acceptodds
Under review as a conference paper at ICLR 2027

Don't Retrain, Remix: Adaptation of Web Agents using Mixture Hypernetworks

44 views

Abstract

The web is not static or uniform. Websites that serve the same purpose can look and navigate very differently. Even within a single site, successive redesigns can change how familiar features are presented and accessed. Humans adapt to this by recognizing familiar interaction patterns and reusing what they already know. We give web agents a similar ability with WebMix, a mixture hypernetwork that adapts them to unseen interfaces by composing previously trained low-rank adapters. WebMix begins with a bank of adapters trained on source interfaces. At deployment, a scout agent explores the target interface and summarizes its observations into a compact manual. A lightweight router uses this manual to predict mixing coefficients over the adapter bank, which combine the existing adapters into a new one for inference. Adapting to a new interface takes only a scouting run and a forward pass, without gradient updates on the target. With Qwen and Llama models, we evaluate *domain transfer* on TimeWarp and WebArena-Lite-v2, and *version transfer* on TimeWarp. In all three settings, WebMix matches or outperforms in-context learning and LoRA baselines, often matching the average success of specialists trained directly on those interfaces. On TimeWarp domain transfer, it gains about 8 percentage points on average over the best baseline with both models. These gains do not come at the cost of seen interfaces, where WebMix keeps its lead, and extend to version transfer. Adapting to a new interface can therefore rely on exploration and composition instead of retraining.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Web Agents, Agent Generalization, Mixture of Adapters, LoRA Adapters

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zc2aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces WebMix, a framework for adapting web agents to unseen websites without additional fine-tuning. WebMix first trains a bank of interface-specific LoRA adapters. A scout agent then explores a target website and generates an interface manual, which a learned router uses to predict mixing coefficients for composing the existing adapters. Experiments on TimeWarp and WebArena-Lite-v2 demonstrate improved performance on unseen domains and website versions compared to existing baselines.

Critical strengths

S1. Interesting Motivation and Approach

The paper addresses an important problem of adapting web agents to unseen interfaces without retraining. The proposed approach of combining interface exploration with adapter composition is intuitive and interesting.

S2. Clear Presentation

The paper is well-written, with a clear explanation of the proposed methodology and overall framework

Critical weaknesses

W1. Limited Practical Scalability and Comparison with General-Purpose Web Agents

While WebMix enables adaptation without target-side fine-tuning, it requires training and maintaining a separate LoRA adapter for each source interface. This raises concerns about scalability as the number of websites increases (e.g., training and storing over 1k lora adapters for each source website?). Also, although the paper compares against a unified LoRA baseline, it does not sufficiently investigate whether adapter composition offers practical advantages over stronger general-purpose web agents trained across diverse websites, such as those developed using VeriEnv or Weblica.

W2. Potential Target-Side Information Leakage and Unfair Baseline Comparisons

In the domain-transfer setting, the scout’s writer has access to reference procedures from target-domain training tasks, even though the final answers are removed. These procedures may provide substantial task-specific supervision, raising concerns about the claimed adaptation without target demonstrations. Moreover, the baselines do not appear to receive equivalent target-side information, making it difficult to disentangle the benefits of adapter composition from those of additional procedural knowledge.

Questions for authors
  1. Could the authors provide an ablation without target-domain reference procedures and compare against baselines using the same scout-generated manual?
  2. How does WebMix scale with more source interfaces, and what are its practical advantages over general-purpose web agents in terms of training, storage, and inference costs?
5 comments

We sincerely thank you for the thoughtful feedback and for recognizing our motivation, approach, and presentation. Your comments have helped us improve our work by analyzing scalability, practical advantages, and adding empirical results on procedure-free scouts.


W1/Q2. Scalability

Thank you for raising this! We have measured the parameter count, training, and inference costs for bank sizes K = 1–10, which goes beyond the previous maximum of five adapters (Table 5).

