acceptodds
Under review as a conference paper at ICLR 2027

ActiveBrain: Benchmarking and Learning Active Cognition in Embodied Agents

11 views

Abstract

Can embodied agents actively determine what remains unresolved throughout an interaction and act accordingly to complete the original instruction? This requires seeking information beyond the current view, distinguishing the instruction-specified target from similar instances, and revising ineffective behavior. We call this Active Cognition (AC): actively updating decisions based on new observations and action consequences while preserving the instructed goal. To evaluate AC, we introduce ActiveBrain-Bench, comprising 1,035 tasks across six tracks with a common policy interface. Its Adapt track separately measures local progress and instruction completion. Zero-shot evaluations of 18 open-weight models and GPT-5.5 reveal fragmented strengths across the six tracks. In Adapt, models may make local progress yet leave the instruction unfinished. To study whether this capability can be learned, ActiveBrain constructs 11.5K replay-verified continuations from generated start states and rollout failure states. Each continuation retains RGB–action history and ends in simulator-verified completion. Our post-training pipeline combines supervised fine-tuning, on-policy distillation, and multi-turn reinforcement learning with environment-derived rewards. ActiveBrain-4B raises Qwen3.5-4B's unweighted six-track mean success rate from 8.1% to 44.0%. ActiveBrain-2B reaches 35.0%, above all evaluated zero-shot models. Across seven general multimodal benchmarks, ActiveBrain-4B records an unweighted mean of 72.6, versus 71.1 for its Qwen3.5-4B initialization.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied agents, Active cognition, Vision-Language Models, Active perception, Benchmarking and Evaluation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CsDTOpenReview ↗
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
  1. 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.
  2. 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
  1. 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.
  2. 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
  1. 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?
  2. 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
AuthorsOpenReview ↗

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

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?

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

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?

Response. We evaluated ActiveBrain-4B and its Qwen3.5-4B base model on EmbodiedBench:

BenchmarkModelAvg.BaseComm.Comp.SpatialVisualLong
EB-ALFREDQwen3.5-4B27.33262836223022
EB-ALFREDActiveBrain-4B35.33364644223430
EB-HabitatQwen3.5-4B28.67661034302012
EB-HabitatActiveBrain-4B32.33622826263814

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.

AuthorsOpenReview ↗

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:

BenchmarkModelAvg.BaseComm.Comp.SpatialVisualLong
EB-ALFREDQwen3.5-4B27.33262836223022
EB-ALFREDActiveBrain-4B35.33364644223430
EB-HabitatQwen3.5-4B28.67661034302012
EB-HabitatActiveBrain-4B32.33622826263814

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.

Reviewer_Ze57OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces ActiveBrain-Bench, a benchmark for evaluating active cognition in embodied agents, focusing on their ability to actively gather information, verify target objects, recover from failures, adapt to ineffective actions, and complete cross-room and long-horizon tasks. Built in AI2-THOR with held-out ProcTHOR-10K scenes, the benchmark contains 1,035 tasks across six tracks. To improve performance, the authors propose ActiveBrain, a training framework that combines supervised fine-tuning (SFT) on verified task-continuation trajectories, on-policy distillation (OPD), and agentic reinforcement learning (ARL). The 4B model achieves a 44.0% average success rate, substantially outperforming the Qwen3.5-4B baseline (8.1%) and surpassing larger models such as Qwen3.5-27B and GPT-5.5 on the benchmark. Ablation studies further highlight the importance of reasoning supervision, failure-state training data, and action-history context. Overall, the paper demonstrates that active cognition can be systematically benchmarked and improved through targeted training.

Critical strengths
  1. This paper introduces ActiveBrain-Bench, a benchmark structured around six capability dimensions: Explore, Verification, Recover, Adapt, Cross-Room, and Long-Horizon. The authors also developed the ActiveBrain model; despite its small 4B parameter size, it delivers performance comparable to models like Qwen-27B, others exceeding 100B parameters, and even the closed-source GPT-5, which is very impressive.

  2. This paper employs an efficient data construction method to create replay-verified task continuation data, enabling the process to resume directly from a failure state toward the final goal.

  3. This paper is logically coherent and easy to follow. It also features comprehensive and extensive experiments and presents a number of significant conclusions.

