{"id":"7b998f9a-c2a4-4dd6-bfba-40921245f471","arxiv_id":"2505.05530","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured survey that classifies 179 recent low-bit quantization papers into 8 categories and 24 sub-categories, with a curated reference list and future directions.","lead":"This paper is a survey of low-bit quantization, a technique that shrinks deep neural networks by storing weights and activations as small integers instead of full-precision numbers. It maps 179 recent papers into eight method families and a curated list, which helps newcomers find their way in a crowded field.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stated scope excludes extreme quantization (1-bit/1.58-bit), yet Section 3 includes binary and 1.99-bit methods, so the survey's coverage boundary is internally inconsistent.","rationale":"The reader's conditional verdict rests on the survey's completeness and faithful organization of 179 papers. The reader identified lack of selection criteria, no inter-annotator agreement, and no quantitative comparison as the weakest assumptions. My stress-test pass found a more concrete, internally checkable problem: the stated scope explicitly excludes 1-bit and 1.58-bit extreme quantization, yet the surveyed corpus contains binary and 1.99-bit methods that are described and categorized as part of the taxonomy. This is not a matter of disagreeing with the field's boundaries; it is a direct contradiction between the Section 1 scope statement and the Section 3 contents. It bears on the central claim because a survey's usefulness depends on a coherent inclusion boundary. The concern is fixable and does not overturn the paper's overall value, so the reader's CONDITIONAL verdict remains appropriate. I would not move the verdict to ACCEPT, because the inconsistency should be resolved before the coverage claim is taken at face value, nor to REJECT, because the taxonomy and discussion are still broadly informative and the error is correctable. Agreement is partial because the reader's weakest assumption was about representativeness and faithful summarization, while my concern is a specific internal violation of the paper's own scope; both point toward the same underlying fragility in the completeness claim.","tokens_in":46050,"tokens_out":4198,"duration_ms":44417,"concrete_test":"Enumerate every surveyed method in Fig. 4 and Section 3 and record its reported weight or activation bit-width. Flag any method with bit-width <= 2, such as BiDM's binary quantizer, BitsFusion's 1.99-bit weights, and QuIP/QuIP#'s 2-bit lattice codebooks. If at least one such method is present, the Section 1 exclusion of 1-bit and 1.58-bit extreme quantization is violated, and the coverage boundary must be restated or the entries must be reconciled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The survey's central value is a comprehensive and correctly delimited map of low-bit quantization methods. Section 1 explicitly states: \"we omit extreme quantization techniques (e.g., 1-bit or 1.58-bit quantization) as they involve substantially different methodologies.\" Section 3, however, includes methods that fall inside that excluded regime. Section 3.7.5 describes BiDM [159] as employing \"a dynamical binary quantizer\" (1-bit), and the reference list gives BitsFusion [161] the title \"1.99 bits weight quantization of diffusion model.\" Section 3.4.3 covers QuIP [104] and QuIP# [105], both 2-bit lattice-codebook methods. These are not peripheral mentions; they are presented as part of the surveyed taxonomy and discussed as representative techniques. The contradiction is internal to the manuscript, not a matter of disagreeing with an external consensus. A reader cannot infer from the stated scope which methods belong in the survey and which do not: some extreme-low-bit methods are included while others are declared out of scope. For a paper whose claim is to have \"surveyed 179 quantization papers and classified them\" into a coherent taxonomy, this weakens the most load-bearing part of the argument, namely that the map is comprehensive and consistently bounded. The fix is straightforward, either relax the exclusion statement to cover only the omitted classes the authors actually intended, or remove or explicitly mark the included extreme-bit methods, but as published the coverage claim is not self-consistent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript surveys low-bit model quantization for deep neural networks over roughly the past five years. Section 2 introduces the quantization formalism, basic quantizer designs, and a foundational taxonomy; Section 3 then proposes an organizing scheme of eight main categories and twenty-four sub-categories and assigns 179 papers to them (Fig. 4). Section 4 lists future research directions, and an accompanying curated repository is advertised. The abstract promises that state-of-the-art methods are discussed and compared; the main text provides qualitative discussion of each category, while the supplementary file discloses that a quantitative performance comparison was attempted but not completed.","tokens_in":46247,"tokens_out":5652,"duration_ms":50623,"significance":"If the coverage boundary and the comparison claim are made consistent, the survey would be a useful reference map for the low-bit quantization community: it aggregates 