acceptodds
Under review as a conference paper at ICLR 2027

DebtMerge: Long-Horizon Expert Evolution for Continual Model Merging

Abstract

Continual model merging (CMM) sequentially integrates task-specific experts without joint retraining. Existing methods mainly make local merge decisions based on the incoming task and current expert state, without accounting for the cumulative effects of past merges. Over long task streams, this can lead to persistent historical perturbation, insufficient integration of incoming knowledge, and growing expert-capacity pressure. Jointly optimizing expert selection and update parameters is also constrained by costly candidate evaluations. To address these, we introduce DebtMerge, a long-horizon optimization framework for expert-managed CMM that jointly addresses long-horizon control and finite-budget merge optimization. DebtMerge formulates CMM as an online problem with time-average targets and uses Lyapunov virtual queues to represent accumulated stability, plasticity, and resource debts. These debts induce a per-arrival objective, while stability and plasticity debts guide geometry initialization without future task information. To optimize this objective under a fixed evaluation budget, DebtMerge uses evolutionary transfer optimization over merge recipes specifying expert structure, knowledge composition, and integration geometry. Compatible historical recipe blocks are adapted to the current expert bank and debt state, with objective feedback guiding their reuse toward promising combinations. We characterize debt-guided directional control, incorporate finite-budget solver error into the long-horizon control bound, and derive sufficient conditions for preserving historical predictions under stated assumptions. Experiments across vision and language models show improved average performance over the compared merging baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Continual Model Merging, Model Merging, Lyapunov Optimization, Evolutionary Transfer Optimization

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_P9d3OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes debtmerge, a method for continual model merging that tracks accumulated violations of forgetting, plasticity, and capacity budgets, and uses them to weight each merge objective. Each merge is chosen by an evolutionary search that reuses parts of earlier merge decisions. Experiments on vision and language models show improved average performance over prior continual merging baselines.

Critical strengths

Strengths

  1. The proposed method shows the best average performance among model merging baselines across vision and language models, up to a 1.5B-scale LLM. Its advantage grows as the task stream gets longer, which supports the paper's core motivation for long-horizon merging.

  2. The ablations clearly separate the contribution of each component, and the extensive analysis throughout the main paper and the appendix provides controlled comparisons that isolate the effect of the proposed debt-based control from simpler control schemes.

Critical weaknesses

Weaknesses

  1. The formulation is not self-contained, which significantly hurts readability. The method builds on Lyapunov drift-plus-penalty optimization from network control domain, but never explains why it suits long-horizon continual merging, making the reader who is unfamiliar to the domain hard to understand the method's core motivation. It is only cited in passing in the introduction and related work. Including this, Many sentences rely on external references without explanation, and the following are only a few examples:
  • "Following the standard virtual-queue construction ... (Neely, 2010)" (Line 216): "virtual queue" is never properly explained. In the current manuscript, virtual queue, debt, and merge-debt state all refer to QtQ_t, in my understandings.
  • "Motivated by the data-free interference signals studied in existing work ..." (Line 160): the related work covers these papers only in short words, so it is unclear what those signals actually are in the paper's notation and how the paper is motivated from those.
  • "DebtMerge employs evolutionary transfer optimization (Tan et al., 2021; Xue et al., 2025)": the paper does not explain what evolutionary transfer optimization is or how it differs from standard evolutionary search. Also, the terms for crossover, mutation, and repair are used but their specific operators are not specified.
  1. There are lots of undefined notations throughout the paper, which require substantial revision. Some examples I found are listed below:
  • Undefined: Δt\Delta_t (Line 226); ht,kh_{t,k} (Line 258); κ,ζ,Qˉt\kappa, \zeta, \bar Q_t (Eq. 8)
  • Defined only in the appendix: d~t\tilde d_t (used in Eq. 7)

As a result, the paper is unnecessarily jargony, also coming with domain-specific terms that are not well explained.

Questions for authors

Questions

  • Are "virtual queue", "debt", and "merge-debt state" the same quantity QtQ_t?
  • What is a "compatible" block, and how is compatibility determined? It is never properly defined, even in Algorithm 1. How does "compatible" differ from "feasible"? Also, what does "adapted to the current expert bank" mean?
  • Comparing Table 1 in the main paper with Table 11 in the appendix, the MADE-IT results differ significantly between its original implementation and in the authors' implementation, making the gap marginal in the controlled table. What causes this difference?
Reviewer_y2DFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DebtMerge for continual model merging with a capacity-constrained expert bank. The paper reports strong results across vision and language benchmarks.

