acceptodds
Under review as a conference paper at ICLR 2027

Personal AI Agent for Camera Roll VQA

19 views

Abstract

We study the personal camera roll visual question answering setting. In this setting, a conversational AI assistant can access a user's personal camera roll and retrieve relevant photos to answer queries, ranging from simple factual questions (e.g., "Name of the food I tried yesterday?") to more open-ended ones (e.g., "Recommend some dishes I have never eaten before"). Given the vast nature of the personal camera roll (i.e., multiple years, hundreds to thousands of photos), a successful AI assistant needs to understand a long-horizon, highly personalized visual content stream in order to navigate and locate the correct and/or relevant information. To support this, we collect and manually annotate questions that mimic real-world usage. The final dataset, camroll, contains 50 users, 31,476 images, and 2,500 QA pairs. We further design camroll-agent, a conversational AI agent equipped with hierarchical memory and a minimal set of tools for efficient navigation over large, personalized visual memory. Experimental results show that camroll-agent outperforms numerous baselines and methods for long-context understanding AI agents system. Together, the camroll dataset and camroll-agent highlight the gap in AI agents' long-context reasoning: personalized visual memory requires different approaches from standard long-context textual memory, especially when consistency, visual details, and user-specific context are present.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI Agents, Personal AI Agent, photo albums, camera roll

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_c1WWOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies VQA over personal camera rolls, where a conversational assistant answers user-specific questions ("What did I eat after the STS-135 launch?") by locating evidence in a long personal visual history. The motivating gap: existing photo products do face/object or keyword search, and RAG-over-captions designs discard pixels, losing the fine-grained cues (identity, relationships, event context) that matter for personal memory. Two contributions follow — camroll, a benchmark of 50 real rolls (31,476 images, 2,500 human-authored questions), and camroll-agent, a training-free reference agent.

Technically the agent (Sec. 4) is a three-level pyramid — raw pixels, profile-conditioned captions C = {c_i}, and event summaries E = {e_j} = (I_j, d_j, m_j) — linked by hashed IDs and an event_id foreign key giving O(1) bidirectional navigation (Sec. 4.1, Fig. 3). Captions condition the captioning MLLM on the user's profile photo plus a look-back window of k = 3 recent images; event segmentation is an incremental ADD / UPDATE / NO_OP decision per image given the current caption, timestamp, recent captions, and the latest event summary (prompt in A.6). Five tools factor along retrieval paradigm × access depth: search (frozen all-MiniLM-L6-v2 over FAISS), grep (BM25 via FTS5), list (metadata filters), get (full record), and view(id, prompt, ≤6 images), which re-examines raw pixels with a VLM. Interaction is a ReAct loop capped at 25 tool calls with ≤5 view calls (A.3).

Results (Table 3, Gemini-2.5-Flash answerer): no context 30.0 MCQ; all captions 63.4; all images 76.5; oracle 86.4. camroll-agent reaches 70.5 MCQ with 88.5 evidence recall and 4.11 judge at ~3.2k context and ~15k total tokens — ~50× fewer than the ~750k all-images setting while retaining ~92% of its accuracy. Baselines span 32.3 (MemOS) to 54.0 (Claude Code, at ~59k tokens), with Mem0 53.2, LightMem 52.7, HippoRAG2 48.5, Self-RAG 46.2, SimpleMem 44.6. Ablations on the in-house split (Table 5, judge only): final 5.04, generic captions 4.01, no events 4.03, no captions 2.29, no search 2.88, no view 3.22. The error analysis (Table 7) assigns 14.7% of failures to wrong evidence, 24.7% to retrieving-but-never-viewing it, 10.5% to budget exhaustion, 21.7% to premature stopping, 17.5% to answering wrongly after viewing the right images.