Bank size K13510
#Bank params21.2M63.7M106.2M212.3M
#Router params00.26M0.26M0.26M
#Total params21.2M63.9M106.4M212.6M
% of Qwen-3.5 4B backbone0.471.412.344.68
Training (H200 GPU-hours)1.75.28.717.4
Decode (tok/s)19.118.417.517.1

We have also added a new research question on scale and cost: What is the overhead in terms of parameters, compute, and inference as the adapter bank grows?

  • Parameters: The bank and router add at most 4.68% of the backbone's parameters.
  • Compute: Cost grows linearly with K, at about 1.7 GPU-hours per adapter.
  • Inference: Exact composition raises the rank from r to Kr, but the routed mixture is merged into the backbone before serving. WebMix therefore has the per-token cost of the base Qwen3.5 4B and needs 8.7–9.8 GiB of weights at every K. The only cost that depends on K is a one-time mixture build at most 7s per site.

W1. Practical advantage over general-purpose web agents

VeriEnv (Chae et al., 2026) and Weblica (Kar et al., 2026) widen the training distribution by cloning or synthesizing websites at scale. However, each new interface then requires retraining on the full corpus. WebMix instead adapts at test time to unfamiliar interfaces by composing existing adapters, with no target demonstrations or target-side fine-tuning. Only the small router (~0.26M) needs retraining, which is much cheaper computationally. The two directions are actually complementary. A stronger general-purpose agent could serve as WebMix's backbone.


W2/Q1. Domain scout using target reference procedures and potential bias

Thank you for the careful observation! We agree that the current scout setup can provide additional task-solving guidance from the domain procedures. To isolate its effect, we removed the procedures and kept only the description of the target domain. The procedure-free baseline (n=0) performs on par with the other scout baselines and leads with the highest average of 34.17% (Table 3). This indicates that the router relies on the domain description instead of the procedure. We have added this discussion in the related works section.

MethodWikiNewsShopAvg.
Scout, n = 0 (procedure-free)23.6657.5827.1634.17
Scout, n = 124.7356.0625.9333.75

We also compare against similar adapter composition baselines, such as uniform composition, MoLE, LoRAHub, and RRR (Table 4). WebMix, using the procedure-free setting, outperforms the other composition baselines.

MethodWikiNewsShopAvg
Uniform Composition20.4345.4517.2826.25
MoLE20.4351.5227.1631.25
LoRAHub23.6648.4828.4032.08
RRR22.5848.4825.9330.83
WebMix23.6657.5827.1634.17

We have added all the results and discussions to the revised manuscript and highlighted them in blue.

Thank you again for the review! Your feedback has helped us shape our work into a much stronger paper.

Thank you for the response. I appreciate the authors' efforts to address my concerns, and some of them have been resolved.

However, I still have two concerns regarding W1.

  1. The additional cost analysis helps clarify the computational overhaed. However, adding a new source interface still requires retraining the router. Although this is inexpensive, it introduces and additional training and maintenance step whenever the expansion happens.

  2. My original concern was about the practical advantage of WebMix over general purpose web agents. The authors explain they are complementary, but this does not directly address my request for an empirical comparison (or joint evaluation). I believe this point is crucial, because it directly relates to the motivation.

AuthorsOpenReview ↗

Reply by the Authors

Thank you for following up on our rebuttal. We are glad that some of your concerns have been resolved! We respond to the remaining points below.


1. Router training and maintenance

We agree that our current approach still requires training a router, and we appreciate that you raised this issue. To remove this requirement, we evaluate a training-free variant, WebMix (prompted). In this version, we prompt the frozen web agent to select the adapter coefficients based on the scout's description of the target domain. This fully bypasses the router training and would only require an adapter bank. We evaluate WebMix (prompted) on TimeWarp domain transfer using the Qwen-3.5 4B model. The results turned out to be surprisingly good.

MethodWikiNewsShopAvg
WebMix (Router)24.73%56.06%25.93%33.75%
WebMix (Prompted)22.58%60.61%33.33%36.67%

The prompted variant raises the average success rate from 33.75% to 36.67% (+2.92 points). It gains on News (+4.55) and Shop (+7.40) and drops slightly on Wiki (−2.15). We investigate this further by inspecting the weight coefficients produced by the trained router and the prompted agent.

