acceptodds
Withdrawn from ICLR 2027

EpiCon: Collective Agent Learning through Co-Evolving Multimodal Memory

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agents can learn from past executions, but enabling different agents to reuse and build on one another’s experience remains challenging. We introduce EpiCon, a shared multimodal memory framework for agent collective learning without updating host model parameters. EpiCon links question-level memory evolution to a persistent experience bank through two independently trained 2B models: a memory controller and a tree self-organizer. The controller jointly refines textual guidance and visual evidence across attempts and selectively includes visual memory. The self-organizer consolidates lessons hierarchically and retrieves experience and rules for new problems. We evaluate EpiCon on eleven benchmarks spanning four multimodal task domains, using two harnesses and multiple backbones. A frozen bank improves other systems even with a single solving attempt. A second harness raises the original system's macro-average score by 2.6 points across eleven benchmarks. Across four host configurations, EpiCon improves macro-average scores by 1.7 to 4.9 points over No Memory and reduces memory-operation time by 67% to 74% relative to backbone-sized memory models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Memory, Multi-Agent Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9ZZiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces EpiCon, a shared multimodal memory framework for collective agent learning that requires no base-agent model updates. Composed of two independently trained 2B components, EpiCon jointly optimizes textual guidance and visual evidence and hierarchically organizes past experiences. Evaluations on 11 multimodal benchmarks verify its advantages in performance, cross-setting experience transfer, and memory efficiency.

Critical strengths

This work makes some key contributions. It integrates multimodal memory evolution and collective learning by combining question-level text–visual memory refinement with hierarchical experience consolidation, facilitating cross-agent knowledge transfer without parameter updates. Besides, comprehensive experiments across diverse agent harnesses, backbones and multimodal benchmarks, alongside cross-system transfer tests and ablation studies, verify the effectiveness of shared experience, hierarchical organization and adaptive visual memory injection.

Critical weaknesses
  1. In the cross-harness memory evolution experiments, the evolved memory bank incorporates an additional 1,000 questions. Therefore, the observed improvements may partially arise from increased memory coverage rather than cross-agent collaboration itself. A controlled comparison might improve the fairness of this experiment.

  2. The experiments primarily consider relatively small construction sets and limited stages of bank evolution. The behavior of the hierarchical memory under substantially larger experience banks remains unclear, particularly regarding retrieval quality, maintenance overhead, and the accumulation of redundant or conflicting experiences.

  3. While the proposed 2B memory modules are trained on approximately 31.5K supervision examples, competing memory systems rely on zero-shot backbone models for memory operations without training. This introduces a training asymmetry that complicates attributing performance gains solely to the proposed memory architecture.

Questions for authors

See Weaknesses.

Reviewer_4Xx2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EpiCon, a shared multimodal memory framework that supports collective learning among multiple agents without requiring updates to host model parameters. The framework implements a two-level memory mechanism through two independently trained 2B-parameter models (the memory controller and the tree-like self-organizer): the co-evolution of text-visual memory at the problem level, and the construction and reuse of a hierarchical experience repository across problems. The study validated performance across 4 types of multimodal tasks, 11 benchmarks, 2 agent frameworks, and various backbone models: Compared to memory-free baselines, EpiCon achieved a 1.7–4.9-point increase in macro-average scores, reduced memory operation time by 67%–74% compared to backbone-level memory models, and supported experience transfer and collective accumulation across frameworks and backbones.

Critical strengths
  1. The paper presents evaluations across 4 categories of multimodal tasks, 11 benchmarks, 2 types of agent harnesses, and multiple model backbones. The frozen experience repository yields improvements of approximately 3.2 and 4.2 macro-average points, respectively, when applied across different backbones and agent harnesses. Furthermore, augmenting the experience repository using a different agent harness boosts the original system's macro-average score by 2.1–3.2 points—a strong validation that shared experiences are indeed transferable.

  2. Compared to performing memory operations using a backbone large model, two independently trained 2B memory models reduced the associated computational time by approximately 67%–74%, while still outperforming the baseline without memory by 1.7–4.9 points across four host configurations. The paper also validated the effectiveness of visual memory evolution, adaptive injection, tree-like organization, and stronger solver models, providing empirical support for the key design choices.

Critical weaknesses
  1. The paper fails to sufficiently articulate the core challenges of the problem it seeks to address. Specifically, while the authors discuss a range of topics—including the evolution of multimodal memory, experience sharing across agents, hierarchical memory organization, and memory management for small-scale models—they do not clearly delineate the distinct technical hurdles associated with each of these scenarios, nor do they systematically explain why existing methods fall short in overcoming them. Consequently, the "Memory Controller," "Tree Self-Organizer," and the shared experience repository appear more like a collection of plausible modules rather than a cohesive solution; it remains unclear which specific challenges each component is designed to tackle, why each is necessary, and where the key innovations lie in relation to existing agent-memory approaches.

