PaperlayerView exact revision on arXiv

Source-linked AI summary

HyperTrace: Hypothesis-Based Preference Tracing for Online LLM Personalization

Jianzhi Shen, Keyu Mao, Minghao Shao, Chuanyang Jin, Yusong Wang, Ailiang Lin, Kotaro Funakoshi, Manabu Okumura, Tianmin Shu, Muhammad Shafique

arXiv:2609.09835v1cs.CL

TL;DR

Personalization must infer preferences that emerge gradually while reconciling long-term patterns with short-term topic-specific needs. HyperTrace addresses this with weighted natural-language hypotheses updated through SMC-style tracing and surrogate choice modeling, and it reports stronger online personalization outcomes than evaluated baselines, with explicit limits on the available feedback and hypothesis reliability.

  • Problem

    Existing training-free personalization methods have limited ability to model uncertainty over user intent and reconcile long-term preferences with short-term topic-specific needs.

  • Method

    HyperTrace maintains weighted natural-language hypotheses for short-term intent and long-term preferences, updating them across turns and sessions with an SMC-style procedure and an LLM-based surrogate choice model.

  • Results

    HyperTrace shows stronger and more robust response alignment and preference prediction than evaluated online baselines, with competitive long-term profile alignment on PRISM and PersonaMem-v2.

  • Takeaways & Limitations

    Combining fine-grained short-term belief updates with long-term memory consolidation supports personalization that is robust, generalizable, and interpretable.

  • Takeaways & Limitations

    HyperTrace assumes explicit comparative choices, while its natural-language hypotheses inherit the underlying LLM’s limits in granularity and reliability.

Abstract

from arXiv · show

Personalized language models aim to adapt responses to individual users, whose preferences are often latent and revealed gradually through interaction. Existing training-free methods rely on stored histories or retrieved memories, but they often struggle to reconcile long- term preferences with short-term topic-specific needs. To address this issue, we propose HyperTrace, a training-free framework that formulates online personalization as latent preference tracing. HyperTrace maintains interpretable natural-language hypotheses over short-term intent and long-term preferences, and updates them through an SMC-style reweight process using an LLM-based surrogate choice model. By updating these hypotheses across turns and sessions, HyperTrace enables personalization without parameter updates. Experiments on PRISM and PersonaMem-v2 show that HyperTrace improves response alignment, preference prediction, and profile consistency over strong online baselines, demonstrating the effectiveness of tracing latent user preferences for robust personalization. Code and scripts are available in the repository: https://github.com/jiseshen/HyperTrace.

1 Introduction

HyperTrace frames online personalization as tracing latent user preferences rather than compressing history into profiles or retrieving memories. It maintains interpretable hypotheses over short-term intent and long-term preferences, updating them from feedback and evaluating alignment, prediction, and profile consistency.

  • Users’ goals, preferences, and expectations emerge gradually, requiring personalization that adapts continually across diverse users.
  • Training-based personalization can be computationally costly and require white-box models, while inference-level methods offer scalability but limited interpretability and weak generalization.
  • HyperTrace represents user information as weighted natural-language hypotheses over latent intent instead of adapted parameters or stored dialogue traces.
  • Short-term hypotheses support rapid topic-specific adaptation, while higher-level hypotheses consolidate stable long-term preferences.
  • The framework evaluates response alignment, preference prediction, and profile alignment on PRISM and PersonaMem-v2, outperforming summary- and memory-based baselines.

2 Related Work

Related work spans personalization benchmarks, training-based adaptation, and inference-time methods. HyperTrace differs from these approaches by weighting multiple natural-language preference hypotheses and tracing actionable signals from user choices.

  • Personalization benchmarks cover user-conditioned tasks, recommendations, generation, long-term memory, dynamic profiles, implicit preferences, and multi-session histories.
  • Preference-feedback datasets expose heterogeneous human preferences through user-specific choices or in-situ feedback on candidate responses.
  • HyperTrace builds an online evaluation setting from pluralistic preference data using response alignment, preference prediction, and profile alignment.
  • Training-based methods learn user representations, PEFT modules, or personalized post-training objectives, while decoding-time steering requires control over sampling.
  • HyperTrace represents user information as a weighted set of natural-language hypotheses rather than a single profile, retrieved context, or memory state.
  • Its update is related to Bayesian hypothesis filtering but focuses on tracing actionable preference signals from user choices.

3 Preliminary

The preliminary formulation models personalization as session-dependent latent preference inference. User feedback over candidate responses supplies observations for a Bradley–Terry choice model and posterior preference updating.