LLM-prompted weights

Held-out \ ExpertWikiNewsShop
Wiki—0.890.11
News0.74—0.26
Shop0.580.42—

Trained router

Held-out \ ExpertWikiNewsShop
Wiki—0.770.23
News0.63—0.37
Shop0.410.59—

The weights are mostly consistent, showing that even without training, a prompted web agent can match the performance of a trained router. This is a very important finding, and we again thank the reviewer for this!


2. Empirical results against general-purpose agents

Thank you for clarifying this point. We compare WebMix, VeriEnv, and a joint training of WebMix with VeriEnv on TimeWarp domain transfer using Qwen-3.5 4B.

MethodWikiNewsShopAvg
VeriEnv21.51%43.94%27.16%29.58%
WebMix24.73%56.06%25.93%33.75%
VeriEnv + WebMix32.26%51.52%34.57%38.33%

On average, WebMix outperforms VeriEnv, but combining both gives the best overall results at +4.58 pp above the base WebMix. The gains are higher on Wiki and Shop, as the synthetic VeriEnv domains are highly similar to TimeWarp's Wiki and Shop domains. Based on the results, combining WebMix with synthetic, scalable approaches such as VeriEnv is a promising way to improve generalization.


Thank you again for following up on our rebuttal! Your feedback is helping us shape our work into a much stronger paper.

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

The paper introduces WEBMIX a mixture hypernetwork that adapts web agents to unseen interfaces by composing already trained low-rank adapters. The architecture consists of (1) an adapter bank that contains a set of domain- or version-specific LoRA adapters trained on source interfaces; (2) a scout that visits an unseen target interface and generates an embedding; (3) a router (hypernetwork) trained on source interfaces to map the embedding of the unseen interface to a set of coefficients to merge the adapters. Evaluation with two models on two benchmarks demonstrate that the proposed method outperforms zero-shot, ICL and LoRA.

Critical strengths
  • Learning to adapt to a new website without requiring model retraining is an important concept.
  • The appendix provides a mathematical formulation of the approach and a theoretical analysis.
  • Unlike other techniques, WEBMIX does not incur additional memory overhead or extensive computation during the inference phase.
Critical weaknesses
  • The paper compares the proposed method only against zero-shot, ICL and standard LoRA. More baselines focused on dynamic LoRA composition and routing (e.g., [1], [2], [3]) should be included for a more comprehensive comparison.
  • No ablation on the scout model is performed.
  • There is a missing discussion of relevant literature in related domains, such as meta-learning.
  • Every time a new source interface is added, the router needs to be retrained, which could become a bottleneck.
  • The paper lacks statistical significance testing for the results. Reporting standard deviations and performing a t-test would strengthen the claims.

[1] Wu, Xun, Shaohan Huang, and Furu Wei. "Mixture of lora experts." arXiv preprint arXiv:2404.13628 (2024).

[2] Celik, Bilge, Anna Vettoruzzo, and Joaquin Vanschoren. "Route, Reuse, Repurpose: Continual Adaptation of LLMs with Bounded Adapter Pools." Continual Adaptation at Scale: Towards Sustainable AI Workshop.

[3] Huang, Chengsong, et al. "Lorahub: Efficient cross-task generalization via dynamic lora composition." arXiv preprint arXiv:2307.13269 (2023).

Questions for authors
  • How robust is the method when the number of source interfaces increases?
  • What happens if the unseen interface is very different than the source interfaces ? Is it possible to incorporate new source interfaces on-the-fly? For example, the method could have a mechanism to detect a novel interface and automatically update or expand the pool of source interfaces.
  • Is WEBMIX restricted to LoRAs or can it be extended also to different adapters?
  • Can the proposed approach be extended to different domains than web?
  • How did the authors verify that an unseen interface was unobserved by the model during its training phase?
  • Why SCOUT n=1 gives the best results in Table 3? My assumption is that more tasks the router sees and more information it has.
  • Would it be possible to compute a distance metric between pairs of interface embeddings and correlate it directly with the learned coefficients assigned to each LoRA?
4 comments
AuthorsOpenReview ↗

Rebuttal by the Authors (1/2)

We sincerely thank you for the thoughtful feedback and for recognizing the concept of our work, the formalization and theoretical analysis, and the efficiency of our method. Your comments have helped us strengthen the paper by adding empirical results on procedure-free scouting and composition baselines, discussion of related literature, a cost and scaling analysis, statistical significance of the results, extension to other types of adapters, and correlation between the learned coefficients and the distance between interface embeddings.


