PaperlayerView exact revision on arXiv

Source-linked AI summary

Quantize by Drift: Label-Free Mixed-Precision Post-Training Quantization for Text Embedders

Hyojung Han, Jongmin Kim, Seung-Hun Jeon

arXiv:2610.09227v1cs.IRcs.CLcs.LG

TL;DR

Text-embedder mixed-precision allocation needs a module sensitivity signal, but measuring retrieval damage directly requires relevance judgments that deployments often lack. This paper uses quantization-induced output drift to allocate bits under a packed-byte budget, finding strong broad-pool ordering but no advantage over uniform precision at a healthy operating point.

  • Problem

    Mixed-precision text-embedder quantization needs a per-module sensitivity signal, but measuring retrieval damage directly requires relevance judgments that deployments usually lack.

    Source: §1 · p. 1

  • Method

    The method measures quantization-induced output-embedding drift per module and bit width, then uses an additive allocation under a hard packed-byte budget without labels or search.

    Source: §1 · p. 2, §2 · p. 3

  • Results

    Configuration drift orders sampled development allocations against held-out retrieval quality at a 0.911 macro Spearman, but on untouched models it beats prior geometry and trails uniform precision at the healthy operating point.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Output drift is a robust coarse sensitivity signal for avoiding harmful precision assignments, but not a universally superior fine-grained allocator.

    Source: §1 · p. 2

    The additive allocation is a local optimum, and a systematic compressive residual can still reorder it.

    Source: §8 · p. 14

Abstract

from arXiv · show

Mixed-precision post-training quantization needs a per-module sensitivity signal; for a text embedder the obvious one -- the retrieval quality a module costs when quantized -- needs relevance labels that deployments rarely have. We measure a label-free substitute: quantization-induced representation drift, obtained by quantizing one module, re-encoding the corpus, and recording how far the output embeddings moved from their full-precision positions. What is specific is the observable: the deployed output representation a dense retriever ranks with. Across five development embedders, configuration-level drift orders sampled mixed-precision plans against held-out retrieval quality at a macro Spearman of 0.911, the sensitivity transports across calibration corpora and retrieval domains in the usable regime, module drifts compose rank-consistently but not numerically, and relevance-derived sensitivity adds no consistent value. The method is one additive allocation under a hard packed-byte budget, with no labels and no search. On three embedders held untouched until method, baselines and hypotheses were frozen and sealed, the pre-registered directional hypothesis against the prior LieQ criterion holds (3/3 at the main budget, no collapse) and drift scores above a two-sided LieQ steelman in 2/3; but at the main budget drift is numerically lower than same-budget uniform precision on all three (-0.99, -0.85, -1.01 points), having reduced module and whole-model drift as designed. Output drift is thus a robust coarse sensitivity signal, not a universally optimal allocation objective: it avoids the catastrophic failures of the transferred signed-geometry adaptation and can remain usable at stressed budgets where uniform collapses, but fine-grained redistribution around a strong uniform operating point remains unresolved.

1 Introduction

The paper proposes output-embedding drift as a label-free, retrieval-facing signal for assigning mixed precision, then tests whether module-level drift can guide allocation across models, budgets, and domains. Its evidence supports drift as a coarse sensitivity signal, not a universally better allocator.

