acceptodds
Withdrawn from ICLR 2027

Anvil: Scaling Embodied Spatial Intelligence through Knowledge Recirculation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied spatial intelligence requires more than visual recognition: models must localize entities, reason about their relations, and express this knowledge in action-relevant forms. Yet joint training can conflate the knowledge needed at different stages because supervision is fragmented across incompatible geometric and semantic outputs. We introduce Anvil, a unified framework for data generation, training, and benchmarking for embodied spatial intelligence. Anvil-GEN is a heterogeneous data flywheel that combines form-specific generation, cross-stage consistency verification, and failure-driven recollection across 11 Ground–Relation–Act forms. Anvil-BEN provides scene-aligned supervision for learning diverse embodied spatial capabilities and scene-linked assessments that distinguish stage competence from error propagation. Anvil-MOPD is a 9B student that consolidates heterogeneous spatial knowledge through distillation from specialist teachers. Semantic-first-fault routing targets teacher supervision to the student's earliest decisive failure, while stage normalization controls the loss scale across native outputs of different lengths. We generate and verify 5.8M records spanning 160K images, reserving 580K verified spatial records for held-out evaluation with human review and using the remaining 5.22M records for training. On average across seven public benchmarks, Anvil-MOPD outperforms Qwen3.5-9B and RynnBrain-8B by 9.29 and 11.03 percentage points, respectively, while reaching 80.67% on VABench-Point. Controlled studies support targeted coverage, selective transfer, and feedback-guided repair as complementary framework components.

Primary area:
applications to robotics, autonomy, planning
Keywords:
embodied spatial intelligence, vision-language models, spatial reasoning, synthetic data generation, multi-teacher on-policy distillation, spatial benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DdS1OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ANVIL, a framework for improving spatial intelligence in VLMs through three components: ANVIL-GEN, which generates and verifies spatial supervision using cross-form consistency and feedback-driven repair; ANVIL-BEN, which organizes 5.8M records across 11 spatial task formats; and ANVIL-MOPD, which introduces semantic-first-fault routing for selective multi-teacher distillation. The authors further propose knowledge recirculation, where the trained student assists subsequent data generation and verification.
Experiments show a 9.29-point average improvement over Qwen3.5-9B across seven benchmarks, alongside component-level evidence for targeted supervision, selective teacher routing, feedback-driven repair, and scene-linked spatial learning.

Critical strengths
    1. Structured spatial supervision for embodied intelligence
      The paper introduces ANVIL-GEN, a unified framework for constructing and verifying Ground–Relation–Action supervision through scene-aligned annotations and cross-form consistency checks. This provides a meaningful approach to organizing heterogeneous spatial knowledge and supporting compositional learning across embodied tasks.
    1. Stage-aware multi-teacher distillation
      ANVIL-MOPD introduces semantic-first-fault routing, which identifies the earliest erroneous stage in student predictions and selectively applies supervision from the corresponding specialist teacher. This task-dependency-aware training strategy is a meaningful methodological contribution, with controlled experiments showing improvements over flat distillation.
Critical weaknesses
    1. Insufficient validation of knowledge recirculation
      Despite being the central claim of the paper, no end-to-end experiments demonstrate that the distilled student improves subsequent data generation or verification, or that the resulting data further improves the student. The reported repair experiments instead use a separate model, leaving the proposed iterative feedback mechanism unvalidated.
    1. Insufficient evidence for the necessity and superiority of ANVIL-MOPD
      Although semantic-first-fault routing improves over flat distillation, its advantage over simpler declared routing is marginal (+0.53 percentage points) and statistically inconclusive. Furthermore, the paper lacks comprehensive ablations against standard SFT and alternative distillation strategies, making it difficult to attribute the final performance gains to the proposed method.
    1. Limited evidence of successful specialist-to-generalist knowledge consolidation
      ANVIL-MOPD aims to integrate Ground, Relation, and Act specialist knowledge into a unified student. However, the paper lacks systematic comparisons between the specialists and the final student across the 11 spatial tasks. Moreover, the linked-supervision experiments reveal substantial degradation on atomic tasks, raising concerns about whether the learned spatial capabilities are genuinely consolidated and transferable rather than adapted to specific output formats.
Questions for authors

