acceptodds
Under review as a conference paper at ICLR 2027

Capek 0.5: An Execution-Centric Vision-Language Model for Robot-Scene Reasoning

Abstract

Vision-language models are increasingly serving as the reasoning core of embodied agents. Robot execution is inherently iterative: each action reshapes the scene and physical state, continually renewing what must be perceived, reasoned about, and verified. Meeting these demands requires complementary capabilities that differ in supervision signals, prediction formats, and verification criteria. Existing approaches typically develop these capabilities against isolated, task-specific objectives, leaving open how they should be organized and integrated around execution as a whole. We present Capek 0.5, an embodied vision-language model built around an execution-centric capability taxonomy. Rather than organizing training by datasets or tasks, the taxonomy groups embodied capabilities according to their functional roles throughout execution and comprises four capability families: Spatial Reasoning, Temporal Understanding, Action Guidance, and State Verification. Each capability is first acquired by a dedicated specialist through reinforcement learning with verifiable rewards from a shared backbone, and the specialists are then consolidated into a single inference-time model through weight-space merging followed by routed policy-space distillation. We instantiate Capek 0.5 at the 2B and 35B-A3B scales and evaluate it from three complementary perspectives: comprehensive benchmark suites including Capek-StateBench, a new benchmark for state verification; a controlled study of capability retention from specialists to the unified model; and closed-loop evaluation in simulated embodied environments. Capek 0.5 improves the large majority of matched benchmark rows over its initialization, retains all four specialized capabilities in one checkpoint with quantified losses, and transfers to closed-loop embodied task execution. The results indicate that execution-centric specialization followed by consolidation is a practical recipe for building unified embodied vision-language models.Spatial reasoning Temporal understanding Action guidance State verification General capability

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied model, RL, State Verification

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_UiVwOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Capek0.5, a VLM focused on the execution-centric capabilities of embodied AI. It first introduces a taxonomy comprising four capability families: Spatial Reasoning, Temporal Understanding, Action Guidance, and State Verification. Guided by this taxonomy, the authors construct a training dataset by integrating multiple data sources, including newly generated data targeting state verification in the BEHAVIOR-1K simulator. Using this dataset, the authors first train 4 specialized expert models as teachers, then merge their checkpoints to initialize a student model, which is further optimized through multi-teacher on-policy distillation. The resulting model is evaluated on a range of general and embodied reasoning benchmarks, as well as in simulated environments such as EmbodiedBench to assess its closed-loop decision-making capabilities. Experimental results demonstrate that Capek0.5 outperforms other models of comparable size.

Critical strengths

This paper presents a comprehensive pipeline encompassing training data construction, benchmark development, and model training, reflecting a substantial amount of work. The evaluation extends beyond conventional QA tasks and offline benchmarks to assess the performance of embodied VLMs in closed-loop decision-making scenarios, providing a more practical and meaningful assessment.

Critical weaknesses
  1. The proposed capability taxonomy and the composition of the training dataset do not appear to offer substantial novelty over prior work. For example, RoboBrain2.5[1] has already considered spatial grounding and temporal understanding, while ERQA[2] and Cosmos-Reason[3] have introduced similar capability taxonomies. Although the authors construct new data for state verification, related efforts have also been explored in previous work, such as RoboBench[4], Progress-Bench[5].
  2. The proposed training pipeline builds upon existing techniques, including GRPO, TIES, and MOPD, but the paper does not sufficiently establish the necessity or effectiveness of each stage. The ablation results in Table 2 suggest that the individual contributions of TIES and MOPD are relatively modest on certain benchmarks. Moreover, these ablations are conducted on only a small selection of benchmarks, making it difficult to assess whether the observed benefits exist across different capabilities and tasks. In addition,l while the authors include a mix-RL baseline, it would also be valuable to consider simpler alternatives, such as mix-SFT, or mix-SFT followed by mix-RL. I think the supervised finetuning objective may lead to smaller gap among capabilities, potentially reducing the need for the proposed merging and distillation procedures.

[1] RoboBrain 2.5: Depth in Sight, Time in Mind

[2] Gemini Robotics: Bringing AI into the Physical World

[3] Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning

[4] Robobench: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models as Embodied Brain

[5] ProgressLM: Towards Progress Reasoning in Vision-Language Models

Questions for authors

Please see weaknesses.

2 comments

Thank you for your valuable comments and for recognizing the breadth of our training pipeline and the practical value of our evaluation. Below we address the raised questions and concerns.

Q1. What does Capek add beyond prior work?

A. RoboBrain 2.5 develops precise spatial reasoning and temporal value estimation; Gemini Robotics demonstrates broad embodied reasoning capabilities and introduces ERQA; and Cosmos-Reason1 uses capability ontologies to guide data construction and post-training. Capek builds on these advances to address how complementary execution capabilities can be developed through focused specialization and retained together in a unified model.

