acceptodds
Under review as a conference paper at ICLR 2027

Remember What Matters: On-Demand Memory for Long-Horizon Robotic Manipulation

Abstract

Memory is essential for long-horizon robotic manipulation: robots must remember task progress and retrieve historical visual details. The key lies in accurately associating the current task with relevant memory. We propose OnCue, a hierarchical memory framework that supports task-guided recall. Language descriptions record task progress, while subtask indices address the corresponding historical visual memories; scene-level and object-level memory preserve scene layout, object appearance, and location. The high-level planner generates the current subtask, selects relevant completed stages for visual-memory retrieval, and predicts bounding boxes around task-relevant objects in the current observation for future memory storage; the low-level executor combines the current image, subtask, and retrieved visual memory to execute manipulation. We evaluate OnCue on RMBench and RoboTwin-MeM, achieving average success rates of 84.0% and 95.8%, respectively, outperforming the strongest baseline by 16.2 and 19.8 percentage points. We also evaluate on three tasks across two real robot platforms, achieving an average success rate of 88.3%. We further conduct ablation studies and demonstrate the importance of memory design for long-horizon robotic manipulation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-horizon robotic manipulation, Vision-Language-Action model, Language memory, Visual memory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WqW3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OnCue handles the case where what the policy needs has already left the frame. Memory is an ordered list of finished subtasks in language, plus a bank of visual evidence keyed by subtask index: the first frame of each subtask, and crops and boxes of the objects the planner flagged. A LoRA-tuned Qwen3-VL planner reads the current image, the goal and that list, then returns the next subtask, which past indices to pull and what to keep. Retrieved features enter a DiT policy through a cross-attention branch that is skipped when nothing is retrieved, and a small classifier decides when a subtask is done. Reported numbers: 84.0% on RMBench, 95.8% on RoboTwin-MeM, 16.2 and 19.8 points over the strongest baseline, 88.3% on three real tasks across two arms. The claim underneath is that indexing evidence by subtask lets the planner pull only what the current step needs.

Critical strengths
  1. The two memories fail on different tasks, which is the right evidence for the complementarity claim. Pick Objects in Order drops to 0 without language memory, Observe and Pick Up to 3 without visual memory.
  2. Coverage is decent for this area: two simulators at 100 trials, real hardware on two platforms, and a failure taxonomy most papers skip.
Critical weaknesses
  1. Table 1 confounds too much. OnCue brings an 8B LoRA-tuned planner, SAM3 and DINOv3, trained per task on supervision nobody else gets: re-segmented subtask boundaries, retrieval indices, object boxes, none of which is explained. The 16.2 points could be the routing idea or could be the annotation. Same planner, same labels, naive memory would settle it.
  2. Selectivity is the thesis and is never ablated. Both variants in 4.3.1 simply delete a memory type. Missing are R_k set to every completed subtask, a fixed recent window, and a random set of matched size. The retrieval labels already exist, so precision and recall on the predicted indices would cost nothing and would show whether the planner routes or has memorised a per-task pattern.
  3. No timing anywhere, for an 8B planner plus SAM3 plus DINOv3 plus a DiT at every boundary, on platforms recorded at 15 and 30 Hz. RoboTwin-MeM episodes run past 1500 steps and every allowed index goes into the prompt, yet there is no curve for success rate against memory size.
  4. Everything is seen at 224 by 224, and the regions worth keeping are chosen by the 8B VLM and then tidied by SAM3. D.2 concedes that Pick Double X Times suffers because the real digits are much smaller than in simulation, which is a sampling problem in the current frame rather than a storage problem. Worth at least discussing spatially dynamic computation here, for instance AdaptiveNN (Wang et al., Emulating human-like adaptive vision for efficient and flexible machine visual perception, Nature Machine Intelligence 2025), which treats perception as coarse-to-fine fixations on task-relevant regions and stops once the evidence suffices, at up to 28x lower inference cost. A fixation policy is another answer to which regions matter, and could supply higher-resolution crops without a per-task 8B model in the loop. Not a condition of acceptance, but active vision is absent from 2.2.
  5. Single seed, no variance, ablations on two or three tasks; 96 against 94 is not a result. The w/o Classifier row at 0 everywhere is a broken control rather than an informative one. Also, Figures 3, 5, 6, 9 and 10 are tables that the text calls Tables, colliding with the appendix numbering; 4.4 ends mid-sentence; Mem-0 and RMBench share a citation.
Questions for authors

See Weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer WqW3 [1/2]

We sincerely thank Reviewer WqW3 for the constructive feedback and detailed suggestions on experimental design, ablation studies, computational efficiency, and evaluation reliability. Following these suggestions, we have strengthened our ablation studies, clarified the experimental settings, and provided additional analyses. Below, we present detailed responses and additional experimental evidence. All changes in the revised paper have been highlighted in red for easy reference.

