acceptodds
Under review as a conference paper at ICLR 2027

UserToolBench: A User-Profile-Hidden Benchmark for Personalized Decision Making in Tool-Use LLMs

Abstract

Tool-use LLMs are increasingly asked to act on users' behalf, but existing benchmarks usually focus on profile recall, style imitation, generic tool use, or response-level personalization. We introduce UserToolBench , a benchmark for personalized decision making in tool-use LLMs. UserToolBench tests whether a model can infer latent user preferences from interaction history, recognize when clarification is needed, and produce user-aligned tool-call trajectories under incomplete information. The benchmark is built from privacy-sanitized real interaction traces and combines structured persona profiles, public API-style tool ecosystems, and long-horizon multi-turn trajectories. It includes 10 user profiles, 36 tool sets, 1,065 turns, 170 unique tools, and evaluation-focused task types covering lack-of-information, single-tool, and multi-tool settings. Experiments with strong tool-use LLMs show that current models still have difficulty with personalized delegation. Multi-tool coordination, missing-constraint inference, and long-horizon behavioral consistency remain major bottlenecks. These results suggest that personalization evaluation should move beyond asking whether outputs sound user-specific and instead ask whether LLMs make correct decisions for the users they represent.

Primary area:
datasets and benchmarks
Keywords:
personalized LLMs, tool-use LLMs, LLM agents, user preference inference, personalized decision making, tool-use benchmark, long-horizon interaction, profile-hidden evaluation

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_6fHUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces UserToolBench, a benchmark for evaluating personalized decision-making in tool-use LLMs. It tests whether models can infer user preferences from interaction history and make appropriate tool calls when the current request lacks important information. The benchmark includes 10 user profiles, 1,065 dialogue turns, and 170 tools. Experiments on nine LLMs show that current models still struggle with personalized tool use, particularly in multi-tool tasks and situations with missing information.

Critical strengths
  1. The paper studies a meaningful problem for personalized AI assistants. The profile-hidden evaluation setting is interesting, as it requires models to infer user preferences from interaction history rather than directly accessing a user profile.

  2. The authors evaluate a range of LLMs and consider different task types. Reporting both exact and relaxed accuracy also helps distinguish strict trajectory matching from task completion.

Critical weaknesses
  1. Both the user simulator and the reference trajectory generator have access to the full user profile. I am concerned about two possible problems. The simulator might reveal the user's preferences too explicitly, making the task too easy. On the other hand, the generated history might not provide enough information to infer the preferences used in the reference trajectory, or several interpretations might be equally reasonable. It is unclear how the authors avoid these cases.

  2. Although LLM-based user simulation is common, I am not fully convinced that these generated conversations reflect how real users interact with assistants. In particular, conditioning the simulator on a fixed profile may produce overly consistent user behavior and unnatural expressions of preferences.

  3. Real users do not always have stable preferences. Their preferences may change over time or depend on the situation. The authors consider this in Appendix D.3, but only through a preliminary experiment. I think the main benchmark still makes a rather strong assumption about preference stability.

  4. Table 1 describes UserToolBench as a long-horizon benchmark. However, the dataset contains 1,065 dialogue turns across 300 topics, averaging only 3.55 turns per topic. I understand that this does not necessarily reflect the full history length, but the paper provides little information about how long these histories actually are or how far back the relevant preferences appear. This makes the long-horizon claim less convincing.

  5. Appendix E.1 describes a human verification interface, but the paper lacks details about the annotation process. For example, how many annotators were involved? Were samples independently reviewed by multiple annotators? What was the inter-annotator agreement? Given the reliance on synthetic data, I would like to see more evidence of the dataset's quality.

  6. I did not find an experiment where the evaluated LLM is given the complete user profile. Without this baseline, it is difficult to tell whether models fail because they cannot infer user preferences from the interaction history, or because they cannot make appropriate tool calls even when those preferences are known.

Questions for authors
  1. How do the authors prevent the user simulator from revealing preferences too explicitly? Have they checked whether the reference decisions can actually be inferred from the interaction history, especially when multiple interpretations are possible?

  2. Have the authors compared the simulated conversations with real user interactions to evaluate how natural and realistic they are?

