acceptodds
Under review as a conference paper at ICLR 2027

VisualClaw: A Real-Time, Personalized Agent for the Physical World

Abstract

Vision language models (VLMs) are serving as general-purpose interfaces for complex multimodal tasks. However, deployment still faces three gaps: VLMs typically incur high latency and cost when processing dense video frames and long prompts, the agent scaffold remains static after deployment, and standard video-QA benchmarks do not test whether agents can use visual evidence inside tool-using workspaces. We present VisualClaw, a self-evolving multimodal agent built around two principles. First, hybrid encoding reduces deployment cost by filtering less informative streaming frames with a cascaded gate and compressing the text skill bank through hot/cold top- injection. Second, skill evolution lets the agent learn from failures: retrieved memories condition an offline evolver either as direct concatenated context or as guided evidence, producing skill-bank updates that help future questions. Across video-QA benchmarks with VLMs (Gemini 3 Flash and GPT-5.2), VisualClaw reduces API cost by versus full-frame upload and by versus offline Uniform-8 under the same skill bank. Under the streaming settings, our VisualClaw improves over the baseline by on average, while reaches a peak boost on EgoSchema with Gemini 3 Flash. To address the benchmark gap, we further curate VisualClawArena, a -scenario multimodal agentic benchmark built through a strict five-stage pipeline; models must use video evidence, documents, dynamic updates, and executable checks inside a workspace. On VisualClawArena, the same self-evolution framework with computer-use agent backends improves macro accuracy by for Codex (GPT-5.5) and for Claude Code (Sonnet 4.6) over no-evolution baselines, with a cost reduction compared to the uniform-sampled baseline. These properties make VisualClaw a natural fit for streaming edge applications such as AI glasses, where the cascade filters continuous video locally and transmits only selected keyframes at interaction, while self-evolution supports adaptive assistance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal, personalized agent, self-evolving, AI application

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_CaT3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a new agentic scaffold to advance multimodal task-solving agents. Current deployments suffer from three main issues that the paper aims to solve: (1) high latency and cost due to dense video processing, which is costly and takes time; (2) the skill set of such agents remains static after deployment; and (3) current evaluations do not assess whether such agents can use visual evidence inside tool-using workspaces. To address these gaps, the authors introduce an agent called VisualClaw (and a corresponding benchmark) that addresses the previously mentioned issues as follows: (1) hybrid encoding filters less informative frames with a CPU-only cascaded gate, reducing the cost of each task; the same hybrid encoding is also used to retrieve only the skills from the skill bank that are useful for the task itself (hot/cold top-k injection); (2) skill evolution allows the agent to improve and learn from previous failures, with an offline LLM evolver that updates the skill bank, optionally conditioned on retrieved memories; (3) they introduce VisualClawArena, a 200-scenario multimodal benchmark where agents are evaluated on tasks in which they need to use video evidence, artifacts and dynamic updates. Results on four video-QA benchmarks (Gemini 3 Flash, GPT-5.2) and on VisualClawArena (Codex, Claude Code) show a highly effective reduction in API cost, while also improving over the same-selector baseline without skills; against uniform frame sampling, however, the streaming version trades accuracy for cost.

Critical strengths
  • S1: Large reduction in API cost while maintaining overall performance.
    −98.1% vs full-frame upload and −25.9% vs Uniform-8 with the same evolved skill bank
Critical weaknesses
  • W1: Each component closely follows prior work.
    • Skills: the evolution loop mirrors MetaClaw
    • Cold catalogue: same idea as Claude Agent Skills' progressive disclosure
    • Multimodal skill evolution: in the introduction the authors say "These systems [adaptive skill approaches] remain text-only", this is factually not true, XSkill from ICML 2026
  • W2: Text-only Rounds in VisualClawArena.
    A scenario is kept if up to 40% of its rounds are text-only solvable, this means that ~26% of rounds are not video-required (18.1 of 24.4, Tab. 2). This might introduce large noise in the evaluation results.
  • W3: No comparison with existing frame selectors.
    The baselines the authors compare VisualClaw are only full-frame @1 fps and Uniform-8. However, there are different frame selectors that might prove to match (or improve) VisualClaw's results. For instance, Reducto
  • W4: Most of the video-QA performance improvement comes from 12 hand-written seed skills that are never shown.
    The seed skill bank explains alone most of the improvements presented in Table 3. However, seed skill contents are not available in the manuscript, so it's not possible to assess if they are benchmark-tailored.
  • W5 (Minor Issue): Notation.
    K is overloaded, it denotes the skill-bank size S = {s1..sK} (Sec. 3.1), the keyframe set (Sec. 3.2, Alg. 1), the seed bank K_seed as both a set (Sec. 3.3) and a count (A.1), and the frame budget K = 8 (Tabs. 4, 12), while lowercase k is the hot top-k. Also Cat(·) (catalogue, Eq. 1) vs Cat. (concatenation, Eq. 2). These examples and other notation choices make the parsing of the paper somewhat hard for a reader to follow.
Questions for authors
  • Q1: How do the authors plan to clarify and restate their contribution with respect to the corresponding existing literature highlighted in W1?
  • Q2: Will the authors extend their results section by ablating the impact that text-only solvable rounds have on the final metrics? If so, how?
  • Q3: Can the seed skill bank be added to the manuscript, to assess the impact that it has on performance and whether such skills are also useful in practice rather than tailored to the benchmark?
10 comments

