PaperlayerView exact revision on arXiv

Source-linked AI summary

MJev-Spatial: System One Spatial Intelligence via Native Attention Matching

Luozheng Qin, Yuhan Zhu, Changhai Zhou, Qian Qiao, Jun Gao, Yuhua Zhou, Harry Zhang, Xiangyu Zeng, Yuxuan Liu, Kaijie Chen, Andrew Chen, Pony Ma

arXiv:2610.22276v1cs.CV

TL;DR

Vision-language models often generate and parse text for compact spatial answers, adding decoding overhead. MJev-Spatial classifies finite candidates using pretrained attention projections, and it achieves the highest average score among eight evaluated models while outperforming its base model on all twelve benchmarks.

  • Problem

    Autoregressive VLMs generate spatial-answer tokens sequentially and require parsing, even when the output is a choice, numerical value, or image location.

    Source: §1 · p. 2

  • Method

    MJev-Spatial classifies supplied options, numerical bins, or image regions using the backbone’s pretrained query and key projections, with refinement stages for numerical and point predictions.

    Source: §1 · p. 3, §3.1 · p. 6

  • Results

    65.64% average score exceeds Qwen3-VL-8B-Instruct’s 62.79%; MJev-Spatial ranks first on seven of twelve benchmarks and outperforms its base model on all twelve.

    Source: §4.2 · p. 9

  • Takeaways & Limitations

    Across the reported benchmarks, MJev-Spatial combines the highest average score with the lowest mean query latency.

    Source: §4.2 · p. 9

    Point refinement costs two additional image encodings per point, and the geometric error bound does not guarantee prediction accuracy.

    Source: §3.1 · p. 6

Abstract

from arXiv · show

Spatial intelligence lets an agent perceive and act in the physical world, and many of its judgments must be made at interaction rates: which spatial relation holds, how far an object is, or where to place it. Vision-language models (VLMs) produce these answers by free-form text generation, paying for autoregressive decoding and answer parsing even though the output space is compact: one of a few options, a scalar, or a 2D point. In light of this, we introduce MJev-Spatial, a System 1 multimodal decision model that tackles all three types of spatial understanding tasks within a unified classification framework. Specifically, MJev-Spatial treats each candidate as a class and scores it using the backbone's pretrained query and key projections, without adding parameters. The candidates are the given options for choice tasks, or value bins and image regions refined through coarse-to-fine classification for numerical and point prediction, respectively. At inference, each classification decision takes a single forward pass and generates no answer tokens. To train on heterogeneous spatial data efficiently, we supervise all coarse-to-fine stages jointly with teacher forcing in one forward pass, and combine this with length-aware global batching and cross-GPU load balancing, raising training throughput by 30.8% over the original batching strategy. Built on Qwen3.5-4B and trained on 12.3M question-answer pairs converted from existing spatial datasets, MJev-Spatial achieves the highest overall score (65.64%) among eight tested models, ranks first on seven of twelve benchmarks, and outperforms its unmodified autoregressive base model on all twelve, while running 2.5 times faster than the fastest autoregressive baseline and 5.6 times faster than the base model.

1 Introduction

MJev-Spatial replaces generated spatial answers with classification over finite candidate sets, reusing pretrained attention projections across choice, numerical, and point tasks. It reports the highest overall score among eight evaluated models and faster inference than autoregressive comparators.

1 Introduction

  • Autoregressive VLMs generate and parse spatial-answer tokens, motivating direct decisions over finite candidate sets for options, values, and locations.The paper identifies sequential decoding and parsing as inference overhead and describes textual formatting as supervision beyond the task-relevant prediction.Source: §1 · p. 2
  • The framework classifies supplied options, numerical bins, or image regions using pretrained query and key projections, without adding prediction heads or trainable parameters.Numerical and point tasks refine selected bins or regions through additional classification stages.Source: §1 · pp. 2–3
  • Joint teacher-forced supervision trains all stages in one forward pass, while length grouping and cross-GPU balancing improve training throughput by 30.8%.The audited conversion pipeline supplies 12.3M QA from existing spatial annotations.Source: §1 · p. 2
  • Across twelve benchmarks, MJev-Spatial achieves the highest overall score among eight evaluated models, leads on seven benchmarks, and outperforms its unmodified autoregressive base model on all twelve.Across all twelve benchmarks, it achieves 2.5× and 5.6× speedups over the fastest autoregressive baseline and the base model, respectively.Source: §1 · p. 2

2 Related Work

Related work spans direct decisions over declared answer spaces, spatial VLMs that generate coordinates as text, and discrete prediction or iterative visual search. MJev-Spatial uses a shared candidate-classification mechanism across spatial output types.