179 papers, organizes them by technique rather than by task, includes recently active areas such as diffusion-model quantization and data-free quantization, and provides an accompanying curated list. The paper is particularly valuable for newcomers who need to locate methodological families. I also credit the authors for explicitly stating in the supplementary material that the attempted quantitative comparison failed rather than hiding the limitation; however, that disclosure is not reflected in the main-text claims.","major_comments":[{"comment":"The stated scope exclusion of extreme quantization is contradicted by the included methods. The Introduction says \"we omit extreme quantization techniques (e.g., 1-bit or 1.58-bit quantization) as they involve substantially different methodologies,\" yet Section 3.7.3 presents BiDM [159] as employing \"a dynamical binary quantizer\" (1-bit), Section 3.7.4 includes BitsFusion [161], whose title is \"1.99 bits weight quantization of diffusion model,\" and Section 3.4.3 includes QuIP [104] and QuIP# [105], which are 2-bit lattice-codebook methods. These are not peripheral mentions: they are described as representative techniques in their subsections. A reader therefore cannot infer from the stated scope which methods belong in the survey. The fix is to relax the exclusion statement to match the actual coverage or to move or explicitly mark the extreme-bit methods as out-of-scope but discussed for contrast.","section":"Section 1 and Section 3 (3.4.3, 3.7.3, 3.7.4)"},{"comment":"The abstract claims that the paper \"discuss[es] and compare[s] the state-of-the-art quantization methods,\" and Section 3 promises a \"comprehensive analysis and discussion,\" but Supplementary Section 1 states: \"We tried to provide the performance comparison of different quantization methods on these benchmarks, but failed. This is because the models, datasets, and quantification schemes adopted by the recent methods are all different.\" No accuracy, latency, or memory comparison table appears in the main text. The comparison actually delivered is qualitative only, so the abstract should either remove the word \"compare\" or the main text should contain a limitations paragraph explicitly stating that no quantitative comparison is provided.","section":"Abstract and Supplementary Section 1"},{"comment":"No selection or classification methodology is reported for the 179 surveyed papers. The text does not state which databases were searched, which keywords or time window were used, what inclusion/exclusion criteria were applied, or how the eight-way assignment was performed and validated. Because the paper's central contribution is a comprehensive and correctly organized map of the field, the absence of this information makes the completeness claim unverifiable. Please add a methodology paragraph describing the paper collection and classification procedure, or rescope the claim to \"a curated selection\" rather than a systematic survey.","section":"Section 3 opening and Fig. 4"}],"minor_comments":[{"comment":"The sentence \"Recent works [12], [43] have indicated that traditional loss functions, such as MSE and Exponential Moving Average (EMA), such as MSE and Cross-Entropy (CE), may not be sufficient\" contains a duplicated \"such as MSE\"; it should be rewritten as a single list.","section":"Section 3.2.2"},{"comment":"Reference [143] is truncated to \"70\" and should be completed, and Supplementary Table 1 contains \"Mixral\" (should be \"Mixtral\"), \"VICUNA-V1.5 []\" with an empty citation, and \"SQAI\" where ScienceQA appears to be meant.","section":"References and Supplementary Table 1"},{"comment":"OWQ appears twice, as [50] and as [193], and Q-BERT appears as both [53] and [194]; these duplicate entries for the same methods should be consolidated into single citations.","section":"Section 3.3.1 and reference list"},{"comment":"The typesetting of the bi-exponent representation \"2en|eo\" is difficult to parse; please rewrite with proper superscripts or parentheses so the shared exponents are unambiguous.","section":"Section 3.6.3, Eq. (15)"},{"comment":"The heading in the text is \"Other\" while Fig. 4 labels the category \"Others\"; the terminology should be made consistent.","section":"Section 3.8 and Fig. 4"},{"comment":"The phrase \"extreme quantization\" appears in Section 2.3b before any definition and in a context that is excluded from the survey's scope; a parenthetical definition or a cross-reference to the scope statement in the Introduction would avoid confusion.","section":"Section 2.3b and Section 1"}],"recommendation":"major_revision","confidential_remarks":"This survey is within the scope of the journal and covers a fast-moving area. I would ask the editor to consider the balance of self-citations: roughly two dozen of the 179 reviewed papers involve the authors' own groups (e.g., QuantSR, DSG, BiDM, MPQ-DM, PassionSR, and Q-DM). The descriptions appear faithful to the published abstracts, and self-inclusion is not by itself a problem, but the concentration is high for a survey and the final version might acknowledge the authors' own contributions explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This survey is worth a look if you need a structured entry point into 2020-2025 low-bit quantization. The eight-category taxonomy with 24 sub-categories, the coverage of LLM/ViT/diffusion, and the maintained GitHub list are genuinely useful. The basic concepts section is fine. The paper does what a survey should at the organizational level.