acceptodds
Under review as a conference paper at ICLR 2027

AC-Scoreq: Amplitude-Coupled Scalable Complex 2-bit Residual Quantization for Efficient Vector Retrieval

Abstract

Efficient vector retrieval requires compact representations that balance storage, ranking quality, and search time. Successive low-bit phase-code applications compensate quantization residuals, but their shared block scales leave local magnitude differences that further phase refinement cannot directly represent. We propose AC-Scoreq, a fixed-codebook complex 2-bit residual quantizer that combines phase encoding, local amplitude adaptation, and cross-stage scale fitting. It applies a fixed four-direction phase codebook to successive residuals; the final application additionally selects a relative-amplitude shape for each pair of adjacent complex coordinates. The encoder alternates discrete-code refinement with joint fitting of the block scales, and each database vector is encoded independently without corpus-level codebook training or prior knowledge of queries. Varying the number of phase-code applications enables trade-offs among retrieval quality, search time, and compression ratio. Experimental results across multiple datasets and vector dimensionalities show that AC-Scoreq maintains competitive retrieval quality and low search latency while substantially reducing vector storage overhead. Its multi-application design further provides flexible accuracy compensation across compression ratios, demonstrating the method’s effectiveness, scalability, and practical deployment value.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Vector Quantization, Efficient Vector Retrieval, Embedding Compression, Fixed-Codebook Quantization, Amplitude-Coupled Quantization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7P6POpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes AC-Scoreq, a fixed-codebook residual quantization method for efficient vector retrieval. It combines coordinate-level 2-bit phase codes with pair-level 2-bit amplitude-shape codes in the final residual stage, using jointly fitted block scales to improve reconstruction. Database vectors are encoded independently, and queries undergo the same orthogonal transform, enabling direct similarity computation from compressed codes for inner product, cosine similarity, and L2 distance. Experiments on FIQA and DBPedia-Entity with three embedding models evaluate retrieval quality, storage, and full-corpus GPU latency.

Critical strengths
  1. The phase–amplitude–scale design is well motivated and cleanly separated. Amplitude coupling is integrated into the final reconstruction rather than applied as post-processing, and the structured scoring identity supports direct similarity computation from packed codes without a full-precision reranking vector.
  2. All codebooks are fixed analytically; database vectors are encoded independently with no corpus-level training or query knowledge required.
  3. The paper reports actual bytes per vector, explicitly frames T=3 as a nearby-budget comparison, averages over 50 seeds with confidence intervals, and includes equal-amplitude and same-budget extra-phase controls.
Critical weaknesses
  1. The specific shape codebook, pairwise coordinate grouping, and final-stage-only amplitude encoding are geometrically motivated but not systematically compared against alternative codebooks, grouping sizes, or stage allocations. Consequently, the benefits of amplitude coupling are more convincing than the necessity of the particular design.

  2. The theoretical analysis provides useful geometric intuition but does not fully characterize the effects of shared or quantized block scales, interactions across residual stages, or errors in inner-product and ranking computations. The connection between the theoretical analysis and downstream retrieval performance therefore remains partly empirical.

  3. Evaluation is limited to FIQA and DBPedia-Entity, without validation on standard large-scale approximate nearest-neighbor benchmarks. Although the paper includes useful ablations, some retrieval differences are small, and more detailed statistical comparisons and effect-size analyses would help establish the robustness of the reported improvements.

Questions for authors
  1. Can the authors provide quality–storage Pareto curves at closely matched budgets, with consistent accounting for scales and auxiliary information? This would clarify whether the observed gains arise from the representation itself or partly from differences in storage allocation.

  2. Have alternative shape codebooks, coordinate grouping sizes, or stage allocations been evaluated? In particular, why is amplitude-shape encoding restricted to adjacent complex-coordinate pairs in the final residual stage?

  3. How sensitive is the alternating scale-fitting and code-refinement procedure to initialization and the number of refinement sweeps?

  4. Do the comparisons against the same-budget extra-phase control show consistent retrieval improvements across embedding models and similarity metrics? Are there cases where lower reconstruction MSE does not translate into better retrieval quality?

6 comments
AuthorsOpenReview ↗

Response to Weaknesses 1–2: Amplitude Design, Scale Fitting, and Retrieval

Thank you for recognizing the method's structure and configurability. To address design choices, theoretical connections, and evaluation scope, we added comparisons of amplitude codebooks and grouping methods, sensitivity analyses of stage placement and initialization, ablations of joint scale fitting, paired query statistics, and results on standard ANN benchmarks and actual storage–quality–latency trade-offs.

(Weakness 1) Evidence for the Amplitude-Codebook Design

We added factorial experiments that separately control the zero shape and asymmetric shapes. As shown in Table 1, adding asymmetric amplitude patterns still substantially reduces reconstruction error when the zero shape is retained.

Table 1. Controlled comparison of amplitude-shape composition. FIQA, mean MSE over three random seeds; lower is better. Encoding fields and actual byte counts are identical within each model.

Shape settingModernBERTQwen-1.5BQwen-7B
Equal amplitude only0.062500032.119877e−59.108570e−6
Zero shape + equal amplitude0.060865332.064419e−58.870571e−6
Asymmetric + equal amplitude, without the zero shape0.049191141.669557e−57.173145e−6
Full AC0.047623011.616152e−56.943650e−6

We also evaluated narrower and wider shapes, four-coordinate grouping, nonadjacent pairing, and different stage placements. The complete results are presented in Tables 6–8 under Question 2 of this response. These controls support local amplitude adaptation as a design mechanism and treat the specific codebook and grouping choices as testable implementation choices.

(Weakness 2) Shared Scales, Quantization Error, and Retrieval

We added conditional analyses at three levels.