2 Related Work

  • Jev predicts over declared answer spaces [TypeSafe AI, 2026], while KEV scores candidates with learned query and key projections; MJev-Spatial uses the backbone’s pretrained projections.Other work studies multimodal decision interfaces, shared-context execution, and skill or action selection.Source: §2 · p. 3
  • AWT enriches class descriptions and weights augmented views, while Self-TPT adapts prompts for new classes to avoid per-image inference optimization.These are complementary routes to efficient vision-language classification.Source: §2 · p. 3
  • SpatialVLM and SpatialRGPT learn spatial relations and numerical answers, while Molmo, RoboPoint, and RoboRefer support pointing through coordinate text.MolmoPoint instead selects visual patches and refines locations with specialized grounding tokens.Source: §2 · p. 3
  • DORN uses logarithmically spaced depth intervals, and visual-search methods refine regions; MJev-Spatial scores regions as classes and re-encodes crops for later decisions.Its length representation uses logarithmic coarse intervals followed by linear refinement.Source: §2 · p. 3

3 Method

MJev-Spatial casts spatial answering as classification over task-specific candidates, unifying options, numerical values, and image locations. Its method combines pretrained attention-based candidate scoring with coarse-to-fine prediction and joint supervision of refinement stages.

3.1 Unified Classification Framework for Spatial Intelligence

  • A shared candidate-classification framework handles supplied options, numerical intervals, and image regions; choices end after one decision, while numerical and point outputs use refinements.Each decision scores current candidates in one forward pass, without answer-token generation or text parsing.Source: §3.1 · p. 4
  • Numerical prediction selects a coarse interval and then a fine cell, using log1p-spaced length intervals and explicit zero or overflow exits.The quantity in the question determines the unit and codebook.Source: §3.1 · p. 5
  • Point prediction makes three nine-way decisions, re-encoding original-image crops after the first two; its 27 × 27 grid’s 1/54-axis bound assumes the correct cell.Coordinates continue to refer to the original image, and an earlier incorrect decision can exclude the correct region.Source: §3.1 · pp. 5–6

3.2 Model Architecture

  • The final attention layer’s pretrained query and key projections score contextualized candidates, using a shared scorer without an added prediction head.The decision state supplies the query and candidate states supply keys; their similarities determine the selected candidate.Source: Figure 3 · p. 6, §3.2 · p. 7

3.3 Training Recipe

  • Training converts existing spatial annotations into 12.3M question–answer pairs and jointly supervises refinement stages through teacher forcing in one backbone pass.Averaging stage losses gives each original question–answer pair equal total weight.Source: §3.3 · pp. 7–8
  • Length grouping and longest-work-first cross-GPU assignment balance padded sequence and visual-token work while preserving optimizer-batch membership and weights.The expanded sequences are grouped by length within each global optimizer batch before microbatches are distributed across GPUs.Source: §3.3 · p. 8

4 Experiments

Experiments compare MJev-Spatial’s spatial quality and latency against seven autoregressive VLMs, then test scoring, refinement, numerical, decoding, and scheduling choices.

4.1 Experimental Setup

  • The evaluation compares spatial prediction quality and latency against seven autoregressive VLMs across twelve benchmarks, alongside controlled design and training-efficiency studies.Model-design ablations use a fixed 72K-QA mixture and common training settings; the scheduling comparison uses a historical 72K corpus.Source: §4 · p. 8, §4.2 · p. 9, §4.1 · p. 9

4.2 Main Result

  • 65.64% average score leads the eight-model comparison, with MJev-Spatial first on seven of twelve benchmarks and faster than both autoregressive references.It also outperforms its base model on all twelve benchmarks; mean query latency is 160.6 ms, with 2.5× and 5.6× speedups over the fastest autoregressive baseline and base model, respectively.Source: §4.2 · p. 9

