM/Quip# often use residual constructions and stronger recovery. Without matched residual-free, same-calibration, same-FT ablations, the attribution (and thus the SOTA headline) is not fully secured. Gaussian SQNR (Fig. 1, Table 4) and angular-uniformity (App. E–F) support the geometry but do not close the LLM comparison gap.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Leech Lattice Vector Quantization (LLVQ) for post-training compression of LLM weights. Building on the Adoul–Barth nearest-neighbour procedure for the Leech lattice (via the extended Golay code), the authors add (i) a hierarchical bijective indexing scheme that maps lattice points to bitstrings without materializing a codebook, (ii) multi-shell angular search supporting shape–gain quantization over unions of shells, and (iii) a fully parallelizable dequantization kernel based on integer arithmetic. They further reinterpret the scale-correction step of shape–gain under GPTQ-style Hessian updates as a retraction onto a product of spheres (“Spherical GPTQ”). Empirically, LLVQ reports the highest SQNR retention among compared lattice/scalar methods on unit Gaussians at 2 bits/dim, and state-of-the-art ~2 BPW PTQ results (WikiText perplexity, MMLU, CSR) on Llama-2/3, Ministral-3 and Qwen-3 models under a unified pipeline, with reduced sensitivity to Hadamard rotations.","tokens_in":21350,"tokens_out":1250,"duration_ms":22902,"significance":"If the practical claims hold, the work is a clear advance for theoretically grounded LLM compression: it makes the highest-dimensional lattice with proven optimal packing (Λ24) usable for codebook-free VQ at LLM scale, supplies the missing indexing and multi-shell angular machinery, and demonstrates that low angular distortion can weaken the need for online Hadamard preprocessing. The Gaussian SQNR/retention numbers (Fig. 1, Table 4) and the own-pipeline multi-model PTQ tables (Table 3, Table 6) are concrete, falsifiable evidence that high-dimensional lattice geometry transfers beyond toy sources. The geometric reading of shape–gain + Hessian correction as Spherical GPTQ is a useful conceptual contribution. These strengths justify publication once comparison and efficiency claims are tightened.","major_comments":[{"comment":"§5.4 and Table 5: The headline SOTA claim versus QTIP, AQLM and Quip# literature numbers is not fully isolated to the Leech codebook. Table 3 (own pipeline) cleanly favors LLVQ, but Table 5 mixes residual constructions, inter-layer/end-to-end fine-tuning, calibration sets and metrics; the manuscript already notes heterogeneity yet still asserts consistent outperformance. Either add residual-free, same-calibration, same-FT ablations against E8P/QTIP-style baselines, or soften the literature SOTA wording to “competitive / superior under matched local GPTQ-style recovery.” Attribution of gains primarily to multi-shell Leech angular codes currently rests on the weakest assumption of the paper.","section":null},{"comment":"§3.3 and contribution (iii): The dequantizer is described as “fully parallelisable” and “maps naturally to GPU execution,” yet the manuscript reports no wall-clock latency, throughput, or memory-bandwidth numbers for encode/decode versus E8P, QTIP or product codes. For a method whose selling point is efficient, codebook-free high-dimensional VQ, at least one controlled microbenchmark (or FLOPs/memory analysis of the integer hierarchy) is load-bearing; without it the efficiency claim remains untested.","section":null},{"comment":"§5.1–5.3 and free parameters: Shell cutoff m (or ball radius) and shape/gain bit allocation (App. F recommends ~1 gain bit at 2 BPW) are free design choices. Sensitivity of LLM perplexity/MMLU to these choices is not reported, only Gaussian sweeps. A short ablation on one model (e.g., Llama-2 7B) varying m and gain bits would confirm that the reported operating point is robust rather than over-tuned to the calibration set.","section":null}],"minor_comments":[{"comment":"Abstract and §1: “outperforming recent methods such as Quip#, QTIP, and PVQ” should be qualified by “under our unified PTQ pipeline” until Table 5 caveats are resolved.","section":null},{"comment":"Table 1: the last two rows (m=13,19) appear truncated/misaligned in the source; verify cumulative counts and bits/dim.","section":null},{"comment":"Eq. (13) and §3.1: the Euclidean–angular equivalence for fixed-norm shells is standard; a one-line pointer that multi-shell Euclidean scoring uses the full ||x−v||^{2} (not just the inner product) would avoid ambiguity for implementers.","section":null},{"comment":"App. D.2: the generalization of GPTQ/LDLQ to vector blocks is useful; explicitly state whether the Cholesky is computed once per layer or refreshed after each group, and whether the same schedule is used for all baselines in Table 3.","section":null},{"comment":"Typos / notation: “V ector” spacing artifacts, “Leven/Lodd” vs “L_even/L_odd”, and occasional “jii” numbering in §3. Clean for camera-ready.","section":null},{"comment":"Fig. 1 and Table 4: report sample size n used for the empirical MSE/SQNR estimates so that retention percentages can be reproduced.