We thank the reviewer for the specific suggestions on positioning, benchmark composition, and seed skills. We address your individual concerns below:

W1/Q1: Components May Follow Prior Work. Thank you for raising the concern. We would like to clarify that our VisualClaw coordinates per-frame filtering, per-query hot/cold skill injection, and post-feedback updates (Secs. 1 and 3). These address redundant uploads, growing skill context and recurring failures within one video-grounded agent. Secs. 1-2 will distinguish this system contribution from MetaClaw and progressive disclosure and explicitly describe XSkill's multimodal learning. We provide additional comparisons with XSkill in General Response G1 and will add them into the next revision.

W2/Q2: Text-Only Solvable Rounds May Introduce Evaluation Noise. Thanks for the question. We clarify that our VisualClawArena is designed to combine explicit video-grounded rounds with downstream artifact and tool operations (Sec. 4.1, Fig. 3). We provide detailed explanations in the General Response G3 and will add them into the final version.

W3: No Comparison with Existing Frame Selectors. Thank you for the suggestion. In General Response G2, we provide the Reducto comparison, where our cascade encoding strategy outperforms Reducto across different frame budgets.

W4/Q3: Most Video-QA Gains May Come from Undisclosed Seed Skills. Thank you for raising the question. We would like to clarify that the 12 seed skills used in our experiments are automatically generated through prior evolution, NOT hand-written (Sec. 3.3 and Appendix A.1). To demonstrate that in more complex scenarios like VisualClawArena, the skill evolution helps more, we conduct additional experiments using codex and claude code (GPT-6.1 sol and Sonnet-5.5) and show their results with scores from Table 5 (FullEvo-Cat. and macro accuracy as the metric).

Benchmark / backendPlainFrozen SeedFullEvo
VisualClawArena / Codex (GPT-5.5)50.21%51.35%54.27%
VisualClawArena / Claude Code (Sonnet-4.6)48.58%49.00%52.16%
VisualClawArena / Codex (GPT-6.1 sol)60.00%61.87%66.92%
VisualClawArena / Claude Code (Sonnet-5.5)58.25%59.95%64.47%

In the more complex VisualClawArena scenarios, giving frozen seed only brings up the macro accuracy to an average of 1.3% across 4 settings (54.26% v.s. 55.54%), while enabling the skill evolution gives us 5.2% boost (54.26% v.s. 59.46). These results validate the use of our skill evolution over just frozen initial skills. We will include these results in the final version.

W5: Ambiguous Notations. We really appreciate your suggestions. We will make these notations clearer in the final version of the manuscript.