4.3 Ablation Studies

  • 51.00% RefSpatial accuracy and 0.3341 m length-estimation MAE are native Q/K’s best results, though no scorer wins consistently across tasks.The position classifier does better on Where2Place and RoboSpatial pointing; native Q/K is retained for its lower parameter count and training cost.Source: §4.3 · p. 9, Table 3 · p. 9
  • 0.3202 m VST-300 dev MAE comes from the 216-cell codebook, versus 0.3277 m with finer bins; conditional expectation raises MAE to 0.3300 m.The study uses fixed coarse intervals and argmax decoding for the candidate-count comparison.Source: §4.3 · p. 10, Table 4 · p. 10
  • 51.00% RefSpatial and 56.84% Point-Bench are crop re-encoding’s highest scores among four visual strategies.Removing crops only at inference drops RefSpatial accuracy from 51.00% to 9.00%; training without crops recovers 46.50%, indicating the importance of matching training to visual input.Source: §4.3 · p. 10, Table 5 · p. 10
  • 32.14% Point-Bench counting success with connected-component decoding exceeds 12.76% with argmax; task-selected decoding raises overall Point-Bench from 50.76% to 54.63%.The hierarchical crop model retains stronger RefSpatial and Point-Bench performance than the one-stage alternative.Source: §4.3 · p. 10
  • 38.78% Point-Bench counting success is reached by replacing repeated continue/stop decisions with a spatial probability tree.Root pruning keeps success unchanged while reducing mean decision nodes from 91 to 51.10 and inference time from 4.716 to 2.638 seconds per question, a 1.79× speedup.Source: §4.3 · p. 11

4.4 Training Speed

  • 30.78% higher throughput than the original strategy is achieved by global length grouping with cross-GPU balancing, reaching 129.54 QA/s.This eight-GPU study measures short-run throughput and does not evaluate downstream quality or convergence.Source: §4.4 · p. 11, Table 6 · p. 11

5 Conclusion

MJev-Spatial applies direct candidate classification to option selection, numerical estimation, and image pointing. The final model leads the reported comparison while delivering faster inference than autoregressive baselines.

5 Conclusion

  • The final Qwen3.5-4B-based model achieves the highest overall score, leads seven of twelve reported metrics, and exceeds its autoregressive base model on all twelve.Its mean batch-one query latency is 160.6 ms across the twelve benchmarks, with 2.5× and 5.6× speedups over the fastest autoregressive baseline and base model under the reported timing protocol.Source: §5 · p. 11
  • MJev-Spatial uses direct classification over candidate sets for options, numerical values, and image locations, with hierarchical refinement for numerical and spatial precision.Tree expansion extends point prediction without further training.Source: §5 · p. 11
  • An audited 12.3M-QA corpus and joint stage supervision support training across the tasks; length grouping and cross-GPU balancing improve throughput by 30.8% in a controlled eight-GPU study.The conclusion presents the results as supporting direct candidate decisions for the spatial tasks studied.Source: §5 · p. 11

A Additional Method Details

The appendix specifies the model’s decision paths, candidate representations and scoring, multi-point decoding, and conversion and validation of spatial training data.

A.1 Autoregressive Reference and Decision Paths

  • The decision formulation replaces autoregressive answer-string generation and parsing with a finite decision tree whose transitions are supervised during training.Reachable states form a tree of depth at most three; each decision supports up to 32 candidates, with padded slots masked before normalization.Source: §A.1 · p. 17

A.2 Input Sequence and Candidate Representations

  • Each question-answer example places images, question, candidate descriptions, and a decision marker in a causal sequence, with later stages teacher-forced in one training pass.A stage’s decision marker precedes later choices and crops, preventing future crops from being shown to earlier decisions; inference constructs each later stage after the current prediction.Source: §A.2 · p. 17

A.3 Candidate Scoring with Pretrained Query and Key Projections

  • Candidate scores reuse the final attention layer’s pretrained query and key projections to compare decision and candidate states, without an additional prediction head.The decision state supplies the query and each candidate state supplies a key; the resulting logits are normalized over legal candidates.Source: §A.3 · pp. 17–18

A.4 Multi-Point Decoding Rules

  • Multi-point decoding expands a three-level conditional probability tree into up to 729 cells, then groups above-threshold connected cells into point predictions.The fixed threshold and root cutoff are set without test labels; unvisited cells remain unknown and omitted probability mass is not renormalized.Source: §A.4 · pp. 18–19