Critical strengths
  • A new, carefully built benchmark with human-authored questions. Unlike LLM-synthesized-query datasets (Deng et al., 2026; Xu et al., 2026), annotators inspect each roll, which is what produces longitudinal items ("Have I ever been to Italy?"). The semantic/episodic split, 80% three-verifier agreement, and evidence-conditioned design (98% of episodic questions carry gold evidence, Table 8) are documented checkably (Sec. 3.1, A.5, A.6).
  • Unusually candid reporting of confounds. Sec. 5.2 names three confounds in Table 3 (profile-conditioned vs. generic captions, the view tool, agentic vs. single-pass retrieval), states that the generic-caption ablation explains part of the gap (5.04 → 4.01), and notes Claude Code's different backbone. A.3 adds that the all-images baseline is "a strong but not strictly upper-bound reference" because images are downscaled to 768px.
  • Ablations are unusually complete for a systems paper. Table 5 changes one component at a time across all five tools and all three memory layers, and the ordering is interpretable: captions matter most (2.29 without them), view second (3.22), search third (2.88) — supporting the "retrieve through abstractions, verify in pixels" thesis rather than asserting it.
  • Trajectory-level error analysis and careful reproducibility. The rule-based taxonomy in A.2 is fully specified (step counts, view cap, ρ = 1, G ∩ V = ∅), making it reproducible, and yields an actionable finding (B and D together are 46.4% of failures). A redistributable YFCC core (20 users, 1,000 questions) ships with exact tool budgets and verbatim prompts (A.3, A.6).
Critical weaknesses
  • The central "modality-preserving" claim is never tested against image-embedding retrieval. Sec. 5.2 admits that disentangling the comparison "would require running baselines with personalized captions and adding image-embedding retrieval baselines (e.g., CLIP or SigLIP top-k images)", and defers this. That baseline is precisely the control for the headline design principle: without it, the gap over Mem0 (53.2) or LightMem (52.7) could come from having pixel access at all, not caption-indexed pixel access. This is the most load-bearing missing experiment.
  • view does not return pixels, which weakens the modality-preservation framing. Sec. 4.2 states view "returns a VLM-generated textual analysis," so the answering model never sees an image and verification is mediated by another model call. With Gemini-2.5-Flash apparently serving both roles, view can confirm rather than correct, and correlated errors go unmeasured.
  • Half the benchmark is unobtainable, and the public half is atypical. The in-house subset (30 users, 1,500 questions) is access-controlled pending legal review; YFCC is 592×514 mean resolution versus 3188×3500 in-house (Table 10), so the released core is not the setting argued for. Table 9 also shows ~59% of released images carry non-commercial licenses.
  • The evaluation instrument is unvalidated and carries the ablations. Free-form scoring is GPT-4o on 0–10 with no human study; the only check is a second judge model (Table 6), which measures judge-family sensitivity, not validity. Tables 4 and 5 rest entirely on it, so "captions are the most influential memory component" is a judge-score claim. The primary MCQ metric is only 3-way (33.3% chance), and no CIs or user-level variance are reported (A.3) despite only 50 clusters.
  • Ablation token accounting is non-monotonic and unexplained. In Table 5, removing grep, list, get, or view increases input tokens (4.0k, 3.7k, 4.3k, 3.5k vs. 2.4k), as do generic captions (4.2k). This is never discussed and undercuts the efficiency framing; trajectory compensation for a lost tool is not isolated.
  • The best configuration is not benchmarked, and the main table is not the targeted regime. Table 4 shows Gemini-3.1-Preview-Pro at 5.80 judge with the same agent, yet Table 3 reports Gemini-2.5-Flash. The motivating claim concerns rolls of "thousands" of photos (1–10M tokens), but every roll is under 1,000 images (median 558; Table 10) — the regime where full context still fits. The paper concedes this (Sec. 1) but never evaluates beyond it.
Questions for authors
  • What happens on a roll that genuinely exceeds the context window (e.g., 10k–50k images)? Without a full-context reference at that scale, how should the 50× efficiency claim be read?
  • Since view returns VLM text rather than pixels, would giving the answering model the raw images directly (within budget) close the 70.5 → 76.5 gap? How often does view contradict the caption it is meant to verify?
  • Table 5 shows every tool removal increasing input tokens. Is this trajectory compensation, and does the total count (including the image tokens Table 5 omits) change the tool ranking?
  • Why is the headline run on Gemini-2.5-Flash (5.04) rather than Gemini-3.1-Preview-Pro (5.80), and does the Table 3 ordering hold with the stronger backbone?
  • Semantic questions have 0% gold evidence (Table 8), so ~20% of the benchmark cannot separate retrieval from answering, contrary to the contribution claim. How are they scored, and would per-user CIs change the significance of 70.5 vs. 76.5?
5 comments
AuthorsOpenReview ↗

Image-Embedding Retrieval Baselines

We thank the reviewer for this suggestion. To disentangle the effects of retrieval representations and raw-image access, we introduce three additional SigLIP-based retrieval baselines:

  1. Personalized Caption (T2T): Text-to-text retrieval using personalized image captions.
  2. Generic Caption (T2T): Text-to-text retrieval using generic image captions.
  3. Image Embedding (T2I): Direct text-to-image retrieval using image embeddings.

