PaperlayerView exact revision on arXiv

Source-linked AI summary

Visual Parallel Search: Learning to Search High-Resolution Images with Parallel Tile Inspection and Adaptive Zoom

Xijia Tao, Yihua Teng, Xinyu Fu, Cheng Gong, Ziru Liu, Xudong Xie, Rui Liu, Lingpeng Kong

arXiv:2609.37002v2cs.CVcs.AI

TL;DR

High-resolution visual question answering requires localized evidence that global resizing can erase, while sequential zooming searches without a reliable overview. VPS combines parallel, question-conditioned tile inspection with adaptive zoom and trains its controller and tile-reader roles. It improves mean accuracy over dedicated zoom-only search in 14 of 15 same-model comparisons and supports a trainable decomposition for visual search.

  • Problem

    High-resolution images require localized evidence that global resizing may erase, while sequential zooming can make inspection an expensive search without a reliable overview.

    Source: §1 · p. 1

  • Method

    VPS has a main agent use parallel, question-conditioned tile readers to gather evidence and then select a precise or merged region for adaptive zoom.

    Source: §1 · p. 1, §3.2 · p. 5

  • Results

    14 of 15 same-model comparisons show higher mean accuracy for VPS than dedicated zoom-only search, and SFT raises observed accuracy on all five benchmark splits.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    The findings support VPS as an inference-time scaffold and a trainable decomposition, while identifying efficient evidence acquisition and reliable termination as controller-optimization targets.

    Source: §6 · p. 9

    The paired GRPO surrogate is a biased, lower-variance semi-gradient relative to the full joint return when weights are shared, and frozen tile prompts omit an occupancy derivative for on-policy tiles.

    Source: §2.6 · p. 5

Abstract

from arXiv · show

High-resolution visual question answering often fails because a multimodal model does not acquire the small, spatially localized evidence needed to answer a question. Sequential zooming can recover detail, but it asks the main model to choose a region before obtaining a reliable overview. We introduce VPS, a visual parallel-search framework in which a main agent first invokes grid_search to inspect image tiles in parallel with question-conditioned sub-agents, and then adaptively invokes zoom_in on a precise or merged region. The same interface supports both training-free inference and post-training of the main and sub-agents. Across five benchmark splits and three model sizes, VPS improves mean accuracy over dedicated zoom-only search in 14 of 15 same-model comparisons, with gains up to 8.0 points and especially strong improvements for smaller main models. ZoomBench retains an approximately 3.2-point gain at every tested size. We further develop a supervision pipeline with hint-free verification and a paired role-specific GRPO surrogate for learning the controller and tile-reader roles. SFT improves observed accuracy on all five benchmark splits, including a 4.17-point gain on HR-Bench 4K. Role-specific RL further reshapes search behavior: main-only RL reduces mean tool use from 2.65 to 2.11 with similar pass@1 in an internal four-response evaluation, while external accuracy changes are mixed. Joint training reveals an asymmetry between local evidence reading and global search control. Together, these results support VPS as an effective inference-time scaffold and a trainable decomposition for visual evidence acquisition.

1 INTRODUCTION

High-resolution visual question answering requires acquiring localized evidence without losing detail or relying on costly sequential search. VPS uses parallel, question-conditioned tile inspection to guide adaptive zoom, and the paper evaluates both its inference-time performance and trainability.

