- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces ActiveBrain, a benchmark and post-training framework for improving active cognition in embodied agents. The authors define active cognition as the ability to adapt decisions based on newly acquired observations and action consequences while maintaining the original task objective. The paper evaluate AC by building benchmarks and study whether this capability can be learned. The post-training pipeline combines supervised fine-tuning, on-policy distillation, and multi-turn reinforcement learning. The authors also give the experimental analyses of the post-training pipeline.
Critical strengths
- This paper is a carefully constructed benchmark for evaluating instruction completion from incomplete embodied states. The proposed benchmark covers several practically relevant challenges in embodied interaction, including active information acquisition, instance verification, and recovery from failed actions.
- This paper shows substantial empirical improvements supported by the proposed verified continuation dataset and detailed behavioral analysis. This forms a useful data contribution which help to substantially improve embodied task completion. The authors also provide informative analyses of action repetition, execution failures, history interventions, and general multimodal performance retention.
Critical weaknesses
- While the overall performance improvements are substantial, the experiments do not sufficiently establish which proposed components are necessary or distinctly effective compared with simpler alternatives. In particular, SFT alone improves the mean success rate from 8.1% to 38.8%, while OPD increases it to 43.1% and the final ARL stage adds only 0.9 percentage points. Moreover, the transition from OPD to ARL produces mixed effects across tracks, improving Recover and Verification while reducing Explore and Long-Horizon performance. The authors could put fine-grained analyses on the effectiveness and necessity of several components, eg, ARL.
- Insufficient evidence of generalizability. In embodied interaction, both training and evaluation are conducted within the AI2-THOR ecosystem. Although the authors use held-out scenes, it remains unclear whether the learned capabilities generalize beyond the specific failure patterns and environment dynamics represented in the benchmark. I would encourage the authors to evaluate more diverse failure scenarios and, ideally, provide cross-environment or cross-simulator evaluations to better substantiate the generalizability of the proposed Active Cognition capabilities.
Questions for authors
- What are the independent benefits of OPD and the proposed ARL components? In particular, how does AC-Reward with Tree-GRPO compare against standard GRPO or PPO with alternative reward specifications?
- Could the authors evaluate more diverse failure scenarios and, ideally, provide cross-environment or cross-simulator evaluations to better substantiate the generalizability of the proposed Active Cognition capabilities?
4 comments
Post-Training Components and Cross-Simulator Generalization
We thank the reviewer for recognizing the benchmark's coverage of practical embodied challenges and the value of our verified continuation data and behavioral analyses. We address each point below.
Weakness 1: Effectiveness of post-training components
While the overall performance improvements are substantial, the experiments do not sufficiently establish which proposed components are necessary or distinctly effective compared with simpler alternatives. In particular, SFT alone improves the mean success rate from 8.1% to 38.8%, while OPD increases it to 43.1% and the final ARL stage adds only 0.9 percentage points. Moreover, the transition from OPD to ARL produces mixed effects across tracks, improving Recover and Verification while reducing Explore and Long-Horizon performance. The authors could put fine-grained analyses on the effectiveness and necessity of several components, eg, ARL.
Response. SFT improves the six-track average success rate from 8.1% for Qwen3.5-4B to 38.8%, and the OPD student reaches 43.1%. On seven general multimodal benchmarks, SFT and OPD average 56.7 and 72.5, respectively, compared with 71.1 for the base model. The OPD student is initialized from the base model and distilled from the SFT teacher; the SFT–OPD difference therefore does not isolate the effect of distillation alone.
Starting from the OPD model, ARL raises the average success rate from 43.1% to 44.0%. Under the same OPD initialization and training budget, standard GRPO with a 0/1 reward reaches 42.9%, versus 44.0% for Tree-GRPO with AC-Reward. Because this comparison changes both the reward and the optimization method, it does not isolate the contribution of either component.
We agree that the overall benefit of ARL is modest and varies across tracks. The current results suggest one feasible path for learning from embodied interaction; further work is needed to improve the reward and optimization design.