W1. Additional composition and routing baselines

Thank you for suggesting this! We added four router and composition baselines in Table 4: a uniform composition baseline that averages over the adapter bank weights, MoLE, LoRAHub, and RRR. WebMix, using the procedure-free setting, outperforms the other composition baselines.

MethodWikiNewsShopAvg
Uniform Composition20.4345.4517.2826.25
MoLE20.4351.5227.1631.25
LoRAHub23.6648.4828.4032.08
RRR22.5848.4825.9330.83
WebMix23.6657.5827.1634.17

W2 Scout model ablation

We clarify that Table 3 includes this ablation. The mixture row is the WebMix hypernetwork without the scout. Adding the scout, without the target domain procedure (n=0), improves every domain and raises the average by 5.0 percentage points.

MethodWikiNewsShopAvg
Mixture (no scout)19.3548.4824.6929.17
+ Scout (n=0)23.6657.5827.1634.17

W3. Related literature

We appreciate your feedback on the related works! We reorganized the related work section, added a paragraph on meta-learning, and positioned WebMix against other contemporary works. WebMix draws on meta-learning methods, such as meta-RL, to gather context by exploring the target interface. It also adapts in a forward pass similar to amortized meta-learning methods. However, unlike either, it generates no new weights. It predicts one mixing coefficient per bank adapter, so the output is a single standard LoRA that any unmodified inference stack can serve.


W4/Q1. Router retraining, robustness, and scaling cost

We thank you for this suggestion! Adding a source interface appends one adapter to the bank and retrains only the router. The router has 0.26M parameters and retrains in under 0.1 H200 GPU-hours, which we don't think will be a bottleneck for compute. Training each source LoRA adapter for the bank requires 1.7 GPU-hours.

We have also added a new research question on scale and cost: What is the overhead in terms of parameters, compute, and inference as the adapter bank grows? We measured the parameter count, training, and inference costs for bank sizes K = 1–10, which goes beyond the previous maximum of five adapters (Table 5).

Bank size K13510
#Bank params21.2M63.7M106.2M212.3M
#Router params00.26M0.26M0.26M
#Total params21.2M63.9M106.4M212.6M
% of Qwen-3.5 4B backbone0.471.412.344.68
Training (H200 GPU-hours)1.75.28.717.4
Decode (tok/s)19.118.417.517.1
  • Parameters: The bank and router add at most 4.68% of the backbone's parameters.
  • Compute: Cost grows linearly with K, at about 1.7 GPU-hours per adapter.
  • Inference: Exact composition raises the rank from r to Kr, but the routed mixture is merged into the backbone before serving. WebMix therefore has the per-token cost of the base Qwen-3.5 4B and needs 8.7–9.8 GiB of weights at every K. The only cost that depends on K is a one-time mixture build at most 7s per site.

W5. Statistical significance

We agree that statistical significance can strengthen our claims, and have added the standard errors for the averages in Tables 1 and 2. We summarize the gains in percentage points over the second-best baseline with standard errors for the unseen interfaces. The gains on unseen domains are statistically significant (>2 SE).

SplitQwen-3.5 4BLlama-3.1 8B
TimeWarp Domains+8.3 ± 3.1+7.5 ± 2.4
TimeWarp Versions+3.2 ± 1.8+1.8 ± 1.6
WebArena Domains+5.2 ± 2.2+4.6 ± 1.7
AuthorsOpenReview ↗

Rebuttal by the Authors (2/2)

Q2. Integrating novel interfaces

We think it's a really interesting direction! We can extend WebMix to detect and integrate interfaces that vary largely from the source.