Our execution-centric organization connects capability grouping directly to specialist training and practical embodied execution. This makes it possible to examine both what each specialist acquires and how those gains change when integrated into one checkpoint. Through comparisons with joint training and alternative consolidation routes, we provide empirical evidence on the benefits and trade-offs of this recipe, alongside its transfer to closed-loop execution.

A further distinction is the role of State Verification. Related work studies state estimation and progress reasoning; Capek brings local physical-state verification and task-level progress assessment together as a dedicated capability family, with aligned supervision and evaluation. This explicitly connects checking execution outcomes with assessing task completion and what should happen next, complementing the spatial, temporal, and guidance capabilities within the same model.

Q2. Effectiveness of each training stage and broader ablation coverage

A. To better demonstrate the effectiveness of each training stage, we expand the original eight-benchmark consolidation study to 21 evaluation rows across all four capability families. The comparisons examine the role of each stage:

  • Capability acquisition through specialist GRPO. The matching specialists improve over the shared Qwen backbone on all eight original signature benchmarks, establishing the capability gains available before consolidation.
  • Weight-space initialization through TIES. Starting MOPD from the TIES-merged checkpoint produces better results than MOPD-only on 16 of 21 rows. Examples include Video-MME, which increases from 72.44 to 74.19, and Where2Place, which increases from 71.14 to 73.52.
  • Policy-space refinement through MOPD. Relative to TIES-only, TIES+MOPD improves all Action Guidance rows and both State Verification rows. PointBench increases from 77.16 to 81.22, StateBench-P from 68.00 to 76.80, and VABench-trace RMSE decreases from 117.49 to 108.15.
BenchmarkQwenMix-RLTIES-onlyMOPD-onlyTIES+MOPD
Spatial Reasoning
CV-Bench88.5688.9590.0490.4389.99
VSI-Bench60.8365.0571.0269.5470.69
OmniSpatial58.8458.5157.9357.2156.36
MindCube58.6760.3858.3868.4869.90
ERQA60.0061.2558.7557.0058.25
OpenEQA67.4867.7070.2368.7569.51
All-Angles-Bench65.4366.5165.4862.1064.12
RoboSpatial-Home63.4671.4069.9070.3172.80
EmbSpatial-Bench82.9483.2184.2984.2084.59
Temporal Understanding
EgoTempo38.2042.8049.4043.2044.60
QVHighlights-TimeLens54.8259.3659.1860.4058.26
LongVideoBench (64 frames)60.4863.3165.1464.4864.89
MVBench (64 frames)68.5572.7471.7671.1671.32
Video-MME (64 frames)71.3372.8174.8972.4474.19
Action Guidance
PointBench76.6680.4977.1679.2281.22
Where2Place57.6163.8972.5771.1473.52
RoboRefIt (test B)84.4585.0985.6886.0085.96
VABench-point54.7158.3357.6759.7859.33
VABench-trace (RMSE ↓)139.28121.80117.49108.83108.15
State Verification
StateBench-P65.8070.4068.0075.8076.80
StateBench-T43.7648.2343.5545.5246.21

Across the complete comparison, TIES+MOPD outperforms both Mix-RL and the original Qwen backbone on most of the benchmarks. Together, these results characterize the distinct roles of the three stages: specialist RL acquires execution-oriented capabilities, TIES provides a merged initialization, and MOPD further refines the consolidated policy. The resulting model combines complementary capabilities in a single checkpoint, although the gains are not uniform across benchmarks. These results support the effectiveness of the evaluated recipe without implying that each stage improves every capability.

Q3. How does Capek compare with simpler supervised alternatives?

A. Thank you for the suggestion. We train a Mix-SFT baseline from the same Qwen3.6-35B-A3B initialization using the pooled training examples from the four specialist families. Capek outperforms Mix-SFT on 16 of the 20 embodied evaluation rows, with advantages spanning spatial reasoning, temporal understanding, and action guidance.

