acceptodds
Under review as a conference paper at ICLR 2027

ERAct: Towards Reasoning over Memorization via Semantic Experience Rationale Informed Planning

Abstract

Vision-Language Models (VLMs) have shown strong potential for embodied planning, yet their performance can degrade substantially under distribution shifts. While accumulated experience provides a scalable source of external knowledge, raw trajectory-based experience provides concrete behavioral references but leaves the decision logic underlying actions implicit, making it difficult to reuse flexibly when current executions diverge from past trajectories. Based on this insight, we propose Experience Rationale Informed Planning (ERAct), a rationale-driven embodied planning framework that transforms past experiences into explicit reasoning knowledge and transfers decision-relevant reasoning cues to novel situations. ERAct consists of a Semantic Experience Rationale Retriever (SERR), which performs coarse-grained retrieval of reasoning-relevant experience instances, and a Rationale-Informed Fusion Network (RIFN), which conducts fine-grained dual-path relevance modeling to extract useful reasoning cues and adaptively inject them through a confidence-aware residual mechanism. Extensive experiments on the EB-ALFRED benchmark within EmbodiedBench demonstrate that ERAct improves both instruction understanding and environmental reasoning, consistently outperforming prior agents. Further analyses show that ERAct exhibits stronger reasoning generalization and more effective experience utilization than direct trajectory reuse.

Primary area:
applications to robotics, autonomy, planning
Keywords:
planning, embodied agent, VLMs

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NkNUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ERAct is a retrieval-augmented framework for embodied planning that encodes past experience as step-level rationales. Its Semantic Experience Rationale Retriever (SERR) selects relevant experiences based on the current instruction and preceding rationale. The Rationale-Informed Fusion Network (RIFN) integrates the retrieved rationale features with textual and visual inputs through separate alignment pathways and gated residual fusion. Training proceeds in three stages: learning from expert trajectories, instruction augmentation, and experience-augmented training. On EB-ALFRED, ERAct achieves 65.0% average success, compared with 61.7% for ExpRAG and 63.3% without RIFN. A comparison between rationale-based and action-based experience further suggests that how experience is represented influences its utility for planning.

Critical strengths
  1. The rationale-versus-action comparison provides useful evidence about experience representation. Table 3 shows that retrieved rationales improve the rationale-trained model from 60.7% to 65.0%, whereas action-only experience reduces its corresponding model’s performance from 62.7% to 60.7%. This comparison supports the idea that the content and organization of experience matter, rather than retrieval being inherently beneficial.

  2. The architecture offers a coherent mechanism for context-dependent experience reuse. SERR conditions retrieval on the instruction and preceding rationale, making experience selection sensitive to inferred task progress. RIFN then uses separate textual and visual pathways to integrate retrieved information with the current situation. This is a reasonable design for addressing mismatches between historical experience and ongoing execution.

  3. The evaluation explicitly separates the experience bank from evaluation tasks. Section 4.1 states that training data and stored experiences come exclusively from the ALFRED training split. The authors also re-evaluate RoboAgent without auxiliary action-success feedback and construct ExpRAG using the same backbone, training data, and experience bank. These choices improve the relevance of the comparisons to the intended evaluation setting.

  4. The reported results show positive aggregate gains and include several component analyses. ERAct achieves the highest reported average success in Table 1, with 65.0% compared with 61.7% for ExpRAG and 63.0% for RoboAgent. Table 2 additionally examines the dual-path design, selector, and complete fusion module. Although the gains require stronger uncertainty analysis, these experiments provide preliminary evidence that the proposed components are useful in combination.

Critical weaknesses
  1. The experiments do not isolate reasoning transfer from richer descriptive information. The rationales include observations, task progress, and action explanations, and are generated with the executed action supplied to the annotator (Appendix A.2). Their advantage over action-only experience could therefore arise from additional state information or action-related cues rather than transferable reasoning specifically. A comparison against state-and-progress descriptions without action justifications, using matched retrieval and input budgets, would better test the central claim.

