Source-linked AI summary
WeaveAgent: A Two-Stage Tool-Routing Agent for Ultra-High-Resolution Remote Sensing Imagery
Zhongyu Pang
TL;DR
UHR remote-sensing images make small objects difficult to resolve at model-native input sizes, while full-resolution visual processing is expensive and vague queries can require expert tools. WeaveAgent routes queries before conditional visual answering or full-resolution tool execution. Alignment training raises extrinsic routing from 0% to 80.75%, while GRPO suppresses intrinsic mis-emissions without changing extrinsic tool selection.
Problem
Resizing UHR scenes can discard informative small objects, while full-resolution visual encoding makes token costs grow out of control.
Source: §1 · p. 1Method
WeaveAgent routes queries before answering intrinsic requests visually or executing tools on original full-resolution imagery and using observations in a second round.
Source: The WeaveAgent two-stage architecture · p. 3, §3.1 · p. 7Results
0% →80.75%: alignment SFT raises extrinsic routing to 80.75% (323/400), while RWA2 GRPO suppresses 9 intrinsic mis-emissions and leaves extrinsic tool selection unchanged.
Source: §8 · p. 32Takeaways & Limitations
The ±-image ablation shows that emission suppression is visually grounded: removing image tokens preserves extrinsic routing but collapses intrinsic routing to 0.022.
Source: §8 · p. 31The evaluated tool pool covers generic detection, segmentation, and counting, but not GIS operations or geospatial workflows.
Source: §7 · p. 31
Abstract
from arXiv · showhide
Problem. Ultra-high-resolution (UHR) remote sensing with vague user intents has two bottlenecks: visual tokens are expensive, and tool calling must be format-reliable (pretrained models emit zero tool calls zero-shot). Method. WeaveAgent, a two-stage tool-routing agent, decouples routing from visual perception. Stage A is routing-first: emission is trained, not elicited. Stage B executes conditionally: intrinsic queries enter visual answering (full-scene thumbnail; a WeaveEarth-style evidence board as an optional fixed-budget, approx. 5k-token compression interface); extrinsic queries execute tool call on original full-resolution imagery, answering from tool observations in a second, observation-masked round. Training: alignment SFT, then GRPO under reward R_WA2. Results. Alignment SFT lifts extrinsic routing from 0% to 80.75% (323/400); GRPO suppresses 9 intrinsic mis-emissions while tool selection is unchanged. The trained 2B system does not beat the zero-shot 8B baseline overall (0.263 vs. 0.250), a diagnostic contribution. Oracle attribution separates two repair ingredients: loading the observation into context lifts extrinsic answer accuracy from 0.025 to 0.425 under marker-free cross-mode returns, and the two-turn SFT stage adds a further +9.3 points to 0.518 at a small routing cost. A +/- image ablation shows emission suppression is visually grounded, and a query-register matrix shows LLM-rewritten queries cost trained checkpoints 2-11 points. Scope. All training and evaluation use the 5,000 / 3,273 / 1,000-record VagueUHR corpus (600 intrinsic + 400 tool-requiring; the base seeds synthesis and is not used for optimization). Single-pass evidence construction runs at 7.31 s per image on an RTX 4090. Code, data, and evaluation protocols will be released.
1 Introduction
UHR imagery makes small objects difficult to resolve after resizing, while processing full-resolution scenes makes visual-token costs prohibitive; vague queries also require precise tool-based measurements. WeaveAgent separates query routing from conditional visual answering or tool execution on original imagery.
1 Introduction
- Resizing UHR scenes can discard informative small objects, while sending full-resolution imagery through a visual encoder makes token costs grow out of control.The evaluated images have a median side length of 6,044 px and a maximum of 27,328 px.Source: §1 · p. 1
- Vague remote-sensing queries such as counting berthed vessels or outlining burned areas call for expert tools applied to the right image region.These measurements are used in disaster-response and urban-monitoring workflows.Source: §1 · p. 2
The WeaveAgent two-stage architecture.