1 Introduction

  • Dense retrievers rank independently encoded query and document vectors, while deployments usually lack the relevance judgments needed to measure quantization damage directly.The output vector’s geometry determines ranking by relative distances, making text-embedder quantization distinct from generative-model quantization.Source: §1 · p. 1
  • Across development plans, configuration drift reaches 0.911 macro Spearman; on untouched models it beats prior geometry but trails healthy-point uniform precision.The paper frames this as a distinction between avoiding harmful precision assignments and resolving small trade-offs among already-good plans.Source: §1 · p. 2
  • The authors distinguish intervention-based allocation by observing the final embedding used for dense-retrieval ranking rather than an internal information-theoretic quantity.Against the prior geometry criterion drift scores better, but not against uniform precision at a healthy operating point.Source: §1 · p. 2
  • The proposed label-free signal measures output-embedding displacement after quantizing a module and re-encoding the calibration corpus.The contribution is the per-module allocation signal, not the correlation between whole-model drift and retrieval quality.Source: §1 · pp. 1–2
  • The study evaluates drift across heterogeneous embedders and packed-byte regimes, including transport across retrieval domains and a pre-registered confirmation on three untouched models.The paper reports that the prior-geometry comparison replicates, while the main-budget comparison against uniform precision does not.Source: §1 · p. 2
  • Module drift composes rank-consistently but not numerically, while relevance supervision and local configuration search add no consistent value.These findings are offered as a mechanistic account of why the allocator can remain additive and simple.Source: §1 · p. 2

2 Method

The method scores each module-and-bit-width choice by output-embedding drift on a calibration corpus, then feeds those scores to a simple additive allocation under a hard packed-byte budget. Output drift is tied to retrieval-score perturbation, but that bound does not guarantee preserved rankings.

2 Method

  • Each module-and-bit-width cell is scored by mean cosine displacement between full-precision calibration embeddings and embeddings after quantizing that module.The calibration quantity includes both documents and queries, with no relevance judgments entering the measurement.Source: §2 · p. 2
  • Cosine drift on unit-normalized output embeddings bounds perturbation of the inner-product retrieval score, linking the measured observable to the score used for ranking.The proof is motivational rather than a guarantee: it bounds score distortion, not ranking inversions, and the measured quantity is a corpus mean.Source: §2 · p. 3
  • The method weights query and document drift equally, a choice motivated by the retrieval-score margin bound and fixed before confirmation models were opened.The balanced form was adopted after development audits found it nearly interchangeable with pooled drift, though not always so on untouched models at the tighter budget.Source: §2 · pp. 2–3
  • The plan minimizes summed per-module drift subject to a hard packed-byte budget that includes group metadata, never exceeding the budget.The solver may leave the realized cost slightly below budget; the integer program is standard, with the contribution in the sensitivity signal rather than the solver.Source: §2 · p. 3

3 Experimental setup

The experiments profile module drift on a probe corpus and evaluate retrieval on disjoint held-out data, using five development embedders and three untouched confirmation embedders. They compare quantization plans across actual packed-byte budgets with a fixed module partition and controlled evaluation pipeline.

3 Experimental setup

  • Five development embedders were used for method development, while three others were held untouched until the method, baselines, budgets, thresholds, and hypotheses were frozen.The confirmation models were run once under the sealed pre-registered protocol.Source: §3 · p. 3
  • Weights use group-wise asymmetric round-to-nearest quantization at 2, 3, or 4 bits, with FP16 scale and zero point; activations remain full precision.Tensors with fewer than two dimensions are not quantized, and two-dimensional tensors kept full precision are charged at 16 bits.Source: §3 · p. 3
  • Retrieval uses each model’s pooling and prompt conventions, normalized embeddings, exact inner-product search, and FP32 full-precision reference weights.All arms for gte-base-en-v1.5 use transformers 4.49.0 because its remote code fails to embed under 5.14.1; the other models use 5.14.1.Source: §3 · pp. 3–4
  • Each encoder is partitioned into twelve units: four attention bands, four feed-forward bands, and four vocabulary-table bands.Vocabulary bands follow tokenizer-ID row order rather than calibration-corpus token frequencies; their sensitivity may still depend on domain.Source: §3 · p. 4
  • Budgets use realized packed bits per weight, including scale and zero-point overhead, and test a stressed midpoint between uniform 2- and 3-bit costs and the uniform 3-bit cost.The early 3-to-4-bit midpoint was never measured and is outside the paper’s fixed scope; table labels B2.5 and B3.0 are budget names, not bit widths.Source: §3 · p. 4
  • Drift was profiled on BEIR ArguAna, while retrieval quality was measured on disjoint BEIR NFCorpus data; no relevance judgments entered drift computation.The probe uses 1,500 documents and 150 queries, and the held-out corpus has 3,633 documents and 323 queries.Source: §3 · p. 4
  • The configuration pool requests 176 plans per model, mixing perturbations around uniform and allocator plans with random feasible plans.Because one quarter of the pool is allocator-derived, the pooled correlation is treated as descriptive of that pool rather than the feasible set.Source: §3 · p. 4

