- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper trains an MLLM agent to construct demonstration contexts for multimodal ICL. During iterative reasoning, the agent identifies its own information gaps, performs targeted retrieval to obtain candidate demonstrations, and then uses the existing context to filter, refine, or further expand the candidate set. Once the agent judges the context to be sufficient, it passes it to an answering model. Training uses a two-stage post-training procedure that jointly optimizes context acquisition and answering. Across 7 benchmarks and 3 MLLMs, the authors report an average improvement of 11.6 percentage points over similarity-based retrieval, and 4.4 points over the strongest baseline.
Critical strengths
- Demonstration selection is a key factor in multimodal ICL performance; the research direction of this paper has practical value.
- The paper is clearly written and well organized, and the experiments include ablations on training stages, the number of demonstrations, and retrieval modes.
Critical weaknesses
- In terms of problem novelty and related-work coverage, the core idea has appeared in published work, and the closest works are not discussed. MetaFlowLLM (NeurIPS 2025) likewise retrieves and reuses the experience of past tasks to assist solving new tasks, and Agent Learning via Early Experience (ICML 2026) likewise leverages the agent's own interaction experience to improve future behavior. The core mechanism of this paper overlaps with both, and the paper does not cite or discuss either of them. In addition, Spinning Straw into Gold (ICLR 2026) constructs successful demonstrations from the model's own rollouts for subsequent training, and Auto-regressive In-context Demonstration Selection (ICML 2026) studies how demonstrations are matched to and organized around a query; neither is mentioned, leaving readers without the information needed to position this paper within this line of work and identify its increment.
- In terms of generality of the method, the acquisition pipeline (retrieve, filter, iteratively retrieve more) and the training pipeline (SFT plus group-relative RL) are combinations of standard components, and all three trained agent backbones in the experiments come from the Qwen family. It remains unverified whether the post-training gains transfer to other model families.
Questions for authors
See Critical Weaknesses. In particular: is there evidence that the post-training gains transfer to backbones outside the Qwen family?
3 comments
Dear Reviewer 2uRJ,
There are only 3 days left in the discussion period. In our response, we discuss the papers you highlighted and provide further evidence of ActICL's generalizability across different model families. We would be grateful if you would consider revisiting your score in light of our responses. If you have other unresolved questions, please let us know. We would be happy to engage in the days that remain. Thank you again.
Dear Reviewer 2uRJ,
We wanted to follow up, as this is the last day of the discussion, and see if our additional experiments and clarifications have addressed your questions. Thank you again.
Rebuttal for Reviewer 2uRJ
Thank you very much for the effort you have put into reviewing our paper, and for your recognition of the significance of our research topic, the quality of our writing, and the thoroughness of our experiments. Below, we provide our responses to the valuable comments you have raised. Changes or additions made in the submission PDF have been highlighted in blue.
Thank you for pointing out these works. We would like to respectfully question the relevance of some of these papers, as their research focuses are not closely aligned with the scope of our work. Specifically, the first three works you mentioned, MetaFlowLLM, Agent Learning via Early Experience, and Spinning Straw into Gold, have a fundamentally different purpose from ours. These methods treat the past execution history of an agent as experience and use it to improve future decisions. Agent Learning via Early Experience and Spinning Straw into Gold use these experiences directly as data to retrain the agent, while MetaFlowLLM trains a rewriter that turns them into code or prompts. In contrast, our work studies the general ICL setting, where demonstrations from an external bank are placed in the input context of the model to support its reasoning. These demonstrations come from pre-defined banks rather than from the agent’s own interaction history. Our pipeline also differs from these methods. They focus on how to collect and effectively use experience, whereas our work focuses on enabling the agent to retrieve suitable demonstrations by itself.
Unlike the first three papers, Auto-regressive In-context Demonstration Selection (AutoSelect) does fall within the same scope as our work. However, it does not address multimodal tasks, and it belongs to one category of our baselines, trained retrievers, as illustrated in Figure 1(b). We have added it to the Related Work section in lines 129--130. Unlike this category, ActICL does not rely on a separately trained retriever. It lets the MLLM itself decide through reasoning what to retrieve, which candidates to inspect, and which demonstrations to keep. In summary, ActICL enables the model itself to reason about and retrieve multimodal demonstrations, and thus the novelty and contributions claimed in our paper do not overlap with these works.
First, we respectfully argue that the three backbones in our experiments already provide good evidence for the generality of ActICL; nevertheless, based on this comment, we have further added a fourth backbone. Within Qwen3-VL, we test both a dense model (Qwen3-VL-8B) and a Mixture-of-Experts model (Qwen3-VL-30B-A3B). Moreover, although Qwen3.5 shares the family name, it differs fundamentally from Qwen3-VL in both architecture and training strategy. In architecture, Qwen3-VL builds on the Qwen3 language model with standard full attention, whereas Qwen3.5 adopts a hybrid design in which three out of every four layers use Gated DeltaNet linear attention and only the remaining layer uses gated full attention. In training, Qwen3-VL starts with a pretrained language model and a vision encoder, trains a projection module to align them, and then jointly trains all components. By contrast, Qwen3.5 trains on interleaved text, image, and video tokens from the outset of pretraining. ActICL brings consistent gains on all three backbones, which shows that its effect does not depend on a specific architecture or training recipe.
To further demonstrate the generality of ActICL, we additionally train ActICL agents on InternVL3.5-8B and report the results in Table 14 of the revised appendix, which is also shown in Table R.6 below. On InternVL3.5-8B, 4-shot ActICL outperforms all baselines on all seven benchmarks, with average gains of 9.0 percentage points over MMICES and 2.9 points over ProMSA. Adaptive acquisition further improves average accuracy by 1.8 points. These results show that the post-training gains of ActICL transfer to other model families.
Table R.6 Comparison of ICL methods on InternVL3.5-8B.