BenchmarkQwenMix-SFTCapek 0.5
Spatial Reasoning
CV-Bench88.5688.0289.99
VSI-Bench60.8362.8570.69
OmniSpatial58.8448.7956.36
MindCube58.6781.4369.90
ERQA60.0048.0058.25
OpenEQA67.4858.6969.51
All-Angles-Bench65.4357.9764.12
RoboSpatial-Home63.4664.2772.80
EmbSpatial-Bench82.9481.5984.59
Temporal Understanding
EgoTempo38.2031.2044.60
QVHighlights-TimeLens54.8250.5758.26
LongVideoBench (64 frames)60.4862.9864.89
MVBench (64 frames)68.5572.1871.32
Video-MME (64 frames)71.3370.0474.19
Action Guidance
PointBench76.6668.9481.22
Where2Place57.6168.1073.52
RoboRefIt (test B)84.4584.1585.96
VABench-point54.7151.2859.33
State Verification
StateBench-P65.8093.2076.80
StateBench-T43.7680.5746.21
General Capabilities
IFEval92.4287.9991.31
LiveCodeBench v677.3159.9172.91
MMLU-Pro84.3983.1284.77
BFCL-v358.75–60.12

Mix-SFT performs substantially better on MindCube and StateBench-P/T, and slightly better on MVBench. This is particularly evident on StateBench, where the supervised targets closely match the evaluation tasks. For physical-state verification, SFT directly learns from state labels. For PVE, it also learns the required structured output through token-level supervision. In Capek, the subsequent GRPO stage instead optimizes format validity and task performance through rewards. This difference in supervision may contribute to the large gap on StateBench-T. We consider these results an important strength of the Mix-SFT baseline.

Despite these advantages, Capek performs better on 16 of the 20 embodied benchmarks, with improvements across spatial reasoning, temporal understanding, and action guidance. It also better preserves the general capabilities of the Qwen backbone. Mix-SFT drops by 4.43 points on IFEval and 17.40 points on LiveCodeBench v6, whereas Capek scores 91.31 and 72.91, respectively. On BFCL-v3, Mix-SFT produces outputs that cannot be parsed under the evaluation protocol, while Capek scores 60.12, compared with 58.75 for Qwen. Thus, Capek achieves broader embodied performance while retaining more of the backbone's instruction-following and coding capabilities.

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

The paper presents Capek 0.5, an embodied VLM post-trained from Qwen3.6-35B-A3B and Qwen3.5-2B. Training data is grouped into four "execution-centric" capability families (Spatial Reasoning, Temporal Understanding, Action Guidance, State Verification); one specialist per family is trained with GRPO from a shared backbone, and the specialists are then consolidated into a single model by TIES merging followed by routed multi-teacher on-policy distillation (MOPD). It also introduces a dual-level State Verification task built from BEHAVIOR-1K traces, with a matching benchmark, Capek-StateBench. Validation covers 34 benchmarks against the protocol-matched Qwen initialization, a consolidation ablation (Mix-RL, TIES, MOPD, TIES+MOPD; Table 2), and closed-loop evaluation on EmbodiedBench and VIGIL.

Critical strengths
  • Large, consistent, and fairly measured gains that carry over to closed-loop execution. Against its own Qwen initialization under the same evaluation protocol, Capek 0.5 improves 28/34 rows at 35B-A3B and 30/34 at 2B, including all 10 Action Guidance rows at both scales (e.g., PixMoPointsEval 62.42 → 74.06, Where2Place 57.61 → 73.52), while largely preserving general ability. The gains also hold on held-out simulated tasks: EB-HAB success rises from 46.0 to 63.0, with long-horizon success rising from 28.0 to 54.0 (Table 3). This matters because it shows, through a fair comparison, that the pipeline yields a stronger embodied model at two scales.
  • Two well-designed, reusable components. First, the consolidation ablation (Table 2) is clean and informative: TIES alone falls below the shared start on MindCube (58.38 vs 58.67) and StateBench-T (43.55 vs 43.76), while using TIES as the starting point for MOPD beats MOPD alone on 7 of 8 rows, a useful finding for anyone merging RL-trained specialists. Second, the State Verification formulation and Capek-StateBench address a real gap in embodied VLM evaluation, with a careful design (future frames hidden from the model, two separately typed progress definitions, next action scored separately).
Critical weaknesses
  • The evidence supports "specialize then consolidate", but not the execution-centric taxonomy that the paper puts at the center of its contribution. The framing and Contribution 1 rest on grouping capabilities by their role in execution ("Rather than organizing training by datasets or tasks…"), yet every experiment uses this one grouping, and there is no comparison with specialists split another way (e.g., by source dataset or by output format) and passed through the same TIES+MOPD pipeline. In the same way, Sec. 4.3 reads the closed-loop gains as showing that "consolidating complementary reasoning capabilities improves the maintenance of task context", but Tables 3 and 4 compare only against the Qwen initialization; neither Mix-RL nor the individual specialists are evaluated in closed loop, so the gains cannot yet be credited to consolidation rather than to the post-training data as a whole. This matters because the taxonomy is the paper's main conceptual claim. One alternative-grouping run (even at 2B) and a closed-loop Mix-RL evaluation would address it; otherwise the claims should be narrowed to the recipe.
Questions for authors
  1. [Clarification] Mix-RL is the only baseline that tests specialization against joint training, but it is described in a single line (Table 2 caption). What data mixture, number of training steps, and compute budget did Mix-RL use, compared with the four GRPO specialists plus the MOPD stage? Knowing that these were matched would strengthen the recipe claim considerably.
  2. [Clarification] Capek-StateBench is built from the same BEHAVIOR-1K traces and output format as the State Verification training data, and App. B.4 says splits are "inherited from the original sources". Are the benchmark's scenes, tasks, or trajectories disjoint from those used in training? If not, the +11.0 gain on StateBench-P should be read as an in-distribution result.
  3. [Suggestion] Results appear to come from single runs at temperature 0.7, and some gaps in Table 2 are small (e.g., VABench-point 57.67–59.78). Variance over repeated runs would show whether these rankings are stable.
2 comments