Q1: Effectiveness of Selective Historical Memory Retrieval

We appreciate the reviewer's constructive feedback on the effectiveness of selective historical memory retrieval. We agree that further evaluation is important for understanding its contribution to OnCue.

To this end, we have conducted additional ablation experiments comparing OnCue with four alternative retrieval strategies: (1) Oracle Source, which uses ground-truth historical source indices, object bounding boxes, and corresponding object crops; (2) Recent-k, which retrieves the most recent k historical memories; (3) Random-k, which randomly selects k historical memories; and (4) All History, which provides all available historical memories without selective retrieval. For these variants, the corresponding Planner and VLA are retrained under their respective retrieval configurations. We set k = 2 for Swap T and k = 3 for Pick Objects in Order.

Retrieval StrategySwap TPick Objects in Order
Oracle Source84%96%
OnCue83%96%
Recent-k10%60%
Random-k40%23%
All History71%45%

As shown in the table, OnCue achieves performance comparable to Oracle Source, whereas Recent-k, Random-k, and All History exhibit substantial performance degradation. These results suggest that the Planner can effectively identify task-relevant historical sources and that selective retrieval provides greater benefits than simply using recent, random, or all available historical memories.

Furthermore, our ablation studies on memory representation, memory granularity, completion-driven updates, and historical retrieval provide complementary evidence supporting the effectiveness of OnCue's key components and design choices.

These additional experiments and analyses have been incorporated into the revised paper, with detailed implementation settings provided in the Appendix.

Q2: Inference Efficiency and Execution Protocol

We appreciate the reviewer's concerns regarding inference efficiency and memory scalability. We provide a component-wise inference time comparison between OnCue and a variant without Visual Memory.

ModuleOnCuew/o Visual Memory
Memory Encoding47.40 ms/call0.47 ms/call
DiT Action Chunk92.97 ms/call68.41 ms/call
Planner (Subtask Transition)2.08 s/call1.79 s/call

As shown in the table, Visual Memory introduces additional encoding and action-prediction overhead. Each DiT inference generates 30 actions, while the Planner is invoked only at subtask transitions rather than at every action prediction. OnCue achieves an execution frequency of 1.93 Hz, compared with 2.38 Hz without Visual Memory.

Regarding memory scalability, OnCue does not create a new memory index at every execution step. Instead, memory indices and updates operate at the subtask level. Therefore, the number of memory entries depends on task decomposition rather than the total execution steps. For example, the number of subtask indices in Press Button does not exceed 20.

We have clarified the inference efficiency, execution protocol, and memory indexing mechanism in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer WqW3 [2/2]

Q3:Clarification of Real-World Perception Challenges and Adaptive Vision

We appreciate the reviewer's insightful suggestion regarding adaptive visual perception. We agree that fine-grained visual recognition and historical memory retrieval address complementary challenges in robotic manipulation.

In Pick Double X Times, the numbered blocks used in our real-world experiments are considerably smaller than those in simulation, making digit recognition more challenging. This scale difference makes fine-grained digit recognition more challenging, independently of historical memory retrieval.

We also acknowledge the relevance of adaptive visual perception methods such as AdaptiveNN, which employ coarse-to-fine processing to focus on task-relevant regions. Such approaches are complementary to OnCue and offer a promising direction for improving fine-grained visual evidence acquisition.

We have clarified this perception challenge and incorporated a discussion of AdaptiveNN in the revised Real-World Evaluation section.

