Paper deep dive
FleetPilot: Fine-Grained Steering Empowers Real-World LLM Personalization
Kounianhua Du, Jianxing Liu, Kangning Zhang, Wenxiang Jiao, Yuan Lu, Jiarui Jin, Jianghao Lin, Weiwen Liu, Yong Yu, Weinan Zhang
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:With the rapid evolution of large language models (LLMs), there is an increasing demand for effective personalization techniques that can adapt model behavior to individual user preferences. Personalized LLMs must infer a user's query-dependent preferences from limited, heterogeneous, and evolving interaction histories. Existing approaches rely primarily on in-context personalization, parametric adaptation, or personalized reward modeling, but respectively incur recurring costs in context management, model optimization and deployment, or multi-candidate generation and scoring. These limitations become particularly restrictive when user observations are sparse and the preference evidence relevant to each query varies across contexts and over time. To address these challenges, we propose \modelname, a fine-grained and instance-tailored steering framework that dynamically generates sample-level interference vectors from user data and injects them into the model's forward pass for personalized adaptation. Our approach introduces: match contrast that produces user-relevant signals, fine-grained steering component that captures nuanced signals by hooking activations from attention and MLP layers, and query-conditioned aggregation module that synthesizes these signals into contextually relevant enhancements. Extensive experiments across diverse scenarios, including short-to-long text generation, and web function calling, validate the effectiveness and compatibility of \modelname. Results show that our method consistently brings significant performance gains, and show extreme superiority in fast-shifting and data-sparse environments. Additionally, \modelname rids us of training cycles and incurs light overhead of storage or inference latency. These make \modelname a strong and real-world deployable method. Implementation is available \footnote{this https URL}.
Tags
Links
- Source: https://arxiv.org/abs/2610.14970v1
- Canonical: https://arxiv.org/abs/2610.14970v1
Trouble viewing inline? Open PDF directly â
Full Text
71,707 characters extracted from source content.
Expand or collapse full text
Front. Comput. Sci., 2026, 0(0): 1â22 https://doi.org/10.1007/sxxxxx-y-z-1 RESEARCH ARTICLE Fleet Pilot: Fine-Grained Steering Empowers Real-World LLM Personalization Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 1 Shanghai Jiao Tong University, 2 Xiaohongshu Inc., Shanghai, China Š Higher Education Press 2026 Abstract With the rapid evolution of large lan- guage models (LLMs), there is an increasing de- mand for effective personalization techniques that can adapt model behavior to individual user prefer- ences. Personalized large language models (LLMs) must infer a userâs query-dependent preferences from limited, heterogeneous, and evolving interac- tion histories. Existing approaches rely primarily on in-context personalization, parametric adaptation, or personalized reward modeling, but respectively incur recurring costs in context management, model optimization and deployment, or multi-candidate generation and scoring. These limitations become particularly restrictive when user observations are sparse and the preference evidence relevant to each query varies across contexts and over time. To address these challenges, we propose Fleet- Pilot, a fine-grained and instance-tailored steering framework that dynamically generates sample-level interference vectors from user data and injects them into the modelâs forward pass for personalized adap- tation. Our approach introduces: match contrast that produces user-relevant signals, a fine-grained steering component that captures nuanced signals by hooking activations from attention and MLP layers, and a query-conditioned aggregation module that synthesizes these signals into contextually relevant enhancements. The method demonstrates high flexi- bility and data efficiency, excelling in fast-changing distribution and high data sparsity scenarios. Extensive experiments across diverse scenarios, including short-to-long text generation, and web function calling, validate the effectiveness and com- patibility of FleetPilot. Results show that our method consistently brings significant performance gains, and show extreme superiority in fast-shifting and data-sparse environments. Additionally, FleetPilot rids us of training cycles and incurs light overhead of storage or inference latency. These make Fleet- Pilot a strong and real-world deployable method. Implementation is available 1 . Keywords LLM, Personalization, Steering 1 https://anonymous.4open.science/r/FleetPilot arXiv:2610.14970v1 [cs.AI] 10 Oct 2026 2Front. Comput. Sci., 2026, 0(0): 1â22 r etr ieve Personal Corpus -Pr ompt Based (I n Context L ear ning) infer ence Query Context few data Fitting failur e Retr aining cycles Stor age over head ... Best of N sampling r ewar d best model r esponse -Per sonalized Rewar d M odels Personal Corpus -Par