Beyond the concerns raised in the weaknesses, I have an additional question regarding the generalizability of the reported improvements. The average gain (in Table 1) is largely driven by Where2Place (+20.03 points) and VABench-Point (+38.34 points), while improvements on several other benchmarks are marginal, and RefSpatial even shows a performance decline. Could the authors clarify whether this imbalance is related to the distribution of training tasks or the alignment between training supervision formats and benchmark output requirements? A more detailed analysis would help determine whether the improvements reflect generalizable spatial reasoning capabilities or are primarily driven by adaptation to specific task formats.

Reviewer_fM2wOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Anvil, which combines data generation, verification, benchmarking, and specialist-to-generalist distillation through shared Ground–Relation–Act representations. It introduces 11 output forms, a 5.8M data pool, and a 9B model trained using semantic-first-fault routing. The final model improves the seven-benchmark average over Qwen3.5-9B by 9.29%, including a substantial pointing gain compared to mentioned methods.

Critical strengths

The use of shared entity identities, coordinate conventions, and dependencies as explicit supervision structure makes cross-task consistency inspectable. The framework appropriately distinguishes consistent outputs from independently supported annotations. Also, the authors study match updates, exposures, or repair calls, and use paired scene-bootstrap intervals. Reporting inconclusive results and failure cases improves transparency.

Critical weaknesses
  1. My major concern is about the limited performance improvement. As illustrated in Tab. 1. the proposed method only achieves best performance among competing methods in Emb-Spatial and VABench Point (besides the self-proposed Anvil-Eval) and its strong average performance mainly comes from VABench Point. However, the use of “published source-protocol results” makes me really confused. The source protocol evaluating metrics for VABench Point is first-point scoring and in this paper the author uses any-point scoring only for some models while leaving the results of other models still using first-point scoring results for comparison. Although the author provides some additional results in Appendix C, I believe the current results in Tab. 1 is an unfair comparison and must be addressed by using the same evaluation metrics for all methods on the same dataset.

  2. The knowledge-recirculation mechanism is not demonstrated end to end. ANVIL-MOPD proposes that a stronger distilled generalist improves subsequent verification and data collection, which then improves later models. However, the experiments primarily evaluate these components separately. The data-repair experiment uses a separate model, Qwen3.6-27B, and shows that specific error feedback increases correctly repaired examples from 11 to 111. This demonstrates better data repair, but the benefit for model training remains uncertain: with equally sized training sets, performance improves by 2.84, with a confidence interval of [−1.41, 6.88], which is not statistically conclusive to improvement. To support the proposed feedback loop, the authors should show that using the newly trained ANVIL to improve data produces a better next-round model than keeping the data-checking model fixed.

  3. The reliability of the proposed benchmark is insufficiently established. The human audit reports 88.79% annotation correctness across 1,463 sampled records from the 580K-record evaluation partition, indicating substantial residual annotation error despite automatic verification. The annotation quality is important as the same supervision framework supplies both training targets and evaluation references.

Questions for authors

Please refer to weaknesses.

Reviewer_bFQ4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ANVIL, a framework combining spatial supervision generation, held-out evaluation, and specialist-to-generalist distillation. ANVIL-Gen aligns eleven Ground–Relation–Act output forms through shared scene entities, verifies candidates against native constraints and available evidence, and uses failure feedback for repair. ANVIL-Ben contains 5.8M records over 160K images, with 5.22M training records and 580K held-out spatial records. ANVIL-MOPD trains a Qwen3.5-9B-based generalist using specialist teachers, semantic-first-fault routing, and stage-normalized supervision. The final model improves the seven-public-benchmark mean from 64.98% to 74.27%. Separate controlled studies examine supervision coverage, selective distillation, feedback-based repair, linked-stage composition, and evidence dependence.