First, with the discrete codes fixed, let CC be the complex basis matrix containing the phase codes of each stage and the final amplitude shapes, and let ss be the vector of real-valued scales. Joint scale fitting satisfies

=\operatorname{Re}(C^{H}z), \qquad \eta=2B\lambda.$$ The off-diagonal terms explicitly describe correlations between stages. This solution minimizes the regularized objective with fixed codes; when the regularization term is nonzero, we do not directly equate this conclusion with minimizing unregularized MSE. Second, if the stored scales are $\widehat{s}$, the additional error introduced by scale quantization satisfies $$\|z-C\widehat{s}\|_2 \le \|z-Cs^{*}\|_2+ \|C\|_2\|\widehat{s}-s^{*}\|_2.$$ This connects scale quantization error to the actual stage basis matrix without requiring different stages to be orthogonal. Finally, let $e_i=\widehat{x}_i-x_i$. For IP retrieval, if the exact score margin satisfies $$q^{\top}(x_i-x_j)> \|q\|_2\|e_i-e_j\|_2,$$ then the relative order of documents $i$ and $j$ is preserved after reconstruction. This is a sufficient condition that depends on the query and score margin. We also added a control that disables joint scale fitting during both initialization and refinement. As shown in Table 2, restoring joint fitting improves exact-neighbor preservation in all nine settings. **Table 2. NN Recall@10 gains of full AC over disabled joint scale fitting. FIQA, three seeds; values are in percentage points, and higher is better.** | Model | IP | Cosine | L2 | |---|---:|---:|---:| | ModernBERT | **+6.816 [6.108, 7.536]** | **+1.590 [0.977, 2.207]** | **+2.459 [1.857, 3.061]** | | Qwen-1.5B | **+3.102 [2.567, 3.642]** | **+1.101 [0.628, 1.574]** | **+1.343 [0.859, 1.836]** | | Qwen-7B | **+1.651 [1.224, 2.094]** | **+0.756 [0.386, 1.137]** | **+1.070 [0.694, 1.456]** | Brackets contain adjusted paired query bootstrap intervals: the three seeds are first averaged within each query, followed by 100,000 resamples. The correction family comprises three models, three similarities, and four controls: zero + equal amplitude, independent 1-bit amplitudes, disabled joint fitting, and extra phase. This gives 36 comparisons with a nominal family confidence level of 95%. **The conditional analyses explain how errors affect scores and rankings, while the independent ablation validates the practical retrieval benefits of joint fitting.**
AuthorsOpenReview ↗

Response to Weakness 3 and Question 1: Benchmarks and Storage–Quality–Latency Trade-offs

(Weakness 3) Benchmark Coverage and Effect-Size Evaluation

We added angular-distance retrieval on NYTimes, with 290,000 database vectors and 10,000 queries. The results are shown in Table 3.

Table 3. Storage, distortion, and exact-neighbor recall on a standard ANN dataset.

DatasetMethodB/vectorMSE ↓NN Recall@10, % ↑
NYTimesAC, T=1T=1581.04693e−365.517
NYTimesAC, T=2T=21003.13131e−481.127
NYTimesAC, T=3T=31429.67977e−589.092

For text retrieval, we further use the official 648 FIQA and 400 DBPedia test queries for paired statistical analysis. The original AC and extra-phase control are evaluated with 50 and three seeds on the two datasets, respectively. The complete intervals are reported in Table 10 of this response.

(Question 1) Quality–Storage–Latency Trade-offs and Storage Accounting

We place the default stage counts and storage-sensitivity configurations in the same table to compare operating points using actual B/vector. T=1,2,3T=1,2,3 are independently encoded configurations; phase codes, final amplitude shapes, and scales jointly determine the payload.

As shown in Table 4, in addition to the stage count, scale precision and block size also change the trade-off between storage and retrieval quality.

Table 4. Actual storage–quality operating points on FIQA. Each cell reports "B/vector / IP NDCG@10." All AC configurations retain amplitude coupling in the final stage.

ConfigurationModernBERTQwen-1.5BQwen-7B
AC, T=1T=1170 / 0.458340338 / 0.530670786 / 0.615047
AC, T=2T=2292 / 0.498751580 / 0.5364941348 / 0.624395
AC, T=3T=3, default414 / 0.502031822 / 0.5390791910 / 0.627198
AC, T=3T=3, 4-bit scales378 / 0.501224750 / 0.5358551742 / 0.624320
AC, T=3T=3, block size 32378 / 0.499455750 / 0.5393461742 / 0.627430

To also show computational cost, Table 5 reports quality and latency produced by the same set of timing runs.

Table 5. Storage–quality–latency operating points corresponding to the same actual retrieval outputs on DBPedia/Qwen-7B. 400 official test queries, A800, document chunks of 50,000 vectors, top-1000; times are means over five runs after two warmup runs.

MethodB/vectorIP NDCG@10 ↑Cosine NDCG@10 ↑IP seconds ↓Cosine seconds ↓
FP32 reference143360.4707760.4707760.8534900.881974
PQ-16x8960.4674100.4713660.2016700.229450
PQ-8x17920.4722640.4702840.2324370.260153
AC, T=1T=17860.4691970.4698960.1980470.226133
AC, T=2T=213480.4703160.4707930.2018810.229622
AC, T=3T=319100.4724700.4721780.2084480.236383

These operating points reflect the method's main purpose: selecting a configuration according to storage, retrieval quality, and query cost, rather than solely pursuing the highest accuracy at a single compression ratio.

AuthorsOpenReview ↗

Response to Questions 2–3: Design Choices and Refinement Sensitivity

(Question 2) Alternative Shapes, Group Sizes, Pairing, and Stage Placement

We compared four types of design choices.