- WeaveAgent separates query-and-catalog routing from conditional execution: intrinsic queries receive visual answers, while extrinsic queries invoke tools on original imagery and use their observations in a second round.The intrinsic path can use a full-scene thumbnail or an optional evidence board; the extrinsic path uses two-turn trajectories with observation masking [4].Source: The WeaveAgent two-stage architecture · pp. 2–3
- The optional evidence board compresses any-size UHR imagery into two views of approximately 5k tokens, but is positioned as a fixed-budget interface rather than an accuracy booster.Its paired deltas against the 2048-px thumbnail baseline are −1.0 points on LRS-VQA, −3.4 on MME-RS, and −2.9 on XLRS.Source: The WeaveAgent two-stage architecture · pp. 3–4
- 0% →80.75%: alignment SFT lifts extrinsic routing to 80.75% (323/400), while subsequent GRPO suppresses 9 intrinsic mis-emissions and leaves extrinsic tool selection unchanged.Extrinsic answer accuracy rises from 0.025 to 0.425 when observations are loaded, then to 0.518 with two-turn SFT under marker-free cross-mode returns.Source: The WeaveAgent two-stage architecture · p. 3
- VagueUHR’s dual-split corpus contains 5,000 template-training synthesis-base records, 3,273 routing-training records (870 tool-call demonstrations), and 1,000 test records, including five task shapes unseen during training.The synthesis base is not used for optimization; the test set contains 600 intrinsic and 400 tool-requiring records.Source: The WeaveAgent two-stage architecture · p. 4
2 Related Work
Prior remote-sensing VLMs answer from pixels, while agent systems add tool use and visual exploration. WeaveAgent combines tool routing with a conditional execution path and adapts ideas from prior evidence-compression and agent-training work.
RS vision-language models and instruction tuning
- Remote-sensing VLMs broaden pixel-based answering but do not decide whether external capabilities are needed or route queries to tool backends.WeaveAgent is complementary: its intrinsic path is a VLM, with the trained router in front.Source: §2 · p. 5
Agentic remote sensing and RL for tool use
- RemoteAgent [1] defines vague-intent routing and reports 95.0% intent-recognition accuracy, but its router uses downscaled images and its GRPO stage leaves tool routing zero-shot.WeaveAgent responds with explicit tool-selection training, two-turn tool execution, an oracle attribution and repair of the observation-to-answer gap, a runnable training stack, and a public corpus.Source: §2 · pp. 5–6
- WeaveAgent adopts OpenEarthAgent [4]’s two-turn trajectories with observation masking and schema-registration integration for its extrinsic path.Observation tokens are excluded from training loss so the policy learns to use observations rather than recite them.Source: §2 · p. 5
UHR remote sensing understanding: passive vs. active perception
- WeaveAgent adopts WeaveEarth ’s evidence pipeline as an optional compression interface and adds two-stage tool routing and training.The evidence pipeline organizes patches into a minimal support set and a topology-preserving board with structured metadata.Source: §2 · pp. 4, 6
3 The WeaveAgent Two-Stage Architecture
WeaveAgent separates routing from conditional visual answering or full-resolution tool execution, then trains routing and observation-to-answer behavior through distinct stages. The architecture preserves image-free tool execution, but its trained router still depends on visual input to suppress calls on intrinsic queries.
3.2 Stage A: Text-Only Router
- Stage A uses query text and a tool catalog to delegate intrinsic queries or select and emit a tool call for extrinsic queries.The unified evaluation prompt still includes visual inputs; the ablation tests whether routing depends on them.Source: §3.1 · p. 6, §3.2 · p. 7
- 0.8375 vs. 0.8075 extrinsic routing is preserved without image tokens, but intrinsic routing falls from 0.985 to 0.0217 and calls occur on 987 of 1,000 records.The image ablation shows emission suppression is visually grounded, so the trained 2B router does not achieve the text-only design target.Source: §3.2 · pp. 7–8
3.3 Stage B (Intrinsic Path): Visual Answering with an Optional Compression Interface
- The intrinsic path answers from a 2048-px thumbnail, with an optional fixed-budget TPEB/SEM evidence board for fine-grained tasks.The two-view interface costs about 5k tokens, but has limited coverage and no demonstrated accuracy improvement over the thumbnail baseline.Source: §3.3 · pp. 8–9
3.4 Stage B (Extrinsic Path): Tool Execution, Observation Backfill, and Two-Turn Trajectories
- Extrinsic calls run on original full-resolution imagery, and their structured tool observations are backfilled for a second-round answer.The tool path avoids visual compression; observation tokens are masked from the training loss so the policy learns to use rather than recite them.Source: §3.4 · p. 9
- With extrinsic routing at 80.75%, single-turn final-answer accuracy remained 0.035–0.045 for detection and 0.005 for segmentation, exposing an observation-to-answer gap.The paper attributes the gap to converting marker-free tool text rather than calling quality and reports a partial repair under protocol caveats.Source: §3.4 · p. 9
3.5 Training Recipe: Alignment SFT →RWA2 GRPO
- Alignment SFT seeds parseable tool emissions, while RWA2 rewards correct routing strongly enough to make routing preferable to direct answering across answer accuracies.RWA2 removes the no-call consolation and raises the tool weight; its correct-route reward is 2.25 versus 0.25 + p_ans for direct answering.Source: §3.5 · p. 11, Table 1 · p. 11
3.6 Two-Turn Trajectory SFT: Repairing the Observation→Answer Conversion
- E17 branches from the E13 aligned base and trains both inherited emission and new observation-to-answer conversion on two-turn trajectories.Its turn 1 matches E13 demonstrations, while turn 2 loads the tool observation in the RemoteAgent inference format.Source: §3.6 · p. 11, Figure 3 · p. 12
4 The VagueUHR Corpus
VagueUHR addresses the absence of a public corpus linking vague UHR intents to tool-call trajectories by synthesizing annotation-grounded records. Its splits test both routing and generalization, with documented limits in query vagueness and one retained annotation error.