4 Results

Across development and untouched confirmation models, output drift is a coarse allocation signal: it predicts sampled-plan quality and transports across domains in usable regimes, but does not beat uniform at B3 on the three confirmation models. Its clearest benefit is resisting collapse under stressed budgets, while finer redistribution remains unresolved.

4.1 Drift predicts retrieval damage

  • Whole-model drift orders sampled plans against held-out retrieval quality, but its correlation weakens among better configurations, making it a coarse rather than fine-grained surrogate.The top-decile figure is shown for shape only, with about 16 plans per model.Source: §4.1 · p. 4, Figure 1 · p. 7

4.2 Allocation quality against budget-matched baselines

  • Among 10 development cells, four were cliff crossings; in all four drift was usable and LieQ was not, showing that the choice of sensitivity criterion can determine whether an allocation stays usable.Three cells were graded and three were dead; cliff crossings are not quality margins.Source: §4.2 · p. 6
  • Removing LieQ's score direction restored usable quality in three collapsed cells, while the unsigned criterion sometimes became byte-identical to uniform and drift remained ahead on two non-uniform usable allocations.The sign reversal was associated with negative attention-projection ∆r; the paper says removing direction removes the catastrophic failures.Source: §4.2 · pp. 6–7

4.3 Confirmation on the untouched models

  • At B3 on three untouched models, drift beat pre-specified LieQ-signed in 3/3 and retained at least three-quarters of full precision on each, but scored 0/3 wins against same-budget uniform, trailing by 0.99, 0.85, and 1.01 points.Source: §4.3 · p. 8
  • At B2↔3 on mxbai-large, drift retained 28.84 versus 5.30 for uniform, while two other untouched models were dead across all tested arms.At that budget, LieQ-signed retained 27.99 on mxbai-large.Source: §4.3 · pp. 8–9
  • On all three re-evaluated confirmation models, drift reduced held-out module-wise and whole-model drift relative to uniform, yet retrieval NDCG was lower.This was the retrieval-mismatch case, not a failure of additive composition.Source: §4.3 · p. 8

4.4 Cross-corpus transport

  • Across five development models, cross-corpus sensitivity ordering had median pairwise Spearman 0.992; among usable cells, 78.6% of probe pairs selected identical plans and worst disagreement was 0.372 points.Across all 10 cells, worst disagreement was 11.425 points; the large disagreement occurred in a collapsed cell.Source: §4.4 · p. 9
  • Cross-corpus transport was measured on development models only; the four-probe measurement was not part of confirmation and was not run on the three untouched models.The authors therefore withhold the broader claim that domain stability is established for the method beyond those development models.Source: §4.4 · p. 9

5 Module drift preserves configuration ordering, not magnitude

Summed module drifts largely preserve configuration ordering but do not predict whole-configuration drift magnitude accurately. Composition error is systematic and can reverse rankings, especially among high-quality neighboring configurations.