1 INTRODUCTION

  • Localized labels, chart readings, and evidence from distant regions can be lost in globally resized images or require expensive sequential zooming.The acquisition challenge is to bring relevant evidence into context at a usable resolution.Source: §1 · p. 1
  • Fixed patches can split evidence across boundaries, while a single sequential agent must trade coverage for detail and entangles perception with search control.The cited prior systems organize search around fixed patches or an initial-thumbnail region proposal (Wu & Xie, 2024; Wang et al., 2025; Shen et al., 2025).Source: §1 · p. 1
  • VPS delegates question-conditioned inspection of a regular grid to tile readers, then uses their reports to decide whether and where to zoom.The main agent aggregates evidence and remains responsible for search control and answer synthesis; tile readers specialize in local perception.Source: §1 · pp. 1–2
  • 14 of 15 same-model comparisons favor VPS over dedicated zoom-only search, with larger gains generally for smaller main agents.At 27B, only ZoomBench and HR-Bench 4K exceed replicate spread.Source: §1 · p. 2
  • SFT improves observed accuracy on all five evaluated benchmark splits, while main-only RL reduces mean tool calls from 2.65 to 2.11 with similar internal pass@1.The RL comparison uses an internal four-response evaluation; external accuracy changes are mixed.Source: §1 · p. 2
  • VPS combines a supervised data recipe for main-agent trajectories and tile-reader ROI judgments with a paired role-specific GRPO surrogate.The surrogate makes gradient accounting explicit and analyzes the distinct effects of training the controller and tile reader.Source: §1 · pp. 2–3

2 VISUAL PARALLEL SEARCH

VPS separates global search control from local evidence reading: parallel grid reports guide adaptive zoom, while supervised and role-specific objectives train the two roles. Its paired GRPO surrogate has an explicit boundary relative to the full joint-policy gradient.

2.1 PROBLEM SETUP

  • A visual-search policy produces a trajectory of observations and tool or answer actions, maximizing answer quality under a bounded number of turns and observations.The policy is factored into main and sub-agent roles, which may share weights but use different prompts and response masks.Source: §2.1 · p. 3

2.2 PARALLEL TILE INSPECTION

  • Grid search partitions the image into an n × n grid and dispatches tiles in parallel; each reader returns a compact evidence record, while tile selection remains with the main agent.The reader does not see the gold ROI or other tiles, and the harness aggregates reports into a model-visible overview.Source: §2.2 · p. 3

2.3 ADAPTIVE ZOOM AND REGION AGGREGATION

  • Zoom-in accepts one region or neighboring boxes and returns their smallest enclosing crop, allowing inspection across grid boundaries and of relational evidence.Grid search supplies high recall and zoom supplies spatial precision; the current system uses one grid-search layer.Source: §2.3 · p. 3

2.4 AGENT LOOP

  • The agent starts from a globally resized view and can answer, scan a grid, or zoom until it answers or reaches the turn budget.The final-turn instruction requests an answer when the budget is exhausted.Source: §2.4 · p. 3

2.5 SUPERVISED POST-TRAINING

  • SFT trains the main role on successful tool-use trajectories and the sub role on verified tile judgments, with the selected SFT policy serving as the cold-start checkpoint for orchestration RL.Main examples teach search, report interpretation, region merging, and stopping; sub examples teach structured local judgments and evidence localization.Source: §2.5 · p. 3

2.6 PAIRED ROLE-SPECIFIC GRPO

  • Paired GRPO applies separate final-answer and local-evidence objectives to prompt-matched main and tile-reader groups, with the two losses updating shared weights.Unlike the full joint return gradient, the surrogate omits global-return credit for nested tile-reader outputs; with shared weights it is a biased, lower-variance semi-gradient.Source: Figure 2 · p. 4, §2.6 · pp. 4–5

3 DATA

The data pipeline builds main-agent trajectories and tile-level judgments from two source pools, verifies potentially unreliable labels, and controls RL prompt difficulty and role mixture. The audited sources differ substantially in box agreement with human annotations.

3.1 MAIN-AGENT TRAJECTORIES

  • The main-agent RL pool contains 1,997 prompts from ZwZ and VisualProbe after conversion and removal of RL-validation overlaps.The SFT corpus contains 2,197 main-agent trajectories and 4,267 sub-agent examples; the 200-prompt RL validation set overlaps earlier SFT data, so external benchmarks provide transfer evaluation.Source: §3.1 · p. 5

3.2 SUB-AGENT ROI DATA

  • Tile supervision combines ZwZ absolute-pixel boxes with VisualProbe boxes annotated by a model, and a hint-free verifier rechecks selected labels before ROI hints are removed from student prompts.The verifier checks suspicious positives, low-coverage answers, zero-overlap positives, and a random audit sample.Source: §3.2 · p. 5