Critical strengths
  1. The paper presents a clearly structured framework. The overall pipeline is easy to follow.

  2. The proposed method combines Lyapunov-based debt control with evolutionary search and historical recipe transfer, supported by theoretical analysis.

  3. Experiments are extensive, covering vision and language tasks, multiple architectures, and a range of baselines. The ablations and sensitivity analyses provide additional insight into the method.

Critical weaknesses

1. Capacity should be measured by actual resource cost.

The paper defines capacity by the number of expert slots (M). However, the actual cost of a slot depends on the expert representation and compression scheme, so the same slot count does not necessarily correspond to the same resource budget. Why not directly use storage or the number of retained parameters as the capacity constraint? For instance, using more slots with stronger compression may yield better performance under the same storage budget. Fixing M may exclude such potentially superior configurations.

2. The source of the main-table improvements remains insufficiently isolated.

In the 8-task / M=8 setting, the slot count permits a separate expert for each task, without sharing or replacement forced by slot shortage. Nevertheless, Table 1 reports gains of 2.89 points over MADE-IT and 4.19 points over MINGLE on ViT-B/32, approaching independent fine-tuning (89.99% versus 90.40%). Which proposed components contribute these gains when slots are sufficient?

The main-table comparison does not sufficiently specify the compression, routing, and hyperparameter settings used for DebtMerge and the baselines. Differences in these settings could contribute to the reported gains independently of the proposed mechanisms. Please clarify the configurations for each method.

3. The benefit of the full action decision space remains unclear.

In Fig. 3b, Retain is barely visible, while Share/Reuse already occur during early arrivals. When free slots remain available, these operations are not required by slot scarcity and may introduce additional interference or approximation error. What benefit justifies choosing them over Allocate? How would a simple policy—“Allocate while slots are free, merge by a fixed rule once full”—compare with the full five-action policy in accuracy and storage? This comparison would clarify the incremental value of the action decisions.

4. The initialization used in the search-budget experiment needs clarification.

Fig. 5 shows that increasing the search budget improves final accuracy. While the paper describes debt-guided geometry anchors, how is the initial population of ten complete recipes constructed in this experiment, particularly the initial action types and target/source expert choices? It therefore remains unclear how much of the observed gain would persist with a reasonable heuristic initialization.

Questions for authors

see weaknesses

3 comments

The authors’ responses have further reinforced my concerns. Prior work has consistently addressed settings where merging interference and forgetting are genuine challenges. In contrast, under the authors’ experimental setup, Allocate-per-task, without using the proposed strategies, already achieves performance close to independent fine-tuning with almost no forgetting. This raises concerns that the underlying setup may have substantially weakened or even circumvented the core difficulties the paper aims to address. Consequently, subsequent experiments conducted under this setup cannot adequately establish the effectiveness of the proposed strategies when these difficulties are actually present, and the current comparisons may overestimate their relative performance advantages.

The capacity–compression trade-off results also show that different configurations can yield different performance under similar actual storage budgets. Therefore, the number of expert slots, M, cannot be directly equated with the actual resource budget. Using slot counts as the primary proxy for resource modeling and cross-method comparisons may not adequately ensure fair comparisons and may confound configuration benefits with methodological contributions.

Furthermore, Retain accounts for only 1.9% and 1.6% of executed actions when free slots are available and when the expert bank is full, respectively, which further raises questions about its necessity. Its independent contribution still needs to be verified through a dedicated ablation that removes Retain.

Thank you for the authors’ further response. However, my central concerns remain insufficiently addressed.

First, the authors attribute the strong performance of Allocate-per-task to independent expert allocation. However, existing multi-expert continual model merging methods also employ similar mechanisms, and independent experts do not inherently eliminate routing interference. Therefore, this explanation remains insufficient to account for why Allocate-per-task achieves performance close to independent fine-tuning with almost no forgetting.

More importantly, regardless of the underlying mechanism responsible for this strong performance, Allocate-per-task already largely avoids performance degradation and forgetting without using the proposed strategies. This suggests that, at least under the experimental setting of Table 1, the core difficulties the paper aims to address have been substantially weakened. Consequently, the performance improvements observed under this setting provide limited evidence that the proposed strategies are effective when these difficulties are genuinely present, while also raising concerns about the fairness of cross-method comparisons and the actual contributions of the proposed strategies.

We thank the reviewer for the continued feedback and for further elaborating on the concerns regarding the evaluation setting and the effectiveness of the proposed strategies. We believe it is important to distinguish between independent expert retention under sufficient capacity and continual knowledge integration under constrained capacity, as these settings involve different optimization challenges.