  • The interaction is modeled as a sequence of sessions with a latent, context-dependent active preference state.
  • The same user may prefer different response styles across intents, such as direct code answers versus broader responses to daily questions.
  • Within a session, user preferences are revealed through feedback on a query’s candidate response set.
  • A Bradley–Terry observation model assigns selection likelihoods using candidate utility under the current query and latent preference state.
  • The posterior preference belief conditions on the observed feedback history, denoted D1:t.

4 Methodology

HyperTrace traces user preferences with weighted natural-language hypotheses, updating them from observed choices within and across sessions. Its evaluation combines response alignment, preference prediction, profile alignment, and controlled LLM-based scoring to assess personalization quality.

  • 4.1 HyperTrace: HyperTrace represents possible preference states as K weighted natural-language hypotheses that retain plausible explanations of user behavior.Examples include preferences for concise implementation details, broader conceptual explanations, or more cautious wording.
  • 4.1 HyperTrace: Belief updates refine hypotheses, filter them against feedback, resample relatively supported hypotheses, and perturb preference axes to avoid premature collapse.The lifecycle independently revises or replaces hypothesis slots before scoring and resampling when effective sample size falls below a threshold.
  • 4.1 HyperTrace: Each filtering step uses an LLM to estimate hypothesis-conditioned utilities, then applies a fixed choice rule to convert chosen-versus-rejected contrasts into particle-weight updates.This separates utility estimation from the comparable weighting rule used for observed feedback.
  • 4.1.2 Cross-Session Preference Consolidation: Cross-session consolidation weights current-session evidence by valid traced turns and belief decisiveness, while topic-labelled storage and retrieval determine where preferences apply.Entropy affects consolidation strength, whereas topic metadata distinguishes stable cross-topic preferences from temporary task-specific needs.
  • 4.2.1 Evaluation Framework: The evaluation measures response alignment, preference prediction, and profile alignment using relative response comparisons, choice recovery, and rubric-based semantic profile matching.Response alignment compares the adapted response with the chosen candidate relative to rejected alternatives; profile alignment evaluates coverage, contradiction avoidance, specificity, and consistency.

5 Experiments

Experiments evaluate HyperTrace across online personalization accuracy, robustness, ablations, cost, and qualitative transfer. HyperTrace generally improves preference prediction and response alignment, with stability across sessions and a cost–accuracy trade-off favoring hybrid deployment.

  • 5.1 Main Comparison: HyperTrace leads preference prediction and response alignment on PRISM and PersonaMem-v2, while profile alignment is strongest on PRISM but only competitive on PersonaMem-v2.The evaluation measures post-turn-20 preference prediction, response alignment, and final profile alignment.
  • 5.1 Main Comparison: 41/50 users transition sessions between turns 18–22, where HyperTrace maintains a relatively stable adaptation trajectory while several baselines lose later-turn gains.CoT and RAG perform better early but do not sustain gains; Dynamic Cheatsheet drops sharply at session changes.
  • 5.1 Main Comparison: Absolute response alignment remains negative for every method on PersonaMem-v2, although the shared protocol still supports relative comparison.The benchmark’s boundary cases motivate safety-aware filtering of hypotheses as a potential extension.
  • 5.2 Method Ablation: Topic grounding and hierarchical consolidation strengthen the tracing system, while removing consolidation makes synchronization of short-term beliefs into long-term memory less reliable.The full hypothesis store also preserves preferences across multiple topics and interaction phases.
  • 5.3 Robustness Analyses: The adaptation pattern persists across evaluators and tracing backbones, although performance is not monotonic in model size.All tested backbones improve preference-prediction accuracy from turn 0, and most retain near-positive or positive late-stage response alignment.
  • 5.4 Cost–Quality Trade-off: 0.5878 Acc>20 at $0.0131 per turn makes HT(hybrid) the practical Pareto point, reducing cost by 53.1% while retaining 95.8% of HT(gpt-5) accuracy.HT(gpt-5) reaches the highest accuracy at 0.6136 Acc>20, so no configuration dominates every criterion.
  • 5.5 Qualitative Analysis: HyperTrace transfers a self-care preference from interpersonal advice to creative-habit guidance, indicating storage as a behavioral tendency rather than a topic-specific memory.A separate qualitative case shows that semantically adjacent positive memories can survive despite a do-not-remember boundary.

6 Conclusion

HyperTrace formulates online personalization as latent preference tracing, using hypothesis updates to address preferences that emerge gradually through interaction. Experiments on PRISM and PersonaMem-v2 show stronger and more robust response alignment and preference prediction than evaluated online baselines, with competitive long-term profile alignment.