Critical strengths
  1. The scene-aligned interface provides a concrete way to connect heterogeneous spatial supervision. Shared entity identities, coordinate conventions, and explicit dependencies link eleven grounding, relation, and action-relevant output forms while preserving their native representations. The admission procedure distinguishes native validity, cross-form consistency, and semantic support. The pot-scene example in Appendix A illustrates why this distinction matters: several outputs can agree on the same physically unsupported contact point. (§3.1–3.2; Appendix A)

  2. The final model achieves substantial improvements on several targeted spatial capabilities. Relative to Qwen3.5-9B, the seven-benchmark mean improves by 9.29 points, including gains of 38.34 points on VABench-Point, 20.03 on Where2Place, and 6.76 on EmbSpatial. Excluding VABench-Point still leaves a 4.45-point mean improvement. Although transfer is uneven, including a 2.67-point regression on RefSpatial, these results demonstrate practical benefits beyond a single benchmark. (§4.2; Table 1)

  3. Several mechanism studies use informative controls, with clear reporting of their scope. Routing comparisons share initialization and maximum rollout budgets across two seeds; repair compares the same 885 candidates with one call per candidate; and tag closure separates missing stage delimiters from errors in predicted content. The appendix also distinguishes study-specific checkpoints from the final model and reports uncertainty for both favorable and inconclusive contrasts, helping readers assess what each experiment supports. (§4.3–4.5; Appendices C–E)

Critical weaknesses
  1. The empirical case does not isolate the proposed mechanisms, and the headline gain is concentrated in two benchmarks. VABench-Point and Where2Place contribute 89.8% of the 9.29-point mean; excluding both, the remaining five benchmarks improve by only 1.33 points on average. The main table compares the final model with the base, coverage and routing use earlier or different checkpoints, and repair uses a 27B model rather than the 9B student. There is no same-corpus, matched-budget SFT-only or flat-distillation baseline, no round-by-round measurement of recirculation, and the routing gain over a simpler declared route is within noise (+0.53, 95% CI [−0.61, 1.63]). VABench-Point is scored with any-point success; the final model’s cardinality and first-point score are not reported, although Table 7 shows the correction is material for the earlier checkpoint (81.12 vs 71.33). The corpus also includes FSD, cited to the same Yuan et al. (2026) work as VABench, and COCO-based QA replay, so image-level overlap with public tests is unverified. (Table 1; Table 5; §4.1–4.3; Appendix C, Table 7)

  2. The 10.09× repair multiplier is driven by cases where the feedback may be answer-equivalent. Of the 100 additional successes, 63 come from depth-order failures, which are only 8.8% of the 885-candidate pool; if these are two-entity rankings, “the order is wrong” already determines the alternative. The named arm is also told that the response failed while the generic arm is not, conflating an explicit failure signal with localization. Reporting task arity and a category-withheld/reversal baseline would clarify the multiplier. (§4.3; Appendix C; Tables 12–13)

Questions for authors
  1. Attribution. On the final 5.22M-record training corpus, how do SFT-only and flat distillation compare with MOPD under matched training budgets? How many student-to-data recirculation rounds contributed to the final checkpoint, and what evidence isolates their incremental benefit? How many final-model training runs were performed, and are any across-seed results available?

  2. Final-checkpoint controls. What are the final model’s VABench-Point first-point and fixed-cardinality scores relative to the base under identical settings, and how many points does each model emit per response? Does the public-test overlap audit cover the complete final training corpus? Please clarify the exact/near-image checking procedure and report results after any identified overlaps are excluded identically for both models.

  3. Repair feedback. How many of the 78 depth-order failures involve exactly two entities, and how does named feedback compare with deterministic reversal on those cases? Could you provide the full prompts for both repair arms and clarify whether both explicitly receive a failure signal? Is there a comparison that supplies this signal while withholding the failed category, to isolate the benefit of failure localization?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.