  2. The improvements over controlled alternatives are modest and lack uncertainty estimates. ERAct improves over the same-data “w/o RIFN” variant by only 1.7 percentage points and over ExpRAG by 3.3 points. The manuscript does not report independent-run variation or confidence intervals. Moreover, removing the dual-path design improves Visual and Spatial success from 50% to 54% and from 46% to 52%, respectively. The average gain is encouraging, but its reliability and the claimed visual-grounding benefit remain insufficiently established.

  3. The analysis provides limited support for the proposed retrieval and confidence mechanisms. Figure 3 reports correlations of −0.058 and 0.061, both close to zero, despite being used to motivate rationale relevance as a more useful retrieval signal. Its vertical axis is labeled “Task Progress,” while the caption refers to success rate. Qualitative attention maps do not resolve this ambiguity or establish that the confidence gate suppresses unreliable experience. Controlled tests with irrelevant or misleading retrieval, together with measurements of gate responses and task performance, would provide more direct evidence.

  4. Important implementation and comparison details remain unclear. Equation 10 introduces G without defining its relationship to S, and the activation in Equation 9 is unspecified. Retrieval K, confidence top-k, temperature, the rationale-generation model, and the construction of Stage 3 trajectories are not clearly documented. The exact retrieval content and context limits for ExpRAG are also needed to assess whether the comparison isolates fusion. In addition, Table 4 reports a Stage 1 average of 44.3%, whereas the six listed category scores average to 36.0%; this discrepancy should be corrected or explained.

Questions for authors
  1. Can you compare full rationales with state-and-progress descriptions that omit the action justification, while holding retrieved instances, fusion architecture, and input budgets fixed? This would help distinguish reasoning content from richer descriptive information. How does SERR compare with instruction-only retrieval under the same planner?

  2. What exactly is held fixed in the ExpRAG and action-trajectory comparisons: retrieval keys, retrieved instance identities, rationale/action content, training targets, and context lengths? For the “w/o RIFN” variant, are retrieved experiences absent entirely or supplied through another interface?

  3. Please provide evaluation counts, run counts, decoding settings, and uncertainty for the main comparisons, particularly the 65.0% versus 63.3% result. Are the improvements consistent across independent training runs? Please also reconcile the dual-path visual-grounding explanation with the higher Visual and Spatial scores obtained without that component.

  4. Please define G in Equation 10 and its relationship to S, specify the activation in Equation 9 and the relevant hyperparameters, and explain how the gate is validated under irrelevant retrieval. Which model generates the rationales, how are the Stage 3 trajectories synthesized, and are retrieval matches from the same underlying source trajectory excluded during training?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 1)

Thank you for your constructive feedback and recognition. Below are our responses, which we hope will address your concerns.

Response to Weakness 1 & Question 1.1: Ablation of Experience Rationales

The experience rationales in ERAct contain two complementary types of reasoning content: State Description and Action Reasoning.

  • We would first like to clarify that the state descriptions contained in retrieved rationales do not provide additional factual information about the current test state. They describe the task progress and environmental state of past ALFRED experiences, whereas evaluation is conducted in different EB-ALFRED environments. Therefore, the utility of both components depends on whether their underlying reasoning patterns can be transferred and re-grounded to the current context.

  • Following the reviewer’s suggestion, we further add a controlled experiment to disentangle the different components of the rationale. The results show that state-oriented and action-oriented reasoning provide complementary transferable cues, and that the advantage of full rationales cannot be attributed simply to richer descriptive information.

Retrieved RationalesAvgBaseCommonComplexVisualSpatialLong
Null60.7747066444664
Only State-Description63.0 (+2.3)747276524460
Only Action-Reasoning63.7 (+3.0)727672465066
Full65.0 (+4.3)787274504670

The ablation of experience rationales has been added to Sec. A.4.1.

Response to Weakness 2.1 & Question 3.1: Independent-Run Variation

EB-ALFRED contains 300 evaluation tasks. We additionally conduct three independent training runs for both ERAct and the w/o RIFN variant and report the variation. During inference, we use deterministic greedy decoding with do_sample=False.

