acceptodds
Withdrawn from ICLR 2027

Action-Informed Merging for Agentic Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agentic models combine reasoning with executable actions to accomplish tasks in interactive environments. Models trained for specific domains often excel within those environments, but their capabilities may not generalize to other domains. Model merging combines the parameters of separately trained models into a single model, offering a way to consolidate their domain-specific strengths into a more broadly capable agent without additional training. However, models trained for different domains may have conflicting parameter updates, making it difficult to broaden domain coverage while retaining specialist-level performance via merging. To combine complementary capabilities while limiting interference, we propose Action-Informed Merging (AIM), which combines action-relevant updates from specialist checkpoints and selects which parameters to modify and how much to change them. Specifically, AIM estimates parameter importance from squared gradients of action-token and full-response-token losses. Updates from specialist checkpoints are weighted according to parameter importance for action prediction, with less weight assigned to conflicting updates. Parameter-block selection then determines which parameters to modify and how much to change them. This selection uses full-response importance to penalize changes to sensitive parameters and constrains estimated loss increases separately for each specialist checkpoint whose capabilities should be retained. The procedure requires *neither further model training nor inference-time routing*. On tool-use benchmarks, AIM reaches a 49.11% equal-weight average of BFCL accuracy and Terminal-Bench success, compared with 47.59% for uniform averaging. Across web, mobile, desktop, and GUI grounding, its four-domain average is 50.58%, a *14.1% relative improvement* over the weighted averaging baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM, Model Merge, GUI agents

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

This paper proposes Action-Informed Merging (AIM), a training-free framework that integrates the domain-specific capabilities of multiple agent specialist models. It is initialized from a shared Qwen3.5-4B model into a single checkpoint without requiring inference-time routing.
It is mainly composed of the following two components: 1. Action-Informed Consensus Construction: It estimates parameter sensitivity (or importance) using the squared gradients of loss on action token spans (L$^act$) and loss on full responses (L$^{full}).Taskvectorsofspecialistmodelsareweightedbyactionsensitivityandcombinedusingaconsensusgateg). Task vectors of specialist models are weighted by action sensitivity and combined using a consensus gate g{_p}$ based on cross-expert support and sign agreement. 2. Sensitivity-Guided Update Selection: Via a diagonal quadratic surrogate objective function, it balances updates driven by action sensitivity toward specialist models with penalties driven by full-response sensitivity against an anchor checkpoint. Candidate parameter blocks are selected under explicit loss budgets (ϵi_i) for each experts.
This method is evaluated in two domains, tool-using and GUI, and both domains get improved.

Critical strengths
  • A sound framework based on action vs. response sensitivity: By decoupling action token loss gradients from full-response loss gradients, it provides an accurate method to maintain execution capabilities while suppressing divergence in general generation capabilities. Through the formulation of a quadratic surrogate objective function, the trade-off between maintaining anchor model performance and acquiring new capabilities from specialist models is explicitly modeled.

  • Broad interactive evaluation and high transparency: Verified across multifaceted interactive environments including function calling, terminal execution, web navigation, mobile operation, desktop OS tasks, and GUI grounding. Clear information disclosures are provided regarding evaluation protocols, prompt templates, calibration data sources, and the use of development sets.

Critical weaknesses
  • Generalisation concerns.

    • The largest reported average improvement is mainly driven by MobileWorld: the final balanced model scores below weighted averaging on web, desktop, and grounding tasks (as in Table 2). This trade-off matters when the goal is to combine several capabilities.

    • Lack of OOD evaluation on GUI benchmarks, which further addresses the concern. In Appendix A.4, authors explicitly stated that the 4-domain computer-use benchmarks were used as development data for model/setting selection. Because there are no results on an independent test set untouched by setting selection, it is unclear to what degree the reported 14.1% relative gain is due to generalizable consensus merging.

  • It is difficult to identify which parts of the method produce the improvements or to reproduce the final model.

    • The final computer-use model combines the proposed merging approach with several additional mixing methods and refinement steps, while the tool-use model follows a different construction. The reported ablations mainly examine later refinements rather than separating the effects of the main proposed components (Table 4; Appendix D). A comparison with a simpler merging approach using the same specialists and comparable tuning would help establish the value of the action-focused design. A clearer end-to-end recipe, including the exact specialist checkpoints and configuration choices, is also needed for independent reproduction.
Questions for authors
  1. Do evaluation results exist on unseen GUI/computer-use tasks that were completely unused during setting selection, anchor choice, and hyperparameter tuning?