Q1. Execution-centric organization and its empirical support

A. This is a good question. We would like to clarify that our objective is to develop an effective organization for learning and integrating complementary embodied capabilities. The four families follow recurring information needs during execution: understanding scene geometry, interpreting temporal changes, producing actionable visual guidance, and verifying physical or task state. Each family brings together related tasks and supervision from multiple datasets and output formats, enabling focused specialist training before consolidation into a unified model. We do not claim to propose an optimal capability partition, and we have revised the Introduction to make this objective explicit.

The empirical comparison evaluates this organization as part of the complete specialization-and-consolidation recipe. Mix-RL trains the same backbone on the same pooled RL data without capability-based specialist separation. We expand the original Table 2 comparison to 21 evaluation rows across all four capability families. Capek outperforms Mix-RL on 16 of 21 rows and improves over the shared backbone on 18 of 21 rows.

BenchmarkQwenMix-RLTIES-onlyMOPD-onlyTIES+MOPD
Spatial Reasoning
CV-Bench88.5688.9590.0490.4389.99
VSI-Bench60.8365.0571.0269.5470.69
OmniSpatial58.8458.5157.9357.2156.36
MindCube58.6760.3858.3868.4869.9
ERQA6061.2558.755758.25
OpenEQA67.4867.770.2368.7569.51
All-Angles-Bench65.4366.5165.4862.164.12
RoboSpatial-Home63.4671.469.970.3172.8
EmbSpatial-Bench82.9483.2184.2984.284.59
Temporal Understanding
EgoTempo38.242.849.443.244.6
QVHighlights-TimeLens54.8259.3659.1860.458.26
LongVideoBench (64 frames)60.4863.3165.1464.4864.89
MVBench (64 frames)68.5572.7471.7671.1671.32
Video-MME (64 frames)71.3372.8174.8972.4474.19
Action Guidance
PointBench76.6680.4977.1679.2281.22
Where2Place57.6163.8972.5771.1473.52
RoboRefIt (test B)84.4585.0985.688685.96
VABench-point54.7158.3357.6759.7859.33
VABench-trace (RMSE ↓)139.28121.8117.49108.83108.15
State Verification
StateBench-P65.870.46875.876.8
StateBench-T43.7648.2343.5545.5246.21

These results support the effectiveness of the instantiated recipe. The four-way organization provides its execution-motivated structure; comparisons with alternative specialist partitions would address the separate question of which organization works best. We will align the abstract, Contribution 1, and experimental discussion around this scope: an execution-motivated capability organization implemented and evaluated through specialist training and consolidation, without implying demonstrated superiority over all partitions.

Q2. Attribution of closed-loop gains

A. Thanks for raising this question. We compare Capek with strong embodied VLM baselines in Tables 3 and 4 to assess its effectiveness in interactive simulated environments. These experiments evaluate whether the capabilities acquired through our complete training recipe translate into successful execution under changing observations and multi-step interaction.

Following the suggestion, we additionally evaluate Mix-RL on EmbodiedBench, providing a direct closed-loop comparison with joint training:

ModelEB-HAB success rate (%)EB-ALF success rate (%)
Qwen3.6-35B-A3B4650.7
Mix-RL46.351.7
Capek 0.5-35B-A3B6355.3

Capek outperforms Mix-RL by 16.7 percentage points on EB-HAB and 3.6 points on EB-ALF. This extends the offline comparison to closed-loop execution and supports the complete specialization-and-consolidation recipe relative to joint RL on the same underlying RL data.

The statement about "consolidating complementary reasoning capabilities" was intended to describe this transfer of the complete recipe to dynamic execution. The offline consolidation study in Table 2 examines different training and integration configurations, while Tables 3 and 4 evaluate the resulting model's closed-loop performance. We have modified the statement to express this distinction directly in the revision.

Q3. Mix-RL configuration and comparison conditions

A. Mix-RL trains the same backbone directly on the same combined training data pool used by the four specialists, using the same task-specific reward functions. Each specialist trains on its corresponding capability subset, whereas Mix-RL jointly learns from the combined pool without specialist learning and consolidation. For Mix-RL and our trained comparison routes, checkpoints are selected by development-set performance among the evaluated checkpoints.