MethodSeed 1Seed 2Seed 3Mean ± SD
ERAct65.063.764.364.3 ± 0.7
w/o RIFN63.361.063.762.7 ± 1.5
Improvement+1.7+2.7+0.6+1.7

These results show that ERAct consistently outperforms the w/o RIFN variant across all three runs.

Response to Weakness 2.2 & Question 3.2: Visual-Grounding Explanation

We thank the reviewer for these valid points.

  • First, we would like to clarify that the Visual and Spatial subsets mainly evaluate whether the agent can plan actions based on textual descriptions of environmental attributes or spatial relations, such as “pick up a yellow object” and “put the plunger under the sink.” Under these settings, ERAct (w/o dual) can perform better.

  • In contrast, the visual-grounding capability discussed in our paper refers specifically to identifying whether a target object is present in the current visual observation. To quantitatively evaluate this capability, we conduct two additional analyses:

  1. We separately evaluate tasks in the EB-ALFRED test set where the agent must recognize that the target object is absent from the current observation and subsequently locate an alternative object_2 to recover from action failure. ERAct achieves a success rate of 50%, compared with 30% for ERAct (w/o dual).

  2. We further sample 220 images from FUNTHOR [2] to evaluate the model’s ability to recognize whether a target object is present in the image. ERAct achieves a success rate of 33.64%, whereas ERAct (w/o dual) achieves 21.82%.

These results support that separately modeling textual and visual features helps reduce entangled interference from cues that are semantically relevant but visually mismatched, thereby improving visual grounding.

[2] Fu Z, Zurbrügg R, Qu K, et al. Funfact: building probabilistic functional 3d scene graphs via factor-graph reasoning[J]. arXiv preprint arXiv:2604.03696, 2026.

AuthorsOpenReview ↗

Response to Reviewer 3 (Part 2)

Response to Weakness 3.1: Fig. 3 correlations close to zero.

We thank the reviewer for this comment. Figure 3 originally presents scatter plots of task progress against the retrieval scores, together with the corresponding Spearman's ρ\rho. Since a task is considered successful only when task progress reaches 1, we further partition the scores into six bins and compute the success rate within each bin. We then analyze the relationship between success rate and the corresponding retrieval scores. The Spearman's ρ\rho between success rate and trajectory similarity is -0.87, whereas that between success rate and rationale relevance is 0.348. These results suggest that rationale relevance provides a more useful retrieval signal. We have revised the description and visualization in Figure 3 accordingly.

Response to Weakness 3.2: do not establish that the confidence gate suppresses unreliable experience.

That is an excellent question. In ERAct, the selector filters mismatched reasoning cues, while the confidence mechanism adaptively controls the strength of experience injection. To validate these mechanisms, we construct a controlled degraded-retrieval setting by reducing the experience bank to one-third of its original size and replacing SERR with instruction-only retrieval. The results show that both fine-grained relevance selection and the confidence gate provide additional robustness by suppressing unreliable experience.

MethodAvgBaseCommonComplexVisualSpatialLong
ERAct63.7747668484868
w/o Confidence63.3786472525064
w/o Selector61.7786074524264

These controlled tests have been added to Sec. A.6.

Response to Weakness 4 & Question 4: Important implementation and comparison details remain unclear.

Thank you for your careful attention to detail, specifically:

  • We split Equation 9 to explicitly define the relationship between GG and SS (see Sec. 3.3.2), and clarify that the activation function used in Equation 9 is the sigmoid function.

  • The retrieval parameter KK is set to 3, and the confidence Top-kk ratio is set to 20%. During inference, we use deterministic decoding with do_sample=False. The rationale-generation model is Qwen2.5-VL-72B-Instruct. Details of the Stage 3 trajectory construction are provided in Sec. A.1.2.

  • ExpRAG and ERAct use the same SERR module and experience bank, with identical retrieval settings. The inputs are not truncated for either method, and the maximum output length is set to 1024 tokens for both.