Dear Reviewer CaT3,

Thank you again for your thoughtful feedback. We would greatly appreciate it if you could review our rebuttal and let us know whether it addresses your concerns :)

I'd like to thank the authors for their response and the additional experiments. Summarizing here the status of my concerns and questions, given the evidence and motivations the author provided:

  • W1 Each component closely follows prior work: I appreciate the XSkill comparison and the planned revisions to Secs. 1–2. In particular, I suggest the authors revise the claim in Sec. 1 that existing skill systems "remain text-only." I would additionally ask the authors to specify here what content the revisions to Secs. 1–2 will involve, since it's unclear to me at the moment whether the positioning of the contribution is the integration of mostly existing things or the introduction of novel tools, and how they are distinguished from existing solutions.
  • W3: No comparison with existing frame selector : I generally find the Reducto comparison not yet conclusive, the claimed improvements are rather mixed. At the moment, it's hard to judge whether Cascade selects frames better than Reducto, also more details on the setup are needed. Were EgoSchema and Video-MME long evaluated in full or on a subset? Additionally, some of the backbone-agent configuration are only evaluated at specific frame budgets (only K=8 and so on), this makes it hard to understand such results.
  • As for W2 (Text-only Rounds in VisualClawArena) and W4 (Seed skills), the responses of the authors partially address my concerns. Nevertheless, I think these aspects still need major revision (and inclusion in the main paper) rather than being part of the appendix. To give a concrete example, I don't think it's clear how the seed skills have been generated and how the video-QA performance is related to them, and the pointers to Sec. 3.3 and Appendix A.1 do not clarify this in any way.

Hoping to have made my current position regarding the responses received clear, I will maintain my score and strongly suggest that the authors improve these, in my opinion, crucial aspects of their contribution.

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

This paper presents VisualClaw, a framework for self-evolving agents that process video inputs and perform complex tasks. The framework comprises an edge cascade that selects video frames, an injector that retrieves skills from a memory bank, and an evolver that updates the memory bank based on failures. To evaluate the framework, the paper also introduces VisualClawArena, a benchmark that combines video inputs with complex tasks such as file editing and workspace maintenance. Evaluations on VideoQA benchmarks and VisualClawArena demonstrate the effectiveness of VisualClaw and its individual modules. Further analysis shows that the edge cascade reduces API costs by selecting fewer frames for processing.

Critical strengths

None

Critical weaknesses
  1. Insufficient comparison against other self-evolving agent frameworks. This paper positions VisualClaw as a self-evolving multimodal agent, while there are many related works on self-evolving agents. Though most of these related works focus on scenarios that do not involve video input, their agent designs are still important baselines for the injector and evolver modules of VisualClaw. This paper does not include these methods as baselines in the experiments, which provides no evidence that the module design of VisualClaw is actually than other self-evolving agents.

  2. Unclear usage scenario regarding the module design. Efficiency is claimed as an important advantage of VisualClaw, and the CPU-only edge cascade is designed to reduce the number of video frames processed. These together suggest a typical edge-device usage scenario. However, the evaluation is carried out using Gemini and GPT models instead of small edge models. This leads to a potential discrepancy. If the target evaluation models are cloud-hosted, then the advantage of the proposed CPU-only frame selection method needs stronger justification through comparisons with alternative methods at comparable accuracy levels.

Questions for authors
  1. I hope that the authors can provide a response to the two points mentioned in Critical Weaknesses.

  2. I'd like to know the authors' perspective on whether combining video input with complex tool-use scenarios provides an appropriate evaluation. Typically, these work scenarios do not require videos as inputs as established in the typical daily usage and also most existing benchmarks.

5 comments

We thank the reviewer for the questions about adding more baselines and deployment.

W1: Insufficient Comparison with Other Self-Evolving Agents. Thank you for raising the concern. We additionally add the requested XSkill [1] comparison across three backbones on VisualClawArena in General Response G1. Our VisualClaw displays overall better accuracy performance. We will add these valuable additional results to our final version.

W2: Unclear usage scenario regarding the module design. Thanks for the question. VisualClaw targets cloud-backed assistants on edge devices (Secs. 3.1-3.2). VLM weights and inference remain in the cloud, so there is no additional on-device VLM hosting requirement. The intended setting includes smart glasses connected through a companion phone/gateway, without requiring the wearable to host a VLM. Gemini and GPT therefore match the intended deployment architecture, rather than replacing an assumed edge VLM. In General Response G2, we also provide the Reducto [2] video frame selector comparisons, and our cascade encoding strategy stays the best practice.

Q2: Is Combining Video Input with Tool-Use Tasks an Appropriate Evaluation? Thanks for bringing up the concern. We would like to pose the view that real-world assistants should connect observation with action: e.g., updating a shopping list from viewed products, organizing inspection images and files, or revising maintenance records after a walkthrough. Visual understanding supplies the evidence; agentic tool use turns it into a completed operation in real-life scenarios. VisualClawArena targets this class of workflows through video-grounded workspace tasks (Secs. 4.1-4.2, Fig. 4). Edge-side visual capture and cloud-backed tool use are complementary; an edge-hosted VLM is not required under our current setting.