3.3 DIFFICULTY AND MIXTURE CONTROL

  • RL rollouts are grouped by prompt into all-wrong, learnable, and all-correct sets, with only mixed-reward groups receiving nonzero within-group advantage.Optional filtering and the easier pretrained sub task can shift the effective mixture of main and sub examples.Source: §3.3 · p. 5

3.4 DATA QUALITY LIMITATIONS

  • VisualProbe boxes had median IoU 0.243 with human boxes and were disjoint in 28.9% of cases, while ZwZ boxes had median IoU 0.887.These source differences motivate verification and separate reporting of annotation quality.Source: §3.4 · pp. 5–6

4 EXPERIMENTAL SETUP

The experiments compare grid-plus-zoom VPS with dedicated zoom-only search across visual-search benchmarks, model sizes, and evaluation protocols. Post-training evaluations examine SFT and role-specific RL under defined data and deployment settings.

4.1 BENCHMARKS

  • Five benchmark splits cover TIR, VSTAR, ZoomBench, and HR-Bench 4K/8K; HR-Bench is also evaluated with CircularEval’s four-rotation criterion.Main-text comparisons use full denominators, and the VSTAR ambiguity-exclusion sensitivity is reported separately.Source: §4.1 · p. 6

4.2 MODELS AND SYSTEMS

  • The study tests Qwen3.5 main agents at 4B and 9B and a Qwen3.6 main agent at 27B, generally paired with a same-size Qwen tile reader.Selected configurations use Gemini 3.5 Flash or a 27B Qwen tile reader.Source: §4.2 · p. 6

4.3 METRICS AND STATISTICAL PROTOCOL

  • Judge-all answer accuracy is primary; paired-policy differences use a 95% bootstrap interval over matched examples, with HR-Bench rotations resampled together.Replicated cells generally report means and sample standard deviations over three runs; post-training role comparisons use unseeded runs.Source: §4.3 · p. 6

4.4 POST-TRAINING SETUP

  • The SFT LoRA model initializes orchestration RL, while role comparisons use distinct training recipes and deployment-role configurations.Main-only RL supervises main-generated tokens, though shared weights can also change tile-reader behavior; joint RL combines main-answer and direct tile-response losses.Source: §4.4 · p. 6

5 RESULTS

VPS outperforms dedicated zoom-only search in most same-model comparisons, with larger gains often appearing for smaller main agents. SFT raises observed accuracy across the evaluated splits, while RL changes tool use and role behavior without establishing consistent external accuracy gains.

5.1 PARALLEL SEARCH VERSUS SEQUENTIAL ZOOM

  • VPS improves mean accuracy over dedicated zoom-only search in 14 of 15 same-model benchmark cells; at 27B, only ZoomBench (+3.2) and HR-Bench 4K (+1.3) gaps exceed both arms’ replicate spread.The remaining 27B gaps, including the −0.8 point TIR difference, lie inside replicate variation.Source: §5.1 · p. 7

5.2 WHERE THE SCAFFOLD HELPS

  • On HR-Bench 4K/8K, the grid-versus-zoom advantage falls from +5.0/+7.1 points at 4B to +1.3/+0.5 at 27B; a 27B tile reader adds +1.1/+2.0 points with a 4B main agent.The authors interpret these patterns as evidence that controller-side evidence acquisition and aggregation remain a bottleneck in the tested systems.Source: §5.2 · p. 7
  • VPS has the higher CircularEval mean in all six model–resolution pairs; at 4B, HR-Bench 4K/8K consistency rises from 48.2/40.2% to 59.0/52.8%.CircularEval requires all four option rotations to be answered correctly; grid-arm row accuracy remains higher than this stricter consistency measure.Source: §5.2 · p. 8, Table 2 · p. 8