  • You are correct; Table 4 has a typo in the manuscript, and we have corrected it in the revised paper.

  • Under irrelevant retrieval, the confidence gate adaptively suppresses the strength of experience injection, thereby improving the robustness of RIFN (see Sec. A.6).

  • Retrieval matches originating from the same source trajectory are not excluded during training. Since our experience bank is organized at the step level, with each instance represented as mt=(lτ,kt,rt,at)m_t=(l_\tau,k_t,r_t,a_t), different steps from the same trajectory correspond to different task progress, reasoning contexts, and decisions, and therefore provide distinct experience information.

Response to Question 1.2: Ablation of SERR

We thank the reviewer for this insightful comment. We have added the ablation studies of SERR in the appendix Sec. A.5. Please see our response for Reviewer 2 Weakness 2.1. The results show that SERR achieves higher average performance than instruction-only retrieval across different experience bank sizes, supporting its effectiveness in retrieving experiences with more relevant underlying reasoning.

Response to Question 2: Details of Variant

We thank the reviewer for this question, which allows us to further clarify the comparisons. Specifically:

  • ExpRAG is kept consistent with ERAct in terms of the retrieval setup.

  • In contrast, the action-trajectory variant removes rationales from the retrieval keys, retrieved instance content: the retrieval key is defined as $ k_i = \pi_e([l_{\tau_i};a_{i-1}]) $ and the retrieved instance is $ m_t = (l_\tau,k_t,a_t) $.

  • For the w/o RIFN variant, retrieved experiences are entirely removed, and the model internalizes the knowledge into its parameters during training.

Thank you for your time and valuable feedback! As the author–reviewer discussion period will end in about one days, we kindly ask you to review our rebuttal and let us know whether it addresses your concerns.

Reviewer_1EFwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ERAct reuses experience for vision-language embodied planning under task/environment shifts. It converts successful trajectories into step-level rationales, retrieves them using the instruction plus prior rationale, and fuses them via visual/textual paths with a confidence-gated residual. Planner: Qwen2.5-VL-7B, three-stage training. On EB-ALFRED: 65.0% avg success vs. 61.7% for ExpRAG and 63.3% without RIFN. Ablations suggest rationale-rich experience helps, while action-only experience can hurt. The memory representation looks useful, but the stronger “transferable reasoning” claim needs more evidence.

Critical strengths
  1. Main comparisons are reasonably controlled. ExpRAG shares backbone, training data, and experience bank with ERAct, so 61.7% → 65.0% is meaningful. Bank excludes eval tasks; RoboAgent is re-evaluated without privileged success feedback.
  2. Memory ablation is useful. In Table 3, experience helps the rationale-trained model (60.7% → 65.0%) but hurts the action-trained one (62.7% → 60.7%). So representation matters, and behavioral examples can backfire.
Critical weaknesses
  1. Rationales mix observations, progress, goal relations, and action explanations. Table 3 changes several info sources at once, and the shared retrieval setup is unclear. Fig. 3 correlations are near zero (−0.058, 0.061), so this supports useful contextual memory more than “reasoning over memorization.”
  2. ExpeL and RAEA already cover related retrieval/fusion. ERAct adds rationale records plus modality-specific gated fusion, but there’s no controlled ablation of SERR’s retrieval key, and the gain over same-data “w/o RIFN” is only 1.7 points. No run variance/uncertainty is reported.
Questions for authors
  1. With retrieved instances and supervision fixed, how does full rationale memory compare with state/progress descriptions without action explanations? Does Table 3 change only retrieved content, or also keys/queries?
  2. How are Fig. 3’s similarity scores and outcome defined? Axis says task progress, caption says success rate. Does rationale-conditioned retrieval beat instruction-based retrieval when memory and fusion are fixed?
  3. What uncertainty supports the gains over ExpRAG and “w/o RIFN”, independent runs or paired outcomes?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 2

Thank you for your constructive feedback and recognition. Below are our responses, which we hope will address your concerns.

