PaperlayerView exact revision on arXiv

Source-linked AI summary

Latent Preference Modeling for Cross-Session Personalized Tool Calling

Yejin Yoon, Minseo Kim, Taeuk Kim

arXiv:2604.17886v1cs.CLcs.AI

TL;DR

Under-specified tool requests require agents to infer missing API arguments from users’ multi-session behavior. The paper introduces MPT to benchmark recall, induction, and transfer, and proposes PREFINE, which refines latent preferences through generate–verify–refine memory. PREFINE improves tool-calling accuracy while using 1.24% of the tokens required by full-history prompting.

  • Problem

    Personalized tool calling must infer missing API arguments from implicit user preferences distributed across noisy, unordered multi-session histories.

  • Method

    MPT benchmarks three preference-modeling challenges, while PREFINE incrementally refines latent preferences into reusable action-level constraints through a generate–verify–refine loop.

  • Results

    PREFINE improves tool-calling accuracy using only 1.24% of the tokens required by full-history prompting.

  • Takeaways & Limitations

    Robust personalization depends on capturing the reasons behind user choices rather than only retrieving the choices themselves.

  • Takeaways & Limitations

    PREFINE can become overly conservative, pruning arguments that are required and trading recall for precision.

Abstract

from arXiv · show

Users often omit essential details in their requests to LLM-based agents, resulting in under-specified inputs for tool use. This poses a fundamental challenge for tool-augmented agents, as API execution typically requires complete arguments, highlighting the need for personalized tool calling. To study this problem, we introduce MPT, a benchmark comprising 265 multi-session dialogues that cover three challenges: Preference Recall, Preference Induction, and Preference Transfer. We also propose PRefine, a test-time memory-augmented method that represents user preferences as evolving hypotheses. Through a generate--verify--refine loop, it extracts reusable constraints from history and improves tool-calling accuracy while using only 1.24% of the tokens required by full-history prompting. These results indicate that robust personalization in agentic systems depends on memory that captures the reasons behind user choices, not just the choices themselves.

1 Introduction

Personalized tool calling addresses under-specified requests by inferring persistent user constraints from multi-session behavior rather than merely retrieving similar actions. MPT evaluates this capability across recall, induction, and transfer, while PREFINE improves accuracy with compact memory.

  • Users often omit essential details, leaving tool calls with incomplete arguments that agents must infer from prior behavior.
  • Latent preferences are persistent, implicit constraints derived from recurring behavioral patterns across contexts.
  • MPT evaluates personalized tool calling through Preference Recall, Preference Induction, and Preference Transfer under intentionally under-specified API arguments.
  • Models that reuse prior decisions effectively for Preference Recall struggle when preferences must be induced across sessions or transferred across domains.
  • 1.24% of full-history prompting tokens are sufficient for PREFINE to improve tool-calling accuracy through test-time memory refinement.

2 Related Work

Prior work studies tool use from current dialogue states, latent preference representations, and long-horizon memory, but this paper focuses on externalizing latent preferences as reusable API constraints. This distinguishes its setting from preference scoring and explicit-memory approaches.

  • Many tool-use benchmarks evaluate planning and execution from the current dialogue state while ignoring prior interactions.
  • Latent-preference research derives user-specific representations from interaction histories to generalize beyond observed choices.
  • This paper externalizes latent preferences as reusable textual constraints that govern unspecified API arguments.
  • Agentic-memory research addresses maintaining, retrieving, and updating information over long horizons, including multi-session collaboration and memory compression.

3 Problem Definition

The task is to generate a valid API call that satisfies explicit query constraints and fills unspecified arguments using latent preferences inferred from prior sessions and API actions. Its difficulty depends on how evidence is distributed across history.

  • The agent selects a valid API call from schema-bounded actions using the current query and accumulated history of sessions and API calls.
  • Schema-aligned preference reasoning identifies persistent argument-level constraints from recurring behavioral patterns rather than performing open-ended preference discovery.
  • Preference Recall: Preference Recall resolves missing arguments by retrieving recurring same-domain argument–value choices.
  • Preference Induction: Preference Induction aggregates evidence across tasks and domains before instantiating an inferred latent preference as concrete argument values.
  • Preference Transfer: Preference Transfer applies a latent preference from other domains when the target domain lacks in-domain evidence.

4 Dataset Construction: MPT

MPT is built by grouping SGD sessions, annotating shared behavioral constraints, and constructing queries that omit preference-sensitive arguments. It supports context-guided and context-free evaluation across three preference-modeling challenges.