5 Module drift preserves configuration ordering, not magnitude

  • 0.939 rank correlation accompanies a 17.5% median relative residual when summed module drifts predict whole-configuration drift.The relative residual has a denominator artifact for low-drift configurations; on scale-free axes, mean absolute residual is 0.69 of the interquartile range.Source: §5 · p. 10
  • β lies in [0.26, 0.77] across cells, indicating that the module sum systematically overstates configuration drift, increasingly for aggressive configurations.The fitted relation has R2 ≈0.84; β and R2 are reported descriptively without a significance test because the configuration pools are structurally generated and not i.i.d.Source: §5 · p. 10
  • 9.9% of configuration pairs are misordered overall, rising to 15.3% in the top retrieval-quality quartile and 12.2% within the additive plan’s neighborhood.A monotone fitted mean trend explains why global ordering is largely preserved, while residuals can reverse individual pairs.Source: §5 · p. 10
  • Composition residuals grow with the number of 2-bit modules and predicted drift, and shrink as nominal bits per weight increase.Interactions are substantial in magnitude but usually insufficient to disrupt global configuration ordering; they can matter in narrow high-quality neighborhoods.Source: §5 · p. 10

6 Analysis: why is the allocator this simple?

The analysis finds that relevance-derived sensitivity adds no consistent predictive value beyond drift, while configuration search can help unevenly and is not adopted as the default. Search gains concentrate where additive composition is poor, but proposed composition-gated search remains an optional, unconfirmed extension.

6.1 Does relevance supervision help?

  • Relevance-derived sensitivity provided no consistent incremental predictive value beyond drift under the tested allocator and calibration protocol.It was the weakest of seven signals on two models, and controlling for drift left its partial correlation with true damage approximately zero; labels remain an ablation, not part of the method.Source: §6.1 · p. 10

6.2 Does local configuration search help?

  • Drift-seeded search beat random seeding in 9 of 10 cells, with Qwen3-0.6B at B2↔3 the exception.In that cell, random candidates happened to include a better plan than the seeded neighborhood.Source: §6.2 · p. 10
  • Search gains were heterogeneous: after nine cells the median gain was positive, but some budgets had no better budget-preserving neighbor of the additive plan.With four cells measured, median Δsearch was negative; Table 21 attributes part of the pattern to small feasible neighborhoods where budgets coincide with uniform grid points.Source: §6.2 · p. 10
  • Across measured cells, Δsearch correlated with composition error at r = 0.95 against nMAE/IQR and r = −0.76 against β.Search gains concentrated where additive composition was poorer, but this relation was exploratory on ten development cells and was not tested by a sealed prediction on untouched models.Source: §6.2 · pp. 10–11
  • Whole-configuration drift curves show a compressive bias in summed module drift, and a tenth of configurations differ by more than a third, supporting rank ordering rather than magnitude claims.The module and whole-configuration drifts were measured on the same corpus, so this spread reflects composition error rather than cross-corpus transfer.Source: §6.2 · p. 11
  • 7 of 10 cells improved, but the median gain of 0.330 pp missed the 0.5 pp threshold and one cell lost by more than 1 pp, so search was excluded from the default method.The inclusion rule required both a minimum median effect and a worst-case floor.Source: §6.2 · p. 11
  • A composition gate estimated from the full pool would require 100–200 whole-configuration evaluations, near the cost of the search it is meant to authorize.A handful of sentinel configurations may be usable, but the proposed gate is optional and unconfirmed; calibrating its threshold against Δsearch uses relevance labels during development.Source: §6.2 · pp. 11–12

7 Related work

Prior allocation methods use geometric, reconstruction, or interaction-aware sensitivity signals, while the paper measures the retrieval-facing consequence of quantization on the embedder’s output. Its comparisons distinguish that observable from prior internal quantities and note limits of the available adaptations.