First, the narrower shapes use 30∘/60∘30^\circ/60^\circ, the default shapes use 22.5∘/67.5∘22.5^\circ/67.5^\circ, and the wider shapes use single-coordinate endpoints at 0∘/90∘0^\circ/90^\circ; all three retain the zero and equal-amplitude shapes.

As shown in Table 6, the default shapes achieve the lowest reconstruction error in this comparison, but the configuration with the highest NDCG is not the same for all models.

Table 6. Sensitivity to the amplitude-shape range. FIQA, seed 42; each cell reports "MSE ↓ / IP NDCG@10 ↑," with identical storage and optimization settings.

Shape rangeModernBERTQwen-1.5BQwen-7B
Narrower, 30∘/60∘30^\circ/60^\circ0.0493293 / 0.5004621.66214e−5 / 0.5385667.10733e−6 / 0.627135
Default, 22.5∘/67.5∘22.5^\circ/67.5^\circ0.0481763 / 0.5020311.62303e−5 / 0.5390796.93949e−6 / 0.627198
Endpoints, 0∘/90∘0^\circ/90^\circ0.0541217 / 0.5003231.82242e−5 / 0.5396707.79336e−6 / 0.627389

Second, under a consistent greedy initialization protocol, we compared placing AC at different residual stages, as well as pairwise shapes versus four-coordinate shapes. The four-coordinate control uses 16 fixed support patterns with 4 bits per group, while the pairwise configuration uses 2 bits per pair, keeping the total shape-field budget the same.

Table 7. Comparison of stage placement and grouping at the same shape budget. FIQA, seed 42, three refinement sweeps; each cell reports "MSE ↓ / IP NDCG@10 ↑."

SettingModernBERTQwen-1.5BQwen-7B
Pairwise AC at stage 10.0625284 / 0.4981802.10056e−5 / 0.5384868.98813e−6 / 0.625520
Pairwise AC at stage 20.0557714 / 0.5005521.87560e−5 / 0.5391958.02267e−6 / 0.626032
Pairwise AC at stage 30.0534314 / 0.5013821.79783e−5 / 0.5379007.68710e−6 / 0.626457
Four-coordinate shapes at stage 30.0592836 / 0.500801.99304e−5 / 0.537848.52514e−6 / 0.62535

Finally, we compared adjacent, fixed-shuffle, and strided pairing, as shown in Table 8.

Table 8. Retrieval sensitivity to coordinate pairing. FIQA, seed 42, IP NDCG@10; higher is better. All other encoding settings are held constant.

Pairing methodModernBERTQwen-1.5BQwen-7B
Adjacent pairing0.5020310.5390790.627198
Pairing after a fixed random shuffle0.5005160.5407340.626885
Strided pairing0.5008720.5395120.627328

These results support the default design as a choice that is simple to implement and requires no additional learning of coordinate pairings, with its scope determined by empirical comparisons.

(Question 3) Initialization and Refinement-Sweep Sensitivity

We compared mean-amplitude initialization, projection initialization, and one, three, and five refinement sweeps. The default encoder selects among initialization candidates based on the reconstruction objective, without using queries or qrels.

As shown in Table 9, increasing the number of sweeps from one to three reduces reconstruction error for all three models; the further improvement from increasing to five sweeps is smaller, and relevance-based retrieval does not improve consistently.

Table 9. Sensitivity to initialization and the number of refinement sweeps. FIQA, seed 42; each cell reports "MSE ↓ / IP NDCG@10 ↑." All configurations have the same storage budget.

ConfigurationModernBERTQwen-1.5BQwen-7B
Mean initialization only, 3 sweeps0.0481771 / 0.5020311.62303e−5 / 0.5390796.93949e−6 / 0.627198
Projection initialization only, 3 sweeps0.0508483 / 0.5003081.71437e−5 / 0.5380167.32697e−6 / 0.625102
Default initialization, 1 sweep0.0498999 / 0.5004241.68134e−5 / 0.5391257.18826e−6 / 0.626474
Default initialization, 3 sweeps0.0481763 / 0.5020311.62303e−5 / 0.5390796.93949e−6 / 0.627198
Default initialization, 5 sweeps0.0478743 / 0.5019301.61279e−5 / 0.5375646.89569e−6 / 0.627279

Three sweeps already capture the main reconstruction gains; further optimization does not guarantee corresponding improvements in retrieval quality, so we retain the same three-sweep configuration.

Reviewer_6ZzmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AC-Scoreq, a training-free quantizer for database embeddings in full-corpus semantic search. Vectors are first preconditioned by a randomized block-Hadamard transform with a fixed permutation and sign flips (applied to queries as well), adjacent real coordinates are paired into complex coordinates, and each coordinate is coded with the fixed four-direction phase codebook {1, i, −1, −i}; blocks of 16 complex coordinates share one real scale. This phase coding is applied T ∈ {1, 2, 3} times to successive residuals, and in the final application each pair of adjacent complex coordinates additionally selects one of four analytic relative-amplitude "shapes" that redistribute the block scale within the pair. Block scales across all applications are jointly fitted by ridge regression, alternated with phase and shape re-selection over three sweeps, and stored as uint8 values with an FP16 base. Scoring uses signed query-component selection, block accumulation and scale weighting; cosine and L2 rankings are derived from cached norms. At d = 768 the three operating points cost 170, 292 and 414 bytes per vector (18.1×, 10.5× and 7.4× compression). On the BEIR FIQA and DBPedia-Entity test sets with ModernBERT, Qwen2-1.5B and Qwen2-7B embeddings, T = 3 is reported close to FP32 and competitive with PQ-8x/16x, OPQ, LOPQ, RSLM4Lite, TurboQuant-4bit, RaBitQ and exRaBitQ-2bit on NDCG@10, Recall@100 and MRR@10 averaged over 50 seeds, with lower end-to-end top-1000 search time than the timed baselines on an A800; ablations compare amplitude coupling against equal-amplitude and same-budget extra-phase controls and vary sweeps, block size, ridge coefficient and scale precision.