Q4: Evaluation Seeds, Classifier Controls, and Presentation Corrections

We thank the reviewer for raising concerns regarding evaluation reliability and the design of our ablation studies. First, we would like to clarify that each simulation task is evaluated across 100 distinct random seeds, following the benchmark-specific evaluation protocols of RMBench and RoboTwin-MeM. We have explicitly emphasized this setting in the revised paper to avoid any misunderstanding.

We also acknowledge that the original completion classifier ablation was not sufficiently informative, as the w/o Classifier variant completely disabled memory updates. To address this limitation, we have replaced it with two more meaningful fixed-interval update strategies: Subtask-specific Median, which uses the median execution length of the closest matching subtask type in the training set as the update interval, and Global Average, which uses a fixed interval computed from the average ratio of total execution steps to the number of subtasks in the training set. Both strategies retain the same Planner and VLA as OnCue, differing only in the memory update triggers. We have updated the corresponding ablation study in the paper and provided further implementation details in the Appendix.

Update StrategyPress ButtonRearrange BlocksCover Blocks
Oracle Completion96%99%99%
OnCue87%93%96%
Subtask-specific Median81%90%95%
Global Average45%70%84%

As shown in the table, the performance gap between Oracle Completion and OnCue highlights the potential for further improvement in completion detection, while OnCue consistently outperforms both fixed-interval update strategies across all three tasks. Notably, Global Average performs substantially worse on Press Button and Rearrange Blocks, where subtask execution lengths vary considerably, making uniform update intervals more prone to premature transitions. These results suggest that completion-driven memory updates provide more appropriate update timing than predefined fixed intervals, thereby improving task performance.

Finally, we thank the reviewer for identifying several presentation and citation issues in our paper. We have corrected the reported formatting inconsistencies, table references, and incomplete sentence in the revised version. We would also like to clarify that Mem-0 was introduced in the RMBench paper, and therefore sharing the same citation is intentional and appropriate. We have clarified this relationship in the revised paper to avoid confusion.

I thank the authors for their detailed responses! After reading them in detail, I decided to maintain my positive score.

Thank you very much for your thoughtful follow-up and constructive feedback throughout the review process. We sincerely appreciate your time, careful consideration, and positive assessment of our work.

Reviewer_SH2QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the challenge of how to add memory to robot decision-making. To address this challenge, the paper proposes a hierarchical framework that includes a high-level planner and a low-level executor. The authors evaluate the approach on two benchmarks and three real-world tasks across two embodiments.

Critical strengths

1: Memory is an important problem in the robot learning literature, but it is underexplored.

2: The designed approach makes sense in separating language memory and visual memory.

3: The proposed approach is evaluated on two benchmarks and outperforms the baselines.

4: The paper also contains real-world tasks and achieves an average success rate of 88.3%.

5: I also like the ablations included in the paper.

Critical weaknesses

1: The introduction definitely can be improved. The current intro starts with long-horizon manipulation problems and then motivates the memory-augmented methods. However, memory is a standalone challenge and can appear even for short-horizon manipulation tasks. So I would suggest directly motivating why we need memory for robots.

2: The current figures have space for further improvements. For Figure 2, I would highlight the hierarchical memory systems better. Also, the figures are overall text-intensive but provide limited additional information that complements the paper.

3: Robot memory has been studied in the literature for several years. Some literature the authors should cite and consider adding as baselines: https://arxiv.org/abs/2602.15010, https://arxiv.org/pdf/2309.15278.

4: The tasks in this paper, especially in the real world, are too simple. I would encourage the authors to try more challenging manipulation tasks that require memory (e.g., how to retrieve a book that I read yesterday from a constrained bookshelf).

5: The paper should also discuss limitations of the proposed approaches.

Questions for authors

1: When you add memory to the VLA, would this affect the efficiency of the VLA during inference? What’s the policy inference frequency of your approach?

2: Would your approach be sensitive to the prompts?

3: Would your approach be robust to noisy perception data (e.g., segmentations)?

4: What’s the maximum size of your memory bank?

3 comments
AuthorsOpenReview ↗

Response to Reviewer SH2Q [1/2]

We sincerely thank Reviewer SH2Q for the constructive suggestions concerning the motivation for robot memory, figure clarity, experimental robustness, and the limitations of OnCue. Below, we present detailed responses and additional experimental evidence. All changes in the revised paper have been highlighted in red for easy reference.