\n\nThe soft spots are real but mostly fixable. First and most important, the scope statement is internally inconsistent. Section 1 says extreme quantization (1-bit or 1.58-bit) is omitted because it involves substantially different methodologies, but Section 3 includes BiDM, described as using a dynamical binary quantizer, and BitsFusion, which is 1.99-bit weight quantization. QuIP and QuIP# (2-bit lattice methods) are also covered as part of the taxonomy. If the intended scope is 'low-bit but not ultra-low-bit,' the exclusion sentence needs to be rewritten or those papers explicitly marked as boundary cases. As is, a reader cannot infer what belongs inside the survey.\n\nSecond, there are no inclusion criteria or inter-annotator agreement for the 179 surveyed papers, so the claim of representativeness rests on the authors' judgment. The authors also admit in the supplementary that quantitative comparison across methods failed because setups differ. That is an honest admission, but it means the abstract's 'discuss and compare' is met only qualitatively.\n\nThird, proofreading: reference [143] is truncated, there are empty citation slots, and the stated 24 sub-categories do not reconcile cleanly with the figure (3.7.1 is missing from the figure while the text has it). These are minor but visible in a survey whose value is organizational.\n\nOn self-citation: roughly two dozen of 179 papers involve the authors. The summaries track the published abstracts, so I don't see a distortion, but with no selection criteria, a reader cannot fully rule out emphasis bias. It is a point to check, not a fatal one.\n\nI would send this to peer review. A good set of reviewers can fix the scope statement, ask for selection criteria, and clean the reference list. The taxonomy itself is a service to the field. Who benefits: a newcomer or a researcher from a neighboring area who needs a map of the territory. I wouldn't cite it in my own work until the boundary issue is resolved, but I might point people to the GitHub list.","headline":"A genuinely useful survey map of 2020-2025 low-bit quantization, softened by an inconsistent scope boundary and the absence of quantitative comparison.","tokens_in":46887,"tokens_out":2615,"would_cite":false,"duration_ms":27088,"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":"The paper claims that 179 papers from 2020–2025 low-bit quantization can be organized into eight methodological families and 24 subfamilies.","keywords":["model quantization","low-bit quantization","post-training quantization","quantization-aware training","mixed precision quantization","data-free quantization","diffusion model quantization","neural network compression"],"falsifier":"Take the 179 cited papers, strip away the survey's own classifications, and have an independent reader assign each paper to one of the eight families using only its abstract and method description; if agreement is no better than chance, or if major methods clearly straddle or fall outside all eight families, the taxonomy is not a stable description of the field. A second check is to count 2020–2025 publications: if extreme 1-bit and 1.58-bit work, which the survey explicitly excludes, dominates the period, then the survey's scope claim omits a major branch of the literature.","tokens_in":45789,"feed_emoji":"🗂️","tokens_out":5638,"duration_ms":57012,"temperature":0.7,"pith_summary":"This survey tries to establish an organizing map of the last five years of low-bit neural-network quantization, covering 179 papers across 8 main families and 24 subfamilies. It argues that nearly all current methods reduce to choices about how to set quantization scales and zero-points, what loss or metric guides the conversion, which layers get which bit-width, how data distributions are reshaped, and which number format is used. If the map is right, a practitioner can locate any new quantization method within a technical family and see which open problems remain. The authors also state plainly in the supplementary file that they could not provide a quantitative comparison across the surveyed methods, because models, datasets, and quantization schemes differ too much, so the survey's value is the organization rather than a leaderboard.","feed_headline":"179 quantization papers, sorted into eight technique families","feed_subtitle":"A 2025 survey gives newcomers a route from bit-widths to diffusion-model quantization.","key_machinery":"The organizing instrument is the taxonomy itself: eight main categories plus 24 subcategories, built on top of the formal quantization operator $X_{\\mathrm{int}} = \\mathrm{Clamp}(\\mathrm{Round}(X_{\\mathrm{FP}}/s)+z, n, p)$ with dequantization $\\hat{X}=s(X_{\\mathrm{int}}-z)$. That operator supplies the survey's vocabulary—bit-width $b$, scale $s$, zero-point $z$—and the taxonomy groups methods by which of these knobs they turn, turning a scattered literature into a decision tree for a practitioner.","core_discovery":"The paper's central discovery is taxonomic: it claims that the recent quantization literature divides along eight recognizable methodological fault lines—scale and zero-point optimization, metrics