ametr ic Adaptation r etr ieve infer ence finetune Query Context Query Context ... ... Layer L Layer L+1 ... Layer L+1 Layer L 1.L inear Repr esentation Hypothesis D LLM Context Window Noisy data L imited model capacity No tr aining, light stor age, high flexibility, high data efficiency 2. Adaptation for Per sonalization X Relevant Context C+ Y X Irrelevant Context C- Y Attention MLP Fine-Gr ained Hook/I nj ect Block Hook/I nj ect Steering Dictionary Query -I nfer ence-Time Steer ing Response Contrast Matched Contrast Whole Layer Hooking Fine-Grained Hooking User-Tailored Instance-Tailored Fig. 1Personalization paradigms and the design of FleetPilot. Left: prompt-based personalization supplies retrieved history through a context window, where noisy evidence and limited model capacity constrain its use. Middle: parametric adaptation and personalized reward modeling introduce optimization or candidate-scoring stages, with challenges in sparse-data fitting, model updates, storage, and best-of-Ninference. Right: inference-time steering modifies intermediate activations in a frozen LLM, motivated by the linear representation hypothesis. FleetPilot adapts this principle to personalization through three design changes: response contrast to matched contrast, whole-layer hooking to fine-grained hooking, and user-tailored to instance-tailored steering. 1 Introduction The advent of large language models (LLMs) Dubey et al. [2024], Yang et al. [2025] has catalyzed trans- formative advances across a multitude of domains. A particularly compelling frontier within this land- scape is the pursuit of personalizationâthe abil- ity to tailor model behavior and outputs to indi- vidual user preferences Cai et al. [2025], Kumar et al. [2024], Liu et al. [2025a], Salemi et al. [2023], Wang et al. [2025]. Achieving effective personal- ization is paramount for deploying LLMs in real- world applications such as personal assistant Guha et al. [2015], Wang and Sun [2025], conversational agents Liu et al. [2025b], Yusuf et al. [2025], and adaptive learning systems Iqbal et al. [2025], Saleem et al. [2025]. A practical personalization system must infer a Received ; accepted E-mail: kounianhuadu@sjtu.edu.cn userâs current, query-dependent preference from a limited and continuously evolving interaction his- tory. Formally, for useruat timet, letH t u =(q i , y i ) i<t denote the observations available before a new query q t . The system must generate an output aligned with the latent preferencep t u (y | q t ), although this dis- tribution is only partially observed, may vary with the query, and may change over time. Production environments expose challenging regimes: ⢠Limited Preference Observations (Data Scarcity). For new or infrequent users,|H t u |can be extremely small, often fewer than ten interactions. The re- sulting cold-start regime Zhang et al. [2025b] pro- vides weak and noisy evidence about individual preferences. An effective system must therefore achieve high statistical efficiency rather than rely on abundant user-specific supervision. â˘Non-Stationary Preferences (Temporal Adapt- ability). The target distribution is generally time- dependent, i.e.,p t+1 u (y | q), p t u (y | q) after a Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot3 change in interests, goals, or context Liu et al. [2024]. Since such changes may occur within days or even hours, personalization must incorpo- rate recent evidence at inference time; adaptation tied to periodic optimization and redeployment can leave the system aligned with stale prefer- ences. â˘Context-Dependent Preference Heterogeneity (Instance Specificity). A single user can exhibit different preferences across queries, domains, and output forms, from short responses and long- form generation to structured tool calls Jiang et al. [2024]. Consequently, the relevant signal is not fully captured by one static, user-level represen- tation. The system must identify which subset of H t u is pertinent toq t and apply the corresponding preference without allowing unrelated or conflict- ing history to dominate. These challenges are coupled: scarce observa- tions make drift difficult to estimate, while tempo- ral and contextual heterogeneity further reduce the amount of history relevant to any individual query. We therefore require a practical method to be data- efficient, rapidly updateable, and instance-specific, while avoiding repeated training or multi-candidate inference. Existing methodologies for LLM personalization can be broadly categorized into three paradigms. Prompt-based methods leverage in-context-learning (ICL) Dong et al. [2022] ability of llms, strategi- cally inserting user profiles or historical interac- tions into the LLM context window for personal- ized output Madaan et al. [2022], Qian et al. [2024], Shi et al. [2025]. Parametric adaptation meth- ods include One-Peft-All-User Lin et al. [2024] that learns collaborative signals within the whole dataset by tuning LLMs on all the usersâ data and One-Peft-Per-User Tan et al. [2024a,b], Zhang et al. [2024] that performs parameter-efficient fine-tuning on user-specific data, tailoring user-specific weights to