B.1 From Spatial Annotations to Classification Supervision

  • Training converts accepted annotations into typed candidates and targets, excluding ambiguous options, unclear numerical answers, and unsupported point or relation labels.Options are retained when exactly one matches the source answer; numerical labels require an unambiguous value and unit, while point labels require verified coordinates.Source: §B.1 · p. 19

B.2 Data Quality Checks

  • Quality checks review converted numerical and option examples, cleanup cases, rendered training examples, and selected pointing questions.The checks include a targeted review of 66,574 QA with known option-conversion issues, recovering 65,902 and excluding 672 ambiguous cases.Source: §B.2 · p. 19

B.3 Duplicate Filtering and Evaluation Separation

The data appendix describes image-level duplicate filtering and reports the final corpus composition, task distribution, and numerical-label ranges.

B.3 Duplicate Filtering and Evaluation Separation

  • Near-duplicate filtering uses hash, aspect-ratio, and thumbnail-color thresholds, but does not establish independence of physical scenes or neighboring video frames.Examples matching public evaluation images or the stated near-duplicate criterion are excluded; examples with conflicting targets are also removed.Source: §B.3 · pp. 19–20

B.4 Final Corpus and One-Pass Distribution

  • The final corpus contains 12.3M QA: 7.1M choices, 3.5M points, 1.6M lengths, 120K angles, and 38K counts.RefSpatial-derived data account for approximately 89.5% of the corpus and provide all current point supervision; the natural task distribution is retained without repeated sampling.Source: §B.4 · p. 20, Table 7 · p. 20, Table 8 · p. 20
  • The length-label maximum is 8520.1 m, and 12K labels exceed the old 200 m limit; count labels reach 1002.The length median and 99th percentile are 1.3381 m and 163.1174 m, while the count 99th percentile is 49.Source: §B.4 · p. 20

C.1 Matched 72K Training Recipe

The appendix details matched small-scale training, the full-corpus run configuration, and scheduling intended to reduce padding and cross-device workload imbalance.

C.1 Matched 72K Training Recipe

  • The matched 72K-QA comparison fixes examples, order, exposure budget, and optimizer settings across three task subsets, with a one-pass recipe taking 1,413 s for the selected native model.The subsets cover 25K choice, 22K length, and 25K pointing questions; the comparison uses eight L20D GPUs and 2,252 optimizer updates.Source: §C.1 · pp. 20–21

C.2 Large-Scale Training Configuration

  • The full-corpus configuration completes 24,032 updates and 12,304,058 real QA exposures in one pass.It uses a flat permutation without task resampling, length and pointing refinement settings, and global length grouping with rank balancing.Source: §C.2 · p. 21, Table 9 · p. 21

C.3 Scheduling and Distributed Execution

  • Global batches are grouped by expanded sequence length, then assigned across eight ranks longest-work-first to reduce padding and workload imbalance.Worker losses are weighted by each microbatch’s share of actual examples so averaging across workers recovers the intended objective.Source: §C.3 · p. 21

D.1 72K Pilot Evaluation Scope and Scoring

The pilot evaluation specifies how the historical benchmark union and local autoregressive baseline were measured, then reports final-model latency and a controlled comparison of candidate readouts. Across twelve benchmarks, MJev-Spatial has lower mean latency than the two reported autoregressive models, while native mean readout improves some metrics and loses on others.

D.1 72K Pilot Evaluation Scope and Scoring

  • The historical pilot union covers eight public datasets in 7.5K execution requests, with SAT’s two-order protocol yielding 150 fewer independent questions than requests.The pilot also evaluated VST-300, a fixed 300-example internal development set for length prediction.Source: §D.1 · p. 21, Table 10 · p. 22
  • Scoring retains original benchmark options and image order except SAT’s two-order protocol, uses ground truth only offline, and leaves invalid outputs in the denominator.Ground-truth values, evaluation masks, and target counts do not select crops, tree branches, component thresholds, or output point counts.Source: §D.1 · p. 22

D.2 Local Autoregressive Baseline for the 72K Pilot

  • The unmodified Qwen3.5-4B baseline uses local BF16 greedy generation with thinking disabled, and incorrect or invalid cases are not retried according to their labels.Choice outputs are parsed as option identifiers; point predictions use integer coordinates in [0, 1000].Source: §D.2 · p. 22