All three baselines retrieve original raw images (rather than captions) and pass them, together with the question, to the same Gemini-2.5-Flash answering model. We evaluate Top-K = 1, 5, and 20.

Results (Free-form Judge Score, 0–10):

Retrieval MethodTop-1Top-5Top-20
Personalized Caption (T2T, SigLIP)1.251.592.01
Generic Caption (T2T, SigLIP)1.241.351.98
Image Embedding (T2I, SigLIP)1.081.201.34

For reference, camroll-agent achieves 4.11 with iterative agentic retrieval rather than fixed Top-K retrieval.

These results show that caption-based retrieval consistently outperforms direct image-embedding retrieval, with personalized captions providing additional improvements. Furthermore, despite having access to raw images, all single-pass retrieval baselines substantially underperform camroll-agent. This supports the value of memory abstractions and agentic search.

AuthorsOpenReview ↗

Direct Visual Access vs. Textual View

We thank the reviewer for this insightful question. To investigate whether textual mediation through the view tool introduces an information bottleneck, we conducted an additional ablation where the agent directly receives the retrieved raw images in its context, rather than VLM-generated textual descriptions.

Interestingly, direct image access achieves 70.2% MCQ accuracy, compared with 70.5% using our original view tool, showing nearly identical performance. This suggests that textual mediation does not introduce a substantial performance bottleneck in our setting. One possible explanation is that modern VLMs can effectively extract relevant visual information, while directly incorporating multiple images into the agent's context may introduce additional irrelevant information.

Moreover, our decoupled view design offers an efficiency advantage by keeping visual tokens outside the main agent's reasoning context. For example, assuming 4k tokens per image, directly processing 10 images requires a context containing 40k visual tokens, whereas separate visual inspection operates on much shorter sequences. This reduces the attention burden on the main reasoning model while maintaining comparable accuracy.

We will incorporate this ablation and clarify that in the revised version.

AuthorsOpenReview ↗

Dataset Availability and Methodological Contribution

We acknowledge the limitations regarding dataset availability and representativeness. As clarified in our previous rebuttal, the YFCC subset (20 users, 1,000 questions) has already been released under its original Creative Commons licenses. For the in-house subset (30 users, 1,500 questions), we are actively consulting with our legal team regarding broader redistribution.

Importantly, our contribution goes beyond introducing a benchmark. We also propose camroll-agent, a task-specific approach that highlights the importance of structured visual memory construction and iterative agentic retrieval for personal camera-roll VQA.

Our central finding is that conventional RAG and memory-layer approaches are insufficient for this task. As shown in Table 3, existing RAG and memory systems achieve only 32.3–53.2% MCQ accuracy, while camroll-agent achieves 70.5%. These results demonstrate the value of combining a dedicated hierarchical memory database with interactive evidence retrieval and visual agent.

Thus, our work not only introduces a new benchmark but also provides empirical evidence for an effective architectural direction for personal visual memory agents.

Reviewer_c8MHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces camroll, a benchmark for question answering over personal photo histories, along with a reference agent that combines text-based memory retrieval with selective image inspection.

