acceptodds
Under review as a conference paper at ICLR 2027

Learned User Embeddings from Model Context Protocol Tool Interactions

Abstract

Model Context Protocol (MCP) is a mechanism that allows Small and Large Language Models to call external data sources in real-time and generate up-to-date and grounded responses for user prompts. MCP servers provide a discrete set of tools which are selected based on reasoning and semantic similarity to the user’s prompt. Due to its discrete nature, use of MCP tools give a unique opportunity to understand user behavior, allowing a Language Model to build personalized experiences such as initial prompt suggestions and customized responses. In this paper, we show that dimension reduction techniques can be applied to build learned representations in the form of fixed size dense embeddings in latent space to represent MCP tool invocations. We apply three different dimensionality reduction techniques on normalized tool invocation counts: Principal Component Analysis (PCA), shared auto-encoder and truncated singular value decomposition. Using a formal approach to the problem we use an experimental data set and show how the learned user embeddings determine similarity between users providing a foundation for personalization related features.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Machine learning, personalization, model context protocol, language models

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_eXz6OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This manuscript proposes a path towards personalization in MCP-based tool calling. The authors formulate the problem as a ranking task where a user may have preference priors for some functions over others. To analyze this, they construct a sparse user–tool matrix that is factorized to produce low-dimensional embeddings used for retrieval or preference alignment.Three dimensionality-reduction approaches are presented: a shared autoencoder with a weighted reconstruction loss (Algorithm 2), truncated SVD with user factors UdΣdU_d\Sigma_d (Algorithm 3), plus standardized PCA and raw tool-count vectors as baselines.

The embeddings are meant to support user similarity (cosine/Euclidean), cohort discovery via k-means, and downstream "personalization features" such as prompt suggestions and customized responses — though no concrete mechanism for feeding the embeddings back into tool ranking is specified.

The authors design a new dataset for this purpose and conduct a straightforward analysis on “canary” users, split into two groups biased toward having centralized and sparse embedding-based clusters, respectively.

Critical strengths

1\ Timely and well-motivated problem framing. Recasting MCP tool invocations as RecSys via user–item matrix where "items" are tools is a novel and sound exploration with privacy-first design at hand.

2\ Reproducibility artifacts: The authors provide a public synthetic dataset, generator code, and the canary-based evaluation protocol that allow readers to reproduce the experiments failrly easily.

Critical weaknesses

While the authors provide an interesting problem setting, there are several key challenges with the paper itself:

1\ It is not entirely clear how the authors intend for these new embeddings to be used in tool-calling ranking.

2\ The technical novelty of the paper is fairly limited to the application domain.

3\ An interesting and potentially more powerful baseline would be to use the mean or median of the embeddings from the textual descriptions of the tools called by a user's LLM.

4\ The rich semantic metadata that MCP tools carry (descriptions, input schemas) is overlooked in favor of counts alone, even though the paper's own Section 1.1 states that tool selection is driven by semantic similarity over descriptions.

5\ It is not entirely clear why the cold-start problem would be a major bottleneck here, since tool calling is naturally a semantically rich process: textual descriptions of both the task and the functionality of each tool are available at all times.

Questions for authors

Please review points in Critical Weaknesses.

Reviewer_ZKm4OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes the idea to personalize LLM assistants using MCP tool-call logs. Each user is represented by how often they call each MCP tool. That count vector is compressed into a small, fixed-size embedding, and distances between embeddings are meant to support personalization: finding similar users, cohorts and prompt suggestions. The authors say this protects privacy because only tool identities are used, never the tool inputs or the prompts. Main contribution: A formal user–tool interaction framework, an embedding-generation pipeline using established dimensionality-reduction methods, and a shared synthetic dataset and experimental code. The main contribution is applying these methods to MCP usage as an implicit personalization signal, rather than introducing a new learning algorithm.

Critical strengths
  1. The paper identifies MCP tool-invocation counts as a structured source of user information and provides a straightforward pipeline for converting them into compact embeddings without processing prompt text or tool arguments. This is consequential because it offers a practical starting point for user modeling in tool-enabled assistants, although downstream personalization benefits remain untested.

  2. Useful scrutiny of misleading clustering metrics. The analysis shows that SVD’s very high silhouette scores arise from isolating a few outliers, while the autoencoder produces less degenerate clusters. Examining cluster sizes alongside aggregate scores strengthens the empirical interpretation and provides a meaningful caution for choosing representations for cohort discovery.