  2. What numerical value was set for κ in Equation (7) during the experiments? Also, what specific algorithm was used to solve the block selection optimization problem in Equation (9)?

  3. In the overall consensus and selection pipeline, if the action mask loss L$^{act}$ is replaced with the regular response loss L$^{full}$ while holding other hyperparameters fixed, how does the performance of AIM change?

Reviewer_aU6BOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a model merging method that uses agent action tokens to determine the merging weights of different checkpoints. Experiments on Qwen3.5-4B-based models cover tool use, terminal execution, and GUI interaction, including web, mobile, desktop, and visual grounding tasks. The method improves aggregate performance over uniform and weighted averaging baselines.

Critical strengths

The idea of using action tokens to guide model merging is interesting and promising. Executable actions are central to an agent’s interaction with its environment, so action-specific sensitivity provides a relevant signal for determining how different specialist checkpoints should contribute to the merged model. Distinguishing action prediction from full-response generation gives the approach a clear motivation: it allows the merging strategy to explicitly account for the behaviors through which agents carry out tasks. This makes the proposed design closely aligned with the goal of consolidating specialized agent capabilities.

Critical weaknesses
  1. Clarity of the method. The presentation and methodological description need improvement. For example, the sample weight ηi,n\eta_{i,n} in Equation (3) influences the merging weights, but its actual setting is not specified.

  2. Limited model scale. All experiments are conducted on 4B models, leaving the applicability to larger models unclear. Since training smaller models on combined datasets is relatively inexpensive, model merging may offer greater practical value at larger scales. Evaluation on models with at least 32B parameters would strengthen the evidence.

  3. Missing comparison with joint training. The authors should compare model merging with directly training on a mixture of the datasets to better establish the advantages of their approach.

Questions for authors
  1. Have you conducted experiments on larger models to assess the scalability of the method?
  2. How are the calibration samples selected?
  3. Does the sample-dependent merging procedure risk overfitting to the calibration data?
Reviewer_bfTUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses training-free consolidation of specialist agent checkpoints: given several models fine-tuned from a shared backbone for different action domains, produce a single deployable checkpoint without routing at inference time. The proposed procedure estimates per-parameter sensitivity from squared gradients of two losses — one over the full response, one restricted to executable action tokens — builds a consensus proposal gated by action relevance, cross-expert support and sign agreement, then selects parameter blocks under separate surrogate-loss budgets for each protected expert, and finally interpolates from an anchor. Experiments cover BFCL and Terminal-Bench 2.1 on the tool-use side, and Web-L2, MobileWorld, OSWorld plus six grounding datasets on the GUI side, the paper pairs a genuinely agent-specific notion of parameter importance with a constrained block-selection procedure, and shows that merging is workable across agent domains. That sentence is accurate, but the evidence delivered supports only its weaker half.

Critical strengths

1. Well-chosen problem, serious experimental scale. Training-free consolidation of agent specialists is a real deployment need, and the study uses tool-use and GUI specialists across BFCL, Terminal-Bench 2.1, Web-L2, MobileWorld, OSWorld and six grounding datasets.

2. Separating action-token from full-response sensitivity is a genuinely agent-specific insight. It identifies two semantically distinct token populations in an agent response with different parameter requirements, rather than renaming an NLP merging trick.

Critical weaknesses
  1. The cleanest ablation of the core signal points the other way. Action-masked Fisher is 0.59 lower in Table 1 (46.42 vs 47.01) and 0.62 lower in Table 3 (67.21 vs 67.83). The paper does not address the contradiction.
  2. TIES, DARE, RAM and AgentPatch are claimed or cited as related works, but no direct comparison. The comparison set is uniform averaging and two Fisher variants only. The paper's premise is that conflicting updates cause merging to fail, so these baseline absence leaves the claim that the proposed agreement treatment beats existing sign election untested.
Questions for authors
  1. The cost of "training-free" is not quantified. The selling point is training-free, yet the procedure requires forward and backward passes over six checkpoints, a constrained selection over entry blocks across layer groups, and staged interpolation with checkpoint writes; no GPU-hour or wall-clock figure is given.

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

Related papers

Discussion (0)

Sign in to comment.