  3. Could the authors discuss how the benchmark handles changes in user preferences, beyond the preliminary experiment in Appendix D.3?

  4. Could the authors report the distribution of history lengths, in both turns and tokens? It would also be useful to know how far back the relevant preference information appears.

  5. Could the authors provide more details about the human annotation process, including the number of annotators, inter-annotator agreement, and sample acceptance rates?

  6. Have the authors considered an oracle-profile baseline where the model directly receives the user profile? I think this comparison would help clarify whether the main difficulty comes from preference inference or tool-use planning.

Reviewer_GKbPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UserToolBench introduces a novel evaluation benchmark for tool-use large language models (Tool-use LLMs), designed to assess models’ personalized decision-making capabilities. Models are required to infer users’ latent preferences from historical interactions, determine when to seek clarification from users, and generate tool-call trajectories aligned with user preferences under incomplete information. The benchmark dataset is constructed from privacy-desensitized real interaction logs.

Critical strengths
  • The paper proposes a testing paradigm with hidden user profiles, which exhibits some novelty.
Critical weaknesses
  • Given that user profiles are available, the standard practice would be to feed them into the system prompt and retain them across the context. Although the design of removing user profiles is an intentional choice by the authors, it appears inconsistent with real-world scenarios. The authors need to justify the necessity of this setting.
  • The tools described in the paper do not seem executable, which is somewhat outdated by current standards.
  • The evaluation methodology is dated. Most contemporary work adopts LLM-based consistency and rubric-based evaluation; rule matching tends to be overly rigid and lacks generalization.
  • No detailed quantitative metrics for data quality (accuracy) are provided, raising concerns about dataset quality.
Questions for authors

Same as weakness

Reviewer_q51YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UserToolBench evaluates personalized tool-use decisions when an explicit user profile is hidden from the evaluated model but available during reference construction. The model receives interaction history, a current request, and tool schemas. The benchmark comprises 10 profiles, 36 toolsets, 170 tools, 1,065 dialogue turns, and 799 evaluation tasks. Nine models are evaluated using exact trajectory accuracy and relaxed task completion, with additional error diagnostics and a preliminary dynamic-preference split.

Critical strengths
  1. The paper addresses an important problem: whether assistants can translate historical user preferences into executable decisions, including tool arguments and clarification behavior. Combining these requirements in a profile-hidden setting offers a potentially useful benchmark.
  2. The evaluation goes beyond a single exact-match leaderboard. Relaxed completion, error categories, severity diagnostics, and a dynamic-preference pilot are useful additions. The paper explicitly acknowledges the limitations of a single reference trajectory, and I appreciate these efforts to qualify its conclusions.
Critical weaknesses
  1. The recoverability of reference decisions from the model-visible history is insufficiently established. Section 3.2 states that missing constraints may be recovered from either the profile or preceding milestones, while the evaluated model cannot access the profile. A profile-consistent reference is not necessarily inferable from the visible history. Although the metric definition and human-verification checklist discuss clarification, the paper does not report an evidence-grounded audit showing that required preference-dependent arguments are supported by preceding history, or that clarification is accepted when they are not. This leaves uncertainty about whether some failures reflect unavailable information rather than model limitations. I am not asserting that such labeling errors are present; the concern is that this central validity condition has not yet been demonstrated.
  2. The interpretation of the results as personalized decision failures is not sufficiently separated from general tool-use errors and reference-trajectory bias. Relaxed scoring changes allowable tool orderings as well as task-completion criteria, so its gap from exact accuracy does not uniquely quantify preference misalignment. Appendix D is helpful, but the annotation procedure, denominators, agreement, and treatment of valid alternative trajectories need clarification. The qualitative URL example also illustrates general grounding and planning failure, rather than independently demonstrating a preference-specific failure. A focused preference-sensitive control would substantially strengthen the benchmark's construct validity.
Questions for authors
  1. Can you verify that preference-dependent reference decisions are supported by the visible history, and that clarification is accepted when evidence is insufficient?
  2. Can you clarify relaxed scoring and provide a small audit or controlled comparison separating preference violations from valid alternatives and general tool-use errors?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.