4 The VagueUHR Corpus
- The corpus fills a gap in public data by linking vague UHR queries to tool-call trajectories whose answers derive from existing annotations.VagueEO provides an evaluation protocol but trains on intrinsic tasks, while WeaveEarth-style benchmarks lack tool traces.Source: §4 · p. 12
4.1 Synthesis Pipeline
- The synthesis pipeline generates deliberately vague queries, builds annotation-grounded intrinsic answers or T_call plans, and validates query–annotation–trajectory consistency.Records failing schema, consistency, or jargon checks are rejected; the shipped corpus passes validation at 100%.Source: §4.1 · p. 13
4.2 Dual Splits and Statistics
- The 5,000-record synthesis base is not optimized on; optimization uses aligned data drawn from 3,273 routing-training records, while a separate 992-record rewritten test split controls query wording.A 1,000-record clear-register split is also derived from the template split.Source: §4.2 · pp. 13–14
4.3 Unseen-Tool and Unseen-Shape Design
- All three trained checkpoints route 0/62 queries to the unseen sm3det tool, showing that catalog registration alone does not yield tool generalization.Five of ten evaluation task shapes are also absent from training.Source: §4.2 · p. 13, Table 2 · p. 13, §4.3 · p. 14
4.4 Annotation Audit
- Template vagueness is fixed at level 0.5 and narrower than open-ended phrasing; the audit also identifies one source-annotation misread retained for the released test set.The record is flagged for correction in the next corpus revision.Source: §4.4 · p. 14
5 Experiments
On VagueUHR-test, the trained 2B system improves tool routing but does not exceed the zero-shot 8B system in overall accuracy. Alignment SFT unlocks routing, while GRPO mainly suppresses intrinsic mis-emissions; final answers remain limited by tool-observation conversion.
5.2 Routing and Tool Metrics on VagueUHR-test
- Alignment SFT unlocks extrinsic routing, lifting it from 0% to 80.75% (323/400; 95% Wilson CI [0.766, 0.843]), with 398 emissions and tool selection 0.905.Evaluations use greedy decoding (temperature 0). Headline numbers are measured on VagueUHR-test; public RS suites lack executable tool backends, so these metrics are not reproducible on them.Source: §5.1 · p. 15, §5.2 · p. 16
- GRPO leaves extrinsic routing at 80.75% while reducing emissions from 398 to 384 and intrinsic mis-emissions by nine; extrinsic tool selection stays 323/400.Overall accuracy moves 0.2430→0.2500, with no significant gain reported for that comparison.Source: §5.2 · p. 16
- Zero-shot baselines bracket the routing problem: the T_call protocol yields zero calls, while the native tools= channel emits broadly without discrimination, with mean routing 0.057.The native-channel arm emits on 395/400 extrinsic and 591/600 intrinsic records; its Mean is not measured.Source: §5.2 · pp. 16–17, Table 3 · p. 16
- 0.2630 for the zero-shot model exceeds E15’s 0.2500 overall accuracy, despite E15 routing correctly on 80.75% of extrinsic records.The section treats Mean as a full-set reference because E13/E15 final answers remain suppressed by the observation-to-answer conversion gap.Source: §5.2 · p. 17
- Argument values are near-perfect for E13/E15 (0.951/0.963), but final answers remain poor (detection 0.035/0.045; segmentation 0.005), locating the break in tool-text conversion.The figure describes five folds of the extrinsic pipeline; E17’s two-turn SFT repairs this conversion fold under the strict-return protocol.Source: §5.2 · p. 17, Figure 4 · p. 18, Table 4 · p. 18
5.3 Compression-Interface Characterization and Component Ablations
The evidence board offers a fixed-token-cost compression interface, but its paired benchmark accuracy is below the thumbnail baseline. Component and budget scans find no positive static-VQA contribution or meaningful accuracy change across tested budgets.