  • For detection, we can use (a) a lightweight classification head on the router to categorize the sample as novel, and (b) a gate on the entropy of the router coefficients, i.e., if coefficients are near uniform, it means no current adapter fits.
  • The novel trajectories can be collected and stored in a replay buffer that can be later used to train new adapters on the fly.
  • To prevent the bank from having too many adapters, we can merge adapters based on the similarity between weight matrices and coefficient patterns.

Q3. Non-LoRA adapters

Yes, WebMix can be extended to non-LoRA adapters, as it only requires adapters that can be combined with scalar coefficients. We evaluate WebMix using QLoRA, DoRA, and (IA)³ with Qwen-3.5 4B on TimeWarp domains, and found that the empirical results for LoRA were the strongest (Fig. 4).

MethodSuccess Rate (%)
LoRA33.75
DoRA30.42
QLoRA31.25
(IA)³30.83

Q4. Extension beyond the web

Yes, WebMix can be applied wherever an agent must act on an interface it was not trained on. For computer use (e.g., OSWorld), one would train source adapters on applications such as VLC and GIMP and evaluate on a held-out one such as LibreOffice Calc. Mobile use (e.g., AndroidWorld) would allow the same split over mobile apps.


Q5. Ensuring unseen interfaces are unobserved

Thank you for raising this! We clarify our experimental setting for the leave-one-out protocol, in which only tasks from the source domain are provided during training, and only the source domain remains active. E.g., for the Wiki target, both the source adapters and the router are trained only on News and Shop tasks. No Wiki tasks, trajectories, or pages are used in training. Wiki is accessed only at inference, through the scout, with no parameter updates.


Q6. Why does n=1 perform best?

During rebuttal, we found that a procedure-free scout (n=0) can slightly outperform a scout using a single target procedure (n=1) (Table 3). This indicates that the router relies on the domain description instead of the procedure. The scout’s role is to characterize the interface (layout, navigation structure), which a single observation largely captures. Further trajectories mainly add task-specific content that says little about which source adapters transfer.

MethodWikiNewsShopAvg.
Scout, n = 0 (procedure-free)23.6657.5827.1634.17
Scout, n = 124.7356.0625.9333.75

Q7. Embedding distance vs. coefficients

Thank you for suggesting this interesting analysis! We added this in Fig. 7, where for each held-out interface u and source vkv_k, we compute the Spearman correlation ρ between the cosine similarity of their encoded manuals (the router’s input) and the router weight cu,kc_{u,k}. Using the distance 1 − cos instead only flips the sign.

  • Domain transfer: the router gives more weight to more similar sources. On TimeWarp domains, ρ = 0.72 for Qwen-3.5 and 0.78 for Llama-3.1, and the most similar source gets the largest weight in 8 of 9 runs. On WebArena-Lite-v2, ρ = 0.52 and 0.62.
  • Not solo transfer: the weights track how well each source adapter transfers on its own only weakly or not at all (ρ = −0.16 and 0.44 on TimeWarp domains). The router is following interface similarity, not picking the adapter that would transfer best alone.
  • TimeWarp versions: the manuals are nearly identical (cosine 0.969–0.978), so similarity carries little signal, and the correlation doesn’t hold (ρ = −0.48 and −0.22).

The results and discussion have been added to the revised manuscript and highlighted in blue.

Thank you again for the review! Your feedback has helped us shape our work into a much stronger paper.

Thank you for all clarifications and additional experiments. I am satisfied with your answers and I increased my score.

AuthorsOpenReview ↗

Reply by the Authors

Thank you very much for reassessing our work! We really appreciate your feedback from the discussion phase.

Reviewer_G7KcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

WebMix adapts a web agent to a new interface by combining adapters trained on familiar interfaces. A scout explores the new site and writes a manual, which a small router uses to choose the mixture. On TimeWarp and WebArena-Lite-v2, it improves transfer over the tested prompting and LoRA baselines without updating parameters on the target interface.