Critical strengths
  1. The paper is explicit about storage accounting. It reports exact bytes per vector for every configuration (Table 19), states that T = 3 (414 B) is 7.8% larger than PQ-8x (384 B), and limits its claims to nearby-budget trade-offs rather than equal-byte dominance.
  2. The ablation design includes the right controls: an equal-amplitude variant that keeps the shape bits but freezes them, and a same-budget variant that spends the shape bits on extra phase information, so the value of relative-amplitude coding is separated from the value of the extra bits (Table 3).
  3. Encoding is training-free and per-vector: fixed analytic codebooks, no corpus-level fitting and no query knowledge, which is a practical property for streaming or distribution-shifting databases.
  4. The evaluation covers two datasets, three embedding models (768–3584 dimensions) and three ranking functions, and reports all three operating points with the same hyperparameters rather than tuning per setting.
  5. T = 1 at 170 B compares favourably with exRaBitQ-2bit at 196 B in most cells, which is the fairest comparison in the paper.
Critical weaknesses
  1. The latency comparison does not hold up as a method comparison. Brute-force FP32 scoring of 400 queries against 4.6M vectors of 3584 dimensions is about 13 TFLOP and 33–66 GB of memory traffic, which an A800 completes in roughly a second; the paper reports 59 s, so the shared pipeline (5,000-vector batches with per-batch top-k) is roughly 50× slower than a plain matrix-multiply scan, and the 1.7 s versus 2.4 s gap over PQ-8x says nothing about the representations. Appendix E also states that each baseline uses its official GPU kernel, while the AI-use statement says exRaBitQ-2bit, OPQ-8x, LOPQ-8x, RSLM4Lite and TurboQuant were implemented with AI assistance; the authors need to say which baselines are re-implementations.

  2. Using relevance labels instead of exact-neighbour recall compresses every method toward the FP32 ceiling (on Qwen-7B all compressed methods are within 0.006 NDCG@10 of FP32), which is why the tables cannot discriminate between methods. A quantization paper should also report recall@k against FP32 neighbours and distortion, at least on one standard ANN benchmark.

  3. The paper's strongest internal evidence for amplitude coupling is corpus MSE (31–37% lower than the equal-amplitude control), but at the retrieval level the gains over the same-budget extra-phase control are 0.0011–0.0052 NDCG@10, below query-sampling noise, and the paper itself notes that retrieval metrics need not follow MSE. The claim that the extra bits are better spent on amplitude than on phase is plausible but not demonstrated where it matters.

Questions for authors
  1. Do the authors agree with the real-coordinate reading of the phase code (per adjacent pair, store the sign and position of the larger-magnitude coordinate and zero the other)? If so, how does the method differ from PolarQuant's random preconditioning plus angle quantization of coordinate pairs, and from LVQ's training-free two-level residual scheme?
  2. Can extended RaBitQ at 3 and 4 bits, RSLM at 2 bits, FAISS SQ4 and LVQ be reported at the same byte budgets as T = 1, 2, 3? What accounts for TurboQuant-4bit's 514 bytes at d = 768, 130 bytes above a 4-bit code? Can paired bootstrap or paired t-test results over queries be reported for the key comparisons (T = 3 versus PQ-8x, RSLM4Lite and TurboQuant-4bit; AC versus the same-budget extra-phase control), rather than the seed-only intervals?
  3. What explains 59 s for the FP32 full scan of 400 queries against 4.6M vectors of 3584 dimensions? Please report the FP32 time with a standard batched GEMM plus top-k (for example FAISS IndexFlatIP on GPU), the PQ/OPQ baselines through FAISS-GPU ADC scoring, and scoring and top-k times separately.
  4. Appendix E states that every baseline uses its official GPU kernel, while the AI-use statement lists exRaBitQ-2bit, OPQ-8x, LOPQ-8x, RSLM4Lite and TurboQuant as implemented with AI assistance. Which baselines run official code and which are re-implementations? Why are RaBitQ and exRaBitQ identical across IP, cosine and L2 on FIQA with Qwen embeddings while FP32 differs?
5 comments
AuthorsOpenReview ↗

Response to Weaknesses 1–2: GPU Latency and Exact-Neighbor Retrieval

Thank you for your detailed examination of the timing protocol, baseline implementation sources, and evidence for bit allocation. We have added controlled timing and component measurements with GPU-resident data, official retrieval implementations, additional low-bit baselines, paired query statistics, and experiments on standard ANN datasets. Our response distinguishes the conclusions supported by the encoding representation, the actual implementation, and retrieval quality.

(Weakness 1) Controlled GPU Latency and Implementation Effects

We checked the previous timing records. The 59.272 seconds came from an earlier pipeline with 467 queries, which performed CPU FP16-to-FP32 conversion, host-to-GPU transfer, and top-kk merging for each batch. It does not represent the efficiency of FP32 exhaustive search over a GPU-resident database.

We reran the evaluation using the official 400 test queries on the same A800, controlling the following conditions:

  • The same 4,635,922 database vectors, with the database and compressed codes resident on the GPU;
  • The same document chunks of 50,000 vectors and sorted top-1000 output;
  • The same GEMM, scoring precision, and top-kk procedure;
  • Inclusion of query preprocessing, decoding, similarity computation, and result selection;
  • Consistent exclusion of the initial data upload and one-time database norm precomputation for both methods.