First, we agree that independent expert allocation does not necessarily eliminate routing interference. In fact, Allocate-per-task and DebtMerge use the same shared router, without access to task identity during inference, as described in Appendix D.3. Therefore, the near-zero BWT of Allocate-per-task is an observed experimental result, rather than an assumption that independent experts eliminate forgetting. In the 8-task setting with M=8, Allocate-per-task can retain a dedicated expert for each task but occupies all expert slots. In contrast, DebtMerge achieves higher ACC, 89.99% versus 89.48%, while using only approximately 72% of the slots, demonstrating the performance and resource-utilization benefits of adaptive expert management even under sufficient capacity.

More importantly, the core challenges addressed in this paper are not circumvented by the sufficient-capacity setting. In the 20-task setting with M=8, independent expert allocation is no longer feasible, requiring cross-task knowledge sharing and integration. The controlled experiments in Appendix D.9, Table 20, show that, under the same expert representation, compression, routing, and evaluation protocol, DebtMerge improves BWT from −0.98 to −0.56 and ACC from 88.18% to 88.93% compared with Allocate-first + Adaptive Merge, while reducing storage overhead from 2.19× to 1.89×. Furthermore, the component ablations in Table 3 show that removing stability debt reduces BWT to −1.52, while removing integration geometry reduces BWT to −1.41, further supporting the contribution of the proposed mechanisms to historical knowledge retention.

Therefore, the low forgetting of Allocate-per-task in the 8-task setting does not imply that the core difficulties of continual model merging have been eliminated across the entire experimental framework. The sufficient-capacity experiment evaluates the trade-off between performance and resource utilization, while the capacity-constrained controlled comparisons and component ablations further demonstrate the effectiveness of DebtMerge under knowledge interference and expert-capacity competition. We also distinguish the complete-method comparisons in Table 1 from the shared-host controlled experiments to avoid attributing cross-implementation performance differences directly to the proposed mechanisms.

We thank the reviewer again for the continued attention and constructive suggestions, and hope these clarifications further address the concerns.

Reviewer_9HA1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies expert-managed continual model merging (CMM), where task-specific models arrive sequentially and must be integrated into a bounded-capacity expert bank with a routing mechanism, without joint retraining or access to historical training data. The authors argue that existing methods make purely local merge decisions informed only by the incoming task and the current expert state, and therefore fail to track the cumulative effects of past merges: persistent historical perturbation, insufficient integration of incoming knowledge, and growing expert-capacity pressure. DebtMerge treats continual model merging as long-horizon online optimization: Lyapunov virtual queues track stability, plasticity, and resource "debts" to form a per-arrival merge objective and anchor integration geometry, while budget-limited evolutionary search optimizes layer-wise merge recipes, reusing semantic blocks from historical recipes with context-and-credit conditioning and feedback. Theory adds solver error to the drift-plus-penalty bound and certifies prediction preservation. On CLIP, RoBERTa, and Qwen2.5 streams, it beats all merging baselines in accuracy.

Critical strengths
  1. Innovate problem framing. Treat expert-managed CMM as a long-horizon online optimization problem — in which accumulated violations, rather than instantaneous interference signals, steer subsequent merge decisions — is a perspective the merging/continual-learning literature has not articulated this clearly. The observation that "the state records what is retained, but does not quantify the cumulative effects of past merges" is crisp and well-motivated.
  2. Consistent improved results. DebtMerge consistently leads in both ACC and BWT across all backbones and horizons, approaching the independent fine-tuning upper bound on 20-task ViT-B/32 (88.93 vs. 89.80).
Critical weaknesses

The headline comparison with MADE-IT conflates mechanism with infrastructure. This is my main concern. Table 1 reports DebtMerge ahead of native MADE-IT by 7.52 / 7.89 / 3.56 ACC points on the 20-task streams. However, the paper's own controlled experiment (Table 11) implements a "MADE-IT-style policy" inside DebtMerge's shared host (same expert factorization, managed modules, router, task heads, compression rule, action executor) and obtains 87.69 ± 0.58, versus DebtMerge's 88.72 ± 0.47 — a mechanism-level delta of only ~1.0 ACC (and 0.38 BWT). That is, roughly 6.3 of the 7.52 headline points appear attributable to host/implementation differences rather than to debts, geometry anchoring, or transfer. The paper honestly footnotes that the controlled analogue "should not be interpreted as the native MADE-IT implementation reported in Table 1," but the main-text narrative ("improves over MADE-IT by 7.52... percentage points") carries no such caveat. Relatedly, Table 1 does not report the native baselines' expert capacity, deployed storage, or merge-time cost, so parity along those axes is unverifiable. Please:

  1. report native MINGLE/MADE-IT at matched capacity (M=8 ), storage, and (ideally) compute;
  2. state explicitly which components differ between native MADE-IT and the shared-host analogue
  3. elevate the controlled comparison to the main text.
Questions for authors

See weaknesses above

Comments on the paper

2 comments
Area_Chair_DLxxOpenReview ↗

Author- Reviewer Discussion

Dear Reviewers, the authors have submitted their response. Kindly review it and update your review where appropriate.