offer personalization. Personalized reward mod- eling methods learn preference-aware scoring func- tions that evaluate multiple candidate generations and select the output most aligned with the target user Chen et al. [2025a], Ryan et al. [2025], Zhuang et al. [2024]. Despite their effectiveness, the three paradigms expose different bottlenecks under these real-world requirements. Prompt-based personalization is training-free and immediately updateable, but its performance is governed by the quality of context content used and the llm capacity Science and Teams [2025]. Long, noisy records can introduce stale, contradictory, or irrelevant evidence that distracts the generator; trun- cation or profile summarization can instead discard minority and recently emerging preferences. Longer prompts also increase prefill latency and repeatedly expose raw personal records to the inference ser- vice Liu et al. [2025a], Vuong et al. [2021]. [Lin et al., 2024] shows that these methods often reach their peak performances at limited context tokens, even far less than the llm capability upper bound. Parametric adaptation encodes preferences more compactly, but converts each preference update into an optimization and deployment problem. A shared adapter must reconcile heterogeneous users and can suffer negative transfer, whereas one-adapter- per-user systems require user-specific supervision and incur training, storage, versioning, and adapter- switching costs that grow with the user popula- tion Lin et al. [2024], Liu et al. [2025a], Tan et al. [2024a,b]. With only a few interactions, parameter estimates are unstable and prone to fitting incidental features; once preferences drift, stale adapters must be retrained or replaced rather than conditioned on 4Front. Comput. Sci., 2026, 0(0): 1â22 the latest interaction at inference time. Personalized reward modeling moves personal- ization into an explicit scoring function, yet sparse pairwise feedback underdetermines a userâs context- dependent utility and may encourage reliance on spurious correlates. Models learned from pooled feedback can blur conflicting tastes, while per-user models face a severe biasâvariance trade-off when feedback is scarce Chen et al. [2025a], Ryan et al. [2025], Zhuang et al. [2024]. Moreover, reranking and rejection-sampling pipelines can only choose among generated candidates: their quality is bounded by candidate coverage, and increasing that coverage through best-of-Nor rejection sampling multiplies generation and scoring cost Gui et al. [2024], Liu et al. [2023]. A reward function estimated offline can also become miscalibrated as preferences or task distributions change, requiring fresh feedback and model updates. Activation steering Rimsky et al. [2024], Turner et al. [2023], Zou et al. [2023] has been used in tox- icity reduction and deception control, where high- level behavioral attributes are hypothesized to be encoded as approximately linear directions in the modelâs residual-stream activation space (the lin- ear representation hypothesis), so that a single con- trastive difference vector suffices to identify each concept. Under this hypothesis, adding or subtract- ing a scaled version of that direction at an appropri- ate mid-to-late layer produces a monotonic, causally faithful shift of the corresponding behavior at infer- ence time. It offers a complementary interface for personalization, which is attractive because it allows behavioral adaptation without gradient-based opti- mization or per-user model checkpoints. However, applying it to realistic personalization requires more than a fixed behavioral direction. Sparse and hetero- geneous histories provide only indirect evidence of user preferences, and the evidence relevant to one query may be inappropriate for another. The central challenge is therefore to extract useful personaliza- tion signals, preserve their effects at appropriate computational interfaces, and select the signals rel- evant to the current query. To address this challenge, we propose FleetPilot, a fine-grained, instance-tailored steering framework built around three coordinated designs. Matched contrast varies the conditioning con- text while holding a historical query and its recorded response fixed. By contrasting activations under user-relevant and user-irrelevant contexts, it cap- tures changes associated with personalized condi- tioning without introducing differences in response content. Fine-grained hooking constructs a pair of steer- ing directions for each historical interaction by con- trasting the modelâs activations under user-relevant and user-irrelevant contexts. Rather than collaps- ing the contrastive signal into a single post-block direction, it separately captures the post-attention residual and the MLP output. These two hook points expose complementary computational path- ways and preserve distinct preference-induced ef- fects that whole-layer steering may conflate. Query-conditioned aggregation then adapts these historical directions to the current query. At infer- ence time, FleetPilot retrieves the most semantically relevant interactions from the userâs steering-vector dictionary