5.3 SUPERVISED FINE-TUNING

  • SFT raises observed accuracy across all five evaluated benchmark splits, with a matched unseeded HR-Bench 4K gain of 4.17 points (95% question-paired interval [+1.54, +6.75]).The other split comparisons are descriptive.Source: §5.3 · p. 8

5.4 MAIN-ONLY RL AND TOOL USE

  • Main-only RL reduces mean tool calls from 2.65 to 2.11 in the internal four-response evaluation, while pass@1 is 71.25% versus 72.50% with an interval spanning zero.External accuracy changes are mixed: the HR-Bench 4K comparison with the SFT tile reader has a 0.88-point difference and an interval spanning zero; cross-benchmark estimates are descriptive.Source: §5.4 · p. 9

5.5 JOINT TRAINING DYNAMICS AND ROLE ASYMMETRY

  • During joint training, aggregate pass@1 rises while sampled pass@8 stays flat; the changing prompts prevent a fixed-set coverage comparison.In a single unseeded deployment-role matrix, the joint RL tile reader’s 1.38-point HR-Bench 4K difference has an interval spanning zero, whereas joint RL as main lowers accuracy and completion; its mean tool calls rise from 2.22 to 4.35 with the SFT sub-agent fixed.Source: §5.5 · p. 9

6 CONCLUSION

The conclusion presents VPS as an inference-time scaffold and trainable decomposition for high-resolution visual search. The related-work discussion positions it as constrained orchestration with geometrically fixed tiles and separately supervised reader and controller roles.

6 CONCLUSION

  • Across five benchmark splits and three model sizes, VPS improves mean accuracy over dedicated zoom-only search in 14 of 15 cells, with larger gains for smaller main agents.SFT raises observed accuracy across all five splits, including a 4.17-point HR-Bench 4K gain.Source: §6 · p. 9

A RELATED WORK

  • VPS uses homogeneous tile readers and a fixed spatial decomposition, while the learned decisions govern fan-out, evidence aggregation, follow-up inspection, and stopping.Its paired role-specific GRPO surrogate gives the main agent and tile reader distinct learning signals, while treating shared-parameter training as a multi-task surrogate rather than exact end-to-end credit assignment.Source: A RELATED WORK · p. 13

B ANALYSIS AND LIMITATIONS

The section qualifies the evidence for VPS by identifying limits in benchmark coverage, model comparisons, evaluation reliability, and role-specific training. It also reports that exploratory RL interventions and the paired training surrogate do not yet establish reliable discovery or cross-role credit assignment.

B ANALYSIS AND LIMITATIONS

  • The tested Qwen3.5 and Qwen3.6 operating points do not isolate parameter-count effects, although smaller main agents show a larger grid-versus-zoom gap on most benchmarks.ZoomBench is an exception, with a persistent grid advantage consistent with broad spatial-scanning tasks.Source: B ANALYSIS AND LIMITATIONS · p. 14
  • Single-rollout validation confounds policy quality with trajectory variance because an early region choice changes later evidence and the final answer.The authors recommend prompt-paired, multi-rollout evaluation that reports mean accuracy and within-prompt dispersion.Source: B ANALYSIS AND LIMITATIONS · p. 14
  • RL pass@1 rises while sampled pass@8 stays flat, a pattern consistent with concentrating on successes but not measuring coverage change on a fixed set.Held-out checks show small accuracy point estimates, a lower high-coverage pass@4 estimate, and fewer tool calls; sparse end-task reward has not reliably discovered new region-selection strategies.Source: B ANALYSIS AND LIMITATIONS · p. 14
  • Joint RL lowers deployment accuracy by 2.38 points with the SFT sub-agent and 3.88 points with the joint RL sub-agent; only the latter paired contrast excludes zero.Completion falls from 789 to 720 rows with the SFT sub-agent and from 790 to 722 with the joint RL sub-agent, with most additional failures reaching iteration or context limits.Source: B ANALYSIS AND LIMITATIONS · p. 14
  • Exploratory grid interventions raised successful grid use from 9.5% to 21.0%, while accuracy across saved policies fell from 71.5% to 64.5%.The changed loop, mixed-role data, missing synchronous no-shaping control, and SFT-exposed validation set make this a mechanism observation, not an external accuracy gain.Source: B ANALYSIS AND LIMITATIONS · p. 14
  • The paired role-masked surrogate is a biased, lower-variance multi-task semi-gradient, not causal credit assignment across the main/sub-agent boundary.Sub-agent ROI rewards do not directly credit the main agent’s grid-call or region-selection decisions, and shared-backbone role losses can change both deployed policies.Source: B ANALYSIS AND LIMITATIONS · p. 14
  • Engineering confounders, including truncated grid reports, bound interpretation of diagnostic runs; one apparent null grid effect was invalidated when 97.7% of reports were truncated.A matched unfinished-reward pilot favored explicit failure penalization in internal accuracy and completion, but its SFT-exposed evaluation and changing intermediate effects limit the result to an internal diagnostic.Source: B ANALYSIS AND LIMITATIONS · p. 15
  • The benchmark suite is dominated by high-resolution VQA, while several task types, replication settings, and ROI labels have limited or noisy coverage.The joint RL role matrix is a single unseeded pass, the internal RL validation set is SFT-exposed, and perceptual overlap checks are sensitivity analyses rather than proof of contamination or its absence.Source: B ANALYSIS AND LIMITATIONS · p. 15