Critical strengths
  • Predicting a few mixture weights gives a practical way to reuse trained adapters. The method combines their actual weight updates, avoiding the ambiguity of averaging adapter factors separately.
  • Experiments test both unseen websites and unseen interface versions with two open models. They also report performance on familiar interfaces.
  • Ablations examine direct weight generation, scouting, exploration budget, and adapter regularization, making the design choices easier to assess.
Critical weaknesses
  1. Domain scouting uses target reference procedures. The writer receives the attempted task's intent and reference procedure with the final answer removed. Although there are no target gradient updates, this supplies task-solving guidance beyond ordinary exploration; a procedure-free comparison is needed to support the claimed adaptation setting.
  2. Simple composition baselines are missing. Appendix D.1 explicitly says the experiments omit the best transferred bank adapter and the uniform mixture. Without these comparisons, it is unclear how much gain comes from the learned router rather than access to the adapter bank.
  3. Matching target-trained specialists is not general across benchmarks. On unseen WebArena-Lite-v2 domains, WebMix reaches 33.77% versus 37.23% for target-trained LoRA with Qwen, and 15.37% versus 22.29% with Llama. The matching result is stronger on TimeWarp, and should be scoped accordingly.
  4. The scaling cost needs measurement. Exact composition increases the adapter rank from r to Kr, and adding a source interface requires retraining the router. Tests use banks of at most five adapters; larger-bank latency, memory, and training measurements are needed to assess the approach at deployment scale.
Questions for authors

see weakness

2 comments
AuthorsOpenReview ↗

Rebuttal by the Authors

We sincerely thank you for the thoughtful feedback and for recognizing the practicality of our method and the design choices behind the experiments. Your comments have helped us strengthen our work by adding empirical results on procedure-free scouting and composition baselines, discussing the benchmark results, and analyzing the scaling cost.


W1. Domain scouting uses target reference procedures

Thank you for the careful observation! We agree that the current scout setup can provide additional task-solving guidance from the domain procedures. To isolate its effect, we removed the procedures and kept only the description of the target domain. The procedure-free baseline (n=0) performs on par with the other scout baselines and leads with the highest average of 34.17% (Table 3). This indicates that the router relies on the domain description instead of the procedure.

MethodWikiNewsShopAvg.
Scout, n = 0 (procedure-free)23.6657.5827.1634.17
Scout, n = 124.7356.0625.9333.75

W2. Composition baselines

Thank you for suggesting this. We added a uniform composition baseline that averages over the adapter bank weights (Table 4). WebMix outperforms the composition baseline and similar adapter-composition methods, such as MoLE, LoRAHub, and RRR.

MethodWikiNewsShopAvg
Uniform Composition20.4345.4517.2826.25
MoLE20.4351.5227.1631.25
LoRAHub23.6648.4828.4032.08
RRR22.5848.4825.9330.83
WebMix23.6657.5827.1634.17

W3. Results are not general across benchmarks

We agree that the gains are stronger in TimeWarp, with WebMix leading over the second-best baseline by 8 pp on average across unseen domains. On WebArena-Lite-v2, WebMix still clearly outperforms methods that do not train on the target domain. On unseen domains, it gains +5.2pp (Qwen3.5 4B) and +4.5pp (Llama-3.1 8B) over the second-best baseline, well above the standard error of 1.7–2.2. It also leads on every domain, by up to 10.2pp. If any other wording seems to overstate the results, we will be glad to change it.


W4. Scaling Cost

We really appreciate this feedback! We have measured the parameter count, training, and inference costs for bank sizes K = 1–10, which goes beyond the previous maximum of five adapters (Table 5).

Bank size K13510
#Bank params21.2M63.7M106.2M212.3M
#Router params00.26M0.26M0.26M
#Total params21.2M63.9M106.4M212.6M
% of Qwen-3.5 4B backbone0.471.412.344.68
Training (H200 GPU-hours)1.75.28.717.4
Decode (tok/s)19.118.417.517.1

We have also added a new research question on scale and cost: What is the overhead in terms of parameters, compute, and inference as the adapter bank grows?