Reference

[1] Jiang et al. XSkill: Continual Learning from Experience and Skills in Multimodal Agents. ICML 2026.

[2] Li et al. Reducto: On-Camera Filtering for Resource-Efficient Real-Time Video Analytics. ACM SIGCOMM 2020.

Dear Reviewer cuHt,

Thank you again for your thoughtful feedback. We would greatly appreciate it if you could review our rebuttal and let us know whether it addresses your concerns :)

Dear Reviewer cuHt,

We have addressed your questions in the rebuttal. As the discussion deadline approaching, we would greatly appreciate it if you could review our response and share any remaining concerns.

Thank you!

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

The paper presents VisualClaw, a self-evolving multimodal agent framework designed to support efficient streaming-video processing and experience-based adaptation. The framework integrates three main components: a cascaded video-filtering mechanism to reduce redundant visual inputs, a hybrid hot/cold skill-injection strategy to reduce prompt overhead, and a memory-guided skill-evolution mechanism that updates reusable skills based on previous successes and failures without retraining the underlying vision-language model. The authors also introduce VisualClawArena, a benchmark containing 200 multimodal agent scenarios involving video understanding, document processing, and tool-based tasks. Experiments across four video question-answering benchmarks and VisualClawArena demonstrate substantial reductions in API costs and improvements in task performance in several evaluated settings. Overall, the paper addresses important challenges in developing efficient and adaptive multimodal agents and demonstrates benefits from integrating visual processing and skill evolution. However, further clarification and experimental validation are needed, particularly regarding end-to-end real-time performance under continuous video streaming and the reliability of the frame-selection mechanism in preserving task-relevant visual information.

Critical strengths
  1. The paper presents a unified framework that combines streaming-video filtering, efficient skill retrieval, and experience-driven skill evolution. The three-timescale design provides a practical approach to managing visual inputs and accumulated knowledge without modifying the underlying VLM.
  2. The memory-guided skill-evolution mechanism allows the agent to update reusable skills based on previous successes and failures without model retraining. This provides a practical approach to adapting agent behavior over time.
  3. The paper evaluates the framework using multiple VLM backbones and video question-answering benchmarks. It also introduces VisualClawArena to assess multimodal agents performing workflows that combine video observations, documents, and tool execution. The experiments provide evidence of both efficiency improvements and potential benefits from skill evolution.
Critical weaknesses
  1. Algorithm 1 compares each incoming frame with previously stored perceptual hashes. Although Section 3.2 mentions a rolling buffer, the paper does not specify how many hashes are stored or when old hashes are removed. It is therefore unclear how the algorithm maintains constant processing time per frame during long video streams.

  2. The first stage removes frames that appear similar to previous frames. However, small visual changes may contain important information for the task. Once these frames are removed, the later stages cannot recover them. The authors should evaluate whether this filtering process causes important visual information to be missed.

  3. Although the paper reports average response latency, it does not fully evaluate the real-time performance of the complete system during continuous video streaming. The total response time also depends on cloud inference, network communication, and tool execution. More evidence is needed to support the real-time claim.

  4. The proposed skill-evolution mechanism relies on knowing whether previous responses are correct or incorrect. However, in real-world applications, the correct answers are often unavailable. The authors should explain how the system obtains reliable feedback to update its skills during actual deployment.

  5. Equation 1 includes retrieved memory as an input to the VLM, but Algorithm 1 does not show this memory being used during inference. The appendix also indicates that memory is used for skill evolution rather than direct inference. The authors should clarify this inconsistency and explain how memory is actually used in the system.

  6. The experiments mainly show that the system improves by learning from previous tasks. However, this does not necessarily demonstrate personalization. It remains unclear whether the system can learn individual users' preferences and adapt its behavior to different users, as suggested by the paper's title.