5.3 Compression-Interface Characterization and Component Ablations
- −1.0 (LRS-VQA), −3.4 (MME-RS), and −2.9 (XLRS-lite) are the paired accuracy deltas for the evidence board versus the 2048-px thumbnail baseline.The board’s token cost is fixed at about 5k tokens and does not grow with source-image size.Source: §5.3 · pp. 18–19, Table 6 · p. 20
- After alignment SFT (E15, green), reward starts high and zero-variance groups stay an order of magnitude lower than in cold-start SFT then RWA GRPO (E12), seeding emission that reward can shape.Source: Figure 5 · p. 19
- On static VQA, the full system’s 0.2940 accuracy does not exceed any leave-one-component variant (0.3000–0.3140).The reported range is within the stated greedy-noise band, so the paper makes no significance claims about single-point orderings.Source: §5.3 · p. 19, Table 7 · p. 20
- Across k ∈{2, 4, 6, 8, 10}, the evidence-budget scan has a maximum accuracy range of 1.6 points, within the noise band.The paper concludes that budget choice should follow task detail needs and token budget, not an accuracy curve.Source: §5.3 · p. 20, Table 8 · p. 20
5.4 Efficiency
Single-pass evidence construction costs 7.31 seconds per sample on an RTX 4090, while query wording and image context affect routing in distinct ways. LLM-rewritten queries lower trained-checkpoint accuracy, and removing images sharply degrades intrinsic routing.
5.4 Efficiency
- 7.31 s per sample is the single-pass evidence-construction cost on an RTX 4090, about 7.4× faster than reported multi-round visual search.The comparison is cross-protocol; the trained agent’s own VagueUHR-test single-pass cost is 6.17 s per record, and the two-turn extrinsic path adds another generation round not separately timed.Source: §5.4 · pp. 20–21, Table 9 · p. 21
5.5 Query-Register Gradient: Clear, Template, and LLM-Rewritten
- LLM-rewritten queries lower overall accuracy by 3.1 points for E15 and 10.9 points for E17, while template vagueness shifts trained-checkpoint accuracy by at most 1.2 points.Extrinsic routing falls by 14 points for both trained checkpoints on the rewritten register.Source: §5.5 · p. 21, Figure 6 · p. 22, Table 10 · p. 22
5.6 The Stage-A ±-Image Ablation: Emission Suppression Is Visually Grounded
- Without image tokens, intrinsic routing falls from 0.985 to 0.0217 and the model emits calls on 987 records, while extrinsic routing is preserved.Per-record routing latency drops 15×, but the text-only policy’s tool selection is 0.348 versus 0.914 with images.Source: §5.6 · pp. 22–23, Figure 7 · p. 23
5.7 Tool-Execution Conversion: Oracle Attribution and Two-Turn Repair
The oracle attribution isolates observation-to-answer conversion as a major bottleneck, and two-turn training improves answer accuracy under both marker-free cross-mode returns and training-distribution returns. The qualitative cases show that conversion errors can persist even when routing and tool execution succeed.