Therefore, the comparison shares the same training data, backbone, checkpoint-selection criterion, and evaluation protocol. It evaluates the complete specialization-and-consolidation recipe against joint training, rather than isolating the taxonomy from the other components. We have clarified this scope in the revised manuscript.

Q4. StateBench separation and construction

A. For state verification, we construct both the training data and StateBench from the BEHAVIOR-1K training trajectories. Before generating any examples, we partition these source trajectories into two non-overlapping pools, one for training-data construction and the other reserved for benchmark construction. State-verification training examples are generated exclusively from the training pool. From the held-out benchmark pool, we select trajectories by task type to construct the physical-state and task-state evaluation tracks, StateBench-P and StateBench-T. This trajectory-level separation ensures that examples derived from the same source trajectory cannot appear in both training and evaluation. We have revised this construction procedure in Appendix B.4.

Q5. Stability across repeated evaluations

A. To quantify evaluation variability, we conducted additional evaluations of the fixed model checkpoints. The tables below combine the original evaluation with additional 4 independent runs using the same checkpoint and evaluation configuration. We report mean ± sample standard deviation, with five runs for each comparison model. Given computational resource constraints, we select benchmarks with small differences in the original results, where repeated evaluations are most informative for assessing ranking stability.

1. Capek versus the Qwen backbone

We report StateBench-T, specifically highlighted in the review, together with six general-capability benchmarks showing small differences in the original table.

BenchmarkQwen: mean ± SDCapek: mean ± SD
StateBench-T43.49 ± 1.4045.84 ± 0.30
MMVet78.53 ± 1.7078.58 ± 0.92
RealWorldQA82.78 ± 0.7383.06 ± 0.37
MMBench-EN92.43 ± 0.1992.50 ± 0.13
IFEval91.82 ± 0.5592.27 ± 0.72
MMLU-Pro84.66 ± 0.2284.69 ± 0.12
BFCL-v358.69 ± 0.6359.50 ± 0.68

On StateBench-T, Capek maintains a mean improvement of 2.36 points, close to the original 2.45-point difference. The general-capability results show closely matched performance, particularly on MMVet, MMBench-EN, and MMLU-Pro.

2. Joint-training and consolidation comparisons

The following table covers benchmarks spanning all four capability families.

CapabilityBenchmarkMix-RLTIES-onlyMOPD-onlyTIES+MOPD
SpatialMindCube61.22 ± 1.1757.57 ± 0.5569.53 ± 1.1970.40 ± 0.54
TemporalLongVideoBench63.69 ± 0.3964.83 ± 0.4064.25 ± 0.1965.11 ± 0.66
GuidanceVABench-point56.06 ± 1.8958.06 ± 0.6659.99 ± 1.1260.37 ± 0.80
StateStateBench-P70.95 ± 0.8568.00 ± 0.6574.85 ± 1.1575.36 ± 1.18
StateStateBench-T48.57 ± 1.5144.28 ± 0.8445.51 ± 0.4745.84 ± 0.30

The aggregated results preserve the main comparison with joint training: TIES+MOPD outperforms Mix-RL on most benchmarks, with StateBench-T remaining the exception. It also achieves better mean performance than TIES-only and MOPD-only.

Reviewer_9KvtOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents Capek 0.5, an embodied VLM post-trained from Qwen3.6-35B-A3B and Qwen3.5-2B. It organizes training around four capability families: Spatial Reasoning, Temporal Understanding, Action Guidance, and State Verification. A specialist for each family is trained with GRPO and verifiable rewards from a shared backbone. The specialists are then merged with TIES, and the merged model is refined with routed multi-teacher on-policy distillation (MOPD). The paper also introduces Capek-StateBench, which has a physical-state track (P) and a task-progress/next-action track (T), both built from BEHAVIOR-1K traces. Validation has three parts:

  • Benchmark tables: the 35B model improves 28/34 rows over its initialization and the 2B model improves 30/34.
  • A consolidation ablation comparing Mix-RL, TIES, MOPD, and TIES+MOPD.
  • Closed-loop evaluation on EmbodiedBench and VIGIL. Gains are notable on EB-HAB (+17 SR) and modest on VIGIL.
Critical strengths

The evaluation is broad and the reproduction is controlled. All baselines are rerun under one harness, and the Capek/Qwen pairs share decoding settings, so within-pair differences are attributable to post-training. Including closed-loop results alongside offline benchmarks makes the claims more credible than benchmark-only embodied VLM papers. The consolidation ablation (Table 2) is informative. It separates acquisition from integration and shows different biases across routes: TIES favors spatial/temporal, MOPD favors guidance/state. TIES+MOPD gives the best overall trade-off. This is a useful practical finding for multi-capability RL post-training.