Questions for authors
  1. The paper describes VisualClaw as a personalized agent, but the experiments mainly demonstrate task-based adaptation. Could the authors clarify how the framework supports user-specific personalization and whether this capability has been evaluated?
  2. Although the paper reports average response latency, how does the complete system perform during continuous video streaming when cloud inference, network communication, and tool execution are included?
  3. The initial filtering stage may discard frames containing small but important visual changes. Have the authors evaluated how often task-relevant information is lost during filtering and how this affects downstream task performance?
3 comments

We thank the reviewer for the detailed evaluation. The General Response collects the new experiments; below we address your individual concerns directly.

W1: Unclear Hash Retention and Constant Processing Time. Thank you for bringing up this question. In our setting, each incoming frame is compared with one reference: the last frame accepted by the hash stage. Acceptance replaces the reference; rejection leaves it unchanged. No growing history is searched. We will specify this in the revision.

W2: Cascade Encoding May not Catch Subtle Event Thank you for raising the concern. To show that VisualClaw can effectively capture small and important visual clues, we conducted human annotations on 50 sampled EgoSchema videos. We asked annotators to annotate every notable event in the video and see how many of them can be detected by VisualClaw, Reducto [1], or the Uniform sampling under the same setting.

SelectorRetained eventsHit rate
VisualClaw66 / 7686.84%
Reducto49 / 7664.47%

The results show that through 76 detected events, the encoding strategy in our VisualClaw catches the most of them with 22.4% more hit rate than Reducto, which demonstrates the use of our method in catching important video scenes.

W3: Insufficient Evidence for End-to-End Real-Time Performance. Thank you for the question. To measure the overall latency and extend results from Table 16, we replay 30 fixed 60-second clips at 4 sampled fps with VisualClaw's encoder and Gemini 3 Flash and present time for getting the answer below:

Query intervalQuery-to-answer p50 / p95
10 s7.21 / 8.93 s
5 s11.29 / 22.22 s

There are a total of 1,440 sampled frames being processed, the cascade encoding takes 6.02/10.04 ms p50/p95. Answer latency is measured on valid tasks where 95% of answers are obtained within 9s when querying the agent every 10s.

The demonstrated capability is continuous streaming perception coupled to cloud/tool interaction. Feedback-driven skill adaptation and self-evolution is a separate capability (Sec. 3.4), which can be applied when users do not query the agent, thus it’s not a definition of real-time response. We will revise the writing to make this clearer in the final version.

W4: Unclear Feedback When Correct Answers Are Unavailable. In the paper, Sec. 3.4 and Algorithm 1 update skills after answering and receiving feedback. The static Video-QA benchmarks use reference answers after prediction; VisualClawArena uses external checks (Sec. 4.2). During deployment, VisualClaw can use explicit user corrections or verifiable outcomes, such as validation of a generated file. Without reliable feedback, the existing bank remains available and feedback-driven updates are not triggered.

W5: Inconsistent Memory Usage across the Equation, Algorithm, and Appendix. Thank you for pointing these out. We clarify that in Eq. (1), Cat denotes the skill catalogue, not retrieved episodic memory. Sec. 3.4 routes memory to skill evolution; Table 3 separately evaluates answer-time +SkillMemCat, and Table 15 distinguishes canonical FullEvo, evolver Cat., and Guide. General Response G1 reports the fresh routing-trace verification. We will revise these vague definitions in the final version.

W6: Task-Based Adaptation Does Not Demonstrate User Personalization. The evaluated capability is task/session adaptation: reusable skills change after feedback (Sec. 3.4) and are applied in video-grounded workflows (Sec. 4.2). We will adapt to use task/session adaptation consistently in the final version of the manuscript.

Dear Reviewer ybcb,

Thank you again for your thoughtful feedback. We would greatly appreciate it if you could review our rebuttal and let us know whether it addresses your concerns :)

Dear Reviewer ybcb,

We have addressed your questions in the rebuttal. As the discussion deadline approaching, we would greatly appreciate it if you could review our response and share any remaining concerns.

Thank you!

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank the reviewers for the constructive feedback. In the general response, we clarify our contribution and add experimental evidence on: (1) comparison with evolving-agent baselines; (2) comparison with video frame selection baselines; and (3) visual dependence in VisualClawArena.

Our VisualClaw coordinates three timescales: per-frame filtering, per-query skill injection, and post-feedback skill updates (Secs. 3.1-3.4 in paper). The contribution is their designed integration for video-grounded agentic workflows. We will distinguish this scope from MetaClaw and explicitly recognize XSkill's multimodal learning.