D.3 Final Inference-Efficiency Protocol

  • 160.6 ms is MJev-Spatial’s mean latency, versus 397.4 ms for Spatial-SSRL-Qwen3VL-4B and 894.1 ms for Qwen3.5-4B across twelve benchmarks.The mean gives each benchmark equal weight; timing uses one NVIDIA B200 at batch size one and excludes warmup and model loading.Source: §D.3 · p. 23, Table 11 · p. 23

E.1 Readout Controls and Optimization Alignment

  • The matched study compares a position classifier, a newly projected KEV pointer, and native mean using pretrained attention projections at an earlier boundary.The KEV pointer adds 1,311,232 parameters; the comparison varies both representation source and downstream computation retained.Source: §E.1 · p. 23
  • Native mean improves CV-Bench by 1.58 points, RoboSpatial VQA by 7.72, and RefSpatial by 2.00, but loses 8.01 points on Where2Place and 3.28 on RoboSpatial pointing.MAE falls by 8.93%; relative to KEV, native mean improves RefSpatial and length error but is not uniformly better on choice tasks.Source: §E.1 · p. 23

E.2 Pilot Quality and Task-Specific Failures

The historical pilot comparison finds native mean close to the autoregressive baseline on some aggregate benchmarks, while category-level results expose substantial task-specific gains and losses. Point-Bench’s counting shortfall accounts for most of its macro-score gap.

E.2 Pilot Quality and Task-Specific Failures

  • Native mean scores 85.38 versus 85.56 for AR on CV-Bench and 63.04 versus 62.98 on BLINK, while improving RefSpatial from 46.00 to 51.00 and RoboSpatial pointing from 50.00 to 62.30.These are results on the historical 72K evaluation union.Source: §E.2 · p. 24
  • Point-Bench counting success is 38.78 versus AR’s 66.84, accounting for 5.61 of the 6.01-point macro gap; the other four categories differ by about 0.50 points on average.The shortfall centers on estimating and separating multiple targets; the training mixture is mostly single-point and the decoder composes conditional single-point decisions.Source: §E.2 · p. 24
  • On RefSpatial, placement rises from 29.00 to 47.00 while location falls from 63.00 to 55.00; RoboSpatial VQA compatibility rises from 35.24 to 52.38.BLINK’s near-equal macro scores conceal a 33.08-point multi-view gain and a 29.17-point counting decrease; MMSI scores 27.70 versus AR’s 34.20.Source: Table 13 · pp. 24–25, §E.2 · p. 24

E.3 Readout Depth and Head Aggregation

The depth and aggregation studies find that spatial readout quality depends on substantial contextual processing, but the best layer and aggregation vary by task. L32 mean is retained as a balanced default, while head analyses show strong agreement alongside measurable effects from single-head removal.

E.3 Readout Depth and Head Aggregation

  • L32 mean is retained as a balanced default: concat has lower MAE (0.3297 versus 0.3341 m) and better MMSI (28.90 versus 27.70), but loses on several other benchmarks.Merge improves two placement-oriented point scores relative to mean but reduces several choice scores and worsens MAE.Source: §E.3 · p. 26
  • Equal-weight late-layer fusion does not consistently transfer layer strengths: L32+L28 raises Where2Place from 59.98 to 65.98 but worsens length MAE from 0.3344 to 0.3875 m.All three fusions worsen MAE, and the three-layer variant reduces RoboSpatial VQA from 68.06 to 65.62.Source: Table 14 · p. 27, §E.3 · p. 27
  • At L4, mean scores 51.41 CV-Bench and 4.50 RefSpatial; at L12 these rise to 79.42 and 37.00, and at L32 to 85.38 and 51.00.Other task optima occur at L28 for RefSpatial, L24 concat for Where2Place, and L28 merge for RoboSpatial pointing.Source: §E.3 · p. 26, Table 15 · p. 27, Table 16 · p. 28

E.4 Head Activity, Redundancy, and Causal Interventions

  • H02 has the strongest RMS and H11 the weakest across task groups, while centered pairwise head-logit cosine averages exceed 0.994 in all three groups.The authors interpret the shared candidate direction but differing amplitudes as more consistent with redundancy than clear task-specific specialization.Source: §E.4 · p. 29, Figure 5 · p. 29
  • No head meets the predeclared near-zero screen: the largest classification change is 0.20 points, Point-Bench can decrease by 0.5092 points, and length MAE changes from −0.003992 to +0.000279 m.Single-head masking can change the logit scale used by thresholded counting maps; joint head pruning was not evaluated.Source: §E.4 · p. 29, Table 17 · p. 30

F.1 Visual Refinement: New Pixels, Replayed Features, or Direct Tokens?