and aggregates their attention- and MLP- side directions separately. The resulting pair of query-specific directions is applied through position- matched sequential injection, enabling the model to condition on the most pertinent preference evi- dence in a single auto-regressive generation. This design supports immediate adaptation to sparse ob- servations and evolving interests without gradient Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot5 updates or per-user model deployment. Overall, FleetPilot constructs a pair of query- specific interference vectors per user query by aggre- gating contextually relevant steering vectors, which are prebuilt from fine-grained contrasting modelâs internal activations under the positive (personal) prompts and the negative (impersonal) prompts. Be- cause the base model remains frozen and the query- specific steering directions are injected only tran- siently into intermediate activations, FleetPilot re- quires neither gradient-state storage nor per-user model checkpoints, while introducing only mod- est inference-time latency. Empirical studies show that FleetPilot outperforms all the competing base- lines across different validation scenarios and show extreme superiority in dynamic and data-sparse en- vironments. Contributions are summarized below: ⢠Comprehensive summarization of existing per- sonalization paradigms and their fragility to- wards realistic challenges. Point out the cor- rectness of inference-time steering as a training- free, flexible, robust, and low-cost paradigm. We analyze existing personalization paradigms, including prompt-based conditioning, paramet- ric adaptation, and personalized reward model- ing, through the lens of data scarcity, temporal preference drift, and query-dependent preference heterogeneity. This analysis identifies their re- spective limitations and motivates inference-time activation steering as a training-free, flexible, and resource-efficient interface for personalization, establishing the design requirements for robust adaptation under realistic conditions. â˘Methodology novelty: We propose FleetPilot, which adapts activation steering to personaliza- tion through three coordinated designs. First, matched contrast holds historical queries and re- sponses fixed while varying user-relevant and user-irrelevant contexts, extracting steering sig- nals associated with personalized conditioning. Second, fine-grained hooking preserves separate directions from post-attention residuals and MLP outputs, while sequential dual-injection applies them at their corresponding computational inter- faces for nuanced control. Third, query-conditioned retrieval and aggregation move beyond fixed concept- , behavior-, or user-level directions to instance- tailored interventions, flexibly selecting the his- torical signals relevant to each query without up- dating model parameters. â˘Empirical validation and analysis: Experiments on various datasets including short content gener- ation, long content generation, and personal web function calling demonstrate FleetPilot âs consis- tent superiority over all competing baselines. Par- ticularly, FleetPilot excels in data-scarce scenario and fast-shifting distribution, with light overhead and no training cost. These justify FleetPilot as a robust and real-world deployable method. 2 Related Work 2.1 Prompt-Based Personalization Prompt-based methods leverage the in-context learn- ing capabilities of LLMs to personalize generation without changing model parameters. They retrieve relevant records from a userâs personal corpus or construct profiles that summarize historical prefer- ences, and supply this information alongside the current query. LaMP Salemi et al. [2023] stud- ies retrieval-augmented personalization across lan- guage tasks, while profile-based approaches com- press interaction histories into more concise pref- erence descriptions Cai et al. [2025]. Memory and retrieval frameworks further investigate how to orga- 6Front. Comput. Sci., 2026, 0(0): 1â22 nize and retrieve contextual evidence Madaan et al. [2022], Qian et al. [2024], Shi et al. [2025]. These methods are flexible and can incorporate new records without retraining. Their effectiveness nevertheless depends on retrieval quality, context composition, and the modelâs ability to use the sup- plied evidence. Finite context windows constrain how much history can be included, and adding noisy or weakly relevant records can reduce performance even within that limit Lin et al. [2024], Science and Teams [2025], Vuong et al. [2021]. FleetPilot is complementary to this paradigm: retrieved text supplies explicit evidence, while query-conditioned steering supplies an additional activation-space in- tervention. 