C.1 TRAINING-FREE REPLICATE PROTOCOLS

The replicate protocols document how training-free runs and their numerators were compiled, including unequal seed conventions and failed trajectories. The Gemini tile-reader results vary by benchmark, while some controls are excluded because their summaries could not be reconciled.

C.1 TRAINING-FREE REPLICATE PROTOCOLS

  • The training-free policies use different prompts and iteration limits, and each grid call dispatches n2 tile-reader requests, so comparisons do not isolate the grid operator or match total inference compute.Grid-plus-zoom uses up to 15 main iterations and image dimensions in the input; dedicated zoom-only uses up to 20 iterations and no dimension injection.Source: §C.1 · p. 15
  • Table 4 archives three run numerators per cell, but seed conventions differ across grid and dedicated zoom-only arms, including the 9B first-run baseline.Failed gold examples remain in the denominator; reported failure counts are not additional errors to subtract from accuracy numerators.Source: §C.1 · p. 15, Table 4 · p. 16

C.2 GEMINI TILE READERS AND DIRECT ANSWERS

  • Gemini 3.5 Flash raises the 27B mean by 1.7 and 2.3 points on TIR and VSTAR, respectively, while ZoomBench changes by about 0.1 point.The V2 entries are three seeded runs; V1 and V2 prompts differ and are not matched comparisons.Source: §C.2 · p. 16
  • No-search TIR controls are excluded because their available summaries disagree on numerators and lack a reconciled scored run; VTC is omitted because its historical aggregate used rule scoring.Source: §C.2 · p. 16

C.3 POST-TRAINING CHECKPOINT LEDGER

The checkpoint ledger distinguishes separate post-training lineages and evaluation passes. The accompanying VSTAR and earlier main-only RL reconciliations show that denominator choices and paired outcomes matter when interpreting small or mixed accuracy differences.

C.3 POST-TRAINING CHECKPOINT LEDGER

  • The selected joint RL policy is a separate, later lineage from the early mixed RL sub-agent, despite sharing an SFT ancestor.The formal role-matrix cells use the first HR-Bench 4K pass; a second pass is reported separately, and the selected joint RL matrix remains a single pass.Source: §C.3 · pp. 16–17
  • The HR-Bench 4K role-matrix repetitions are distinct evaluation passes: the formal cells use pass one, the second pass reports another set of scores, and pass three repeats only the early mixed RL matrix.Source: §C.3 · p. 17