The AC and PQ decoders used the same development-stage tuning procedure, after which their configurations were fixed. Table 1 reports the means and standard deviations of seven measured runs following three warmup runs.

Table 1. End-to-end time with the shared retrieval pipeline controlled. DBPedia/Qwen-7B, 400 queries; times are in seconds, and lower is better. Both input precisions produce FP32 scores; TF32 is disabled for FP32 inputs.

GEMM input precisionSimilarityAC, T=3T=3PQ-8x
FP16IP0.198750 ± 0.0002700.232632 ± 0.000145
FP16Cosine0.226550 ± 0.0002150.260921 ± 0.000270
FP32IP0.935511 ± 0.0023220.968959 ± 0.000539
FP32Cosine0.962336 ± 0.0009020.996860 ± 0.000757

To identify the source of the difference, we recorded component times separately using CUDA events. As shown in Table 2, GEMM and top-kk times are similar at the same precision, with the main difference occurring in decoding.

Table 2. IP retrieval time breakdown under the same configurations as Table 1. Times are in seconds.

Input precisionMethodQuery preprocessingDecodingScoringTop-kk
FP16AC, T=3T=30.0007410.0602000.0677810.070687
FP16PQ-8x0.0001130.0946870.0681840.070858
FP32AC, T=3T=30.0006720.0707620.7962870.071364
FP32PQ-8x0.0000980.1094000.7893250.070838

With the shared scoring and result-selection procedures controlled, AC's structured decoding still provides measurable end-to-end gains.

(Weakness 2) Exact-Neighbor Recall and Reconstruction Distortion

We added NN Recall@1, @10, and @100 against the FP32 full-corpus exact-neighbor sets, together with corpus reconstruction MSE, and evaluated them on the standard NYTimes ANN dataset.

NN Recall@kk is the intersection size of the top-kk neighbor sets from compressed retrieval and FP32 exact retrieval, divided by kk and then averaged over queries. MSE measures the mean squared error between the reconstructed vectors and their corresponding uncompressed vectors.

NYTimes contains 290,000 database vectors and uses angular distance, with 10,000 queries. As shown in Table 3, we report reconstruction distortion and exact-neighbor recall for three residual-stage configurations.

Table 3. Reconstruction distortion and exact-neighbor recall on a standard ANN dataset. Original encoder, seed 42. Lower MSE is better; NN Recall is reported in %, and higher is better. The FP32 row is the evaluation reference by definition.

DatasetMethodCorpus MSE ↓NN Recall@1 ↑NN Recall@10 ↑NN Recall@100 ↑
NYTimesFP32 (reference)0100100100
NYTimesAC, T=1T=11.04693e−371.28065.51757.821
NYTimesAC, T=2T=23.13131e−483.23081.12777.464
NYTimesAC, T=3T=39.67977e−589.66089.09287.504

As the number of residual stages increases, reconstruction distortion decreases and exact-neighbor recall improves. These results supplement the evaluation of quantization quality beyond relevance-based metrics.

AuthorsOpenReview ↗

Response to Weakness 3: Bit Allocation and Query-Level Statistical Evidence

(Weakness 3) Bit Allocation and Query-Level Statistical Evidence

We performed paired reruns of the original AC and the same-budget extra-phase control. Both use the same six encoding bits per coordinate pair in the final stage: AC allocates them to four-direction phases and pairwise amplitude shapes, whereas the extra-phase control uses eight-direction phases for each of the two coordinates. The scale format and optimization effort are held constant.

FIQA uses 50 random seeds that were actually run, and DBPedia uses three seeds. We first average the seed results within each query, then perform 100,000 paired query bootstrap resamples. Table 4 reports NDCG@10 differences across all models and similarities.

Table 4. NDCG@10 differences (original AC minus the same-budget extra-phase control), with pointwise 95% paired query intervals. Positive values favor AC. FIQA: 648 queries and 50 seeds; DBPedia: 400 queries and three seeds.

DatasetModelIPCosineL2
FIQAModernBERT−0.000302 [−0.001782, 0.001135]−0.000925 [−0.002374, 0.000528]−0.000772 [−0.002172, 0.000630]
FIQAQwen-1.5B+0.000494 [−0.000457, 0.001443]+0.000629 [−0.000349, 0.001612]+0.000405 [−0.000534, 0.001341]
FIQAQwen-7B+0.000375 [−0.000317, 0.001111]+0.000424 [−0.000257, 0.001120]+0.000274 [−0.000390, 0.000950]
DBPediaModernBERT+0.000070 [−0.002376, 0.002524]−0.001297 [−0.003660, 0.000959]+0.000942 [−0.001372, 0.003369]
DBPediaQwen-1.5B−0.000955 [−0.002410, 0.000492]−0.000548 [−0.002255, 0.001122]−0.000137 [−0.001876, 0.001565]
DBPediaQwen-7B+0.000661 [−0.000493, 0.001792]+0.000587 [−0.000810, 0.001949]+0.000716 [−0.000719, 0.002121]

These intervals are insufficient to support the claim that the original AC outperforms extra phase across the board. We therefore distinguish two further questions: whether amplitude adaptation has an independent effect, and whether the current amplitude codebook outperforms all same-budget alternatives.

For the first question, we fix the four-direction phase structure and add only asymmetric amplitude shapes while retaining the zero shape. As shown in Table 5, the benefits appear in both reconstruction and exact-neighbor retrieval.

Table 5. Independent benefits of asymmetric amplitude adaptation. FIQA, three seeds; the control is "zero shape + equal amplitude." MSE reductions are relative percentages; NN Recall@10 gains and intervals are in percentage points.

