FleetPilot: Fine-Grained Steering Empowers Real-World LLM Personalization

1 Introduction

• Personalization must use limited, evolving interaction histories to infer preferences that vary with the current query and output type.Histories may contain fewer than ten interactions, preferences can shift within days or hours, and one user can prefer different outputs across queries and domains.Source: §1 · pp. 2–3
• Prompt-based, parametric, and reward-model approaches face different costs or constraints under sparse observations, preference drift, and query-dependent heterogeneity.Their cited bottlenecks include noisy or truncated context, optimization and deployment costs, and multi-candidate generation and scoring.Source: §1 · pp. 3–4
• FleetPilot combines matched contrast, separate attention- and MLP-side steering directions, and query-conditioned retrieval and aggregation.Matched contrast varies user-relevant context while holding the historical query and response fixed; the retrieved directions are applied sequentially during generation.Source: §1 · pp. 4–5
• FleetPilot outperforms competing baselines across validation scenarios, with particularly strong performance in dynamic and data-sparse environments.The framework builds query-specific steering vectors and requires neither gradient-state storage nor per-user model checkpoints, with modest inference-time latency.Source: §1 · p. 5
• The paper positions inference-time steering as a training-free, flexible interface for personalization under data scarcity, preference drift, and query-dependent heterogeneity.Its contribution summary frames these conditions as design requirements and reports validation on short- and long-form generation and web function calling.Source: §1 · p. 5

2.1 Prompt-Based Personalization

• Prompt-based personalization retrieves records or summarizes preferences for the prompt, but depends on retrieval and context quality and is constrained by finite context windows.The paper describes FleetPilot as complementary: retrieved text supplies explicit evidence while steering adds an activation-space intervention.Source: §2.1 · pp. 5–6

2.2 Parametric Adaptation

• Parametric methods use shared or user-specific adaptations, but still incur optimization and evidence requirements as well as storage, loading, or maintenance costs.FleetPilot instead represents personal evidence as a retrievable dictionary of activation directions that can be updated through vector preparation.Source: §2.2 · p. 6

2.3 Personalized Reward Modeling

• Personalized reward models or judges score candidate responses, so candidate-selection quality depends on available candidates and costs rise when more candidates are generated and evaluated.FleetPilot steers generation without requiring explicit pairwise preference labels or multiple completed candidates for each query; the paper limits this contrast to the illustrated inference pipeline.Source: §2.3 · p. 6

2.4 Inference-Time Activation Steering

• Activation-steering methods vary in their contrasts, extraction and injection interfaces, and whether directions vary by input; FleetPilot uses matched contexts, separate post-attention and MLP directions, and query-based retrieval.These choices distinguish it from response-contrast, whole-layer, and fixed user-level alternatives.Source: §2.4 · pp. 6–7

3.1 Personalization Setting

• The personalization setting uses prior query-response interactions to generate outputs aligned with current, query-dependent preferences, while recognizing that history is incomplete and its relevance changes over time.Under the temporal protocol, only information available before the evaluated interaction may be used; the current target response is unavailable for retrieval or direction construction.Source: §3.1 · p. 7

3.2 Personalization Interfaces

• The section distinguishes prompt conditioning, parameter changes, and reward-based selection among generated candidates as interfaces for using personal evidence.A prompted judge or learned scorer can rank candidates, but the stated expression describes candidate selection rather than every use of reward models.Source: §3.2 · pp. 7–8

3.3 Activation Steering

• Activation steering adds a direction to an internal activation while leaving model parameters unchanged; the contrast's meaning depends on the paired inputs and pooled activations.FleetPilot keeps the historical query and response fixed while varying context, and uses query-conditioned directions; the local approximation does not establish global linearity or monotonic behavior.Source: §3.3 · p. 8

4.1 Overview

• FleetPilot contrasts model activations under user-relevant and user-irrelevant contexts to create steering vectors for personalized adaptation.Relevant context comes from the user’s personal corpus; irrelevant context is sampled from other users’ histories.Source: Figure 2 · p. 9, §4.1 · p. 9
• At inference, semantic similarity between the query and historical sample pairs selects top-k vectors, which are aggregated and injected into the model.The method uses sequential dual injection at a predefined layer.Source: Figure 2 · p. 9, §4.1 · p. 9
• Fine-Grained Hooking separately extracts steering signals from attention and MLP layers for each user’s historical interactions.The vectors are prepared offline from historical interactions.Source: §4.1 · p. 9
• FleetPilot is designed for dynamic, data-sparse settings and claims minimal memory and latency overhead.The framework performs personalization without model training.Source: §4.1 · pp. 9–10

4.2 Steering Vectors Preparation

• Preparation uses old user logs in the time-splitting setting to build steering vectors.The preparation stage is based on historical interactions.Source: §4.2 · p. 10

4.2.1 Contrastive Prompt Construction

• Contrastive prompts hold the historical query and response fixed while varying context between relevant user history and other users’ histories.The same query and response are supplied under both contextual conditions, so the construction examines representation changes for a fixed response.Source: §4.2.1 · p. 10

4.2.2 Fine-Grained Activation Hooking

• Fine-grained hooking retains separate attention and MLP contrastive signals, but contextual factors may contribute systematically to the resulting directions.The method does not claim that every empirical steering direction isolates an unbiased preference effect.Source: §4.2.2 · pp. 10–11