  • Parameters: The bank and router add at most 4.68% of the backbone's parameters.
  • Compute: Cost grows linearly with K, at about 1.7 GPU-hours per adapter.
  • Inference: Exact composition raises the rank from r to Kr, but the routed mixture is merged into the backbone before serving. WebMix therefore has the per-token cost of the base Qwen3.5 4B and needs 8.7–9.8 GiB of weights at every K. The only cost that depends on K is a one-time mixture build at most 7s per site.

We have added all the results and discussions to the revised manuscript and highlighted them in blue.

Thank you again for the review! Your feedback has helped us shape our work into a much stronger paper.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer,

We thank you again for your feedback on our work! Your suggestions were very valuable in improving our submission. As we are near the end of the discussion period, we would greatly appreciate your feedback on our rebuttal.

We tried our best to address your concerns on (W1) domain scouting, (W2) composition baselines, (W3) result variations across benchmarks, and (W4) scaling cost.

We would be really grateful to know whether these address your concerns. We are more than happy to address any remaining questions or run further experiments if that would be helpful!

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of reviewers' feedback and authors response

Dear Reviewers and Area Chairs,

We sincerely thank all the reviewers for their constructive feedback! We are grateful that the reviewers found several strengths in our submission, which affirm the value of our research. The comments have been invaluable in helping us improve our manuscript.


Strengths Highlighted by Reviewers

  • Important problem: adapting web agents to unseen interfaces without retraining (ygAv, zc2a).
  • Intuitive and practical method: combining interface exploration with adapter composition is intuitive (zc2a). Predicting a few mixture weights is a practical way to reuse trained adapters, and composing their actual weight updates avoids the ambiguity of averaging LoRA factors (G7Kc).
  • Efficient inference with no additional memory overhead or extensive computation (ygAv).
  • Evaluation on unseen websites and unseen versions with two open models, while also reporting performance on seen interfaces (G7Kc).
  • Ablations on direct weight generation, scouting, exploration budget, and adapter regularization (G7Kc).
  • Formalization and theoretical analysis in the appendix (ygAv).
  • Clear presentation (zc2a).

Key Changes in the Revision

  1. Procedure-free scouting (Tab. 3): With only a description of the target domain and no reference procedures, the scout gives the best average (34.17% vs. 33.75% with one procedure).
  2. Composition baselines (Tab. 4): WebMix outperforms uniform composition, MoLE, LoRAHub, and RRR (33.75% vs. 32.08% for the strongest, few-shot LoRAHub).
  3. Training-free variant (Sec. D.2): Prompting the frozen web agent to choose the mixing weights removes router training and raises the average to 36.67%, better than other composition methods.
  4. Comparison with general-purpose training (Sec. D.3, Tab. 7): WebMix outperforms VeriEnv (33.75% vs. 29.58%), and joint training VeriEnv with WebMix performs best (38.33%).
  5. Bank size and cost (Tab. 5): A bank of 10 adapters adds at most 4.68% of the backbone's parameters, and training cost grows linearly at ~1.7 GPU-hours per adapter.
  6. Statistical significance (Tabs. 1, 2): Over three seeds, the gains on unseen domains exceed two standard errors for both models on both benchmarks (e.g., +8.3 ± 3.1 pp on TimeWarp with Qwen-3.5 4B).
  7. Router analysis (Fig. 7, Sec. D.1): The router gives more weight to sources whose manuals are more similar (Spearman ρ = 0.72–0.78 on TimeWarp domains, 0.52–0.62 on WebArena-Lite-v2).
  8. Other adapter types (Fig. 4): WebMix also works with DoRA, QLoRA, and (IA)³, though LoRA remains the strongest.
  9. Related work and clarifications: a new meta-learning paragraph (Sec. 2), the leave-one-out protocol (App. B.6), and extensions beyond the web (App. A).

We have already added the changes to the revised manuscript and highlighted them in blue. We are encouraged that Reviewers ygAv and zc2a found their concerns addressed and raised their scores.

We again thank all the reviewers for their critiques of our work. Their feedback led directly to these revisions, and we believe our paper is much stronger as a result. We are more than happy to clarify anything further!

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.