ModelMSE reductionIP NN Recall@10 gainCosine gainL2 gain
ModernBERT21.76%+0.988 [0.576, 1.415]+1.137 [0.689, 1.584]+1.085 [0.658, 1.528]
Qwen-1.5B21.71%+0.612 [0.262, 0.960]+0.566 [0.190, 0.952]+0.520 [0.170, 0.864]
Qwen-7B21.72%+0.520 [0.213, 0.828]+0.453 [0.149, 0.751]+0.422 [0.118, 0.730]

The intervals in Table 5 are adjusted for 36 comparisons: three models, three similarities, and four controls comprising zero + equal amplitude, independent 1-bit amplitudes, disabled joint fitting, and extra phase. The correction applies to this family of mechanism comparisons and uses paired query bootstrap.

We also completed a supplementary design study explicitly distinguished from the original method. AC64 uses redundant codewords in the original six-bit field to add single-coordinate endpoints; the extra-phase encoding used for comparison employs the same block-direction selection and scale bit widths. This variant uses database mean centering, with the shared mean accounted for separately.

Table 6. Exact NN Recall@10 of the supplementary AC64 variant and extra-phase encoding with the same format. Three seeds; recall is in %, while differences and intervals are in percentage points. Intervals are adjusted for four comparisons across two datasets and two budgets.

DatasetScale/direction formatB/vector for bothAC64Extra phaseAC64 gain and adjusted interval
NYTimes7/7/8-bit scales + block direction14293.21292.979+0.233 [0.108, 0.356]
NYTimes4-bit scales + block direction13292.92892.770+0.158 [0.034, 0.284]
Fashion-MNIST7/7/8-bit scales + block direction48294.91894.456+0.462 [0.363, 0.562]
Fashion-MNIST4-bit scales + block direction44793.53193.070+0.461 [0.354, 0.572]

The evidence supports the independent value of amplitude adaptation and shows that an appropriate analytic codebook design can improve neighbor preservation at the same payload.

AuthorsOpenReview ↗

Response to Questions 1–2: Method Design, Baselines, and Statistical Comparisons

(Question 1) Complex Phase Coding, PolarQuant, and LVQ

Yes. At an ordinary stage, four-direction phase selection can be interpreted as recording the axis and sign of the larger-magnitude component among the real and imaginary parts of a complex residual. The role of the complex representation is to unify phase selection, residual updates, amplitude coupling, and similarity computation within one compact algebraic formulation. For example, the four-direction phases correspond to signed query-component selection, followed by within-block accumulation and scale weighting.

The three methods have different design focuses:

  • PolarQuant uses recursive polar transformations and quantizes angles; AC-Scoreq uses repeated residual superposition with a fixed four-direction codebook and introduces local relative amplitudes in the final application.
  • LVQ uses per-vector scaling and scalar quantization, and can encode scalar residuals in a second level; AC-Scoreq uses analytic amplitude shapes across coordinate pairs and joint fitting of block scales across residual stages.

Our contribution lies in how this phase-amplitude-scale structure supports configurable compression and retrieval computation without codebook training.

(Question 2) Additional Low-Bit Baselines, Actual Storage, and Paired Comparisons

We added official multibit RaBitQ configurations at 1–4 bits, RSLM configurations at 2/3 bits, FAISS SQ4, and an LVQ reference implementation based on the published equations.

As shown in Table 7, we select Qwen-1.5B and report two metrics separately:

  • B/vector: the actual number of bytes occupied by each vector; lower is better.
  • IP NDCG@10: inner-product retrieval quality; higher is better.

Storage includes the scales and auxiliary fields saved with each vector by each method; shared parameters and additional norm caches are accounted for separately. The following results come from the same supplementary evaluation as the paired statistics in Table 8.

Table 7. Storage and retrieval quality on FIQA / Qwen-1.5B. Bold marks the AC configurations.

MethodStorage: B/vector ↓Retrieval quality: IP NDCG@10 ↑
AC, T=1T=13380.530670
AC, T=2T=25800.536494
AC, T=3T=38220.539079
PQ-8x7680.538057
Official RaBitQ, 1 bit2040.531629
Official RaBitQ, 2 bit3960.537359
Official RaBitQ, 3 bit5880.535231
Official RaBitQ, 4 bit7800.538248
RSLM, 2 bit3860.534335
RSLM, 3 bit5780.538241
RSLM4Lite7680.536660
TurboQuant-4bit adapted implementation10260.537067
FAISS SQ47680.536604
LVQ 2×0, reference implementation3880.450796
LVQ 3×0, reference implementation5800.530515
LVQ 4×0, reference implementation7720.536512
LVQ 2×2, reference implementation7720.534068

TurboQuant's 514 B at 768 dimensions comes from padding in this adapted implementation: vectors are padded to 1024 dimensions, the four-bit code occupies 512 B, and an additional 2 B norm is stored. This is the actual format overhead of this implementation, not an inherent overhead of all TurboQuant implementations.

We also performed paired query tests against the key baselines. As shown in Table 8, the difference is defined as AC's NDCG@10 minus the control method's NDCG@10: a positive value means that AC has a higher mean. The confidence interval assesses the uncertainty of this difference under query sampling.

Table 8. IP NDCG@10 differences between AC, T=3T=3, and the main baselines, with 95% paired query bootstrap intervals. FIQA, 648 queries.

ModelControl methodNDCG@10 difference: AC − control95% confidence interval
Qwen-1.5BPQ-8x+0.001022[−0.002747, 0.004958]
Qwen-1.5BRSLM4Lite+0.002418[−0.000982, 0.005977]
Qwen-1.5BTurboQuant-4bit+0.002012[−0.001213, 0.005315]

The multi-seed paired results for the original AC and extra phase are reported in full in Table 4. These results show how AC provides different storage–retrieval-quality choices through its stage count, while actual byte counts and paired statistics clarify the scope of comparison for each configuration.