Q1: Memory Motivation and Related Literature

We thank the reviewer for this valuable suggestion. We agree that memory is a fundamental requirement arising from partial observability and historical dependence, rather than merely from long task duration.

To address this concern, we have revised the Introduction and Figure 1 to better emphasize the necessity of memory in robotic manipulation, while also incorporating the suggested memory-related literature.

Q2: Improving Figure Clarity and Presentation

We agree that the figures should convey the key ideas more intuitively while avoiding excessive textual descriptions.

So, we have refined the details of the overview figures, reduced redundant text annotations, and improved the visual presentation to make the overall framework more intuitive and easier to understand.

Q3: Method Limitations

We thank the reviewer for highlighting this important aspect. We agree that further discussion of the limitations of our approach is important. Accordingly, we have added a dedicated Limitations section to the revised paper, primarily discussing: (1) the computational and storage overhead introduced by the Planner and memory modules; (2) the dependence on accurate historical retrieval and visual perception; and (3) the remaining challenges in low-level manipulation accuracy and generalization to more complex environments.

Q4: Policy Inference Efficiency

We appreciate the reviewer's question regarding inference efficiency. To evaluate the computational overhead of Visual Memory, we compare OnCue with a variant without Visual Memory on Swap T.

ModuleOnCuew/o Visual Memory
Memory Encoding47.40 ms/call0.47 ms/call
DiT Action Chunk92.97 ms/call68.41 ms/call
Planner (Subtask Transition)2.08 s/call1.79 s/call

The results show that Visual Memory introduces additional computational overhead. However, each DiT inference generates 30 actions, and the Planner is invoked only at subtask transitions rather than at every action prediction. OnCue achieves an execution frequency of 1.93 Hz, compared with 2.38 Hz without Visual Memory.

We have included the corresponding efficiency analysis in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer SH2Q [2/2]

Q5: Sensitivity to Planner Prompts

We thank the reviewer for raising the important question of prompt sensitivity. We agree that evaluating the robustness of the Planner to different prompt formulations is valuable.

To investigate this, we keep the trained Planner and VLA weights fixed and evaluate three inference-time prompt settings: (1) Original, using the original training prompt; (2) Reduced, removing the first two introductory paragraphs; and (3) Simplified, removing detailed field descriptions, including those for related_task_indices, while preserving the essential instructions and output schema.

Prompt SettingSwap TPick Objects in OrderPress Button
Original83%96%87%
Reduced83%94%87%
Simplified80%90%84%

The results indicate that the LoRA-tuned Planner maintains a degree of robustness to prompt variations, while further simplification results in a modest performance decline. This may be related to fine-tuning on a fixed prompt template, which encourages adaptation to its instruction structure. Therefore, we recommend maintaining consistent prompt formats between training and inference, while incorporating diverse templates during fine-tuning could improve adaptability to broader prompt variations.

We have included these additional experiments and analyses in the revised Appendix.

Q6: Robustness to Noisy Perception

We thank the reviewer for highlighting the importance of robustness to perception inaccuracies. We agree that the quality of object localization and segmentation directly affects the reliability of historical visual memory.

To mitigate this issue, OnCue employs SAM3 to refine the coarse bounding boxes predicted by the Planner. We have evaluated the effectiveness of this design through an ablation study comparing SAM3 Refinement with Raw Crops, as detailed in Appendix E.2.

MethodObserve and Pick UpSwap T
SAM3 Refinement74%83%
Raw Crops63%71%

As shown in the table, SAM3 Refinement improves success rates by 11 and 12 percentage points on Observe and Pick Up and Swap T, respectively. These results demonstrate that more accurate object boundaries and reduced background interference improve the quality of historical visual evidence and subsequent task performance.

Q7: Memory Bank Capacity

We thank the reviewer for raising this important question regarding Memory Bank capacity. We have clarified the corresponding implementation details in the revised Appendix.

Specifically, OnCue employs a dynamically expanding Memory Bank that accumulates historical visual evidence as subtasks are completed, without a predefined limit on the number of stored entries. To avoid introducing excessive irrelevant historical information, the Executor typically uses no more than three historical memory entries per action prediction, selected through task-dependent selective retrieval.

Dear Reviewer SH2Q,

We sincerely thank you again for your careful evaluation and constructive feedback. We greatly appreciate the opportunity for further discussion.