2.2 Parametric Adaptation Parametric personalization incorporates user infor- mation through model optimization, commonly us- ing parameter-efficient fine-tuning such as Low- Rank Adaptation (LoRA) Hu et al. [2022]. One line of work trains a shared adapter using multi- ple usersâ histories, allowing the model to learn collaborative signals and task-specific response pat- terns. ReLLa Lin et al. [2024] exemplifies this ap- proach. Another line maintains user-specific adapta- tions. OPPU Tan et al. [2024b] combines personal PEFT parameters with non-parametric knowledge from retrieval and profiles, while PER-PCS Tan et al. [2024a] composes shared personal parameter components to construct an adaptation suited to a target userâs history. Compared with full fine-tuning, these methods reduce the number of trainable parameters, but adap- tation still depends on optimization and sufficient training evidence. Shared adapters must accom- modate heterogeneous users, whereas user-specific adapters introduce storage, loading, and mainte- nance costs. Changes in user preferences can re- quire renewed fitting or adapter recomposition. In contrast, FleetPilot leaves the base model fixed and represents personal evidence as a retrievable dictio- nary of activation directions. Updating this dictio- nary requires vector preparation rather than per-user gradient-based training. 2.3 Personalized Reward Modeling Personalized reward modeling represents user pref- erences through a scoring function over candidate responses. Such scores can guide rejection sam- pling Liu et al. [2023] or best-of-Nselection Gui et al. [2024]. Hydra Zhuang et al. [2024], for ex- ample, learns user-specific scoring adapters from positive and negative examples. Other approaches use an LLM as a personalized judge, conditioning its evaluation on user information rather than neces- sarily training a separate reward model Ryan et al. [2025]. Learned reward models depend on the availabil- ity and coverage of preference feedback, while both learned and prompted judges must distinguish user- specific suitability across diverse outputs. In candidate- selection pipelines, quality is bounded by the avail- able candidates, and increasing their number raises generation and evaluation costs. FleetPilot instead steers the generation process itself, without requir- ing explicit pairwise preference labels or multi- ple completed candidates for each inference query. This distinction concerns the illustrated inference pipeline rather than a claim that all uses of reward models require best-of-N decoding. 2.4 Inference-Time Activation Steering Activation steering modifies intermediate represen- tations to influence a modelâs behavior without up- dating its parameters. Prior work studies additive Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot7 activation interventions and contrastive direction construction Frandsen et al. [2022], Rimsky et al. [2024], Turner et al. [2023], while representation engineering explores the monitoring and control of high-level model representations Zou et al. [2023]. Persona- and style-oriented approaches further in- vestigate steering directions associated with behav- ioral traits or personalized generation Chen et al. [2025b], Zhang et al. [2025a]. Within this family, methods differ in how con- trasts are constructed, where directions are extracted and injected, and whether the intervention varies across inputs. Figure 1 organizes the design of Fleet- Pilot along these three axes. It contrasts contexts while holding a historical query and response fixed, preserves separate post-attention and MLP-output directions, and retrieves historical directions accord- ing to the current query. These choices distinguish its matched, fine-grained, instance-tailored formula- tion from response-contrast, whole-layer, or fixed user-level alternatives. The basic steering formula- tion is introduced next; the concrete construction and inference procedure are given in Section 4. 3 Preliminary 3.1 Personalization Setting LetM θ denote a base language model with parame- ters θ. For user u, let D u =(X i , Y i ) n u i=1 (1) denote the historical interactions available before the current queryX, whereX i is a historical query andY i is its recorded response. The outputYmay be natural-language text or a task-specific textual seri- alization of a structured action. The personalization objective is to generate a response aligned with the userâs current, query-dependent preferences using the evidence in D u . The history provides observations rather than a complete specification of those preferences. Its rele- vance can vary across queries and over time. Under the temporal evaluation protocol, only information available before the evaluated interaction may be used for personalization. We writeD u without an explicit time index for simplicity. Historical re- sponses are available during preparation; the un- known response to the current query is not available for retrieval or direction construction. 3.2 Personalization Interfaces The paradigms in Figure 1 differ in the interface through which user evidence influences generation. LetĎ(X, D u ) retrieve or summarize the personal con- text relevant toX, and letâĽdenote textual concate- nation. Prompt-based personalization conditions the model on P u (X) = XâĽĎ(X, D u ),Y âźM θ (¡| P u (X)). (2) Actual inputs use the modelâs formatting template; the concatenation notation abstracts away those for- matting details. For user-specific parametric adaptation, personal evidence is encoded in a parameter change δ u : Y âźM θ+δ u (¡| P u (X)).