7 Related work

  • Reconstruction methods such as AdaRound [23], BRECQ [18], and GPTQ [8], and allocation methods such as HAWQ [7] [6] and HAQ [30], address tensor quantization or precision distribution through different objectives.The paper uses plain round-to-nearest quantization to isolate its allocation signal.Source: §7 · p. 12
  • Encoder quantization studies Q8BERT [37], Q-BERT [27], and I-BERT [15] evaluate classification or extraction with a task head, unlike an embedder whose pooled output is the product.Dense-retrieval compression often targets output vectors or the index; encoder weight-level mixed precision is complementary to those approaches [38] [39] [12] [17] [14].Source: §7 · pp. 12–13
  • LieQ [33] assigns higher precision to layers with greater training-induced spectral concentration, using a full-precision geometric criterion rather than quantization intervention.The paper compares the criteria under an identical allocator and budget.Source: §7 · p. 12
  • InfoQ [1] quantizes layers and scores downstream mutual-information change, but its class-target branch has no canonical one-to-one port to dense retrieval.The paper evaluates the transferable information-flow component as InfoQ-X, not as a reproduction.Source: §7 · p. 12
  • CoopQ [42] models layer interactions with Shapley values, while MixQuant [21] marginalizes distortion over quantized upstream configurations.The paper reports large, compressive residual interactions that rarely reorder global configurations but can matter locally; it does not compare against an interaction-aware allocator.Source: §7 · pp. 12–13
  • The paper measures quantization’s effect on final embedding displacement, the output representation from which retrieval metrics are computed.This differs from prior methods that chiefly use intrinsic geometry, reconstruction error, or internal information-flow proxies.Source: §7 · p. 13

8 Limitations

The main boundaries are that the untouched-model results do not beat uniform precision at the main budget, while several comparisons and proposed extensions have restricted scope. The composition-risk gate remains exploratory, and deployment performance beyond packed-byte accounting is unreported.

8 Limitations

  • The InfoQ-X comparison is unavailable because its estimator adaptation failed the pre-specified stability criterion, leaving one controlled prior-criterion comparison.The adaptation reached the original work’s 6,000-sample scale; its retrieval metric was not computed, and the static-geometry post-hoc column is an upper bound rather than a competitor.Source: §8 · p. 13
  • LieQ-adapt transfers LieQ’s geometry criterion into a shared twelve-unit packed-byte framework, not LieQ’s native per-layer deployment format.LieQ keeps precision uniform within each layer for standard kernels.Source: §8 · p. 13
  • At the main budget, drift scores below uniform precision on all three untouched models, by about a point.The pre-registered hypotheses against LieQ hold, but the comparison against uniform goes the other way.Source: §8 · p. 13
  • Domain stability was measured on the five development models, not the three untouched models.Across development probes, 78.6% of probe pairs agree on a byte-identical plan and worst held-out disagreement is 0.372 points.Source: §8 · p. 13
  • The development correlation is not wholly independent of the signal because one quarter of each pool perturbs allocator outputs.Random-only pools still give ρ = 0.872–0.929 on development models and ρ = 0.611–0.908 on untouched models, above the pre-registered 0.6 threshold on all three.Source: §8 · p. 13
  • The composition-risk gate is unconfirmed: its threshold is supervised, fitted on only four development endpoints, and was not tested on untouched models.It remains an optional exploratory extension rather than part of the method.Source: §8 · p. 14
  • Uncertainty reporting is limited: only the untouched B3 drift-versus-uniform comparison has paired bootstrap intervals, and search curves are single trajectories.Other held-out NDCG values are point estimates; four of five frozen development budgets coincide with uniform 3-bit packed size, weakening permutation nulls.Source: §8 · p. 14
  • Per-band embedding-table precision assumes row-blocked packing, and no kernel-level latency or throughput is reported.Rank-consistent composition does not establish that additive allocation is a local optimum because compressive residuals can reorder plans.Source: §8 · p. 14

9 Conclusion

Quantization-induced output drift is a stable, retrieval-facing coarse sensitivity signal, but its strengths and limits differ across operating regimes. Untouched-model results support avoiding severe misallocations, not superiority to uniform precision at the healthy 3-bit point.