Reviewer_h99MOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

AC-Scoreq compresses vectors using a fixed four-direction complex phase codebook, pairwise relative-amplitude shapes in the final application, and joint scale fitting across residual applications. Direct packed-code scoring supports inner-product, cosine, and L2-based retrieval. Experiments on FIQA and DBPedia-Entity report competitive semantic retrieval quality and storage–latency–quality trade-offs.

Critical strengths
  • The fixed phase codebook turns similarity computation into signed query-component selection, block accumulation, and scale weighting. This connects the compressed representation directly to packed-code scoring without corpus-level codebook training.

  • AC reduces corpus reconstruction MSE relative to equal-amplitude coding across all three embedding models on both datasets. The same-budget extra-phase comparison also favors AC across the reported FIQA retrieval metrics, supporting the complete AC construction.

Critical weaknesses
  • The equal-amplitude (no shape refinement) control removes both asymmetric shapes and the zero shape (0,0). AC’s gains over this control therefore cannot distinguish the benefit of relative-amplitude redistribution from suppressing an entire pair. The central explanation based on unequal local magnitudes remains only partially supported.

  • The no phase refinement, refinement-sweep, and ridge-coefficient variants all retain joint scale fitting. These comparisons do not isolate the contribution of fitting scales across applications, leaving the benefit of this central component entangled with the rest of the phase–amplitude–scale construction.

  • Reconstruction error, used for scale fitting and initialization selection, is an imperfect proxy for inner-product retrieval, a limitation also studied in Norm-Explicit Quantization. On FIQA/ModernBERT, block size 8 lowers MSE from 0.04818 to 0.03536 but reduces IP NDCG@10 from 0.50295 to 0.50026. The paper acknowledges this nonmonotonicity, which limits how reliably its optimization objective tracks retrieval quality.

  • NDCG@10, Recall@100, and MRR@10 are evaluated against relevance labels, so scores close to FP32 do not show how well the original FP32 rankings are preserved. Changes among equally relevant or unjudged documents can remain invisible to these metrics. The results support semantic retrieval quality, but leave the claimed ranking fidelity insufficiently characterized.

Questions for authors

Please see my "weaknesses" comments.

4 comments
AuthorsOpenReview ↗

Response to Weakness 1: Amplitude Redistribution versus Pair Suppression

Thank you for pointing out the connections among amplitude redistribution, joint scale fitting, and ranking fidelity that require further validation. To address these issues, we added mechanism-isolation experiments, measurements of actual inner-product score errors, exact-neighbor recall, and retrieval evaluations on standard ANN datasets and under shared candidate sets to directly examine the role of each component.

(Weakness 1) Amplitude Redistribution versus Pair Suppression

We added four controls with the same storage layout, separately controlling the zero shape and asymmetric amplitude shapes. All configurations use the same phase fields, scale precision, and number of optimization sweeps. Thus, comparing "zero + equal amplitude" with full AC isolates the effect of asymmetric amplitude allocation while retaining the ability to zero out an entire pair.

As shown in Table 1, adding asymmetric shapes further reduces reconstruction error for all three models.

Table 1. Factorial controls for the zero shape and asymmetric amplitude shapes. FIQA; mean corpus MSE over three random seeds, lower is better. Storage remains identical within each model.

Amplitude-shape settingModernBERTQwen-1.5BQwen-7B
Equal amplitude only0.062500032.119877e−59.108570e−6
Zero + equal amplitude0.060865332.064419e−58.870571e−6
Asymmetric + equal amplitude, without the zero shape0.049191141.669557e−57.173145e−6
Full AC0.047623011.616152e−56.943650e−6
Error reduction of full AC relative to "zero + equal amplitude"21.76%21.71%21.72%

We further examined whether these benefits translate into exact-neighbor retrieval. Table 2 reports the NN Recall@10 gains of full AC over "zero + equal amplitude," alongside the joint scale-fitting results corresponding to the next issue.

Table 2. Exact NN Recall@10 gains of the original AC and intervals adjusted for multiple comparisons. FIQA, three seeds; units are percentage points, higher is better. Brackets contain bootstrap intervals adjusted over the family of 36 comparisons, with a nominal familywise confidence level of 95%.

ModelSimilarityRelative to "zero + equal amplitude"Relative to "no joint scale fitting"
ModernBERTIP+0.988 [0.576, 1.415]+6.816 [6.108, 7.536]
ModernBERTCosine+1.137 [0.689, 1.584]+1.590 [0.977, 2.207]
ModernBERTL2+1.085 [0.658, 1.528]+2.459 [1.857, 3.061]
Qwen-1.5BIP+0.612 [0.262, 0.960]+3.102 [2.567, 3.642]
Qwen-1.5BCosine+0.566 [0.190, 0.952]+1.101 [0.628, 1.574]
Qwen-1.5BL2+0.520 [0.170, 0.864]+1.343 [0.859, 1.836]
Qwen-7BIP+0.520 [0.213, 0.828]+1.651 [1.224, 2.094]
Qwen-7BCosine+0.453 [0.149, 0.751]+0.756 [0.386, 1.137]
Qwen-7BL2+0.422 [0.118, 0.730]+1.070 [0.694, 1.456]

The nine settings are the combinations of three embedding models and three similarity measures. The 36 comparisons covered by the statistical correction are:

3 models×3 similarity measures×4 controls.3\text{ models}\times3\text{ similarity measures}\times4\text{ controls}.