(3) The prompt may be the query alone or may addi- tionally contain retrieved context. Shared-adapter methods instead learn an adaptation across users and use conditioning information to distinguish their re- quests. For reward-based candidate selection, letY ( j) N j=1 be candidate responses generated for the current query and letr u (X, Y ( j) ) be a user-conditioned score. 8Front. Comput. Sci., 2026, 0(0): 1â22 Best-of-N selection takes the form Y â = arg max Y ( j) , 1⤠jâ¤N r u (X, Y ( j) ).(4) The scorer may be learned from preference data or implemented by a prompted judge. This expression describes the candidate-selection setting, not every possible use of a reward model. 3.3 Activation Steering Activation steering introduces personalization at an internal computational interface instead of requiring a parameter update. At a selected layerL, leth L t â R d denote the activation at token positiont. A basic additive intervention is e h L t = h L t + Îłâ L ,(5) where â L â R d is a steering direction andÎł â Rcon- trols its strength. The modified activation is passed to the remaining computation whileθremains un- changed. The layer, interface, token positions, and intervention strength are design choices rather than universal constants. A common construction uses a contrastive pair of complete inputs Z + and Z â : â L = h L (Z + )â h L (Z â ),(6) whereh L denotes an activation pooled over a spec- ified token set. The meaning of this direction de- pends on what differs between the paired inputs and which activations are pooled. In a response contrast, the continuations differ; in the matched contrast used by FleetPilot, the historical query and response remain fixed while the conditioning con- texts differ. Section 4 specifies the context construc- tion, response-token pooling, and two extraction interfaces. The linear representation hypothesis motivates additive steering by suggesting that some high-level attributes are approximately encoded as directions in activation space Rimsky et al. [2024], Turner et al. [2023]. Separately, a local approximation characterizes how a sufficiently small intervention propagates. LetF L map the activation at the selected interface to downstream logits, with the other inputs fixed. IfF L is twice continuously differentiable with bounded second derivative in a neighborhood ofh, then for a perturbation δ within that neighborhood, F L (h + δ)â F L (h) = J F L (h)δ +O(âĽÎ´âĽ 2 2 ),(7) whereJ F L (h) is the local Jacobian. This approxi- mation does not establish global linearity, guaran- tee a monotonic behavioral change, or show that a contrast isolates user preference. Steering quality still depends on the evidence, extraction interface, intervention strength, and downstream nonlinear computation. For personalization, a fixed user-level direction can be replaced by a query-conditioned direction â L u (X) obtained from historical evidence. FleetPi- lot instantiates this idea with a dictionary of paired, fine-grained directions and query-dependent retrieval. It also retains retrieved textual context in the gen- eration prompt, so activation-space steering and prompt-based conditioning act as complementary information channels. The sequential dual-injection procedure is detailed in Section 4.3.2. 4 Methodology 4.1 Overview Despite the increasing demand for personalized LLMs services, existing methods face critical prob- lems in dynamic and data-sparse environments. Lim- itations of existing methods include structural in- terference, slow adaptation, and context window Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot9 Personal Corpus Logs from Other Users Query X ... Retriever ... ... Query Corrupted Context Query Corrupted Context Query Relevant Context Contr astive Pr ompts Constr uction K ey: Value: Steer ing Vector s Fine-Gr ained Hooking Attention MLP ... Pooling Quer y-Conditioned Aggr egation Stage I : Steer ing Vector s Pr epar ation Stage I I : I nstance-Tailor ed Per sonalized Adaption and I nfer ence ... ... Attention MLP Desired Output Top K Pulse Re-Pulse Steer ing Dictionar y Constuction Sequential Dual-I nj ection Response Response Response Query X Query X Fig. 2Overview of FleetPilot. 1) Steering Vectors Preparation. During this stage, we construct contrastive prompts from user logs, where relevant context retrieved from personal corpus is concatenated with target query to form the positive sample and irrelevant context sampled from other users is concatenated with target query to form the negative sample. For each positive sample, we generateKnegative samples. Each pair is then fed into the LLM in the teacher forcing mode, during which the response tokensâ activations of attention and MLP blocks are hooked. We store the difference between two sample activations of a pair to serve as the steering vector, with the text of each pair being the key for indexing convenience. 