9 Conclusion

  • Drift avoids the severe misallocations of the transferred signed-geometry criterion and never collapses at the main B3 point on three untouched embedders.At that point drift is numerically lower than uniform precision on all three models, having reduced module, held-out, and whole-model drift as designed.Source: §9 · p. 14
  • On development models, output drift reproduces under calibration resampling, transports across retrieval domains in the usable regime, and orders sampled allocations against held-out quality.The configuration-level relationship is reported as a macro Spearman of 0.911 over the broad sampled pool.Source: §9 · p. 14
  • Drift is stronger at avoiding dangerous precision mistakes than at resolving small trade-offs among already-good allocations.Fine-grained, ranking-aware or configuration-level refinement remains an open next problem.Source: §9 · p. 14
  • The conclusion is a coarse sensitivity signal, not a universally optimal allocation objective.It can remain usable where uniform collapses, while fine-grained redistribution around a strong uniform operating point remains unresolved.Source: §9 · p. 14

Reproducibility

The appendix reports ledger-backed reproducibility, a proof step for the proposition, validation and rejection of an InfoQ-X estimator adaptation, and a check of query-document pooling choices.

Reproducibility

  • Reported values, tables, and plots come from a snapshot-validated experiment ledger, and missing required cells fail the build.Source: Reproducibility · p. 14

A Proof of Proposition 1

  • The proof derives the margin bound by applying Equation 2 to query-positive and query-negative pairs and using the triangle inequality.The derivation adds and subtracts cross terms and applies Cauchy–Schwarz with unit norms.Source: A Proof of Proposition 1 · p. 17

B InfoQ-X adaptation: estimator stability validation

  • The InfoQ-X adaptation failed its pre-specified estimator-stability criterion through 6,000 samples, so unstable retrieval comparisons are not reported.Implementation audits corrected seed handling and standardized projections, but neither correction met the criterion or reproduced a plan; the rejection applies to this adaptation, not the estimator family generally.Source: B InfoQ-X adaptation: estimator… · pp. 17–18

C Pooling the query and document sides

  • Across ten development cells, balanced, pooled, and document-only drift select the same plan, while query-only differs in two cells.At B2↔3, balanced and pooled plans differ on two of three untouched models, with balanced ahead by up to 0.79 points; the frozen confirmation uses balanced pooling.Source: C Pooling the query and document sides · p. 18

D Reproducibility of our own signal

Calibration resampling reproduced the selected plan on every model, while null comparisons found the drift allocation consistently beat random plans but not always the permuted-signal null.

D Reproducibility of our own signal

  • Every model reproduced its selected plan across three independently drawn calibration subsets, bounding calibration-sampling effects but not estimator variance.The audit is not a controlled estimator-variance comparison.Source: D Reproducibility of our own signal · p. 18

E Does the gain come from the signal or from the allocator?

  • Against random plans, drift allocation ranked at the 100.0th percentile on all five development models; against permuted sensitivities, it was strong on four and ambiguous on one.The permutation null preserves the sensitivity values and allocation setup but randomly reassigns scores to modules.Source: E Does the gain come from the signal… · p. 18
  • On EmbeddingGemma, even scrambled sensitivities beat uniform allocation, so heterogeneous spending itself explains some gain; drift still ranked at the 89.5th percentile.The paper does not count this model as evidence that the criterion itself is doing the work.Source: E Does the gain come from the signal… · p. 18

F How much score agreement does a reproducible allocation need?

The study calibrated how score agreement translates into repeatable plans, separating allocator brittleness from sensitivity-estimator instability. Plan agreement falls sharply when score correlation drops below roughly 0.9.

F How much score agreement does a reproducible allocation need?

  • At rank correlation 0.875, independently perturbed scores produced identical plans 56.2% of the time; at 0.730, exact agreement fell below one in ten.The test perturbed measured drift scores from four development cells and ran both perturbations through the unchanged allocator.Source: F How much score agreement does a… · pp. 18–19
  • The control places the allocator’s agreement knee near ρ ≈0.9, where score agreement stops reliably determining an allocation.The authors conclude that InfoQ-X’s zero exact-plan rate is not distinctive: its best rank agreement was 0.530, near a control that agreed on only 6.2% of plan pairs.Source: F How much score agreement does a… · pp. 18–19

G Is twelve allocation units an arbitrary choice?