Critical weaknesses
  1. In the experiment section, although models with varying parameters were evaluated, the focus remained primarily on the Qwen series. The scope could be broadened to include a wider variety of models, for instance, closed-source models such as GPT-6-Astra, Gemini, and Claude, as well as open-source models like DeepSeek-V4[1] and Cosmos3[2]. Furthermore, incorporating more embodied AI models, such as RxBrain[3], HY-Embodied-0.5[4], and RoboBrain[5], would help make the benchmark more comprehensive and reliable.

  2. This benchmark currently relies on closed-loop testing conducted within a simulation environment. Given that there is a sim-to-real gap, would it be possible to incorporate testing in a real-world environment? It would be valuable to see whether similar conclusions hold true when experiments are performed under real-world conditions.

[1] Xu, Anyi, et al. "Deepseek-v4: Towards highly efficient million-token context intelligence." arXiv preprint arXiv:2606.19348 (2026). [2] Agarwal, Niket, et al. "Cosmos 3: Omnimodal world models for physical ai." arXiv preprint arXiv:2606.02800 (2026). [3] Liang, Haotian, et al. "RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination." arXiv preprint arXiv:2607.14187 (2026). [4] Team, H. Y., et al. "HY-Embodied-0.5: Embodied Foundation Models for Real-World Agents." arXiv preprint arXiv:2604.07430 (2026). [5] Team, BAAI RoboBrain, et al. "Robobrain 2.0 technical report." arXiv preprint arXiv:2507.02029 (2025).

Questions for authors
  1. The AI2THOR simulation environment is somewhat outdated. Why was it still chosen to build this simulation setup?
3 comments
AuthorsOpenReview ↗

Expanded Baselines, Real-world evaluation, and Simulator Choice

We thank the reviewer for recognizing the breadth of our benchmark, the performance of ActiveBrain, and the value of our replay-verified task continuations. We address each point below.

Weakness 1: Diversity of evaluated models

In the experiment section, although models with varying parameters were evaluated, the focus remained primarily on the Qwen series. The scope could be broadened to include a wider variety of models, for instance, closed-source models such as GPT-6-Astra, Gemini, and Claude, as well as open-source models like DeepSeek-V4[1] and Cosmos3[2]. Furthermore, incorporating more embodied AI models, such as RxBrain[3], HY-Embodied-0.5[4], and RoboBrain[5], would help make the benchmark more comprehensive and reliable.

Response. Following your suggestion, we expanded the evaluation to include a wider range of models. In addition to the Qwen family, we evaluated RynnBrain and RoboBrain models, GPT-6 Astra and Luna, DeepSeek 4.1 Flash, HY4, and Cosmos3 Nano. We also evaluated HY-Embodied-0.5 and RxBrain; both achieved 0% success and frequently failed to follow the required output format.

The results on ActiveBrain-Bench are:

ModelExp.Ver.Rec.Adp.CRLHSR
Qwen3.5-4B15.01.517.01.05.98.08.1
Qwen3.5-27B55.015.538.039.016.534.033.0
RynnBrain-8B0.00.07.00.00.00.01.2
RynnBrain-30B-A3B7.53.56.00.04.77.34.8
RynnBrain1.1-9B19.02.514.07.04.712.710.0
RoboBrain2.5-4B0.51.011.00.00.01.32.3
RoboBrain2.5-8B-NV0.00.03.50.00.00.00.6
DeepSeek 4.1 Flash56.520.039.568.018.849.342.0
HY462.526.051.571.034.150.649.3
GPT-5.552.019.041.039.516.534.033.7
GPT-6 Astra70.045.062.078.036.556.758.0
GPT-6 Luna56.528.530.074.024.722.739.4
Cosmos3 Nano0.51.013.526.50.00.06.9
ActiveBrain-4B63.015.553.042.535.354.744.0

GPT-6 Astra achieves the highest scores across all six tracks shown in the table, with an average SR of 58.0%. The models also show different strengths: HY4, DeepSeek 4.1 Flash, and GPT-6 Luna perform strongly on Adapt, while ActiveBrain-4B outperforms these three models on Recover and Long-Horizon.

Exp., Ver., Rec., Adp., CR, and LH denote Explore, Verification, Recover, Adapt, Cross-Room, and Long-Horizon. Scores are percentages; Adapt reports Goal@20, and SR is the unweighted mean of the six track scores. The listed baselines are evaluated zero-shot, while ActiveBrain-4B is post-trained with our trajectory data.