and training mechanisms, mixed precision, redistribution of weights and activations, data-free quantization, advanced numeric formats, diffusion-model-specific methods, and a residual 'other' bucket—with 24 finer subcategories. It formalizes quantization as mapping floating-point tensors through clamp, round, scale, and zero-point operations, and presents post-training quantization (PTQ) and quantization-aware training (QAT) as a spectrum rather than a strict dichotomy. The paper also reports that the field has moved beyond plain integer formats into float formats, fixed-point formats, learned rotations, and adaptive rounding, while deliberately excluding 1-bit and 1.58-bit methods as a methodologically separate branch.","pith_inferences":["If the taxonomy is accurate, a useful next step is to use it as a checklist for designing new methods: a paper that combines scale optimization, rotation-based outlier removal, and adaptive bit allocation would span three families, which the survey itself permits since some papers appear in multiple categories.","The absence of comparative benchmarks also suggests an opportunity: building a standardized quantization benchmark with fixed models, calibration sets, and bit-width budgets would let future surveys replace prose organization with reproducible measurement.","The eight-way split may over-weight recent LLM and diffusion-model work, since the field's center of gravity moved there in 2023–2025; a reader should expect older CNN-only methods to be concentrated in earlier families such as 'better s and z' and 'redistribution.'"],"forward_implications":["A newcomer can use the taxonomy to locate any major 2020–2025 quantization method and identify its core technique without reading the full literature.","Because PTQ and QAT are presented as a continuum, the boundary between calibration-only methods and retraining-based methods is expected to keep blurring.","The four future directions named by the paper—multimodal deployment, combining quantization with pruning and low-rank compression, software-hardware co-optimization, and task-specific quantization—are where the paper expects the next progress.","The survey implies that extreme 1-bit and 1.58-bit quantization is a separate research vein, not a subcase of the eight families, so its conclusions should not be read as covering that line of work."],"supporting_citations":[{"why":"Supplies the quantization problem setup and the motivation of reducing memory and computation cost.","marker":"[9]"},{"why":"Supplies the formal clamp, round, scale, and zero-point operator used throughout the survey.","marker":"[10]"},{"why":"Introduces the straight-through estimator that defines the QAT half of the PTQ/QAT spectrum.","marker":"[11]"},{"why":"Exemplifies the 'better s and z' category with dual uniform quantization for vision transformers.","marker":"[12]"},{"why":"Exemplifies GPTQ-style post-training weight quantization for large language models.","marker":"[45]"},{"why":"Exemplifies outlier-processing redistribution via activation-weight scaling for LLMs.","marker":"[99]"},{"why":"Supplies adaptive rounding, the basis of a redistribution subcategory.","marker":"[117]"},{"why":"Exemplifies data-free quantization through weight equalization and bias correction.","marker":"[121]"},{"why":"Exemplifies the advanced-format category with NormalFloat for low-bit fine-tuning.","marker":"[148]"},{"why":"Exemplifies diffusion-model quantization with step-aware calibration and activation-distribution handling.","marker":"[153]"}],"fun_headline_variants":["Eight families map the quantization landscape","Low-bit quantization taxonomy: 179 methods","From PTQ to QAT: 8 quantization approaches","Surveying 179 quantization papers, 24 subcategories","Quantization demystified: 8 technique families"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole map is only as good as the selection and classification of the 179 papers; if the papers were chosen with a hidden bias or assigned to the wrong families, the survey would mislead rather than orient.","fun_headline_variants_meta":{"raw":{"variants":["Eight families map the quantization landscape","Low-bit quantization taxonomy: 179 methods","From PTQ to QAT: 8 quantization approaches","Surveying 179 quantization papers, 24 subcategories","Quantization demystified: 8 technique families"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1711,"prompt_tokens":898,"completion_tokens":813,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":741}},"tokens_in":514,"tokens_out":813,"duration_ms":7854,"temperature":1.0,"reasoning_tokens":741,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:09:28.361308+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the 179 cited papers, strip away the survey's own classifications, and have an independent reader assign each paper to one of the eight families using only its abstract and method description; if agreement is no better than chance, or if major methods clearly straddle or fall outside all eight families, the taxonomy is not a stable description of the field. A second check is to count 2020–2025 publications: if extreme 1-bit and 1.58-bit work, which the survey explicitly excludes, dominates the period, then the survey's scope claim omits a major branch of the literature.","supporting_citations":[],"review_version":1}