  • MPT pairs multi-session interaction histories with queries containing intentionally under-specified API arguments and three defined problem types.
  • Multi-Session Grouping: Multiple SGD sessions are grouped per user, with accumulated API calls preserving the action traces from which latent preferences are inferred.
  • Preference Annotation: Related API arguments are manually grouped into higher-level preference categories because SGD supplies domain–slot–value triples without preference labels.
  • Preference Annotation: 19 annotators produced agreement of 89.7% for budget groups and 97.4% for travel groups.
  • Query Construction: Query templates omit one or more preference-related arguments, producing target queries paired with interaction histories.
  • Query Construction: Context-guided queries partially specify constraints, whereas context-free queries provide no such information and require preference modeling alone.
  • Dataset Statistics: MPT contains 265 multi-session dialogues, 2,020 sessions, 39,884 turns, and 332 Recall, 293 Induction, and 472 Transfer instances.

5 Proposed Method: PREFINE

PREFINE treats personalized tool calling as an abstraction problem: it infers reusable latent constraints from behavior and applies them to under-specified API arguments. Its generate–verify–refine memory loop retains generalizable preferences and grounds them to changing tool schemas at inference time.

  • Motivation: Full interaction history does not reliably distinguish reusable user constraints from local or situational decisions.The method therefore frames personalized tool calling as abstraction rather than retrieval.
  • Latent Preference Modeling: PREFINE represents latent preferences as implicit, persistent constraints on API argument selection derived from recurring behavioral patterns.These patterns may recur within one domain or span flights, restaurants, and hotels.
  • PREFINE System: PREFINE is the only compared memory-augmented method that stores latent preferences and refines them iteratively.The comparison concerns both memory content and memory-update mechanisms.
  • Generate–Verify–Refine Loop: PREFINE updates a single accepted preference hypothesis through a generate–verify–refine loop across sessions.Subsequent sessions can refine, broaden, or overturn earlier hypotheses.
  • Generate–Verify–Refine Loop: The verifier retains hypotheses supported by multiple interactions, abstract beyond one-off events, actionable for future API arguments, and temporally consistent.These four conditions govern whether a candidate becomes preference memory.
  • Schema-Agnostic Preference Memory: PREFINE stores abstract, schema-agnostic constraints and grounds them to the current API schema when making tool-call decisions.This defers schema grounding to inference, allowing memory to remain usable across different tool interfaces.

6 Experimental Setup

The experiments evaluate test-time personalized tool calling without additional training, comparing PREFINE with prompting and memory baselines across query settings and metrics. The setup varies memory-construction and inference models while limiting PREFINE’s refinement loop to three iterations.

  • Methods and Models: All methods are evaluated without additional training against Base prompting, RAG, Mem0, and LangMem.The baselines represent full-dialogue, retrieval-based, summary-based, and agentic memory approaches.
  • Methods and Models: PREFINE memories are constructed with four base LLMs and evaluated with eight inference LLMs across all 4×8 combinations.Table 3 reports results averaged over the four memory-construction models for each inference LLM.
  • Evaluation: Table 3 compares baselines and PREFINE under context-guided and context-free settings, using shading to indicate the magnitude and direction of performance changes.Green denotes gains, red denotes losses, and intensity ranges from light to very strong.
  • Methods and Models: PREFINE’s generate–verify–refine loop is capped at three iterations, while ten iterations provide no consistent gain despite higher inference cost.The larger budget is evaluated in Appendix C.5.
  • Metrics: P-EM measures preference-driven unspecified arguments, EA-F1 measures explicitly specified arguments, and OA-F1 measures all arguments.Context-guided queries require both extracting explicit arguments and filling unspecified ones.

7 Experimental Results

PREFINE addresses the gap between direct preference recall and latent preference induction or transfer, improving argument calibration, transfer, and memory efficiency while retaining schema robustness. Its benefits involve predictable precision–recall trade-offs for some models.

  • 7.1 Existing Baselines Recover Observations but Not Latent Preferences: Existing baselines are relatively strong on Preference Recall but degrade on Preference Induction and Preference Transfer.Base prompting average F1 falls from 53.19% in Recall to 43.00% in Induction and 16.26% in Transfer.
  • 7.2 How PREFINE Improves Tool Use: PREFINE improves preference-driven and explicit-argument prediction by providing compact, directly applicable preference guidance.The method reduces reliance on long dialogue history during inference.
  • 7.2 How PREFINE Improves Tool Use: 0.77 to 0.56 and 1.08 to 0.77 are the mean absolute deviations after PREFINE in context-guided and context-free settings, respectively.These correspond to reductions of 28.1% and 28.7%, indicating closer alignment with the ground-truth argument count.
  • 7.3 Memory Efficiency, Scalability, and Utility: 23.28 tokens per dialogue equals 1.24% of full dialogue history, while memory remains around 20–25 tokens after ten sessions.Figure 5 compares average retrieved tokens and memory growth across accumulated sessions.
  • 7.4 When and Where PREFINE Helps Most: PREFINE improves Preference Transfer for most inference LLMs, but transfer varies with each backbone’s ability to apply stored preferences in new contexts.The method supports both preference abstraction and inference-time preference application.
  • 7.4 When and Where PREFINE Helps Most: Stronger action-space control can reduce recall when required arguments are pruned, as shown by R1-Distill-Llama-8B’s 3.34 → 2.85 argument count change.These cases are described as predictable precision–recall trade-offs rather than contradictions to the overall trend.
  • 7.5 PREFINE Supports Dynamic Schema: PREFINE retains gains under schema mismatch, with GPT-5 context-guided P-EM rising from 3.75% to 47.00%.The evaluation uses unseen API domains whose argument names and values differ from those used during memory construction.