5.7 Tool-Execution Conversion: Oracle Attribution and Two-Turn Repair
- E17’s Eval A uses strict marker-free oracle returns with fallback rate 1.0, so its cross-mode result is not an end-to-end measurement against live services.Live end-to-end evaluation remains future work.Source: §5.7 · pp. 23–25, Table 13 · p. 26
- 0.505 strict versus 0.840 embedded overall on 100 routed-correct records; the gap concentrates in segmentation, at 0.339 versus 0.758.Both E16 arms use oracle returns and are upper bounds; the 100 routed-correct records are a self-selected, easier subset of the 400 tool-bearing rows.Source: §5.7 · p. 24, Table 12 · p. 24, Figure 8 · p. 24
- 0.025 → 0.425 with strict returns loaded, then +9.25 points to 0.5175 after two-turn SFT on the 400 tool-bearing rows; the E17 ladder reaches 0.965 with training-distribution returns.The controls without two-turn training score 0.4225 (E13) and 0.4250 (E15); E17’s cross-mode point is an upper-bound evaluation using oracle returns.Source: §5.7 · p. 25, Figure 9 · p. 25, Table 13 · p. 26
6 Design Rationale and Analysis
The analysis traces routing behavior to prompt alignment and reward structure, then identifies concrete limits on the evidence-board interface. It also records implementation and prompt-format differences as protocol-alignment issues.
6.1 Why Alignment SFT Is Necessary: Two Mechanisms
- 85–90% of base-model GRPO rollout groups have zero reward variance, leaving no learning signal when all four rollouts answer directly; alignment SFT seeds emission.With the evaluation-format prompt, the same unaligned SFT weights emit on 0/400 tool tasks, versus 0.65 under the training-format prompt; evaluation-identical prompts lift extrinsic routing from 0% to 80.75%.Source: §6.1 · p. 28
6.2 Why the Incentive Must Be Routing-First: Reward Economics
- RWA2 makes correct routed calls score 2.25 versus 0.25–1.25 for direct answering, and E15 raises overall routing from 0.905 to 0.914.Its design removes the no-call consolation, raises routing weight, and splits the format term; the design conclusion is that it stabilizes emission and suppresses intrinsic mis-emissions rather than creating the initial routing behavior.Source: §6.2 · pp. 28–29
6.3 The Boundary of the Compression Interface: the 1.31× Ceiling and Conditions of Applicability
- On 660 spatial-ground-truth records, retrieval recall was 0.291 at top-8 anchors and 0.230 at the final top-6 board; MME-RS position questions lost 10.9 paired points.Source: §6.3 · p. 29
- No repair significantly beats the thumbnail baseline (all p ∈[0.08, 1.0]); routing spatial questions to the global view is the only stably positive component, at about 2 points.The authors position TPEB+SEM as a fixed-budget compression interface for fine-grained task types, not a default accuracy booster.Source: §6.3 · pp. 29–30
6.4 Implementation Differences and Prompt Sensitivity (Neutral Records)
- The released script gains 2.67 points under the paired greedy protocol, while the paper reports 6.70 points on the full set; the paper’s baseline configuration is undocumented.A paired control finds the Category task-type line changes accuracy by −1.7 points (p = 0.51), refuting the inferred answer-information effect; option-instruction order versus SEM placement also shifts accuracy by about 10.5 points in an indicative audit.Source: §6.4 · p. 30
7 Limitations
The reported system remains bounded by its model scale, query and prompt distributions, oracle-based conversion evaluation, and the tested tool and data scope.
Model scale and routing interface
- All two-stage training and evaluation use Qwen3-VL-2B; scaling the alignment pipeline to 8B remains future work, and the trained policy does not yet operate as a text-only router.The image-free decision stage is estimated at 15× lower latency, but a genuinely text-only router is not yet attained.Source: §7 · p. 30
Tool execution and routing trade-off
- E16/E17 conversion results use oracle or annotation-grounded returns, and Eval A has fallback rate 1.0; end-to-end success against live expert backends is unmeasured.The repair also carries a routing give-back: tool selection changes from 0.914 to 0.868 while calls rise from 384 to 433.Source: §7 · pp. 30–31
Tool and task scope
- All three trained checkpoints route 0/62 on training-unseen sm3det records, and the region parameter p of T_call is unevaluated.The evaluated tool pool is limited to three tools; the generic-CV pool excludes GIS operations, while temporal reasoning is outside the evaluated scope.Source: §7 · p. 31
Query and prompt coverage
- Training queries are templated, and multi-turn, real-user phrasing remains unmeasured despite testing an LLM-rewritten register.The rewrite channel is not enabled for training; the prompt audit also finds routing can fall from 0.65 emission under the training-format prompt to 0/400 under the evaluation-format prompt.Source: §7 · pp. 30–31
Data and evaluation scope
- Public-benchmark pretraining exposure cannot be excluded, geographic coverage is limited, and most benchmark results are stratified estimates from a single seed.The XLRS sample covers 7 of 8 subcategories; the SR subcategory is unavailable.Source: §7 · p. 31
8 Conclusion
WeaveAgent separates query-and-catalog routing from conditional visual answering or full-resolution tool execution. On VagueUHR, alignment SFT supplies the main routing gain, while the conclusion also reports answer-conversion repairs, wording sensitivity, and open scope boundaries.