Critical weaknesses

Possible in-distribution evaluation, with no variance estimates. Several training sources overlap with evaluation benchmarks: VSI-590K/VSI-Bench, MindCube, EmbSpatial-SFT/Bench, PixMo/PixMoPointsEval, RoboPoint/Where2Place, ShareRobot, and BEHAVIOR-1K/StateBench. Many of the largest gains are on exactly these rows. Deduplication is mentioned, but there is no contamination or split-overlap analysis. In addition, results come from single runs at temperature 0.7, and EmbodiedBench subsets have 50 episodes each. Many differences are therefore within plausible noise, e.g., VIGIL W 37.4 to 38.6, StateBench-T +2.45, and several general-retention rows.

Limited attribution of gains to the proposed design. TIES and MOPD are existing methods, so the main conceptual contribution is the execution-centric taxonomy, and the taxonomy itself is not ablated. There is no comparison to dataset- or format-based specialist groupings, or to a different number of specialists. It is also unclear whether Mix-RL received matched compute; total rollouts and steps should be compared against 4 specialists plus MOPD. Mix-RL outperforms TIES+MOPD on StateBench-T. Retention gaps relative to the experts are also non-trivial: EgoTempo 50.8 to 44.6 and StateBench-P 80.0 to 76.8.

Questions for authors

What is the overlap, in scenes, tasks, or source episodes, between training data and each benchmark? This matters most for VSI-Bench, MindCube, EmbSpatial, Where2Place, ShareRobot, and the BEHAVIOR-1K-derived StateBench splits. How was the StateBench held-out split constructed? Was Mix-RL compute- and data-matched to the specialist+MOPD pipeline? Can you report variance (seeds or bootstrap CIs) for the key rows and for the closed-loop results? Is there any evidence that the four-way taxonomy matters, as opposed to an arbitrary or dataset-based partition with the same consolidation recipe? Figure 1 averages heterogeneous metrics within domains, while Section 4.1 says averaging is avoided. Please reconcile these. The StateBench-T qualitative example lists “six” goal conditions but enumerates four, and its caption calls the prediction correct even though the progress estimate is wrong. The title differs between the PDF and OpenReview. There are duplicated references (Wang 2026a/b, Plizzari 2025a/b, Fu 2024/2025).

4 comments

Thank you for the detailed review and for recognizing the breadth of our evaluation, the controlled reproduction protocol, and the practical findings of the consolidation study. Below, we address the raised questions and concerns.

Q1. Training–evaluation separation and StateBench construction

A. We would like to clarify that all public datasets follow their original train–test splits, with no overlap between training and evaluation examples. Evaluation examples are excluded from both specialist training and consolidation. Scene- and task-level separation follows each dataset’s original protocol.

For state verification, we construct both the training data and StateBench from the BEHAVIOR-1K training trajectories. Before generating any examples, we partition these source trajectories into two non-overlapping pools, one for training-data construction and the other reserved for benchmark construction. State-verification training examples are generated exclusively from the training pool. From the held-out benchmark pool, we select trajectories by task type to construct the physical-state and task-state evaluation tracks, StateBench-P and StateBench-T. This trajectory-level separation ensures that examples derived from the same source trajectory cannot appear in both training and evaluation. We have revised this construction procedure in Appendix B.4.

We also like to note that the observed improvements extend beyond benchmarks associated with the training sources. The specialists are explicitly trained to acquire spatial, temporal, guidance, and verification capabilities, so evaluating these capabilities on the corresponding held-out benchmarks directly measures whether specialization is effective. The broader evaluation then tests how these gains extend across benchmarks and into execution. Capek 0.5-35B-A3B improves most of benchmark rows, including benchmarks beyond the related-source evaluations and five of eight general-capability controls. It also improves closed-loop performance on EmbodiedBench and VIGIL, neither of which is used for task-specific training. Together, these results support broader embodied capability and transfer to interactive execution, beyond improvements confined to related-source benchmarks.

Q2. Regarding Mix-RL training data and compute

A. Mix-RL trains the shared backbone directly on the same pooled training data used in our specialist-and-consolidation pipeline, with the same task-specific reward functions as specialist RL, to ensure a fair comparison. The difference is that it jointly learns the four capability families without specialist checkpoints or MOPD. For our method and the trained baselines, checkpoints are selected using the same development-set performance criterion.

The comparison is therefore data-matched, but not end-to-end compute-matched: our pipeline additionally includes MOPD student rollouts and teacher scoring, and the training routes differ in sequence lengths and execution costs. Table 2 compares the resulting recipes under a common backbone, training-data pool, checkpoint-selection criterion, and evaluation protocol. We have clarified these in the revision.