8 Conclusion

The paper frames personalized tool calling as latent-constraint modeling over multi-session behavior rather than simple retrieval. It introduces MPT and PREFINE, while identifying richer personalization and noisier long-horizon interactions as future directions.

  • 8 Conclusion: MPT evaluates Preference Recall, Preference Induction, and Preference Transfer for under-specified API arguments.The benchmark exposes a gap between naïve pattern matching and true latent preference modeling.
  • 8 Conclusion: PREFINE represents preferences as revisable hypotheses and generates, verifies, and refines reusable constraints for personalized tool calling.The method remains effective under dynamic schema.

Ethics Statement

The supplied passages describe MPT’s dataset structure, preference mappings, interaction scale, and visualization conventions. They do not provide an ethics statement or ethics-specific findings.

  • Dataset and Preference Mapping: MPT covers 11 preference-sensitive domain–argument pairs across 8 domains, excluding sparse or ambiguous preference categories.The BUDGET mapping omits mid_cost, and TRAVEL retains solo_usage while excluding group_usage.
  • Dataset Scale: Each multi-session dialogue averages 7.6 sessions and 19.7 turns per session across 265 dialogues.The benchmark contains nearly 40k turns for long-horizon preference modeling.
  • Dataset Distributions: Figure 6 reports preference-group and API-call distributions, noting that counts are not mutually exclusive.A single example may contain multiple preference groups and multiple API calls.
  • Evaluation Settings: Context-guided and context-free queries target the same preference-sensitive argument, but context-guided queries include additional in-session information.The supplied table description identifies U as User and A as Agent.
  • Evidence Representation: MPT represents evidence through preference groups with explicit counts and argument provenance.Figure 8 describes a machine-readable aggregation of cross-session preference signals.

A.6 Human Validation of Preference Grouping

The study validates that the benchmark’s preference categories reflect broadly shared behavioral intuitions, while identifying ambiguity in the travel-group boundary.

  • Agreement Results: Annotators agreed with the grouping in 89.7% of budget cases and 97.4% of travel cases.Fleiss’ κ was 0.701 for budget and 0.880 for travel.
  • Interpretation: Substantial agreement supports using the preference categories as a task-grounded evaluation scaffold rather than arbitrary schema choices.The discussion attributes lower budget agreement to the heterogeneous forms of budget-related behavior.
  • Scope Boundary: Parties of two may indicate couple travel rather than group travel, creating ambiguity at the boundary between solo_usage and group_usage.The benchmark conservatively retains only solo_usage and excludes group_usage from preference evidence.

B.1 Detailed Experimental Settings

The experiments use fixed LLM backbones and compare PREFINE with several memory-augmented baselines across preference-modeling query settings. Results are reported with query-level macro-averaging and detailed backbone-specific and metric breakdowns.

  • All models are evaluated on the same fixed query–history pairs without stochastic sampling or reranking, using query-level macro-averaged metrics.
  • The RAG baseline retrieves the top-5 utterances by cosine similarity and appends them to prompts for Gemini-3-Flash and GPT-5 inference.
  • Mem0 and LangMem retrieve top-5 user-scoped memory snippets or contents before tool-calling inference with Gemini-3-Flash and GPT-5.
  • Tables 11–13 provide per-backbone results, including precision, recall, and F1 breakdowns across context-guided and context-free preference query types.

C.4 RAG, Mem0, LangMem Backbone LLM-Specific Results

The additional experiments examine backbone stability, refinement budgets, schema mismatch, and prompt design. Across these settings, memory methods remain competitive in recall, while PREFINE shows stronger robustness to schema changes.

  • Backbone-specific results: Memory-augmented methods remain competitive in Preference Recall across backbones, but their gains diminish for Preference Induction.
  • Refinement budget: Extending PREFINE from three to ten refinement iterations yields little consistent benefit, with changes ranging from −0.006 to +0.034 across settings.The only noticeable gain is +0.034 for context-guided Preference Transfer; context-guided Recall and Induction each decrease by −0.006.
  • Dynamic schemas: Mem0 is omitted under schema change because lexical mismatch causes its retrieval API to return no memories for test queries.
  • Dynamic schemas: With GPT-5 under schema mismatch, PREFINE raises context-guided P-EM from 3.75% to 47.00% and context-free F1 from 36.39% to 51.45%.RAG and LangMem show sharp drops relative to their in-schema performance.
  • Prompt design: PREFINE separates preference abstraction from verification, while base prompting directly infers preferences and generates an API call from full dialogue history.
Loading 2604.17886v1…