Following your suggestion, we have also cited DeepSeek-V4, Cosmos 3, RxBrain, HY-Embodied-0.5, and RoboBrain 2.0 in the revised paper.

Weakness 2: Real-world evaluation

This benchmark currently relies on closed-loop testing conducted within a simulation environment. Given that there is a sim-to-real gap, would it be possible to incorporate testing in a real-world environment? It would be valuable to see whether similar conclusions hold true when experiments are performed under real-world conditions.

Response. We conducted qualitative physical-robot Explore and Verification experiments, showing that the protocol can run in a hardware loop. Fair quantitative comparison remains challenging because real-world outcomes can also depend on execution components, such as VLA models and low-level control.

Question: Choice of AI2-THOR

The AI2THOR simulation environment is somewhat outdated. Why was it still chosen to build this simulation setup?

Response. AI2-THOR provides directly executable navigation, viewpoint, and object-interaction actions. This reduces dependence on external VLA models and low-level robot control, allowing us to focus on the agent's closed-loop decisions based on RGB observations and action history.

We also evaluated ActiveBrain-4B and its Qwen3.5-4B base model on EmbodiedBench:

BenchmarkModelAvg.BaseComm.Comp.SpatialVisualLong
EB-ALFREDQwen3.5-4B27.33262836223022
EB-ALFREDActiveBrain-4B35.33364644223430
EB-HabitatQwen3.5-4B28.67661034302012
EB-HabitatActiveBrain-4B32.33622826263814

Scores are percentages; Avg. is the unweighted mean of the six categories. EB-Habitat provides a cross-simulator evaluation.

I appreciate the author’s detailed rebuttal; they conducted numerous experiments that addressed my concerns, so I have decided to maintain my score.

Thank you very much for your kind reply and positive assessment. We are grateful for your suggestions on model comparisons and the applicability of our work, and we are glad to hear that our response and additional experiments addressed your concerns. If you feel that these additions strengthen our paper, we would be grateful if you could take them 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.

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

This paper aims to evaluate and improve the active-information-acquiring ability of embodied agents. The ability called Active Cognition is defined as the agents' capability of actively updating the decisions based on observations and action consequences, while keeping the original goal.

A benchmark called ActiveBrain-Bench is proposed to quantitatively measure this ability, where the 1,035 tasks from 6 tracks are generated by the simulator AI2-THOR from held-out ProcTHOR-10K test scenes. Evaluation on 18 open models and GPT-5.5 across 5 model families shows that the models have scattered strengths across different aspects.

A model called ActiveBrain is further developed by training on 11.5K replay-verified trajectories, together with supervised fine-tuning, on-policy distillation, and multi-turn reinforcement learning. Results on ActiveBrain-Bench show the trained model significantly improves the performance and beats all the baselines including GPT-5.5. Ablations show that in the SFT dataset, the history of past frames and actions, written rationales, and failure states are all important for this ability.

Critical strengths

This paper is well written and easy to follow. The authors try to define and develop an important ability in embodied AI, where the AI is expected to actively explore and acquire new information when the task gets stuck. A new benchmark, dataset, and model family are developed to reach this goal. Comprehensive ablations are also presented to help understand the key factors behind this ability.

Critical weaknesses

The major concern is the quality part.

  1. The training distribution closely matches the benchmark distribution, which raises concerns about overfitting and its generalizability to other scenarios. To address this concern, I would recommend that the authors test the model on other benchmarks as well, such as [1].

  2. More evaluations of the latest agents are encouraged, such as GPT-6.1-Sol, Claude Opus 5.5, GPT-6 Astra.

References

[1]: Yang, Rui, et al. "Embodiedbench: Comprehensive benchmarking multi-modal large language models for vision-driven embodied agents." ICML 2025.

Questions for authors
  • Are there any plans for releasing the 11.5K training dataset?
  • I am curious about the latest models' performance, such as Astra's, on the proposed benchmark.
  • In the most "cognitive" track, Verification, ActiveBrain-4B still underperforms GPT-5.5. It would be nice to have a case study for that to help understand the reason.
3 comments
AuthorsOpenReview ↗

Generalization, New Baselines, and Verification Case Study

We thank the reviewer for recognizing the clarity of our paper and the importance of studying active cognition in embodied agents. We address each point below.