G1. Lacking Multimodal Self-Evolving Baseline

We compare VisualClaw with XSkill [1] on the VisualClawArena and report the accuracy and costs. We make sure that the model configurations and testing settings are the same, where both VisualClaw and XSkill evolves within each scenario. The table below reports scenario-macro round accuracy.

BackboneVisualClawXSkill
Gemini 3 Flash47.80%45.20%
GPT-6.1-Sol67.60%59.00%
Gemini Argon75.00%74.00%

VisualClaw gains an average of 4.07% across models. Specifically, the Sol gain is significant under paired scenario analysis (p=0.0014). As for cost, the VisualClaw with Gemini 3 Flash cost 63.5% less than XSkill ($28.98 v.s. $79.46). Together, VisualClaw shows more promising performance and efficiency than the XSkill baseline.

G2. Lacking Video Frame Selection Baseline

To demonstrate the use of our cascade encoding, we compare our cascade frame selection with Reducto [2], while keeping the same evolving workflow in VisualClaw. In detail, we use Gemini 3 Flash on EgoSchema and Video-MME Long for assessing the efficiency of the selection methods. To make a fair comparison, we control K maximum selected frames within each video. We compare our Cascade encoding with Reducto and Uniform sampling baselines with 3 model backbones and 3 different video frame settings:

BackboneBenchmarkKCascade (VisualClaw)ReductoUniform
Gemini 3 FlashEgoSchema874.00%72.00%64.00%
GPT-6.1-SolEgoSchema882.00%74.00%78.00%
Gemini 4 ArgonEgoSchema886.00%78.00%82.00%
Gemini 3 FlashVideo-MME long861.67%61.67%55.00%
GPT-6.1-SolVideo-MME long870.00%71.67%65.00%
Gemini 4 ArgonVideo-MME long872.00%71.67%66.67%
Gemini 3 FlashEgoSchema1678.00%74.00%76.00%
Gemini 3 FlashEgoSchema3280.00%80.00%72.00%
Gemini 3 FlashVideo-MME long1668.33%65.00%53.33%
Gemini 3 FlashVideo-MME long3266.67%68.33%61.67%

At K=8, our Cascade encoding exceeds Reducto selection by average 6.00% and 0.45% on EgoSchema and Video-MME, respectively. For larger K settings, when K=16, our encoding strategy still outperforms Reducto by average 3.7% across two benchmarks, while stays comparable with the Reducto when the frame selection space gets wider (K=32). These results again support the performance advantage of our cascade encoding strategy.

G3. Text-Solvable Rounds May Distort Evaluation in VisualClawArena

We would like to clarify that each task in VisualClawArena is designed to combine explicit video-grounded rounds with downstream artifact and tool operations. As demonstrated in Sec. 4.1 and Figs. 3-4, visual rounds ask for agent actions; later rounds in the same task may use that evidence to take further actions — a later operation does not necessarily request the same visual observation again to belong to a video-grounded workflow.

In the VisualClawArena curation workflow (Sec. 4.1), we design explicit text-only leakage gate evaluates candidates with the Gemini 3.1 Pro, with the video hidden or available. It distinguishes visual-required and document-only rounds, and flags excessive textual answer leakage for being repaired or dropped. As a real-life agentic scenario, it does not require every intermediate round to inspect video.

We also provide the visual scoring analysis by restricting scoring to rounds only labeled as visual-required, VisualClaw retains gains over the frozen-skill setting of +3.01 on Claude Code (2,330 rounds) and +2.20 on Codex (1,679 rounds), comparable to the respective overall gains of +3.16 and +2.05. This means our VisualClawArena tasks require video content to be solved.

References

[1] Jiang et al. XSkill: Continual Learning from Experience and Skills in Multimodal Agents. ICML 2026.

[2] Li et al. Reducto: On-Camera Filtering for Resource-Efficient Real-Time Video Analytics. ACM SIGCOMM 2020.

Area_Chair_oVQVOpenReview ↗

Request to participate in discussion

Dear Reviewers,

We are approaching the deadline of the discussion period. Please go through the authors' rebuttal at your earliest convenience, engage in the discussion, indicate any points the rebuttal missed as follow-up comments, and update your rating based on the rebuttal. As the paper received mixed reviews, your additional input is really important for this paper.

Thank you for your time and participation

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.