TIES+MOPD outperforms Mix-RL on seven of eight signature benchmarks, and the results demonstrate the practical effectiveness of our specialist training followed by consolidation across the evaluated capability suite.

Q3. Evaluation variability and the stability of small differences

A. We conducted additional evaluations of the fixed model checkpoints. The following tables combine the original evaluation with all additional independent runs using the same checkpoint and evaluation configuration. We report mean ± sample standard deviation, with five runs for Capek/TIES+MOPD and each comparison model.

  • Capek versus the Qwen backbone: We focus the presentation on StateBench-T, specifically highlighted in the review, and six general-capability benchmarks with small differences in the original table.
BenchmarkQwen: Mean ± SDCapek: Mean ± SD
StateBench-T43.49 ± 1.4045.84 ± 0.30
MMVet78.53 ± 1.7078.58 ± 0.92
RealWorldQA82.78 ± 0.7383.06 ± 0.37
MMBench-EN92.43 ± 0.1992.50 ± 0.13
IFEval91.82 ± 0.5592.27 ± 0.72
MMLU-Pro84.66 ± 0.2284.69 ± 0.12
BFCL-v358.69 ± 0.6359.50 ± 0.68

On StateBench-T, Capek maintains a mean improvement of 2.36 points, close to the original 2.45-point difference. The general-capability results show closely matched performance, particularly on MMVet, MMBench-EN, and MMLU-Pro.

  • Joint-training and consolidation comparisons: The following table covers benchmarks spanning all four capability families.
CapabilityBenchmarkMix-RLTIES-onlyMOPD-onlyTIES+MOPD
SpatialMindCube61.22 ± 1.1757.57 ± 0.5569.53 ± 1.1970.40 ± 0.54
TemporalLongVideoBench63.69 ± 0.3964.83 ± 0.4064.25 ± 0.1965.11 ± 0.66
GuidanceVABench-point56.06 ± 1.8958.06 ± 0.6659.99 ± 1.1260.37 ± 0.80
StateStateBench-P70.95 ± 0.8568.00 ± 0.6574.85 ± 1.1575.36 ± 1.18
StateStateBench-T48.57 ± 1.5144.28 ± 0.8445.51 ± 0.4745.84 ± 0.30

The aggregated results preserve the main comparison with joint training: TIES+MOPD outperforms Mix-RL on most benchmarks, with StateBench-T remaining the exception. It also achieves better mean performance than TIES-only and MOPD-only.

Q4. Execution-centric organization and its empirical support

A. This is a good point. Our goal is to identify an effective capability organization that facilitates learning complementary embodied skills. We derive the four families from recurring information needs during execution. Each family brings together supervision from multiple datasets and output formats that serves a related execution role. This organization is intended to facilitate focused learning within each specialist before integrating their complementary capabilities.

The taxonomy is therefore an execution-motivated design, rather than a claim that four families are optimal partition. Our empirical question is whether this organization, implemented through specialist training and consolidation, provides an effective route to a unified embodied model.

Mix-RL provides a joint-training reference: it learns from the same pooled RL data without capability-based specialist separation. In Table 2, our recipe outperforms Mix-RL on seven of eight signature benchmarks. We further expand the comparison to 21 evaluation rows across all four capability families, where TIES+MOPD outperforms Mix-RL on most rows and also improves over the shared backbone (in point 6). These comparisons support the effectiveness of the proposed organization within our training recipe.

Q5. Results of Mix-RL on EmbodiedBench

EmbodiedBench. We additionally evaluate Mix-RL on EB-HAB and EB-ALF to compare joint RL with our complete specialization-and-consolidation recipe in closed-loop execution.

ModelEB-HAB SR (%)EB-ALF SR (%)
Qwen3.6-35B-A3B4650.7
Mix-RL46.351.7
Capek 0.5-35B-A3B6355.3

Capek outperforms Mix-RL by 16.7 percentage points on EB-HAB and 3.6 points on EB-ALF. This comparison extends the offline consolidation analysis to closed-loop execution and supports the effectiveness of the complete specialization-and-consolidation recipe relative to joint RL.

Q6. Capability retention in a single model

A. We would like to point out that our retention objective is to obtain one single model that performs well across complementary embodied capabilities while preserving general capabilities, with no need to select or maintain four specialist models at inference time. Each specialist’s in-domain performance provides a reference for measuring how much of its acquired capability survives consolidation. We further extend the original eight-metric comparison to additional evaluation rows across all four capability families. In the expanded comparison below, Capek improves over the shared backbone on most of rows with available backbone results.