Weakness 1: Generalization

The training distribution closely matches the benchmark distribution, which raises concerns about overfitting and its generalizability to other scenarios. To address this concern, I would recommend that the authors test the model on other benchmarks as well, such as [1].

Response. To reduce overlap between training and evaluation, we use separate ProcTHOR scene splits: training uses scenes from the 10K training split, while ActiveBrain-Bench uses scenes from the held-out 1K test split. This ensures that training and evaluation do not share scenes and tests generalization to unseen scenes.

Following your suggestion, we further 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:

BenchmarkModelAvg.BaseComm.Comp.SpatialVisualLong
EB-ALFREDQwen3.5-4B27.33262836223022
EB-ALFREDActiveBrain-4B35.33364644223430
EB-HabitatQwen3.5-4B28.67661034302012
EB-HabitatActiveBrain-4B32.33622826263814

Scores are percentages; Avg. is the unweighted mean of the six categories. EB-Habitat also provides a cross-simulator evaluation. The revised paper reports these results in Table 6.

Weakness 2: Evaluation of recent models

More evaluations of the latest agents are encouraged, such as GPT-6.1-Sol, Claude Opus 5.5, GPT-6 Astra.

Response. Following your suggestion, we additionally evaluated GPT-6 Astra, GPT-6 Luna, DeepSeek 4.1 Flash, HY4, and Cosmos3 Nano. The results are shown below:

ModelExp.Ver.Rec.Adp.CRLHSR
GPT-5.552.019.041.039.516.534.033.7
DeepSeek 4.1 Flash56.520.039.568.018.849.342.0
HY462.526.051.571.034.150.649.3
GPT-6 Astra70.045.062.078.036.556.758.0
GPT-6 Luna56.528.530.074.024.722.739.4
Cosmos3 Nano0.51.013.526.50.00.06.9
ActiveBrain-4B63.015.553.042.535.354.744.0

GPT-6 Astra achieves the highest scores across all six tracks shown in the table, with an average SR of 58.0%. The models also show different strengths: HY4, DeepSeek 4.1 Flash, and GPT-6 Luna perform strongly on Adapt, while ActiveBrain-4B outperforms these three models on Recover and Long-Horizon.

Exp., Ver., Rec., Adp., CR, and LH denote Explore, Verification, Recover, Adapt, Cross-Room, and Long-Horizon. Scores are percentages; Adapt reports Goal@20, and SR is the unweighted mean of the six track scores. We report the updated baseline comparisons in Table 1 of the revised paper.

Question 1: Release of the training dataset

Are there any plans for releasing the 11.5K training dataset?

Response. Yes. We plan to release the 11.5K replay-verified training continuation trajectories.

Question 2: Astra's performance

I am curious about the latest models' performance, such as Astra's, on the proposed benchmark.

Response. GPT-6 Astra achieves an average SR of 58.0% on ActiveBrain-Bench, compared with 44.0% for ActiveBrain-4B and 33.7% for GPT-5.5. The per-track results are shown in the table above and in Table 1 of the revised paper.

Question 3: Verification case study

In the most "cognitive" track, Verification, ActiveBrain-4B still underperforms GPT-5.5. It would be nice to have a case study for that to help understand the reason.

Response. ActiveBrain-4B improves Verification from 1.5% for Qwen3.5-4B to 15.5%, compared with 19.0% for GPT-5.5 and 45.0% for GPT-6 Astra. Although ActiveBrain-4B improves active cognition, successful Verification still depends on the model's basic perception and target-grounding capabilities.

In a task to open a book on the dining table, ActiveBrain-4B attempted to open a coffee machine. Although it later recognized that it was in the wrong location, it failed to reach the target. In contrast, GPT-5.5 initially failed to open the book because it was too far away, but moved closer and succeeded.

We include the case and its trajectory visualization in Section 5.6 (Fig. 7) of the revised paper.

Thanks for the additional experiments, which address all of my concerns. It is impressive that ActiveBrain-4B surpasses general-purpose 500B+ models in EmbodiedBench. I would like to maintain my current positive rating.

Thank you very much for your kind reply and positive assessment. We are grateful for your suggestions on generalization and comparisons with recent models, and we are glad to hear that our additional experiments addressed all of your concerns. If you feel that these additions strengthen our paper, we would be grateful if you could take them 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.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.