4.2.3 Steering Dictionary Construction

• The dictionary associates each record with a retrieval representation of its positive-context sequence, distinct from the language model’s activation-extraction role.The negative context estimates the activation contrast but is excluded from the retrieval key; the historical response is available during preparation, not current-query retrieval.Source: §4.2.3 · p. 11

4.3.1 Query-Conditioned Retrieval and Aggregation

• For each query, FleetPilot retrieves semantically relevant steering entries, aggregates attention- and MLP-side directions separately, and uses a distinct prompt budget for textual history.The steering retrieval budget K and prompt budget B govern different information channels, and their retrieved records need not coincide.Source: §4.3.1 · p. 12

4.3.2 Sequential Dual-Injection

• At layer L, sequential dual-injection applies each aggregated direction at the computational interface from which it was extracted.The attention-side perturbation changes the residual entering the MLP, which is recomputed before the MLP-side direction is added.Source: §4.3.2 · p. 12
• Coefficients α and β independently control the attention- and MLP-side interventions, and setting both to zero recovers the unsteered computation.Setting either coefficient to zero yields the corresponding single-site intervention.Source: §4.3.2 · p. 12
• In the cached dual-site configuration, interventions apply during prompt processing and autoregressive decoding, without requiring identical token ranges for direction preparation and inference.Extraction and injection use matching interfaces, not necessarily matching token ranges.Source: §4.3.2 · p. 13

5.1.1 Datasets

• The evaluation uses short personalized content generation, abstract writing, and personal web function calling datasets, each containing user IDs, profiles, and interaction histories.The study samples 200 users and splits earlier question-answer pairs for training and later pairs for testing.Source: §5.1.1 · p. 13

5.1.2 Baselines

• Baselines include in-context learning, One-Peft-All-User, One-Peft-Per-User, personalized reward models, and steering-based methods.Table 2 describes the main results for content generation and personal web function calling, with baselines using n-shot demonstrations.Source: §5.1.2 · p. 13, Table 2 · p. 14

5.1.3 Metric

• Generation quality is measured with Rouge and METEOR, while personal web function-calling quality is measured with Overall Accuracy (ACC).Table 3 reports performance for the hooking-strategy comparison.Source: §5.1.3 · p. 13, Table 3 · p. 14

• RQ1 asks whether FleetPilot outperforms competing baselines, and RQ2 asks whether fine-grained hooking improves performance.Source: §5.2 · p. 14
• RQ3 examines robustness to heterogeneous or fast-changing distributions, while RQ4 examines data efficiency, especially under high data sparsity.Source: §5.2 · p. 14
• RQ5 asks about FleetPilot’s overhead and real-world feasibility, while RQ6 asks about sensitivity to hyperparameters.Source: §5.2 · p. 14

• FleetPilot outperforms all other baselines; the reported comparison also finds parametric methods outperform prompt-based methods.The authors report this as evidence for steering-based personalization over traditional personalized methods.Source: §5.3 · p. 14
• The compared hooking strategies include post-attention, post-MLP, post-block, and hybrid sequential injection at the attention and MLP interfaces.Source: §5.3 · p. 15
• The Hybrid Strategy improves content-generation metrics by 0.36%–1.64% over the strongest single-site or post-block alternative (p ≤0.001).Source: §5.3 · p. 15

5.4.1 (RQ3) Robustness to Drifting Distribution

• FleetPilot outperforms all methods in every tested time range under realistic temporal drift.The authors attribute the contrast with parametric methods to their sluggish offline training loops, whereas FleetPilot adapts at inference time for each instance.Source: §5.4.1 · p. 15
• Illustrations of one user's samples show variation in content themes, audience-engagement strategies, and emotional tones across question-answer pairs.The samples are from user 10000051.Source: Figure 3 · p. 15, §5.4.1 · p. 16

5.4.2 (RQ4) Robustness to Data Scarcity and Data Efficiency Discussion

• FleetPilot achieves the highest ROUGE-1 across all four data fractions and remains superior to parametric methods under extreme sparsity (<5 shots).Performance increases with additional evidence, while preserving much of its level under substantial history reduction.Source: §5.4.2 · p. 16, Table 5 · p. 16, Figure 4 · p. 16

5.4.3 (RQ5) Overhead Discussion and Deployment Analysis

• Per-user storage is below 1 MB for FleetPilot, compared with approximately 18 MB for personalized LoRA.FleetPilot also incurs light inference overhead; retrieving relevant steering vectors is its most time-consuming step.Source: §5.4.3 · p. 17

5.4.4 (RQ6) Hyperparameters Analysis

• Performance variances are 0.000019 on abstract writing, 0.000006 on headline generation, and 0.000092 on PersonalWAB; α is the only systematically impactful parameter, with α ≤ 1.1 as a general guideline.The authors report low variance across the tested parameter range.Source: §5.4.4 · p. 18, Figure 5 · p. 18, Figure 6 · p. 18

• FleetPilot combines matched contrast, separate attention- and MLP-layer steering signals, and query-conditioned retrieval of historical vectors for instance-level adaptation without retraining.The framework is presented as addressing data efficiency and adaptability in LLM personalization.Source: §6 · p. 19
• FleetPilot outperforms both prompt-based and parameter-efficient fine-tuning baselines.Source: §6 · p. 19
• FleetPilot maintains strong performance with as few as 5–6 examples per user and under rapidly shifting user distributions, where traditional methods show significant performance degradation.Source: §6 · p. 19