2) Instance-Tailored Personalized Adaptation and Inference. During this stage, we sample from target userâs steering vectors set to interfere model for personalized output. Concretely, we rank the similarity between query text and sample pair text to select the top-k steering vectors, which are then aggregated and injected into LLM for personalized adaptation. constraints, which hinder their applicability in real- world scenarios with fast-changing user interests and limited interaction data. In this paper, we pro- pose FleetPilot, a novel inference-time personal- ization framework that treats personalization as a sample-level activation shift rather than a paramet- ric update. FleetPilot operates through a two-stage pipeline designed for instance-aware personaliza- tion without model training: â˘Offline Steering Vectors Preparation: For each user, we construct a personalized dictionary of steering vectors from historical interactions. Each vector is derived by contrasting model activations under user-relevant and user-irrelevant contexts, using a Fine-Grained Hooking mechanism that separately extracts signals from attention and MLP layers. â˘Online Inference with Instance-Tailored Dual Steering: During inference, for a target query, FleetPilot dynamically selects and aggregates the most relevant steering vectors from the userâs dic- tionary based on semantic similarity. The aggre- gated vectors are then injected into the modelâs forward pass at a predefined layer via a Sequen- tial Dual-Injection mechanism, enabling real- time, query-specific personalization. This ap- proach ensures that each query receives a tailored activation shift, offering instance-level adapta- tion. The position-matched injection offers rein- forced interference with subtle signals from each component, under the constraints of geometric consistency and linear compatibility. These designs enable FleetPilot to adapt instantly with fine-grained signals to meet evolving user in- terests, with minimal memory and latency overhead, making it particularly suitable for dynamic and data- 10Front. Comput. Sci., 2026, 0(0): 1â22 sparse environments where traditional parametric methods struggle. 4.2 Steering Vectors Preparation Based on the time-splitting setting, we use old user logs for steering vectors preparation. 4.2.1 Contrastive Prompt Construction We construct contrastive prompt by varying the con- ditioning context while holding the historical query and target response fixed. For useru, let (X i , Y i ) denote a historical interaction, whereX i is the query andY i is its recorded response. A relevant context is retrieved from the userâs history, excluding the current interaction, whereas a negative context is sampled from other usersâ histories: C + i = Retrieve X i , D u \(X i , Y i ) ,(8) C â i = Sample(D ).(9) Here,D denotes the available histories of users other thanu. The positive context supplies user- relevant evidence, while the negative context serves as a user-irrelevant contrast. Excluding the current interaction prevents its target response from being directly retrieved into the conditioning history. All context candidates must be restricted to the infor- mation available under the evaluation protocol. Using the same formatting template, we construct Z + i = Format(X i , C + i , Y i ),(10) Z â i = Format(X i , C â i , Y i ).(11) The superscripts + andârefer to the contextual conditions, not to different responses: the same X i andY i are supplied in both inputs. For struc- tured outputs,Y i denotes the recorded response in its task-specific textual serialization, which is also kept identical across the pair. This construction examines how relevant versus irrelevant conditioning changes the modelâs repre- sentation of a fixed response, rather than comparing independently generated continuations. 4.2.2 Fine-Grained Activation Hooking Consider a pre-normalized Transformer block at a selected layerLfor a complete input sequenceZ. Letr L t (Z)â R d denote the incoming residual state at tokent, and letr L â¤t (Z) denote the corresponding causal prefix. The block computes a L t (Z) = Attn L LN L a (r L â¤t (Z)) t ,(12) u L t (Z) = r L t (Z) + a L t (Z),(13) m L t (Z) = g L (u L t (Z)),(14) v L t (Z) = u L t (Z) + m L t (Z),(15) whereg L =MLP L ⌠LN L m . These equations distin- guish three interfaces: the post-attention residualu L t , the MLP outputm L t before residual addition, and the post-block residualv L t . FleetPilot extracts directions fromu L t andm L t , preserving the signal entering the MLP separately from the signal it contributes to the residual stream. We process each complete inputZ Âą i with the frozen language model in teacher-forcing mode: the recorded responseY i is supplied in both passes rather than sampled from the model. The fine-grained hook- ing vectors are then captured from the matched contrast mode. â˘Matched Contrast. LetT + i andT â i denote the nonempty position sets corresponding toY i in the respective fully formatted and tokenized in- puts. Only response-token activations are pooled; query, context, formatting-only, and padding po- sitions are excluded. Our working hypothesis is that, for a fixed query and response, contrasting relevant and irrelevant Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot11 contexts reveals an activation displacement asso- ciated with user-specific conditioning. Holding the query and response fixed removes differences in their content from the pair construction. In- terpreting the remaining contrast as a preference- related displacement additionally requires other contextual effects to be balanced in expectation. For a fixed user, layer, and interface, consider the idealized decomposition h L,Âą p,i = Îź L p (X i , Y i ) + Ď L,Âą p,i + Îľ L,Âą p,i ,(16) where Îź L p is a shared queryâresponse component, Ď L,Âą p,i is the preference-related contribution under each context, andÎľ L,Âą p,i collects remaining contex- tual effects. If E h Îľ L,+ p,i â Îľ L,â p,i | X i , Y i i = 0,(17) then the contrastive estimator satisfies E[â L p,i | X i , Y i ] = E h Ď L,+ p,i â Ď L,â p,i | X i , Y i i .