BenchmarkQwenMix-RLTIES-onlyMOPD-onlyTIES+MOPD
Spatial Reasoning
CV-Bench88.5688.9590.0490.4389.99
VSI-Bench60.8365.0571.0269.5470.69
OmniSpatial58.8458.5157.9357.2156.36
MindCube58.6760.3858.3868.4869.9
ERQA6061.2558.755758.25
OpenEQA67.4867.770.2368.7569.51
All-Angles-Bench65.4366.5165.4862.164.12
RoboSpatial-Home63.4671.469.970.3172.8
EmbSpatial-Bench82.9483.2184.2984.284.59
Temporal Understanding
EgoTempo38.242.849.443.244.6
QVHighlights-TimeLens54.8259.3659.1860.458.26
LongVideoBench (64 frames)60.4863.3165.1464.4864.89
MVBench (64 frames)68.5572.7471.7671.1671.32
Video-MME (64 frames)71.3372.8174.8972.4474.19
Action Guidance
PointBench76.6680.4977.1679.2281.22
Where2Place57.6163.8972.5771.1473.52
RoboRefIt (test B)84.4585.0985.688685.96
VABench-point54.7158.3357.6759.7859.33
VABench-trace (RMSE ↓)139.28121.8117.49108.83108.15
State Verification
StateBench-P65.870.46875.876.8
StateBench-T43.7648.2343.5545.5246.21

Beyond the four embodied capability families, Table 1 shows improvements on five of eight general-capability controls for the 35B-A3B model, supporting the largely preserved general performance. Tables 3 and 4 additionally show that the unified model transfers to closed-loop embodied execution. Together, these results demonstrate that substantial gains across multiple capabilities can coexist in a single inference-time checkpoint, with the specialist-to-unified differences explicitly quantifying the consolidation trade-offs.

On the two benchmarks highlighted by the reviewer, the matching-specialist and unified-model scores in Table 2 are 50.80 versus 44.60 on EgoTempo and 80.00 versus 76.80 on StateBench-P. Capek remains 6.40 and 11.00 points above the shared backbone, respectively. These results quantify the capability-specific trade-offs of consolidation: our objective is to combine strong performance across complementary embodied capabilities in a single inference-time checkpoint while largely preserving general capabilities.

Q7. Regarding the presentation issues

A. Thank you for highlighting these points. Figure 1 and Table 1 serve different purposes. Figure 1 uses within-domain averages to provide a compact visual overview of the models’ capability profiles. NaviTrace, VABench-trace, and ShareRobot-Trajectory are excluded because their trajectory metrics have different scales or directions. Table 1 supports detailed benchmark-by-benchmark analysis, preserving each task’s native metric without adding an aggregate average that could obscure task-specific differences. We have revised the paper to better clarify this.

The StateBench-T example illustrates a correct next-action prediction alongside an incorrect progress estimate (75% versus 83.33%). The six-versus-four inconsistency occurs in the model’s original reasoning. We will label the example as partial success and explain the reference progress calculation while preserving the original output. We have also fixed the title and the duplicated reference issues.

I appreciate the thoroughness in the responses. Solid details. I increased the score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive feedback and valuable suggestions. We appreciate their recognition of the comprehensive and controlled evaluation (9Kvt, aBpW, UiVw), practical insights from specialist training and capability consolidation (9Kvt, aBpW), effectiveness in closed-loop embodied execution (9Kvt, aBpW, UiVw), and the dual-level State Verification formulation and benchmark (aBpW, UiVw).

To address the raised concerns, we have conducted additional experiments and clarified our claims:

  • Expanded consolidation ablations: We extended the original 8-benchmark study to 21 evaluation rows across all four capability families. Capek outperforms Mix-RL on 16/21 benchmarks, providing broader evidence for the complete specialization-and-consolidation recipe.
  • Additional baselines: We evaluated Mix-RL on closed-loop EmbodiedBench, where Capek improves success rates by 16.7 points on EB-HAB and 3.6 points on EB-ALF. We also introduced a Mix-SFT baseline, with Capek outperforming it on 16/20 embodied benchmarks, while better preserving general capabilities.
  • Evaluation stability: We conducted additional evaluations with independent runs per fixed checkpoint, reporting mean and standard deviation for selected benchmarks and consolidation comparisons.
  • Data separation and reproducibility: We clarified training–evaluation separation, including trajectory-level disjointness for Capek-StateBench, and documented the data-matched but not compute-matched comparison with Mix-RL.
  • Clarified contribution scope and presentation: We refined the execution-centric taxonomy claims, distinguishing the effectiveness of the complete recipe from the optimality of the capability partition, and corrected the identified presentation issues.

We believe these additional experiments and revisions strengthen the empirical evidence and address the reviewers' major concerns. We sincerely thank all reviewers again for their time and constructive feedback. We would also be happy to address any further questions or concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.