Critical strengths
  1. The benchmark addresses a relevant task for personal AI assistants, requiring reasoning across multiple images and events.
  2. Human-annotated questions from 50 users, including supporting evidence for episodic questions, provide a useful evaluation resource.
  3. The proposed agent demonstrates a reasonable balance between answering accuracy and token efficiency.
Critical weaknesses
  1. The additional evaluation value of camroll over existing personal photo retrieval and memory-QA benchmarks is not sufficiently demonstrated. More direct comparisons would help clarify its unique contributions.
  2. The evaluation of recent models is limited to an internal subset of the dataset, with key results on the full benchmark missing. This makes comprehensive performance comparisons difficult.
  3. The dataset includes only 50 users and relatively small photo collections, limiting its representativeness and the evaluation of scalability to larger personal photo histories.
  4. The annotation quality requires further validation, particularly for subjective personal preference questions and whole-album reasoning, where ground-truth reliability may be difficult to establish.
Questions for authors
  1. How does camroll differ from existing personal photo retrieval and memory-QA benchmarks, and what additional evaluation insights does it provide?
  2. Can the recent models in Table 4 be evaluated on the full dataset, including MCQ accuracy, to enable more comprehensive comparisons?
  3. How are answers to personal preference and whole-album questions validated, and how consistent is model performance across different users and data sources?
4 comments
AuthorsOpenReview ↗

W1 & Q1: Novelty and additional evaluation value of camroll

We thank the reviewer for this important question. We agree that the value of a benchmark lies not only in its scale, but also in the distinct capabilities it evaluates and the insights it enables.

1. Differences from existing benchmarks.

Existing benchmarks make important contributions to personalized multimodal understanding, but differ from camroll in several fundamental ways:

BenchmarkKey distinction from camroll
Memory-QA (2025)Relies on user-initiated memory recordings (e.g., "remember this"), followed by retrieval and QA over explicitly saved memories. In contrast, camroll requires automatically constructing and organizing memories from raw photo histories without explicit user instructions about what to remember.
OmniQuery (2024)Evaluates relatively short personal memory histories (1–4 months, 2.3 months on average), whereas camroll spans 1.6–9.4 years per user, enabling evaluation of long-horizon visual reasoning.
ATM-Bench (2026)Evaluates long-term multimodal memory QA from a single user's personal history, whereas camroll covers 50 distinct users, providing more diverse personal histories for evaluation.
PhotoBench (2026)Focuses on retrieving photos satisfying complex user intents, whereas camroll evaluates end-to-end VQA, requiring models to retrieve relevant images and reason over visual evidence to generate answers.
DISBench (2026)Uses VLM-generated queries with human verification, potentially biasing question diversity toward patterns recognizable by the generating model. In contrast, camroll uses human-authored questions and evaluates downstream QA beyond image retrieval.

2. Additional evaluation insights.

Beyond benchmark design, camroll enables evaluation of longitudinal visual reasoning, with 32.2% of questions requiring multiple images and 20.0% requiring whole-album reasoning. With gold evidence annotations for episodic questions, it further enables separate analysis of evidence retrieval and answer correctness. Our results reveal that evidence access is a major bottleneck (Table 7), while demonstrating the accuracy–efficiency trade-off of selective visual memory access (70.5% MCQ accuracy with ~15k tokens vs. 76.5% with ~750k tokens for full-image context).

In summary, camroll complements existing benchmarks by evaluating how effectively and efficiently AI assistants construct, retrieve, and reason over memories from raw, longitudinal personal visual histories. We will strengthen the related-work comparison and clarify these distinctive evaluation capabilities in the revised manuscript.

AuthorsOpenReview ↗

W2 & Q2: Full-benchmark evaluation of recent models

We thank the reviewer for this suggestion. We would like to clarify the different purposes of Tables 3 and 4.

Table 3 focuses on system-level comparisons, rather than backbone performance. Our goal is to demonstrate the importance of a task-specific memory architecture and tool harness for personal camera-roll VQA, compared with existing retrieval methods, memory layers, and general-purpose coding agents. Most methods share the same answering backbone (Gemini-2.5-Flash), allowing us to focus on system-level differences. Even with this relatively lightweight model, camroll-agent substantially outperforms existing approaches, highlighting the importance of how visual memories are organized and accessed, rather than simply relying on stronger models.