  • HyperTrace treats user preferences as latent variables gradually revealed through interaction and formulates personalization as latent preference tracing.
  • The framework combines fine-grained short-term belief updates with long-term memory consolidation.
  • Experiments on PRISM and PersonaMem-v2 show stronger and more robust response alignment and preference prediction than evaluated online baselines.
  • The experiments also show competitive long-term profile alignment.

Limitations

HyperTrace assumes explicit comparative choices as its main observation source and does not directly infer preferences from verbal critiques or unobserved implicit behavior.

  • HyperTrace assumes explicit comparative choices and does not directly infer preferences from verbal critiques or unobserved implicit behavior.
  • Extending the observation model without losing inspectability is identified as an important direction.
  • Natural-language hypotheses inherit the underlying LLM’s limits in granularity and reliability.
  • The paper motivates structured representations, user-editable memory, and safety-aware verification as future directions.

Ethical Considerations

The study uses public data, models, and APIs under released terms, with aggregate research-only evaluation and no redistribution or personally identifying information. It acknowledges privacy, profiling, bias, and persuasion risks while describing controlled evaluation and future inspection controls as mitigations.

  • The study uses public datasets, models, and APIs under their released terms.
  • The authors do not redistribute source data or model weights, identify users, or add personally identifying information.
  • Evaluation is aggregate and research-only.
  • Personalization can leak sensitive information, misprofile users, amplify bias, or become overly persuasive.
  • Controlled offline and aggregate evaluation limits exposure, while explicit external traces support future inspection, editing, deletion, and retention controls.

A Implementation Details

The implementation uses a modular, schema-constrained tracing pipeline with five natural-language hypotheses, skip gating, hierarchical retrieval and consolidation, and model-routing controls. Ablations and robustness checks examine structural components, initialization, scoring, propagation, and hypothesis diversification.

  • The tracer maintains five active natural-language hypotheses and skips low-signal turns when a majority vote indicates insufficient preference evidence.
  • Initialization creates five topic-labelled hypotheses, while later turns revise or replace slots before weighting, resampling, and near-duplicate grouping.
  • Cross-session memory uses a FAISS inner-product index with 1536-dimensional text-embedding-3-small representations and topic metadata.
  • Structured outputs use typed JSON schemas, malformed outputs are rejected, and model variants differ through routing or system configuration while prompt semantics remain fixed.
  • Each online turn builds or retrieves a response-time profile, generates an adapted response, and updates beliefs from chosen-versus-rejected comparisons.
  • The flat-slot ablation removes hierarchical retrieval and consolidation, leaving stale hypotheses unrepaired.
  • Replacing axis-level diversification with paraphrases of collapsed hypotheses reduces all three metrics in Table 9.

D More Qualitative Examples

The qualitative traces illustrate how HyperTrace separates visible candidate contrasts into interpretable preference hypotheses across recipe, chronic-health support, and FPS recommendation scenarios. These examples refine response-style, task, and domain preferences without unsupported assumptions about the user.

  • Qualitative tracing: The traces use five-slot candidate contrasts to expose response preferences without treating the cases as annotated evaluations.The preprocessing and tracing outputs are presented as examples of visible candidate differences rather than formal annotations.
  • Recipe request: Recipe tracing separates concise baked preparation, complete ingredient use, brief serving suggestions, and rejection of verbose or off-topic alternatives.The selected candidate uses the requested ingredients in a concise baked recipe, while alternatives are verbose, taco-style, or unrelated.
  • Chronic-health support: Chronic-health tracing captures validating, practical, collaborative support that adapts to good and bad days without inferring new medical facts.The hypotheses also recognize proactive health management and a support network without patronizing language.
  • FPS refinement: FPS tracing narrows a broad competitive-game preference into deathmatch, low time-to-kill, precise hitscan gunplay, limited progression, and strong anti-cheat.Later explicit rejections refine the earlier genre hypothesis into factors intended to guide subsequent recommendations.

E Detailed Turn-Level Results

The paper reports detailed turn-level online evaluation curves for PRISM and PersonaMem-v2, complementing PRISM adaptation curves with SEM and aggregate statistics. Figure 8 plots the complete curves at each interaction turn.

  • Turn-level evaluation: The detailed evaluation provides complete turn-level curves for both PRISM and PersonaMem-v2.These curves supplement the main text's PRISM adaptation curves and aggregate statistics.
  • Turn-level evaluation: Figure 8 plots online evaluation results at each interaction turn for PRISM and PersonaMem-v2.The figure presents the per-turn trajectories rather than only aggregate summaries.
Loading 2609.09835v1…