Response to Weakness 1.1 & Question 1.1: Ablation of Experience Rationales

Thanks for your advice, we have added the ablation studies of rationales in the appendix Sec. A.4.1. Please see our response for Reviewer 3 Weakness 1. The results show that ERAct benefits from re-contextualizing transferable state-oriented and action-oriented reasoning cues underlying past experiences rather than simply from memorizing more contextual descriptions.

Response to Weakness 1.2 & Question 1.2: Retrieval setup of table 3 is unclear.

We thank the reviewer for this comment, which allows us to clarify our experimental design.

The action-trajectory variant in Table 3 removes rationales from the retrieval keys, retrieved instance content. Specifically, the retrieval key(queries) is defined as $ k_i = \pi_e([l_{\tau_i};a_{i-1}]) $ and the retrieved instance is $ m_t = (l_\tau,k_t,a_t) $.

Response to Weakness 1.3 & Question 2.1: Fig. 3 correlations are near zero.

We thank the reviewer for this comment. We use a semantic encoder to compute cosine similarity as the retrieval score. Task progress is provided by the simulator, and a task is considered successful when its task progress reaches 1. We further partition the scores into six bins and compute the success rate within each bin. We have revised the description and visualization in Figure 3. The Spearman's ρ\rho between success rate and trajectory similarity is -0.87, whereas that between success rate and rationale relevance is 0.348. These results support the use of rationale relevance as a more useful retrieval signal for identifying experiences that contain relevant reasoning knowledge.

Response to Weakness 2.1 & Question 2.2: Ablation of SERR

We thank the reviewer for this question. We keep the semantic encoder, Top-KK, and planner fixed, and conduct a controlled comparison against instruction-only retrieval across different experience bank sizes. The results show that incorporating decision-level reasoning enables SERR to retrieve experiences that are more relevant in terms of underlying decision logic.

Bank SizesRetrievalAvgBaseCommonComplexVisualSpatialLong
FullInstruction-only64.3727276465268
FullSERR65.0 (+0.7)787274504670
1/3Instruction-only63.7747668484868
1/3SERR64.7 (+1.0)767674485064

The ablation of SERR has been added to Sec. A.5.

Response to Weakness 2.2 & Question 3: Run variance/Uncertainty

Please see our response to Reviewer 3 Weakness 2.1.

Thank you for your time and valuable feedback! As the author–reviewer discussion period will end in about one days, we kindly ask you to review our rebuttal and let us know whether it addresses your concerns.

Thanks for the thorough response. The state-description control, the three training runs, and the confirmation that ExpRAG shares the same retrieval setup substantially address my main experimental concerns and put the 3.3-point gain over ExpRAG on firmer footing.

My remaining concern is the overall contribution. The new ablations support combining state descriptions and action reasoning. However, the specific benefits of SERR and the confidence gate remain small: 0.7–1.0 points over instruction-only retrieval and 0.4 points in the degraded-retrieval setting, respectively. Although the matched ExpRAG comparison supports the value of learned fusion, I still view the approach as an incremental extension of existing experience-retrieval and fusion methods around a useful memory representation.

I also remain cautious about the stronger "transferable reasoning" claim. The component ablations confirm the value of rationale-rich memory, but do not clearly distinguish that interpretation from more effective contextual experience reuse. The correlation over six score bins provides limited additional evidence for this distinction.

The response improves my confidence in the experimental results, but I remain unconvinced that the overall contribution is substantial enough for acceptance. My rating stays at 2.

Reviewer_BTZXOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ERAct, which turns successful trajectories into step-level rationales and learns to retrieve relevant rationales during inference. A learned fusion module then selects useful information from the retrieved rationales, and integrates it with the current visual and textual inputs.

Critical strengths

The method is intuitive and practically motivated.