Table 4 instead studies backbone generalization across proprietary and open-source models. As an ablation, it was originally evaluated on the in-house subset (30 users, 1,500 questions). Nevertheless, we agree that full-benchmark results would improve comparability. We have therefore extended all 11 configurations to the full benchmark (50 users, 2,500 questions), reporting both MCQ accuracy and free-form scores:

Base ModelBuild ModelMCQ Acc. ↑Free-form Judge ↑
Gemini-2.5-FlashGemini-2.5-Flash72.64.65
Gemini-3.1-Preview-ProGemini-2.5-Flash78.95.11
GPT-4oGemini-2.5-Flash56.03.55
GPT-5.2Gemini-2.5-Flash77.74.89
Gemini-2.5-FlashQwen3-VL-8B-Instruct56.13.41
Gemini-3.1-Preview-ProQwen3-VL-8B-Instruct75.45.08
GPT-4oQwen3-VL-8B-Instruct58.93.25
GPT-5.2Qwen3-VL-8B-Instruct72.04.38
Qwen3-VL-8B-InstructQwen3-VL-8B-Instruct48.52.01
Qwen3-Coder-30B-A3BQwen3-VL-8B-Instruct58.43.62
GLM-4.7-FlashQwen3-VL-8B-Instruct51.02.25

The results confirm the trends observed on the in-house subset, with stronger models further improving performance. Together, Table 3 demonstrates the importance of task-specific memory and tools, while Table 4 demonstrates their effectiveness across different backbones. We will clarify this distinction and incorporate the expanded results in the revision.

AuthorsOpenReview ↗

W3 & W4: Dataset scale and annotation reliability

We thank the reviewer for raising these concerns. We acknowledge that dataset scale and annotation reliability are important challenges for personal visual memory benchmarks.

1. Dataset scale and representativeness (W3).

Collecting real personal camera rolls is inherently challenging due to costs. As discussed in our response to W1, existing benchmarks often rely on a single user (e.g., ATM-Bench) or VLM-generated questions (e.g., DISBench). To our knowledge, camroll is the largest human-authored VQA benchmark over longitudinal personal camera rolls, comprising 50 users, 31,476 images, and 2,500 QA pairs spanning 1.6–9.4 years per user.

Moreover, camroll covers two distinct data sources (YFCC and contemporary smartphone photos), providing complementary capture patterns and temporal distributions. Our agent is also training-free, with memory constructed independently for each user, making the architecture applicable to larger collections without requiring retraining. We acknowledge that evaluating substantially larger photo histories remains important future work.

2. Annotation quality and ground-truth reliability (W4).

Ideally, personal questions would be annotated by photo owners themselves, but owner-based annotation is difficult to scale across multi-year histories. We therefore use trained human annotators with calibrated guidelines and independent verification.

We would like to clarify that the reported 80% agreement measures consistency among three independent verifiers in assessing whether candidate QA pairs follow the general annotation protocol, rather than gold-answer correctness. Importantly, all QA pairs retained in the final benchmark were unanimously approved by the verifiers.

The authors clarified the benchmark's contributions and provided additional results with newer models, which addressed my main concerns. I decide to raise my score to 3. I would not give 4 as I remain concerned that a benchmark covering only 50 users is still relatively small given the nature of the task.

Reviewer_2cVSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents "camroll", an eval of 50 real camera rolls with over 30k images and 2.5k questions+gold answer based on these camera rolls that can reasonably be asked by the owners of each camera roll (written by trained annotators to emulate user Q&A behavior). It also introduces a camroll-agent which is an agent with a heirarchical memory system (memory being the main innovation here) and achieves over 70% MCQ accuracy which is close to 76.5% by putting all images in context while being substantially more token efficient.

Critical strengths
  • The benchmark itself is new and well designed. The database of real user camera rolls along with human written questions is a valuable addition to testing q&a on camera roll. Moreover, the problem is well defined. The need and usability of a camera roll eval is clear.
  • The paper does a very good job of separating various parts of the agentic pipeline e.g. retrieval from answering (providing perfect retrieval setup, all images setup etc.) from a measurement standpoint
  • The ablations clearly show that pixel access and captions matter. Removing captions (5.04 → 2.29), search (→ 2.88) or view (→ 3.22) causes large drops, which directly supports the paper's idea of keeping raw images around instead of relying on text alone.
Critical weaknesses
  • The design of the benchmark rests heavily on the semantic/episodic split. This framing itself however is not very well discussed or defended. Yet the question ratio and design (both of question and golden response criteria) depends on this.
  • The design of the camroll-agent itself mirrors camroll and is customized to the benchmark (in how the hierarchies are defined) and therefore it is further more important that the semantic/episodic framing is valid
  • The case for eval is a real, practical use-case. Yet the paper does not discuss addition of new content (user taking new photos) either from an eval standpoint, or from a camroll-agent standpoint where presumably the memory would have to be updated. there
  • I find the benchmark the main contribution of the paper with camroll-agent being a reasonable reference baseline. However, several details are lacking normally present in a benchmarking paper such as more detailed annotation guidelines, quality control protocol, judge prompt etc..
Questions for authors
  • What was the main motivation behind organizing the benchmark around semantic/episodic split (when other dimensions/splits are possible and popular?
  • Events in camroll-agent are defined as episodic memory unit. Was the agent heirarchy designed before or after question taxonomy?
1 comment