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The core algorithmic contribution (indexing + multi-shell angular search on Λ24) is solid and under-sold relative to the SOTA marketing. I would accept after the comparison wording and a minimal efficiency/sensitivity check; a hard reject would be disproportionate. Fit for a top ML venue is good if the authors keep the geometric story and do not over-claim literature SOTA."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is that they turned the Leech lattice into a practical, codebook-free 24-D quantizer for LLM weights: bijective hierarchical indexing over shells/classes/Golay placements, multi-shell angular search for shape-gain, and a parallel dequant sketch. That is real systems work on top of Adoul–Barth and classical lattice geometry, not a rehash of E8 product codes.\n\nWhat they do well: Gaussian SQNR/retention (Fig. 1, Table 4) and the shell-union vs single-shell angular study (App. E–F) cleanly support the geometry claim—multi-shell Leech shape-gain sits closer to the Shannon line than E8P/Quip# under their setup. Own-pipeline PTQ (Table 3) consistently favors LLVQ over E8P at 2 BPW across Llama-2/3, Ministral-3, and Qwen-3, including a strong Hadamard-free shape-gain regime (Table 6). The Spherical-GPTQ framing (scale correction as retraction onto product of spheres) is a clean geometric reading of something people already do. Citations to packing theory, Golay constructions, Quip#, QTIP, PVQ, and GPTQ are appropriate; self-cites to prior PVQ/GPTVQ are minor and on-topic.\n\nSoft spots, in proportion: the SOTA headline against literature (Table 5) mixes pipelines, residual constructions, and recovery strength, which the authors themselves flag. Attribution of gains purely to the Leech codebook is only fully clean inside their matched local-Hessian + optional shared-scale FT setup; without residual-free same-calibration ablations against QTIP/AQLM-style recovery, the cross-paper claim is softer than the abstract. No code, no multi-seed error bars, no dequant latency numbers—standard for a preprint, but it caps confidence. Free parameters (shell cutoff, gain bits, 6100-seq DCLM-edu, light FT) are disclosed and not hidden.\n\nThis is for people who care about ultra-low-bit PTQ and lattice VQ engineering. The central argument holds under the authors’ controlled pipeline; the geometry is not circular. I would send it to peer review. Engage if you work on 2-bit deployment or high-D codes; skim the appendices if you only need the Gaussian/angular evidence.","headline":"Solid engineering of multi-shell Leech VQ into a usable LLM PTQ method; own-pipeline wins look real, SOTA vs literature is only partly secured.","tokens_in":21999,"tokens_out":589,"would_cite":true,"duration_ms":6445,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"The Leech lattice in 24 dimensions yields practical codebook-free vector quantization that beats prior 2-bit LLM compression methods.","keywords":["Leech lattice","vector quantization","LLM compression","shape-gain quantization","post-training quantization","Golay code","spherical codes","lattice packing"],"falsifier":"Re-run the identical calibration, Hessian-correction and (optional) scale-only fine-tuning pipeline on Llama-2/3, Ministral-3 and Qwen-3 models while swapping only the codebook (Leech multi-shell versus E8P/Quip#, QTIP trellis, AQLM, PVQ); if LLVQ no longer yields lower WikiText perplexity and higher MMLU/CSR at matched 2 bits per weight, the central claim fails.","tokens_in":21851,"feed_emoji":"📐","tokens_out":1018,"duration_ms":11661,"temperature":0.7,"pith_summary":"Scalar quantization of large language models is information-theoretically limited; encoding parameters in blocks can do better, but only if the codebook never has to be stored or searched explicitly. This paper turns the 24-dimensional Leech lattice—the highest-dimensional lattice known to achieve optimal sphere packing—into a usable quantizer called LLVQ. It extends a Golay-code search algorithm so that lattice points can be indexed and reconstructed from bitstrings without materializing any table, supports angular nearest-neighbour search over unions of lattice shells for shape-gain coding, and supplies a fully parallel dequantization kernel. On both ideal Gaussian