C.4 VSTAR DENOMINATOR RECONCILIATION

  • Restoring both ambiguous VSTAR rows gives 168/191 (88.0%) for both SFT and main-only RL; applying the archived filter gives 167/189 (88.4%) for each.The zero net difference holds under both denominator conventions, though RL has two iteration-limit failures among the retained rows.Source: §C.4 · p. 17

C.5 EARLIER MAIN-ONLY RL CHECKPOINTS

  • An earlier main-only RL checkpoint is below its paired SFT baseline on HR-Bench 4K (81.69 vs. 82.97) and VSTAR (87.30 vs. 88.36).Pairing that checkpoint with its own sub-agent yields a +1.41 point HR-Bench 4K estimate with McNemar p = 0.32; VSTAR remains below the SFT baseline.Source: §C.5 · p. 17

C.6 SFT COMPARISON AND UNCERTAINTY

The HR-Bench 4K SFT comparison shows a positive paired gain, while the clean HR-Bench 8K comparison with main-only RL is descriptive. The shared-actor evaluation separately compares four responses per prompt.

C.6 SFT COMPARISON AND UNCERTAINTY

  • Main-only RL with the SFT reader scores 662/800 (82.75%) versus 658/800 (82.25%) for clean SFT on HR-Bench 8K.The +0.50-point difference is descriptive, with no paired uncertainty analysis; the older 567/800 SFT observation is excluded because of service failures and unresolved reader deployment state.Source: §C.6 · p. 18
  • +4.1667 percentage points is the mean paired HR-Bench 4K SFT gain over three pretrained-policy runs.The 95% question-paired interval is [+1.54, +6.75] points; common-completion comparisons yield gains of 3.28, 1.70, and 4.62 points.Source: §C.6 · pp. 18–19
  • The HR-Bench 4K paired interval describes sampled-example uncertainty conditional on the evaluated runs, not variation across SFT runs.The clean HR-Bench 8K reference remains unchanged under exact decoded-RGB and perceptual-overlap exclusions, but this does not establish other benchmarks are overlap-free.Source: §C.6 · p. 18

C.7 SHARED-ACTOR FOUR-RESPONSE EVALUATION

  • The shared-actor evaluation samples four responses per prompt and reports pass@1, pass@4, and tool-call contrasts with prompt-paired intervals.The table caption identifies accuracy differences in percentage points, tool-call differences in calls per trajectory, and shading for the tool-use contrast; it supplies no outcome values here.Source: §C.7 · p. 19, Table 10 · p. 19

C.8 JOINT-TRAINING DIAGNOSTICS

Joint training stayed near its warm-start value on internal main-agent validation, while selected-policy evaluations showed different behavior across training and validation measures.

C.8 JOINT-TRAINING DIAGNOSTICS

  • Two saved joint-stage evaluations scored 0.789 and 0.793, and none of 14 paired checkpoint tests had p ≤0.11.The authors characterize this as no resolved change, not evidence of equivalence; selected-policy main mean@4 reached 0.8325, while training pass@1 rose from 0.4538 to 0.5205 as sampled pass@8 stayed at 0.7727 over a prolonged interval.Source: §C.8 · p. 19

C.9 ROLE-MATRIX CONTRASTS

The role-matrix analyses show mixed accuracy evidence and substantial completion effects, alongside diagnostics of evidence consumption, tool behavior, and the limits of the paired RL surrogate. The results are conditional on evaluated runs and do not establish stable gains for the selected joint policy.

C.9 ROLE-MATRIX CONTRASTS

  • Across three passes, main-agent directions were positive under both readers, but repeated comparisons did not establish a stable gain or replicate the selected joint RL matrix.With the joint sub-agent fixed, the main-agent drop had exact McNemar p = 0.0022; an exploratory early mixed-reader swap was nonsignificant, and its confidence AUC changed from 0.438 to 0.573.Source: §C.9 · pp. 19–20, Table 11 · p. 20