Weakness 2: Generalization
Insufficient evidence of generalizability. In embodied interaction, both training and evaluation are conducted within the AI2-THOR ecosystem. Although the authors use held-out scenes, it remains unclear whether the learned capabilities generalize beyond the specific failure patterns and environment dynamics represented in the benchmark. I would encourage the authors to evaluate more diverse failure scenarios and, ideally, provide cross-environment or cross-simulator evaluations to better substantiate the generalizability of the proposed Active Cognition capabilities.
Response. The training and evaluation scenes and task trajectories in ActiveBrain-Bench do not overlap. To evaluate generalization beyond ActiveBrain-Bench, we tested ActiveBrain-4B and its Qwen3.5-4B base model on EmbodiedBench:
| Benchmark | Model | Avg. | Base | Comm. | Comp. | Spatial | Visual | Long |
|---|---|---|---|---|---|---|---|---|
| EB-ALFRED | Qwen3.5-4B | 27.33 | 26 | 28 | 36 | 22 | 30 | 22 |
| EB-ALFRED | ActiveBrain-4B | 35.33 | 36 | 46 | 44 | 22 | 34 | 30 |
| EB-Habitat | Qwen3.5-4B | 28.67 | 66 | 10 | 34 | 30 | 20 | 12 |
| EB-Habitat | ActiveBrain-4B | 32.33 | 62 | 28 | 26 | 26 | 38 | 14 |
ActiveBrain-4B improves the average by 8.0 percentage points on EB-ALFRED and 3.7 percentage points on EB-Habitat. Scores are percentages; Avg. is the unweighted mean of the six categories. EB-Habitat also provides a cross-simulator evaluation.
The existing Recover track covers blocked navigation (90 cases), premature interaction (84), search stagnation (19), and rejected view control (7), with failure-state handoffs from four source models (Fig. A1).
Thank you for the responses and clarifications. I appreciate the authors' efforts, which address my concern. I maintain my positive assessment and my score of 3 (weak acceptance).
Thank you very much for your kind reply and positive assessment. We are grateful for your suggestions on generalization and the post-training components, and we are glad to hear that our responses and clarifications addressed your concern. If you feel that the additional evidence strengthens our paper, we would be grateful if you could take it into account in your final assessment. We sincerely appreciate the care and dedication you have shown throughout the review process, and we respect your judgment.
Post-Training Components and Cross-Simulator Generalization
We thank the reviewer for recognizing the benchmark's coverage of practical embodied challenges and the value of our verified continuation data and behavioral analyses. We address each point below.
Question 1: Independent benefits of OPD and ARL
Response. The OPD student reaches 43.1% embodied success, compared with 38.8% after SFT, and averages 72.5 on seven general multimodal benchmarks, compared with 71.1 for the base model. The OPD student is initialized from the base model and distilled from the SFT teacher; the SFT–OPD difference therefore does not isolate the effect of distillation alone.
We agree that SFT accounts for the largest improvement in our pipeline. Starting from the OPD model, ARL increases the average success rate from 43.1% to 44.0%. Verification improves from 9.5% to 15.5% and Recover from 43.0% to 53.0%, while Explore decreases from 70.5% to 63.0% and Long-Horizon from 64.7% to 54.7%.
Using the same OPD initialization and training budget, standard GRPO with a 0/1 reward achieves 42.9%, while Tree-GRPO with AC-Reward reaches 44.0%. Because this comparison changes both the reward and the optimization method, it does not isolate the contribution of either component. These results suggest that ARL provides one feasible path for learning from embodied interaction, but its overall benefit is modest and varies across tracks. Further work is needed to improve the reward and optimization design.
Question 2: Diverse failures and cross-simulator evaluation
Response. We evaluated ActiveBrain-4B and its Qwen3.5-4B base model on EmbodiedBench:
ActiveBrain-4B improves the average by 8.0 percentage points on EB-ALFRED and 3.7 percentage points on EB-Habitat. Scores are percentages; Avg. is the unweighted mean of the six categories. EB-Habitat provides a cross-simulator evaluation.
The existing Recover track covers blocked navigation (90 cases), premature interaction (84), search stagnation (19), and rejected view control (7), with failure-state handoffs from four source models (Fig. A1). Table A14 reports SFT and OPD success rates by failure type.