acceptodds
Withdrawn from ICLR 2027

Knowledge Has No Fixed Home: Dynamic Verified Placement Across External and Parametric Memory

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Self-evolving agents must decide whether acquired knowledge should be maintained externally or internalized in model parameters. Neither choice is universally preferable: their value depends on future use, updates, executor costs, and reliability. We formulate *dynamic knowledge placement* and show that preferred representations change with lifecycle conditions, earlier decisions can expire, and average executor success hides knowledge-specific risk. Our method, *Dynamic Verified Placement (DVP)*, filters executors by capability and risk, compares their expected lifecycle value, revisits earlier placements, and verifies new representations before use. On LightShop, external and parametric memory solve complementary tasks. Under fixed budgets, DVP completes 23.78 and 63.32 more correct tasks than always external on the primary and shifted workloads. It also completes 3.80 and 9.38 more tasks than a matched adaptive-amortization policy that uses the same forecasts, executor cards, and transaction protocol but optimizes only serving-time savings. Under the exact benchmark verifier, no incorrect internal output was released.

Primary area:
foundation or frontier models, including LLMs
Keywords:
agent memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tQBDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how self-evolving AI agents can decide whether to store acquired knowledge in external memory or integrate it into model parameters. The authors highlight that the better choice depends on factors such as how often the knowledge is used or updated, along with the cost and reliability of each memory mechanism. To address this problem, they propose Dynamic Verified Placement (DVP), which makes placement decisions based on expected long-term benefits, updates these decisions as conditions change, and verifies new knowledge representations before using them. The experiments on LightShop show that external and parametric memory have complementary strengths. Under fixed execution budgets, DVP completes 23.78 and 63.32 more correct tasks than the external-memory baseline on the primary and shifted workloads, respectively. It also improves over the adaptive-amortization baseline by 3.80 and 9.38 correct tasks.

Critical strengths

1- This paper addresses an interesting problem in self-evolving agents: deciding where to store acquired knowledge and when to reconsider previous storage decisions. The idea that external and parametric memory can be beneficial in different situations, depending on how frequently knowledge is reused or updated and the associated execution costs, is interesting and practically relevant.

2- Instead of making a one-time decision, the proposed approach considers the expected future benefits of each memory mechanism and revisits placement decisions as conditions change. Considering execution reliability, verification, and recovery costs when making these decisions is another useful aspect of the framework.

3- The paper evaluates DVP under different workload conditions and compares it with fixed placement, one-step placement, adaptive amortization, and other memory-management strategies. Using matched workloads and executor outcomes makes these comparisons more meaningful. The additional experiments on cold-start behavior, verification, and computational overhead further support the effectiveness of the proposed approach and strengthen the overall evaluation.

Critical weaknesses

1- The paper emphasizes verifying memory representations before using them. However, the main LightShop experiments rely on an exact benchmark verifier that can directly determine whether the resulting task state is correct. In practical agent settings, such verification may not always be available, making it difficult to assess the correctness of an output or action. The authors also provide an evaluation using an independent trace-based verifier, which is useful, but some incorrect outputs still pass verification. Although the paper includes experiments with imperfect verification, it would be helpful to further analyze how DVP performs under more realistic verification conditions, particularly the trade-off between task completion, verification errors, and recovery costs. Such an analysis would provide stronger evidence of the approach's practical applicability.

2- The proposed approach relies heavily on executor reliability estimates, as its placement decisions depend on executor success rates, activation probabilities, and verification risks. However, obtaining reliable estimates for newly acquired knowledge can be challenging, especially when very limited historical information is available. In the main LightShop experiments, the method uses executor-specific information estimated from separate development tasks. Although the authors provide additional cold-start experiments, these also show that collecting such information can be costly. Therefore, it would be useful to further evaluate DVP in more realistic settings where little or no historical information is available, particularly in a zero-shot scenario where the agent continuously encounters new knowledge. This would help demonstrate how effectively DVP can adapt and evolve without relying on previously collected executor information.

3- The proposed approach is mainly evaluated in controlled environments with a limited range of knowledge types, memory mechanisms, and workload conditions. In particular, the LightShop experiments use a small set of predefined rules and version changes, while the ZsRE experiments mainly focus on factual knowledge updates and different external-access costs. However, in real-world scenarios, self-evolving agents may need to handle much larger and more diverse knowledge, including dependencies between stored information, unexpected changes, and interactions between parameter updates. Therefore, I would recommend evaluating the proposed approach in a more complex and realistic agent setting to better demonstrate whether its advantages extend beyond the controlled workloads considered in the paper.

Questions for authors

Please address my all concerns raised in the weaknesses section.

Reviewer_sD5TOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to dynamically manage knowledge between external and parametric memory. It proposes Dynamic Verified Placement, which considers future usage, update frequency, cost, and reliability to decide where knowledge should be stored. The method also verifies memory before use. Experiments on LightShop and ZsRE show improvements over several baselines.

Critical strengths
  1. The paper studies a valuable problem of managing external and parametric memory in self-evolving agents.
  2. The proposed method considers multiple factors, including future usage, update frequency, cost, and reliability, when making memory placement decisions.
  3. The paper includes comparisons with several baselines and ablation studies to evaluate the effectiveness of different components.
Critical weaknesses
  1. The main experiments on LightShop involve only four rules. This scale is too small to convincingly evaluate memory management. It is unclear whether the method can work well with a large number of knowledge units.
  2. Although the method includes maintenance costs in its objective, the experiments do not account for the storage and switching costs of LoRA adapters. When scaling to many knowledge units, each with its own adapter, these costs could become significant. Ignoring them may make parametric memory appear more beneficial than it actually is.
Questions for authors

Please see the weaknesses above.

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

This paper studies dynamic knowledge placement for self-evolving agents, addressing whether learned knowledge should reside in external stores (retrieval) or model parameters (model editing). The authors propose Dynamic Verified Placement (DVP), an adaptive framework that evaluates the expected lifecycle value of each placement, updates decisions as workloads evolve, and verifies parametric changes before serving. On LightShop and ZsRE under fixed budgets, DVP outperforms static external and one-step placement policies while preventing unverified parametric errors. However, although the concept of the life cycle system is novel, the effectiveness and underlying mechanisms still need more analysis.

Critical strengths

Generally, the proposed lifecycle formulation is novel. The authors propose an adaptive tradeoff between reuse frequency, update costs, and serving overhead, providing an alternative to static storage.

The second strength is that they have risk certificates and pre-serving verification, which can prevent model edits that could be potentially harmful to the models.

Lastly, the experiments show the effectiveness of the methods.

Critical weaknesses

The weaknesses, I personally think, are:

  1. Reliance on verifier quality: the safety and efficiency of this parameter internalization depend on this verifier. If the verification has issues, this can harm the expected cost savings and risk guarantees.
  2. Calculating the expected lifecycle values requires estimating the future query frequency, which sometimes cannot be robustly estimated.
  3. The experiments are restricted to these two datasets. I would suggest providing more open-ended, agentic environments and interactive settings.
Questions for authors

So, how sensitive is this proposed method to verifier problems, issues, or mistakes?

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

Related papers

Discussion (0)

Sign in to comment.