Critical weaknesses
  1. The novelty is limited because prior work, for example ERA (https://arxiv.org/abs/2510.12693), already generates action-conditioned step-level reasoning from embodied trajectories; ERAct mainly adds retrieval and a specialized fusion module, which looks incremental.
  2. The evaluation omits strong EB-ALFRED baselines such as ERA, ORBIT (https://aclanthology.org/2026.acl-long.1822), and ELITE (https://arxiv.org/abs/2603.24018), which are some SOTAs in EB-Alfred according to my understanding.
Questions for authors
  1. Could the authors evaluate on task categories or environments that are completely held out during training, as done in ERA? The rationale learning process looks much like on-policy self-distillation in LLM, which is well known to overfit, and the current train-val split doesn't test if it learns valid reasoning pattern that can be generalized in the same environment but different task categories.
  2. Why wouldn't the summarized rationales be post-hoc, for example the action is 'pick up a knife' and the rationale is 'I want to cut something', which isn't reasonable as a real hint to yield an action. How can the authors validate that this doesn't happen, and is there a way to prevent this?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 1 (Part 1)

We sincerely appreciate the reviewer’s constructive feedback and recognition. We hope the following responses address your concerns effectively. For clarity, all major revisions and newly added results in the updated PDF have been highlighted in blue.

Response to Weakness 1: The novelty of action-conditioned step-level reasoning from embodied trajectories.

We respectfully clarify that while our action-conditioned step-level rationale generation is an established technique, its mechanism for utilizing and objective are fundamentally different from prior art.

1. Mechanism for utilizing (vs. ERA)

ERA generates step-level reasoning from raw trajectories as Trajectory-Augmented Priors for SFT, using them as fine-grained supervision. In contrast, ERAct treats past rationales as explicit external reasoning memory that remains available at inference time.

2. Objective (vs. ERA)

ERA internalizes reasoning into the model parameters, whereas ERAct selectively retrieves and integrates transferable reasoning cues into the current decision context.

The central contribution of ERAct is therefore not rationale generation itself, but our finding that rationale-form experience provides more effective reasoning transfer than action-form experience. This motivates SERR and RIFN as central components of the framework rather than auxiliary additions.

Response to Weakness 2: The evaluation omits strong EB-ALFRED baselines.

The key insight of ERAct is to leverage retrieved rationale-based experiences to extract transferable reasoning cues for enhancing reasoning and planning in the current context. To directly demonstrate this capability under cross-dataset transfer, both our training data and experience bank are constructed exclusively from the ALFRED training split, with no access to any EB-ALFRED evaluation tasks. In contrast, ERA and ORBIT use EB-ALFRED subsets during training. These differences motivated us to initially focus on baselines under more comparable data constraints.

Nevertheless, we agree that adding more baseline comparisons could provide a more accurate evaluation. We therefore additionally re-evaluate the publicly released ERA and ORBIT models available through the official repositories linked by the papers under the same closed-loop protocol as ERAct, where each action decision is made without access to auxiliary privileged feedback. For a comparison aligned with our closed-loop protocol, we also include the unsupervised online results of the training-free ELITE method. For consistency, we report Avg-5 over Base, Common, Complex, Visual, and Spatial, following these prior works, as well as Avg-6 that additionally includes Long-Horizon. These results, which have been added to Table 1, show that ERAct retains a clear performance advantage.

MethodAvg-6Avg-5BaseCommonComplexVisualSpatialLong
ERAct65.064.0787274504670
ERA47.756.868527450402
ORBIT34.740.052424832268
ELITE--60----62
AuthorsOpenReview ↗

Response to Reviewer 1 (Part 2)

Response to Question 1: Unseen Task Generalization

Thanks for your valuable advice. We would like to clarify that ERAct does not perform on-policy self-distillation. ERA directly uses EB-ALFRED subsets during training for online reinforcement learning, whereas ERAct is trained exclusively on the offline ALFRED training split, with no access to any EB-ALFRED evaluation tasks.