Dear PC, SAC, AC, and Reviewers,

We thank the AC for facilitating the discussion and inviting reviewers to reassess our responses. We appreciate the reviewers' valuable evaluations and feedback. Our multi-round discussions prompted targeted experiments, controlled comparisons, and clearer explanations.


Recognition from Reviewers

  • Reviewer 9HA1 (Rating: 3): Recognized the novel long-horizon online optimization perspective and consistent improvements in ACC and BWT across backbones and task horizons.
  • Reviewer y2DF (Rating: 2): Highlighted the structured framework combining Lyapunov debt control, evolutionary search, and historical transfer, supported by theory, broad evaluation, and ablations.
  • Reviewer P9d3 (Rating: 2): Recognized the leading aggregate performance on vision and language benchmarks, especially longer streams, and controlled ablations isolating debt-based control. Following our revisions, the reviewer acknowledged improved presentation and raised the rating; we have since supplied further shared-host results addressing the remaining attribution concern and await reassessment.

Main Contributions

  1. Long-horizon control beyond local merging. Unlike methods driven primarily by the incoming task and current expert state, DebtMerge uses Lyapunov queues to track accumulated stability, plasticity, and resource debts. They coordinate preservation, integration, and capacity without future tasks or historical training data.
  2. Finite-budget optimization with historical transfer. DebtMerge jointly optimizes expert structure, knowledge composition, and integration geometry. Debt-guided initialization and context-/credit-conditioned reuse of historical recipe blocks improve search quality under limited evaluations, rather than restarting local search at each arrival.
  3. Theory and cross-model validation. We characterize debt-guided directional control, account for finite-budget solver error in the long-horizon bound, and derive sufficient conditions for historical prediction preservation. Across three CLIP backbones, RoBERTa, and Qwen2.5-1.5B, DebtMerge achieves the highest aggregate scores among compared merging methods, alongside retention and resource-efficiency results.

Key Revisions and Responses

  • Reviewer 9HA1 — Attribution and fairness. We moved the shared-host policy comparison to the main text and extended it to 8/14/20-task streams. With expert representation, routing, compression, and execution fixed, DebtMerge improves ACC over MADE-IT-style by 2.62/3.78/1.03 points. Under a common eight-slot, 350-MiB inference-memory cap, DebtMerge reaches 81.76% ACC versus 78.86% for independently implemented MADE-IT and 76.92% for MINGLE-FIFO. We also added comparisons at matched 309/327-MiB footprints and clarified differences between native MADE-IT and the shared-host analogue.

  • Reviewer y2DF — Evaluation difficulty, resource fairness, and mechanism necessity. Through multiple rounds of discussion, we clarified the shared routing mechanism and distinguished sufficient-capacity settings from capacity-constrained continual merging. We added controlled Allocate-per-task and Allocate-first comparisons, component ablations, capacity–compression and matched-storage evaluations, initialization robustness tests, and a preliminary w/o Retain ablation. These additional experiments provide evidence for the effectiveness of DebtMerge in mitigating forgetting, improving expert utilization, and balancing performance with resource costs under capacity constraints.

  • Reviewer P9d3 — Clarity and contribution validation. We expanded the motivation for Lyapunov control, unified notation, and specified recipe operators, compatibility, feasibility, and evaluation-budget accounting. Following the reviewer's partial acknowledgment and rating increase, we added 8/14-task shared-host evaluations showing that controlled gains are not uniformly limited to about one point. Component ablations further distinguish debt control, geometry, and historical transfer.


Overall, the discussions have strengthened mechanism isolation, baseline transparency, and evidence under capacity and memory constraints. DebtMerge's central advance is to move beyond local merging through accumulated debt control and transferable optimization experience, balancing knowledge preservation, integration, and resource efficiency.

We sincerely hope that the AC and reviewers will consider the additional evidence and substantive revisions in their final assessment.

Thank you again for your time and consideration.

Sincerely,

The Authors of DebtMerge

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.