Changing allocation granularity barely changed the sensitivity ordering, but retention depended on the number of units. Twelve units were the coarsest tested setting that improved on uniform allocation.

G Is twelve allocation units an arbitrary choice?

  • Pairwise sensitivity rank correlations across six, twelve, and twenty-four units were at least 0.996 after projection to a common frame.The check held other settings fixed and was run on E5-base, not the untouched confirmation models.Source: G Is twelve allocation units an… · p. 19
  • At six, twelve, and twenty-four units, retention gains over uniform were -0.072, 0.397, and 0.276 points, respectively.Twenty-four units cost 3.7× the six-unit probe time; twelve was the coarsest tested setting fine enough to help, not a tuned optimum.Source: G Is twelve allocation units an… · p. 19
  • Probe-to-probe transport exceeded 0.99 at all three granularities, while six units was too coarse to beat uniform allocation.The authors attribute the smaller gain at twenty-four units to spreading the budget across units too small for the integer program to use effectively.Source: G Is twelve allocation units an… · p. 19

H Implementation audit

The implementation audit corrected rank-statistic bugs and recomputed stored correlations without crossing preregistered thresholds. Supporting tables document confirmation, transport, composition, and allocation comparisons under specified protocols.

H Implementation audit

  • Three rank-statistic bugs were corrected and 146 stored correlations recomputed; the largest change was 0.0008 in ρ, with no preregistered threshold crossed.InfoQ-X seed-pairwise correlations could not be recomputed because only summary statistics, not per-seed vectors, were stored.Source: H Implementation audit · p. 19

I Supporting tables

  • Table 4 reports the LieQ-adapt arm selected in advance, while the best-of-fifteen held-out choice is reported separately as a post-hoc upper bound.The pre-specified arm uses an isotropic untrained reference, LieQ’s per-head split, and layer-band scoring with vocabulary bands uniform.Source: Table 4 · p. 20, I Supporting tables · p. 20
  • Table 5 describes frozen-protocol confirmation on three untouched embedders, comparing held-out NDCG@10 at the same packed-byte budget and classifying usability by a 0.5× threshold.The strata are graded, cliff, or dead according to whether both, one, or neither arm remains usable.Source: Table 5 · p. 21
  • Table 6 independently re-evaluates untouched B3 cells on another machine and reports score differences, drift measures, and paired-bootstrap drift-minus-uniform NDCG@10.Its bootstrap uses 5,000 resamples over 323 queries.Source: Table 6 · p. 21
  • Tables 7 and 8 specify transport tests across probe and held-out corpora; Table 8 separately marks collapsed cells rather than treating their plan disagreement as evidence of stability.The four-probe test evaluates plans on SciDocs, absent from the probe set.Source: Table 8 · p. 21, Table 7 · p. 21
  • Table 9 evaluates additive drift composition using inversion rates and calibration residuals, while Table 20 reports correlation and relative residual measures.The relative residual can be inflated by a small denominator; Table 9’s nMAE/IQR is scale-free.Source: Table 9 · p. 22, Table 20 · p. 25, I Supporting tables · p. 25
  • Tables 10–15 report held-out NDCG@10 for search at a 100-configuration budget, InfoQ-X stability, pooling, calibration-resampling sensitivity, allocator-structure nulls at B3, and drift-score perturbation response across four development cells at B3.Table 14 compares permuted-sensitivity gains over uniform with the drift allocation’s position within each null distribution. Table 15 compares plans from two independent perturbations of real scores at target rank correlations.Source: Table 10 · p. 22, Table 11 · p. 22, Table 12 · p. 23
  • Tables 17–22 document LieQ and static-geometry variants, quality-stratified drift correlations, budget-preserving neighbours, and the compression frontier.Table 22 reports realized packed bits per weight because nominal budget names cost different amounts across models; collapse and saturation are separately defined.Source: Table 17 · p. 24, Table 18 · p. 25, Table 19 · p. 25
Loading 2610.09227v1…