The visual-refinement experiments favor crop re-encoding over removing crops, replaying features, or direct visual-token candidates, with performance depending on training and decoding choices. For multi-point counting, probability-tree decoding outperforms continue/stop decisions, while root pruning reduces computation with little omitted probability mass.

F.1 Visual Refinement: New Pixels, Replayed Features, or Direct Tokens?

  • Removing crops from fixed crop-trained weights drops RefSpatial from 51.00 to 9.00, Where2Place from 59.98 to 32.00, and RoboSpatial pointing from 62.30 to 20.49.An original-image hybrid prefix cache leaves these aggregate scores unchanged; the comparison also changes the training-time input distribution.Source: Table 18 · p. 30, §F.1 · p. 30
  • Matched original-image training recovers RefSpatial to 46.50 and Point-Bench macro to 52.55, but does not fully replace newly resolved local visual information in this setting.This variant removes crop placeholders in training and inference while retaining the selected rectangle and region options in text.Source: §F.1 · p. 31
  • Neither feature-replay scheme closes the crop re-encoding gap: ViT replay reaches 47.00 RefSpatial and 52.87 Point-Bench, while KV replay raises Where2Place to 68.95 but lowers RoboSpatial pointing to 54.92.KV replay also increases length MAE to 0.4097 m; the schemes expose region-focused representations without new pixel encoding.Source: §F.1 · p. 31
  • Direct visual-token prediction uses one decision over merged image-grid cells, so its spatial resolution depends on image aspect ratio and token budget and cannot separate instances sharing one cell.Ground-truth points map to occupied cells with equal target mass across distinct cells.Source: §F.1 · p. 31
  • Crop re-encoding remains the final representation: hierarchical mean scores 51.00 RefSpatial and 56.74 Point-Bench, exceeding task-routed visual tokens (45.50 and 54.63).Task-based routing restores single-point categories after components hurt them, but does not close the gap to hierarchical mean.Source: Table 20 · p. 32, §F.1 · p. 32

F.2 Multi-Point Decoding: Probability Trees versus Stop Decisions

  • The continue/stop extension reaches 19.39% success and count MAE 11.64 on 196 counting questions, with long overproduced sequences inflating mean predicted count to 10.48.Fixed peak thresholds also change success: 30.10, 38.78, and 37.24% at 5%, 10%, and 20%, respectively.Source: §F.2 · pp. 32–33, Table 22 · p. 33
  • Probability trees reach 38.78% Point-Bench counting success and 46.43% exact-count accuracy; root pruning reduces mean nodes from 91 to 51.10 and time from 4.716 to 2.638 seconds.Pruning omits 0.588% probability mass on average, a 1.79× time ratio under this execution setup; the paired table uses fixed L32 mean weights without retraining.Source: Table 21 · p. 32, §F.2 · p. 32, Figure 6 · p. 33
  • The later consolidated rerun has slightly different full-tree scores and timings from the paired pruning control, so the records are retained separately.In the paired run, predicted counts agree on 195 of 196 questions and success agrees on all 196; the count discrepancy is associated with changed numerical execution rather than root-region omission alone.Source: §F.2 · pp. 32–33

G.1 Numerical Design: Coverage, Granularity, and Decoding

The numerical codebook experiments trade coverage and resolution against candidate count, input cost, and prediction difficulty. The 4/8 argmax model reports lower MAE than the compared codebooks, while decoding and learned-model results expose trade-offs that oracle reconstruction alone does not capture.