1. Semantic/Episodic Framing

Thank you for the comment. As discussed in Sec. 3.1 (L222–230), our semantic/episodic distinction is inspired by cognitive science but defined operationally rather than as a strict taxonomy. Semantic questions assess cross-event understanding of user attributes and preferences, while episodic questions focus on specific past experiences. The 10:40 ratio is a deliberate design choice to emphasize event-grounded reasoning, rather than an estimate of real-world query frequencies.

To further validate this framing, we provide category-wise free-form judge scores for representative baselines:

MethodSemanticEpisodicOverall
All Captions4.663.613.82
All Images5.854.805.01
HippoRAG23.122.452.58
LightMem2.602.402.44
Mem02.802.652.68
Claude Code4.623.563.77
camroll-agent (Ours)4.853.934.11

We observe consistent trends across semantic and episodic questions: methods performing well on one category generally also perform well on the other, with identical relative rankings among the reported methods. This is expected, as both categories require understanding user-specific visual histories. These results suggest that our main conclusions are not driven by the particular semantic/episodic composition of the benchmark.

2. Camroll-agent Design

We would like to clarify that camroll-agent's hierarchical memory is organized by information granularity, rather than the benchmark's semantic/episodic question categories. Specifically, the three levels—raw images, image captions, and event summaries—provide progressively more abstract representations of the same visual history. The memory is constructed independently of the benchmark questions, answers, and category labels.

Such multi-granularity representations are also widely explored in long-video understanding, where hierarchical organization enables efficient coarse-to-fine retrieval and reasoning (e.g., VideoTree, CVPR 2025; VideoAgent, ECCV 2024). Our design follows a similar general principle, adapted to personal camera rolls rather than tailored to the benchmark's question taxonomy.

3. Dynamic Memory Updates

Thank you for the comment. We clarify that camroll-agent inherently supports streaming memory construction, as described in Sec. 4.1 (L314–322). Each incoming image is processed incrementally through ADD, UPDATE, or NO_OP, using only previously observed information. Thus, the agent does not require the complete camera roll in advance or full memory reconstruction when new photos arrive.

This follows the established paradigm of incremental memory updates in NLP, as explored in MemoryBank (AAAI 2024), and LightMem (ICLR 2026). While our benchmark evaluates QA on memory snapshots, the underlying memory construction is inherently streaming and naturally supports continuously growing camera rolls.

4. Annotation Guidelines and Quality Control

Thank you for the comment. We will provide more detailed annotation guidelines and the complete judge prompt in the revised appendix. Each annotator is assigned one camera roll and first familiarizes themselves with the user's history, including writing diary entries for days with more than five photos. They then formulate natural questions with plausible, semantically similar distractors. For quality control, approximately 50% of questions per roll are randomly sampled and reviewed by three independent annotators. Unreasonable questions are rejected and re-annotated, while gold answers must be unanimously verified as correct and answerable. Finally, our primary metric is directly computed MCQ accuracy, independent of the LLM judge, which is only used for supplementary free-form evaluation. The exact judge prompt will be included in the revised version.

Comments on the paper

1 comment
Area_Chair_nR4ZOpenReview ↗

Discussion phase will end on Nov 18

Dear Reviewers,

This is a kind reminder that discussion phase will end on Nov. 18. As the authors have provided the rebuttal, could you please check authors' rebuttal and update your thoughts? Your early feedback would make the discussion more smooth and informative.

Thank you.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.