The four controls are "zero + equal amplitude," "independent 1-bit amplitude," "no joint scale fitting," and "same-budget extra phase." We first averaged the three seeds within each query, then performed 100,000 paired query bootstrap resamples and applied a Bonferroni correction over these 36 comparisons. Table 2 shows the two controls directly relevant to this issue and Weakness 2.

These results show that asymmetric amplitude allocation provides independent benefits for reconstruction and neighbor preservation even when zeroing out an entire pair is already allowed.

AuthorsOpenReview ↗

Response to Weaknesses 2 and 3: Joint Scale Fitting and Retrieval Quality

(Weakness 2) Isolating Cross-Stage Joint Scale Fitting

We added a control that genuinely disables joint scale fitting: cross-stage joint fitting is disabled during both initialization and subsequent refinement, while retaining per-stage scale handling, discrete code updates, and the original storage format.

As shown in Table 3, restoring joint scale fitting improves IP NDCG@10 for all three models, with paired-query intervals entirely above zero. Table 2 further shows that these benefits also occur in exact-neighbor recall across all nine settings.

Table 3. Effect of joint scale fitting on semantic retrieval quality. FIQA; mean IP NDCG@10 over three seeds, higher is better. Intervals are pointwise 95% paired query bootstrap intervals.

ModelNo joint scale fittingFull ACDifference and 95% interval
ModernBERT0.4911270.500730+0.009603 [0.004969, 0.014365]
Qwen-1.5B0.5353440.538494+0.003150 [0.000804, 0.005509]
Qwen-7B0.6240280.626653+0.002625 [0.000550, 0.004791]

Joint fitting rebalances the contributions of different stages to the same vector and accounts for correlations between stages. Thus, the benefits of cross-stage scale fitting are verified through an independent ablation and are reflected in both relevance-based retrieval and exact-neighbor preservation.

(Weakness 3) Reconstruction Error, Score Accuracy, and Retrieval Quality

We agree that lower average reconstruction error does not necessarily correspond to higher NDCG. To examine whether reconstruction improvements affect actual retrieval, we kept the codes fixed, directly measured inner-product score errors between the test queries and the database, and evaluated exact-neighbor recall.

As shown in Table 4, adding asymmetric amplitude shapes improves reconstruction error, actual inner-product score error, and exact-neighbor preservation on the same set of codes. To ensure consistency, all three metrics in the table come from the same encoding run with seed 42.

Table 4. Corresponding measurements from reconstruction to scoring and neighbor retrieval. FIQA, seed 42; each cell shows "zero + equal amplitude → full AC." Relative IP RMSE is normalized by the root mean square of the FP32 inner-product scores.

ModelCorpus MSE ↓Relative IP RMSE ↓IP NN Recall@10, % ↑
ModernBERT0.06154171 → 0.048176270.033734 → 0.02691990.586 → 91.497
Qwen-1.5B2.072520e−5 → 1.623027e−50.037391 → 0.03034394.120 → 94.907
Qwen-7B8.865349e−6 → 6.939486e−60.037013 → 0.03000496.713 → 97.145

This controlled experiment shows that the benefits of amplitude adaptation translate into actual scoring and neighbor preservation.

AuthorsOpenReview ↗

Response to Weakness 4: Exact-Neighbor Preservation under Compression

(Weakness 4) Exact-Neighbor Preservation under Compression

We added NN Recall@1, @10, and @100 against exhaustive full-corpus FP32 results. This metric measures the overlap between the compressed retrieval results and the exact top-kk set, and is reported separately from relevance-based Recall computed using qrels.

As shown in Table 5, the original AC maintains high exact-neighbor recall across two text retrieval datasets, three embedding models, and three similarity measures.

Table 5. Exact-neighbor preservation of the original AC with T=3T=3. Each cell lists NN Recall@1 / @10 / @100, in %, higher is better. FIQA uses the official 648 test queries and 50 seeds; DBPedia uses the official 400 test queries and three seeds.

DatasetModelIPCosineL2
FIQAModernBERT90.358 / 91.603 / 92.38591.537 / 92.026 / 92.66091.392 / 91.907 / 92.762
FIQAQwen-1.5B94.920 / 94.797 / 95.70695.170 / 94.946 / 95.79095.185 / 94.972 / 95.813
FIQAQwen-7B97.364 / 97.059 / 97.66797.543 / 97.155 / 97.70797.583 / 97.151 / 97.713
DBPediaModernBERT91.167 / 90.758 / 91.39593.333 / 91.050 / 91.51293.000 / 90.850 / 91.414
DBPediaQwen-1.5B96.000 / 94.367 / 95.09796.000 / 94.725 / 95.21395.917 / 94.750 / 95.214
DBPediaQwen-7B97.833 / 96.317 / 96.92597.833 / 96.400 / 96.98597.917 / 96.408 / 96.994

In addition, we placed the different encodings into the same IVF inverted lists, fixed the candidate set for each query, and did not use FP32 reranking. As shown in Table 6, the exact-neighbor improvements from additional residual stages remain present.

Table 6. NN Recall@10 under shared IVF candidate sets. NYTimes, 10,000 queries; Recall is reported in %.

nprobeAverage candidatesFP32 (.ref)AC, T=1T=1AC, T=2T=2AC, T=3T=3
11,423.350.93239.75045.65748.259
45,335.965.96049.51158.06461.865
1619,661.378.00356.71467.74972.775
6474,939.890.92362.55076.50183.219

These supplementary results directly characterize neighbor preservation and validate the quality choices provided by multistage encoding at different storage costs.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, with approximately one day remaining, we would like to ensure that we have adequately addressed all your concerns. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we would be happy to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper. We wish you a wonderful day!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.