Critical weaknesses

All experiments use synthetic logs, with similarity checked primarily through planted identical or slightly perturbed histories. Mapping identical inputs to identical embeddings is expected for a shared deterministic encoder; nearest-neighbor lists and clustering scores provide no independent evidence that the representations recover meaningful preferences. Without real-world validation or a downstream personalization task, the paper’s central practical benefit remains unsubstantiated.

The methodological contribution is limited, and the comparisons do not establish an advantage over standard methods. The approach applies established autoencoder and truncated-SVD techniques to a user–tool count matrix. Demonstrating the value of this application therefore depends heavily on convincing empirical comparisons. However, preprocessing differs across methods, and reconstruction losses compare weighted MSE for the autoencoder against unweighted MSE for SVD. These differences make the reported results insufficient to establish which method produces better representations or why the proposed pipeline should be preferred.

Questions for authors

Does the synthetic generator encode persistent user-specific preferences beyond the canary groups? If so, how, and do the embeddings recover those preferences? This would determine whether the experiments demonstrate meaningful user modeling beyond preserving identical or near-identical inputs.

Is there existing held-out or downstream evidence that the embeddings improve prediction or personalization over raw counts and PCA? If so, please report it; this could materially strengthen the paper’s central claim.

1 comment
AuthorsOpenReview ↗

Any remaining questions ?

Dear Reviewer ZKm4, thank you again for your thoughtful review. As the discussion period progresses, we wanted to check whether our response and addresses your concerns, in particular your concerns on synthetic data and downstream evidence?

Reviewer_rdp2OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a method for learning user embeddings from Model Context Protocol (MCP) tool invocation records by constructing a user–tool interaction matrix based on tool-call frequencies, applying normalization, and employing several dimensionality-reduction techniques, including PCA, a shared Autoencoder, and Truncated SVD, to derive low-dimensional user representations. To evaluate whether the learned embeddings preserve user similarities, the authors introduce canary users with artificially constructed interaction histories. The results indicate that the Autoencoder produces more balanced clustering outcomes, while SVD achieves lower reconstruction errors but suffers from degenerate clustering behavior. Overall, the paper explores an interesting direction by leveraging MCP tool interactions as behavioral signals for user representation. However, the proposed approach relies heavily on conventional dimensionality-reduction techniques and synthetic data, and, more importantly, fails to demonstrate whether the learned embeddings can effectively benefit downstream LLM personalization tasks.

Critical strengths
  1. Comparison of multiple embedding methods. The paper evaluates several dimensionality-reduction techniques, including PCA, Autoencoder, and Truncated SVD. It also discusses the impact of different methods on clustering behavior, providing some insights into the trade-offs between reconstruction quality and embedding distribution.
Critical weaknesses
  1. The novelty is insufficient. The proposed approach mainly applies conventional dimensionality-reduction methods, such as autoencoders and SVD, to a user–tool interaction matrix. This is closely related to traditional collaborative filtering and matrix factorization approaches. It seems that the paper does not introduce a substantially new learning objective, model architecture, or optimization strategy. More importantly, it does not sufficiently explain what unique challenges arise from MCP-based user representation and why existing user modeling or collaborative filtering methods are inadequate for this setting. Consequently, the methodological contribution appears limited.

  2. The paper has significant structural and presentation issues. The writing and presentation need improvement; for example, the definition of the problem model and the articulation of the contributions are not sufficiently professional. Moreover, more than half of the main content in the main text relies on references to the appendix (including the main pseudocode and conclusions, most of the main experimental results, etc.), which makes the main text very inadequate. Under the submission requirements, reviewers are not required to read the appendix. Authors should ensure that the main paper contains sufficient details and key experimental evidence to support its core claims, and should use supplementary material for additional analyses, supplementary experiments, and implementation details.

Questions for authors
  1. What is the fundamental methodological novelty of the proposed approach compared with traditional collaborative filtering and matrix factorization methods?

  2. The paper lacks analysis of the experimental results. Is there any further theoretical analysis or explanation for the experimental results and the methods presented in the paper? For example, why does truncated SVD achieve lower reconstruction error but produce highly degenerate clustering results, while the nonlinear autoencoder yields more balanced clusters?

Comments on the paper

1 comment
Area_Chair_QAPtOpenReview ↗

Discussion

Dear All,

Thank you for reviewing the paper. Please read the authors' response and leave a comment accordingly.

Yours,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.