  2. After reading this paper, I am inclined to view it primarily as an innovation in the system framework. However, I would like the authors to further elaborate on the core innovative aspects, as my primary concern remains the significance—or necessity—of "shared memory." If the innovation lies in the system framework, the authors need to more clearly distinguish EpiCon from existing approaches involving hierarchical memory, multimodal memory, and shared memory. Furthermore, if the innovation centers on the two pre-trained 2B models, the main text should provide a comprehensive description of the training methodology and the data generation mechanism. Additionally, it is worth addressing whether training 2B models—as opposed to using memory distillation from larger models—introduces any disadvantages or biases regarding generalization.

  3. EpiCon undergoes specialized training, whereas the baseline relies on zero-shot memory manipulation. This comparison should complement the performance evaluation of EpiCon (2B model) without additional training, as well as that of other memory frameworks—also using 2B models—thereby better isolating the specific contributions of each module.

  4. The tasks used to validate the shared memory approach are too heavily skewed toward MLLMs; it is essential to include additional tasks where the use of shared memory demonstrably improves performance for MAS.

Questions for authors
  1. Compared to the independent memory of sub-agents—which ensures that context remains uncontaminated—how is the importance of shared memory manifested? Is it truly necessary for experiences to be reused by other agents? In other words, what challenges arise when attempting to facilitate such reuse using traditional memory architectures?

  2. Are the tasks selected by the authors—such as VQA—truly in need of a multi-agent approach? Could VQA not be handled more effectively by a MLLM alone? Arguably, a multi-agent framework should be reserved for more complex tasks—those that require the use of specific tools or skills to interact with the digital or physical world.

  3. The author mentions in the ablation study that the "no-memory" condition represents a single attempt, whereas the MC approach involves five attempts. Does the performance gain in this case stem from the inclusion of memory, or simply from the fact that multiple attempts were made?

  4. It appears that shared memory can lead to significant performance degradation, depending on the specific harness architecture. Is it fair to conclude that the use of shared memory carries potential risks and does not always yield a net performance benefit?

Reviewer_1yiaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EpiCon is a shared multimodal memory for multi-agent systems built from two LoRA-tuned 2B models. A Memory Controller co-evolves textual guidance and image-region evidence across attempts and gates visual injection. A tree self-organizer places, merges, splits, and consolidates lessons into a hierarchical bank and retrieves rules or experiences. On 11 benchmarks with two harnesses and two backbones, the 2B variant improves macro-average by 1.7–4.9 points over No Memory and cuts memor operation time by 67–74% relative to backbone-sized memory models. The paper also reports harness bank transfer and gains after a second harness evolves the bank.

Critical strengths
  1. Broad evaluation of 11 benchmarks in four domains, two harnesses, two open backbones, and an API solver.
  2. Fair comparison of every memory method uses the same single-attempt protocol with disjoint construction/evaluation splits.
  3. Frozen-bank transfer across backbones (+3.2) and harnesses (+4.2) is a substantive result for shared, system-agnostic memory.
  4. Separate reporting of MAS vs. memory time and tokens is more transparent than is typical in agent-memory work.
  5. Joint text–visual memory with adaptive injection is a sensible design and also supports both evolving visual memory and selective injection.
Critical weaknesses
  1. Collective learning gain is confounded, the evolved bank uses 2,000 construction questions vs. 1,000 for the original.
  2. No seeds or confidence intervals, despite tiny test sets (DocVQA2026, n = 70) and many 1–2-point differences.
  3. Unexplained baseline anomalies for all methods on DeepSeek and Gemma.
  4. The efficiency claim covers time only, and the memory models are distilled from the same backbone and harnesses used at test time, so "transfer" may be closer to in-distribution than claimed.
Questions for authors

Please answer these questions:

  1. Can authors add the control where the original harness expands its own bank with the same extra 1,000 questions?
  2. Since authors used the controller only as a visual gate, can they report the bank with question-level co-evolution enabled during construction vs. disabled, to show co-evolution improves the stored lessons?
  3. During bank construction, what criterion decides that a lesson is "accepted", and is it the same for baselines?
  4. Have authors checked overlap between training sources (DocVQA, MathV360K) and evaluation items such as MP-DocVQA?
  5. Why does EpiCon-2B lower MP-DocVQA from 79.5 to 66.1 on Qwen?

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

Related papers

Discussion (0)

Sign in to comment.