8 Conclusion
- The two-stage design routes intrinsic queries to thumbnail-based visual answering and extrinsic queries to tools on original full-resolution imagery, then answers from observations in a second round.The intrinsic path may use a fixed-budget evidence-compression interface.Source: §8 · pp. 31–32
- Alignment SFT lifts extrinsic routing from 0% to 80.75% (323/400), while GRPO leaves tool selection unchanged and suppresses 9 intrinsic mis-emissions.GRPO moves overall accuracy from 0.2430 to 0.2500, within the noise band.Source: §8 · p. 32
- Loading tool observations raises extrinsic answer accuracy from 0.025 to 0.425, and the two-turn SFT stage raises it further to 0.518 at a small routing cost.These results are under marker-free cross-mode returns.Source: §8 · p. 32
- LLM-rewritten phrasing costs 3.1 points for E15 and 10.9 points for E17 overall, with an approximately 14-point extrinsic-routing give-back.Template-level vagueness moves overall accuracy by at most 1.2 points.Source: §8 · p. 32
- The study is scoped to the VagueUHR corpus, and the author identifies 8B-scale alignment, live-service conversion repair, broader query-register coverage, and a truly text-only router as open problems.The conclusion says code, data, and evaluation protocols will be released; imagery files are not redistributed and are obtained from source benchmarks.Source: §8 · p. 32
A Training and Prompt Configuration
Appendix A collects a prompt-format audit and a record of prompt-family inputs. It also catalogs experiment configurations and result pointers.
A Training and Prompt Configuration
- This appendix provides a register of every experiment ID referenced in the paper, listing each experiment’s configuration and a pointer to the table or figure reporting it (Table S3).Source: A Training and Prompt Configuration · p. 33
- The byte-level prompt audit documents non-image differences between the unaligned training prompt and evaluation prompt, while confirming that the evidence board is injected at evaluation.The audit is presented as ruling out “evidence not given” as an explanation.Source: Table S1 · p. 33
- Table S2 consolidates, from disclosures in Table S1, Sections 3.5 and 6.1, and Table S7, where the thumbnail, TPEB board, and SEM block appear across prompt families used in the paper, introducing no new measurements.Source: A Training and Prompt Configuration · p. 33
B Reward Scores and Neutral Implementation Records
Appendix B records the measured RWA scoring configuration and explains how its total combines answer, tool, and format components.
B Reward Scores and Neutral Implementation Records
- The appendix collects the full RWA per-case score table, identified as Table S4 and associated with Sections 3.5 and 6.2, plus neutral implementation-difference records from Section 6.4.Source: B Reward Scores and Neutral… · p. 33
- The table states that the scoring structure makes bare T_call behavior nearly worthless relative to direct answering.The passage connects this structure to what SFT teaches.Source: B Reward Scores and Neutral… · p. 33
C Supplementary Measurements and Qualitative Cases
Appendix C assembles supplementary wording, category-line, retrieval-recall, and qualitative-case materials. The appendix descriptions specify the scope and protocols of these records.
C Supplementary Measurements and Qualitative Cases
- The appendix also catalogs the prompt-family audit, experiment configurations, RWA scores, implementation records, GCC retrieval recall, and four E17 two-turn qualitative cases.The retrieval-recall table covers 660 spatial-ground-truth records and excludes counting from its uniformity statement; the qualitative cases use greedy decoding and Eval A.Source: Table S2 · p. 34, Table S3 · p. 34, Table S4 · p. 35
- The wording-sensitivity figure compares per-task-type accuracy on template and LLM-visual-rewrite splits across 992 paired records, describing the movement as a reshuffle rather than systematic damage.It reports the zero-shot Qwen3-VL-8B greedy setting and points to Table 11.Source: Figure S1 · p. 36
- The Category-line paired control reports an overall effect of −1.7 points (p = 0.51), with per-type changes of +13.3 classification, −8.3 detection, and +8.3 reasoning points.It covers 300 records, 60 per task type, at E15 with greedy decoding; the movement is characterized as a reshuffle, not a systematic leak.Source: Figure S2 · p. 36