sources and real LLMs the resulting codes retain more signal per bit than E8, trellis, and pyramid alternatives, reach state-of-the-art 2-bit post-training quantization accuracy, and reduce the need for expensive Hadamard rotations. The result is a concrete demonstration that high-dimensional lattices can deliver scalable, theoretically grounded model compression.","feed_headline":"Leech lattice hits SOTA 2-bit LLM compression","feed_subtitle":"Codebook-free 24-D packing beats E8, trellis and pyramid methods and needs fewer rotations","key_machinery":"Leech Lattice Vector Quantization (LLVQ): an extension of the Adoul–Barth Golay-code nearest-neighbour procedure that adds hierarchical indexing of shells/classes/symmetries, angular search over unions of shells for shape-gain coding, and a parallel integer-arithmetic dequantizer, together with a geometric view of scale-corrected GPTQ as spherical retraction onto product-of-spheres directions.","core_discovery":"A practical, codebook-free vector quantizer built on the Leech lattice (LLVQ) achieves state-of-the-art roughly 2-bit-per-weight post-training quantization of large language models, consistently outperforming E8-based, trellis-based and pyramid-based methods on perplexity and downstream tasks while remaining fully indexable and parallelizable.","pith_inferences":["Because the dequantizer is pure integer arithmetic and embarrassingly parallel, the same kernel could be fused into matrix-multiplication units, making 24-dimensional VQ competitive with scalar integer kernels on hardware.","If the angular-uniformity advantage of shell unions persists at higher rates, the same construction may close the gap to the Shannon bound for other high-dimensional sources beyond LLM weights.","The geometric retraction view of scale correction suggests that any lattice whose shells form good spherical codes can be paired with GPTQ-style updates without inventing a new residual scheme."],"forward_implications":["At 2 bits per weight, multi-shell Leech codes can replace residual-vector-quantization ladders used by lower-dimensional lattices to reach usable bit-rates.","Low angular distortion of union-of-shells Leech codes reduces or eliminates the need for online Hadamard rotations, cutting inference latency when rotations cannot be fused.","Shape-gain quantization with Leech spherical codes plus closed-form scale correction acts as a Spherical GPTQ that primarily corrects directions, simplifying the recovery stage.","The same indexable multi-shell construction immediately supplies a continuous family of bit-widths without redesigning the codebook.","High-dimensional lattice packing density and shell uniformity translate from ideal Gaussian sources into measurable gains on real LLM weight matrices."],"fun_headline_variants":["Leech lattice VQ reaches SOTA 2-bit LLM compression codebook-free","24-D Leech packing beats E8 trellis and pyramid for LLM PTQ","Indexable Leech LLVQ tops Quip# QTIP PVQ on LLM perplexity","Codebook-free Leech lattice enables strong Hadamard-free LLM PTQ","High-dim Leech VQ sets new bar for efficient 2-bit LLM compression"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Accuracy differences versus earlier lattice and trellis methods can be attributed mainly to the Leech codebook itself once the same layer-local Hessian corrections and optional shared input-scale fine-tuning are applied to every method.","fun_headline_variants_meta":{"raw":{"variants":["Leech lattice VQ reaches SOTA 2-bit LLM compression codebook-free","24-D Leech packing beats E8 trellis and pyramid for LLM PTQ","Indexable Leech LLVQ tops Quip# QTIP PVQ on LLM perplexity","Codebook-free Leech lattice enables strong Hadamard-free LLM PTQ","High-dim Leech VQ sets new bar for efficient 2-bit LLM compression"]},"model":"grok-4.5","effort":"low","cost_usd":0.006594,"raw_usage":{"total_tokens":1681,"prompt_tokens":825,"num_sources_used":0,"completion_tokens":114,"cost_in_usd_ticks":65940000,"prompt_tokens_details":{"text_tokens":825,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":742,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":825,"tokens_out":114,"duration_ms":6301,"temperature":1.0,"reasoning_tokens":742,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T23:09:40.559962+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the identical calibration, Hessian-correction and (optional) scale-only fine-tuning pipeline on Llama-2/3, Ministral-3 and Qwen-3 models while swapping only the codebook (Leech multi-shell versus E8P/Quip#, QTIP trellis, AQLM, PVQ); if LLVQ no longer yields lower WikiText perplexity and higher MMLU/CSR at matched 2 bits per weight, the central claim fails.","supporting_citations":[],"review_version":1}