Four geometry interfaces
BoxDetectionMaskPoint
DATA → SPATIAL INTELLIGENCE
AnvilScaling Embodied Spatial Intelligence
through Knowledge Recirculation
CUHK1The Chinese University of Hong KongKnowin.AI2Knowin.AIHKUST(GZ)3The Hong Kong University of Science
and Technology (Guangzhou)
and Technology (Guangzhou)
Connect spatial supervision, verification and learning through Anvil-Gen data generation, Anvil-Ben benchmarking and Anvil-MOPD specialist distillation.
SUPERVISION COVERAGE+30.25 pts
Where2Place · mixed vs. replay
SELECTIVE DISTILLATION+2.13 ptsGround → Relation · uniform vs. flat
FEEDBACK-GUIDED REPAIR10.09×Correct recoveries · equal call budget
ANVIL-GEN → ANVIL-BEN → ANVIL-MOPD → KNOWLEDGE RETURNClick figure to enlarge ↗
5.8Msource-grounded records constructed
160Kunique canonical images
11native spatial formats
580Kheld-out spatial records
5.22M training records, including 464K QA replay records, plus 580K held-out spatial records. The portable package has passed full record, image and byte checks, which are not a human semantic-accuracy estimate. The first public release includes the 464K CLEVR QA records, complete file-level inventory and image index; the remaining spatial records and original image bundle are not yet public.
Four relational interfaces
Depth orderMetric comparisonDirectionRelation graph
Three interaction interfaces
AffordanceGripper configurationAction plan
01 / ANVIL-GEN
Different sources.
Connected supervision.
Anvil-Gen preserves each source's useful structure, connects entities across tasks and turns failed checks into targeted generation feedback.
Keep native structure
Combine image, RGB-D and action-related sources through shared entities and typed Ground–Relation–Act dependencies.
Verify, then refine
Check format, geometry and source consistency. A named failure tells the generator what to repair—without supplying the target answer.
Teach where it matters
Route a student's earliest decisive error to a specialist and a response span. Keep the deployed student independent of teacher calls.
02 / EFFICIENCY BY CONSTRUCTION
More useful data.
Less repeated work.
Reuse grounded source structure before spending on generation. Spend the next model call on a specific failure.
See what feedback repairs ↓linked records constructed
Eight CPU workers, cached RGB-D frames, zero model calls. This timing covers linked-record construction, not the entire data pipeline.
source-correct records recovered
Named feedback vs. repeat generation at the same 885-call repair budget. 2.82 vs. 2.77 inference GPU-hours; no new human annotation.
03 / SPATIAL REASONING, MADE VISIBLE
From a scene
to a usable answer.
Inspect native supervision, measured placement changes and the feedback that repairs a generated spatial relation.
SOURCE → NATIVE TARGET
Supervision with structure.
Entity references connect grounding, relations and action-relevant fields. Each output keeps a task-specific contract instead of becoming a generic caption.
SAME SCENE → DIFFERENT PREDICTION
Restore the missing supervision.
Targeted placement continuation moves points into valid regions in two selected cases—and regresses in another. The figure pairs these cases with full-split, two-seed results.
FAILED CHECK → TARGETED REPAIR
Fix the relation, not the whole scene.
Named feedback corrects a coffee-maker/sink depth ordering while retaining the predicted boxes. This is a training-data generation example, not a held-out student prediction.
04 / ANVIL-MOPD
A targeted lesson.
A standalone student.
First-decisive-fault routing selects a specialist and correction span. Stage-balanced reference learning supervises active stages, while span-normalized distillation focuses correction. The standalone 9B student can return auxiliary verification signals under evidence-backed acceptance.
Find the first decisive fault
The student attempts active stages. Shared references distinguish a correct, incorrect or uncertain response.
Select the teacher and span
Stage-balanced reference learning and span-normalized distillation retain each native output's structure.
Assist the next verification round
The generalist flags inconsistencies; independent evidence retains acceptance authority. Held-out targets stay outside the loop.
05 / RESULTS & MECHANISMS
Spatial transfer.
Targeted correction.
The table presents manuscript-reported final-model results. The downloadable checkpoint is still awaiting an exact weight-to-evaluation binding. Coverage and routing studies below use separate controlled configurations; their gains are not added to the final model's scores.
COVERAGE CONTROL / WHERE2PLACE
Match the updates. Change the supervision.
Two-seed mean; 303 updates and 7,264 row exposures per arm. Mixed improves placement by 30.25 points over replay while retaining more pointing performance than vacancy-only. Image and token counts are not matched.
ROUTING CONTROL / STRICT G → S
Make the correction selective.
282 images, 50 scenes, two seeds. Uniform vs. flat: +2.13 points, 95% scene-bootstrap interval [0.39, 3.90]. Bars start at zero. Separate controlled configurations, not the final released checkpoint.
Manuscript-reported final-model comparison
Score (%) · higher is better| Model | CV-Bench | EmbSpatial | RefSpatial | Where2Place | BLINK depth | ERQA | VABench-Point | Mean |
|---|---|---|---|---|---|---|---|---|
| Qwen3.5-9B · base | 87.15 | 75.58 | 61.67 | 47.97 | 90.32 | 49.83 | 42.33 | 64.98 |
| Anvil-MOPD · reported final | 88.21 | 82.34 | 59.00 | 68.00 | 91.13 | 50.52 | 80.67 | 74.27 |
Values follow the manuscript, not a new evaluation of released weights. The mean equally weights seven public benchmarks; VABench-Point uses any-point success. Bold compares only these two rows, not a global leaderboard. Anvil-Eval is not included in this mean; its release version and scoring manifest remain pending.
Placement and routing: protocols & limitations
Source coverage, not extra updates alone
Vacancy + replay reaches 68.00 / 66.71 on Where2Place, compared with replay-only 38.00 / 36.22 (seeds 43/44). All arms use 303 updates and 7,264 row exposures. Unique images and token counts are not matched. Some reasoning and general-retention scores decrease.
Route a decisive error
Weighted semantic routing reaches 71.81% strict G→S success vs. 70.21% for flat distillation on 282 images across 50 scenes, averaged over seeds 17/23. The difference's 95% scene interval is [0.41, 2.80] points. It ties declared routing; uniform semantic routing reaches 72.34%.
06 / PROJECT RESOURCES
Explore the project.
Current release status · 15 October 2026. Public resources and private staging are distinguished below.
Paper
Author draft · not yet on arXivAuthor layout and supporting figures are being finalized. A public paper URL will appear here after release.
Code ↗
Uploaded · private accessGeneration, verification, native training and evaluation. MIT applies to project-owned code; upstream notices remain intact.
Anvil-Ben ↗
Public · staged data release464K CLEVR QA replay records, full file-level inventory, image index and replay reconstruction script. Other spatial records and original images remain private pending source-specific clearance.
Anvil-MOPD-9B ↗
Public model card · weights pendingThe final model repository is reserved, with no weights yet. Four SFT preparation adapters are separately staged privately; they are not the final MOPD checkpoint.
Private repository links require authorized access. Public authorship is restored with author approval; anonymous peer-review materials are maintained separately. Institutional names above are text identifiers, not official logos.