We believe our responses have substantially addressed your main concerns. In particular, we revised the Introduction and Figure 1 to better motivate the need for robot memory, incorporated the suggested literature, and added a discussion of limitations. We also conducted additional experiments on inference efficiency, prompt sensitivity, and segmentation refinement, and clarified the memory bank capacity and selective retrieval mechanism.

We would be grateful for any further feedback or questions you may have. We hope that our additional experiments and clarifications can provide useful context for your final re-evaluation of our work.

Reviewer_X8NFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes OnCue, a hybrid memory-based planner and executor for addressing manipulation tasks that require memory. Tasks come from prior work like RMBench and RoboTwin-MeM that have published baselines against them. The high-level planner and language memory play a global role, figuring out subtask stage and extracting relevant older states. Then the visual memory system extracts relevant scene and object-level memories from marked segments to create an observable state for the next action. The system also has a completion classifier which detects when a subtask has finished and triggers memory update. OnCue exhibits marked improvement upon chosen baselines and ablations show that language and visual memory as well as the scene-level and object memory play complementary roles in solving the tasks.

Note: For RMBench other contemporary baselines exist which perform-at-par or outperform proposed system (MemoryWAM, BridgeVLA++), needing citations.

Critical strengths
  • Ablations showing contributions of the design choices in the system's performance
  • Sound global and local as well as language and visual memory retrieval system
  • Analysis of failure modes
  • Good performance over baselines
Critical weaknesses
  • Baseline numbers are really low for RoboTwin-MeM benchmark begging the question if these are the right baselines. Also majority of baselines are borrowed over from EventVLA work.
  • No baseline showing oracle-memory with planner and executor, i.e. how much of the error is from retrieving the right segments versus choosing right snippets from it
  • Unclear ablation setups
Questions for authors
  1. Please add the requested baseline.
  2. I am unclear on how you implemented the high-level planner without a language memory. How were relevant segments selected or subtask states reasoner over without this indexing?
  3. Your ablation with classifier seems to be (a) classifier as is and (b) no memory update ever. Is this the right read? If yes, I don't think this is a good setup. There needs to be a midway where memory updates occur on a fixed schedule and now the problem becomes if you don't have a good classifier does memory overwhelm your retrieval system? Much more interesting question to study than having no memory at all.
4 comments
AuthorsOpenReview ↗

Response to Reviewer X8NF [1/2]

We sincerely thank Reviewer X8NF for the constructive feedback and detailed suggestions. Following these suggestions, we have clarified the positioning of OnCue relative to concurrent memory-augmented approaches, strengthened the evaluation of historical memory retrieval, improved the ablation designs, and provided clearer experimental details. Below, we present detailed responses and additional experimental evidence. All changes in the revised paper have been highlighted in red for easy reference.

Q1: Concurrent Memory-Based Methods

We appreciate the reviewer's suggestion to consider relevant concurrent works, including MemoryWAM and BridgeVLA++. As stated in the ICLR 2027 Reviewer Guidelines, “We consider papers contemporaneous if they are published within the last two months.” Following this guidance, we have incorporated these works into the revised Related Work section and further discussed their methodological differences from OnCue:

Concurrent works further explore memory-augmented robotic manipulation. MemoryWAM combines recent frames, anchor frames, and compressed gist tokens to efficiently preserve long-term historical information, while BridgeVLA++ integrates temporal and spatial memory to capture interaction history and spatial context. While these approaches emphasize effective historical representation and utilization, OnCue focuses on identifying and retrieving task-relevant evidence according to the current task progress.

Regarding the RoboTwin-MeM baselines, we have also clarified the sources of the reported results in the revised paper.

Q2: Oracle Memory and Historical Source Retrieval

We thank the reviewer for this valuable suggestion. We agree that evaluating the Planner's ability to identify relevant historical sources and the effectiveness of selective retrieval is essential to understanding the contributions of OnCue.

To further evaluate the contribution of selective historical memory retrieval, we have conducted additional ablation experiments comparing OnCue with four alternative retrieval strategies: (1) Oracle Source, which uses ground-truth historical source indices, object bounding boxes, and corresponding object crops instead of Planner predictions; (2) Recent-k, which retrieves the most recent k historical memory entries; (3) Random-k, which randomly selects k historical entries; and (4) All History, which provides all available historical memories without selective retrieval.