G.1 Numerical Design: Coverage, Granularity, and Decoding

  • The historical codebook covers all old 72K length labels but overflows on 12K expanded-corpus labels, so coverage is a constraint before finer precision.The oracle audit compares codebooks on 1.6M accepted lengths and assumes correct cell selection.Source: §G.1 · p. 33
  • On training-label reconstruction, the 30-by-32 proposal reaches 0.0391 m oracle MAE and 21.39 m maximum error, versus 1.6548 and 8320.10 m for the historical codebook.These are oracle comparisons assuming every classification decision is correct; task-specific routing is not adopted because it adds question-type routing and has poorly supported rare categories.Source: §G.1 · pp. 33–34
  • The 30-by-32 learned candidate text adds 527.43 tokens on average, while its oracle MAE falls from 0.006591 to 0.003891 m and P99 error from 0.045625 to 0.023542 m.Worst-case error rises from 0.7475 to 0.9800 m, and the longest checked scalar sequence is 2,068 tokens within the 4,096-token budget.Source: §G.1 · p. 34
  • The learned L32 model lowers VST-300 MAE from 0.335420 to 0.327732 m, only 2.29%, while lowering all five classification benchmark scores.The larger oracle gain does not translate directly into prediction accuracy, motivating reduced fine-stage difficulty rather than still more resolution.Source: §G.1 · p. 34
  • The 4/8 codebook reduces positive leaves from 960 to 216 and meets the specified oracle tolerance on every audited label without overflow, while increasing oracle MAE from 0.039102 to 0.154041 m.Its maximum oracle error rises from 21.39 to 59.55 m; 108K positive labels use four fine candidates, 1.5M use eight, and 382 terminate at zero.Source: §G.1 · p. 34
  • 0.320207 m MAE is 4.54% below the historical control and 2.30% below 30-by-32; 10 cm accuracy is 59.00%.All nine public primary scores improve over 30-by-32; against the historical codebook, SAT and four pointing scores improve while four classification scores remain slightly lower. Mean relative error favors the historical codebook: 24.64% versus 29.03%.Source: §G.1 · p. 35
  • Fine-stage conditional expectation raises MAE to 0.329955 m and reduces 10 cm accuracy to 55.67%, despite a 70.81% hit rate within 25% relative error.The same-logit comparison leaves non-length outputs unchanged; the final model uses argmax at both length stages.Source: §G.1 · p. 35

H.1 Length Grouping Separates Padding from Rank Imbalance

The scheduling study finds that reducing padding alone can worsen rank imbalance, while global grouping with balancing improves measured throughput. Expanded-data profiling then selects a feasible global batch and microbatch recipe, with additional accumulation providing a modest throughput increase.

H.1 Length Grouping Separates Padding from Rank Imbalance

  • 129.54 QA/s from global grouping plus rank balancing is a 30.78% throughput increase over original order, with padding reduced to 1.77%.The controlled study used a fixed global batch of 256, microbatch 8, four accumulation steps, eight L20D GPUs, and the old 72K data/codebook; timing excluded eight warmup updates.Source: §H.1 · pp. 35–36, Figure 7 · p. 36
  • Local grouping raises throughput from 99.06 to 116.91 QA/s and cuts padding from 35.17% to 12.36%, but increases rank ready-time spread from 115.0 to 306.6 ms.Tail communication and wait time also rise from 39.5 to 157.6 ms, illustrating that per-rank padding reduction can leave unequal workloads at synchronization.Source: §H.1 · p. 36
  • The 128-QA buffer lowers padding to 1.03% and achieves 129.39 QA/s, but raises step-time coefficient of variation from 6.07% to 19.77%, indicating diminishing returns from further packing.When pairs of updates are aggregated into equal 512-QA units, variation is 6.30% versus 4.61% without the buffer; no downstream accuracy evaluation was performed in these short runs.Source: §H.1 · p. 36

H.2 Microbatch Feasibility and Global Batch Size

  • Microbatch 16 and 32 fail with OOM in both tested global256/512 conditions; successful configurations use approximately 242 decimal GB per rank.The additional global384/micro12 trial also fails.Source: Table 26 · p. 36, Table 27 · p. 37, §H.2 · p. 37
  • 139.78 QA/s at global768 exceeds global512's 135.48 QA/s by 3.17% over ten measured updates.The final recipe selects global512/micro8/accumulation8; all 16 executed global768 updates fell within the zero-backbone-LR period, although backward computation and Adam updates still ran.Source: §H.2 · pp. 36–37

Related papers

Loading 2610.22276v1…