C.10 JOINT-POLICY COMPLETION AND TOOL USE

  • Joint RL’s end-to-end net change was −19 correct answers with the SFT reader and −20 when deployed in both roles, as losses on SFT-only completions outweighed shared-row gains.The table distinguishes end-to-end accuracy from conditional accuracy on each policy’s completed rows; the latter uses different denominators.Source: §C.10 · p. 20, Table 12 · p. 20

C.11 ORACLE-EVIDENCE DIAGNOSTIC

  • +9.80 points separated gold-ROI from model-selected crops on injected examples, supporting evidence-consumption diagnosis; the intervention did not measure the causal contribution of grid evidence.The non-injected gain was 1.67 points across repeated update-question observations and fell to 0.88 after excluding budget-hit trajectories; these are not independent-question sample sizes.Source: §C.11 · pp. 20–21

C.12 FORCED-GRID AND VALID-GRID DIAGNOSTICS

  • Forced-grid accuracy rose from 44.0% to a maximum of 67.5% on the same internal inputs, while unfinished trajectories fell from 47.0% to 9.0%.In a later free-choice continuation, successful grid use rose from 9.5% to 21.0% while accuracy fell to 64.5%; the changing setup does not isolate a causal grid-reward effect.Source: §C.12 · p. 21, Table 13 · p. 21

C.13 JOINT-SYSTEM GRADIENT AND THE IMPLEMENTED TRUNCATION

  • The implemented paired surrogate retains global main-agent and direct local sub-agent terms but drops system-sub/global credit, making the shared-weight update a biased, lower-variance semi-gradient.Grid-derived replay also omits the occupancy derivative; separate role groups and refreshed behavior snapshots are required by the described procedure.Source: §C.13 · p. 22

C.14 PAIRED RL PSEUDOCODE

  • The paired RL procedure samples main trajectories and tile responses separately, normalizes rewards within identical-prompt role groups, masks copied and nested outputs, then applies task weights in one update.When configured, it filters zero-variance groups and refreshes the behavior snapshot after the update.Source: §C.14 · p. 22

C.15 COMPLETION-AWARE REWARD PILOT

  • At the final completion-aware reward pilot evaluation, the intervention scored 151/200 versus 139/200, but paired tests did not resolve a difference at the 5% level.The intervention marked 105 of 2,560 training trajectories unfinished; broad truncation indicators were much more common than semantic non-completion, so the diagnostics remain distinct.Source: §C.15 · p. 23

C.16 UTILIZATION OF RELATIVE-REWARD GROUPS

  • In the unfiltered main-only policy, 478 of 800 training prompt groups were mixed, while filtering variants do not isolate a speedup or held-out accuracy benefit.Filtering also changes the training distribution, and online rejection adds generation work.Source: §C.16 · p. 23

C.17 DATA SNAPSHOT AND ANNOTATION AUDIT

The audit documents the training and validation data, ROI-labeling scope, overlap checks, and recovered training configurations. It also records gaps in annotation accounting and historical run details that limit reproduction.

C.17 DATA SNAPSHOT AND ANNOTATION AUDIT

  • The SFT snapshot has 6,464 examples, but no separate test set or role/source split is documented, and ROI agreement was not measured.Two annotators each labeled 500 images without double annotation; the tile-labeling, verifier, and rewrite totals do not reconcile to a complete final export manifest.Source: §C.17 · p. 24
  • Exact decoded-RGB matches in audited pools were confined to ZoomBench; perceptual candidates occurred on ZoomBench and TIR, with none detected on VSTAR or either HR-Bench split.Across 25 audited evaluations, the largest absolute accuracy shift was 0.28 points; these checks do not cover every form of training exposure.Source: §C.17 · pp. 24–25, Table 16 · p. 25

C.18 RECOVERED TRAINING CONFIGURATION

  • Historical joint-run reward and loss details, exact prompts, service revisions, and complete visual-token and resizing policies remain unavailable, limiting exact reproduction.The selected joint stage used 893 training and 264 validation examples and followed multiple warm-started stages; it did not share the main-only run’s initialization or data mixture.Source: §C.18 · p. 25

Related papers

Loading 2609.37002v2…