Retrieval StrategySwap TPick Objects in Order
Oracle Source84%96%
OnCue83%96%
Recent-k10%60%
Random-k40%23%
All History71%45%

As shown in the table, OnCue achieves performance comparable to Oracle Source on both tasks. In contrast, Recent-k, Random-k, and All History exhibit substantial performance degradation. These results demonstrate the effectiveness of task-dependent selective retrieval, with OnCue achieving performance close to Oracle Source despite relying on predicted historical evidence.

The retrieval size k is task-dependent, with k = 2 for Swap T and k = 3 for Pick Objects in Order. We have incorporated these additional experiments and analyses into the revised paper, with detailed implementation settings provided in the Appendix.

AuthorsOpenReview ↗

Response to Reviewer X8NF [2/2]

Q3: Clarity of Ablation Settings

We thank the reviewer for pointing out the insufficient clarity in our ablation settings. We agree that this was an oversight in our presentation, and we have expanded the ablation analyses in the paper and clarified the corresponding implementation details in the revised Appendix.

Specifically, we provide detailed configurations for (1) Memory Representation, clarifying how removing Language or Visual Memory affects the Planner and Executor inputs; (2) Memory Granularity, specifying the removal of scene-level or object-level memory and the corresponding retraining; (3) Completion Classifier, explaining the simulation-based Oracle Completion criteria and fixed-interval update strategies under the same Planner and VLA; and (4) Historical Memory Retrieval, clarifying how different retrieval strategies are configured and implemented, including Oracle Source, Recent-k, Random-k, and All History, along with the corresponding retraining protocol.

Q4: Completion Classifier and Fixed-Interval Memory Updates

We appreciate the reviewer's constructive feedback on the design of our original completion classifier ablation. We agree that disabling memory updates entirely represents an overly extreme setting. Therefore, we have replaced this comparison with two fixed-interval update strategies: Subtask-specific Median and Global Average. Both strategies use the same Planner and VLA as OnCue, differing only in the memory update triggers.

Specifically, we compare OnCue with three variants: (1) Oracle Completion, which uses ground-truth completion signals derived from simulator physical states; (2) Subtask-specific Median, which determines the update interval based on the median execution length of the closest matching subtask type in the training set; and (3) Global Average, which uses a fixed interval computed from the average ratio of total execution steps to the number of subtasks in the training set.

MethodPress ButtonRearrange BlocksCover Blocks
Oracle Completion96%99%99%
OnCue87%93%96%
Subtask-specific Median81%90%95%
Global Average45%70%84%

As shown in the table, the performance gap between Oracle Completion and OnCue highlights the potential for further improvement in completion detection, while OnCue consistently outperforms both fixed-interval update strategies across all three tasks. Notably, Global Average exhibits substantial performance degradation on Press Button and Rearrange Blocks, as some subtasks require considerably longer execution times than others, making a uniform update interval prone to premature memory updates and task failures. These results demonstrate the effectiveness of completion-driven memory updates compared with predefined fixed intervals.

We have updated the corresponding ablation study in the revised paper and provided further implementation details in the Appendix.

Thank you for the additional experiments and clarifications. I like these quite a bit and believe they make your paper more sound. I will note that the subtask-specific median is a very strong heuristic competing closely with your classifier; even though there is room to gain I don't believe the classifier by itself is a significant contributor to your algorithm. Please soften the claims around it.

I will adjust my rating to "Weak Acceptance".

Thank you for your positive feedback and for raising your rating. We agree that Subtask-specific Median is a competitive heuristic and that the contribution of the completion classifier should not be overstated. Accordingly, we have revised the discussion in the paper to acknowledge the modest performance gains and emphasize its practical role in enabling adaptive memory updates based on actual execution progress.

Comments on the paper

1 comment
Area_Chair_gyyEOpenReview ↗

reviewer discussion

Dear Reviewers,

The authors have submitted their rebuttal. Please read their response carefully and engage in the discussion as soon as possible. Let us know whether the rebuttal addresses your concerns and update your final rating accordingly.

Thank you for your time and service to ICLR.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.