Nevertheless, we agree that evaluating generalization to unseen task categories is meaningful. We therefore categorize tasks according to their underlying action trajectories and conduct a category-held-out analysis on the Cool and Heat task types (Sec. 4.5). The performance on the Cool and Heat tasks in EB-ALFRED is shown below:

Training DataTest-Time Experience BankSR
IncludedIncluded80 (+7)
IncludedExcluded77 (+4)
IncludedNULL73
ExcludedIncluded68 (+11)
ExcludedExcluded64 (+7)
ExcludedNULL57

The results demonstrate that reasoning cues retrieved exclusively from other task categories can still benefit the held-out categories, and that this capability can generalize to unseen task types. This generalization analysis has been added to Sec. 4.5.

Response to Question 2: Rationales would be post-hoc.

We follow WAP [1] and adopt context-conditioned retrospective rationale generation rather than unconstrained action explanation. Specifically, rationale generation is conditioned on the task instruction, current observation, preceding reasoning history, and current action, and the prompt explicitly requires the generated rationale to be grounded in its relevance to the task goal, current task progress, and environmental state. Moreover, rationales are generated only from successful expert trajectories, providing reliable execution contexts for retrospective reasoning.

To mitigate the post-hoc issue, we additionally introduce a rationale verification mechanism. LLMs assess whether the generated rationale is supported by the available pre-action context and correctly reflects reasoning information. Rationales that fail to reach the verification threshold are regenerated. We have added this verification procedure in Sec. A.2.

[1] Shi J, Fei Z, Wang S, et al. World-aware planning narratives enhance large vision-language model planner[J]. Advances in Neural Information Processing Systems, 2026, 38: 50557-50583.

Thank you for your time and valuable feedback! As the author–reviewer discussion period will end in about one days, we kindly ask you to review our rebuttal and let us know whether it addresses your concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary for Submission 57671 (Part I)

Dear Area Chair,

To facilitate your assessment, we would like to provide a brief summary of the review and rebuttal process for our paper, “ERAct: Towards Reasoning over Memorization via Semantic Experience Rationale Informed Planning.”

During the discussion period, we conducted an extensive suite of additional experiments, clarified potential misunderstandings, and substantially revised our manuscript to address all concerns raised by the reviewers.

We are grateful that all three reviewers acknowledged the value of our responses and clarifications. Specifically, Reviewer BTZX stated that they “do feel more positive about the paper (compared with before reading the authors' response),” while Reviewer 1EFw noted that “The response improves my confidence in the experimental results.” Reviewer NkNU also acknowledged our additional evaluation of visual grounding and increased their rating.

However, Reviewers BTZX and 1EFw subsequently raised further concerns. Unfortunately, the discussion period ended before we had the opportunity to engage in further exchanges. Nevertheless, we have submitted additional responses addressing these remaining concerns to ensure that our clarifications and supporting evidence are available for your consideration.

AuthorsOpenReview ↗

Rebuttal Summary for Submission 57671 (Part II)

Our rebuttal and manuscript revisions primarily focused on the following experiments:

  • Strong EB-ALFRED Baselines in Sec. 4.2: In response to [Reviewer BTZX W2].

  • Unseen Task Generalization in Sec. 4.5: In response to [Reviewer BTZX Q1, Reviewer 1EFw C2].

  • Ablation of Experience Rationales in A.4.1: In response to [Reviewer 1EFw W1.1, NkNU W1].

  • Fig. 3 correlations in Sec.3.2: In response to [Reviewer 1EFw W1.3, NkNU W3.1].

  • Ablation of SERR in A.5: In response to [Reviewer 1EFw W2.1, NkNU Q1.2].

  • Transferable Reasoning Claim in A.4.2: In response to [Reviewer 1EFw C2].

  • Visual-Grounding Explanation in A.7: In response to [Reviewer NkNU W2.2].

  • Ablation of RIFN in A.6: In response to [Reviewer NkNU W3.2].

We would like to once again express our sincere gratitude to all reviewers for their time and constructive guidance. Your feedback has been invaluable in helping us significantly improve the quality and clarity of our work. The revised manuscript, which incorporates all the changes and additional experiments mentioned above, has been uploaded.

Thank you for your time and dedication to the review process.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.