(18) The expectation is over context-pair construction for the fixed interaction. This is a conditional identification statement, not a guarantee of the construction procedure: context topic, length, fac- tual content, and induced positional shifts may contribute systematically to the contrast. Accord- ingly, the method uses empirical steering direc- tions without claiming that every direction iso- lates an unbiased preference effect. ⢠Fine-Grained Hooking. The term fine-grained refers to retaining two within-block contrastive signals rather than collapsing them into a single post-block difference. Attention steering modu- lates the query-key-value interactions (capturing stylistic focus patterns), while MLP steering ad- justs the feed-forward transformations (capturing lexical and semantic preferences). These oper- ate on distinct representational subspaces. We therefore capture these complementary signals. Concretely, for pâattn, ml p, define h L,Âą p,i = 1 |T Âą i | X tâT Âą i h L p,t (Z Âą i ),(19) whereh L attn,t =u L t andh L ml p,t =m L t . The paired steering directions are â L attn,i =h L,+ attn,i âh L,â attn,i ,(20) â L ml p,i =h L,+ ml p,i âh L,â ml p,i .(21) The two response representations are averaged separately before subtraction. 4.2.3 Steering Dictionary Construction To support query-conditioned retrieval, each record is associated with the representation e i = Z + i ,k i = Ď(e i ),(22) whereĎis the retrieval encoder, distinct from the frozen language model used for activation extrac- tion. The retrieval text is the positive-context se- quence, containingX i ,C + i , and the historical re- sponseY i . The negative context is used to estimate the activation contrast but is not included in this retrieval key. The historical response is available during preparation; the target response to the cur- rent inference query is never used for retrieval. At layer L, the user-specific dictionary is S u = n k i , â L attn,i , â L ml p,i o N u i=1 .(23) Thus, each retained contrastive record provides both a semantic retrieval key and a pair of activation- space directions. This separation allows the histori- cal evidence to be indexed in the encoderâs represen- tation space while its behavioral effect is expressed in the base modelâs hidden space. For clarity, the subsequent formulation uses one selected layer; the same construction applies separately to each layer when multiple layers are used. 12Front. Comput. Sci., 2026, 0(0): 1â22 4.3 Input-Aware Personalized Inference 4.3.1 Query-Conditioned Retrieval and Aggrega- tion Given a queryX, FleetPilot selects relevant entries fromS u rather than applying a fixed average over the userâs entire history. Let Ě Xdenote the task- specific retrieval text extracted fromX. We use a Contriever encoder with mean pooling over non- padding tokens and score each entry by the unnor- malized inner product s i (X) = Ď( Ě X) ⤠k i .(24) For a nonempty dictionary and a positive retrieval budgetK, letK u =min(K, N u ). The selected index set is I u (X) = TopK iâ1,...,N u s i (X), K u ,(25) whereTopKreturns the indices of theK u highest scores. The selected directions are aggregated inde- pendently at the two interfaces: s L p (X) = 1 K u X iâI u (X) â L p,i ,pâattn, ml p.(26) The generation prompt additionally includes retrieved textual history: P u (X) = Format X, H u (X; B) ,(27) whereH u (X;B) denotes history retrieved with a prompt budgetB. The budgetsBandKgovern dif- ferent information channels: the former determines explicit contextual evidence, whereas the latter de- termines activation-space steering. Their retrieved records need not coincide. This distinction makes the role of steering explicit even when it is com- bined with retrieval-augmented prompting. 4.3.2 Sequential Dual-Injection Afterwards, we perform a sequential dual-injection for inference-time enhancement. Overall, this position- matched injection ensures: ⢠Geometric consistency: The intervention occurs at the same representational locus where the steer- ing signal was derived. â˘Linear compatibility: Per the linear represen- tation hypothesis, additive vectors in the same latent space propagate without nonlinear distor- tion through subsequent layers. Concretely, at layerL, each aggregated direction is applied at the same computational interface from which it was extracted. For a token positiontre- ceiving an intervention, sequential dual-injection computes e u L t = u L t + Îąs L attn (X),(28) e m L t = g L ( e u L t ) + βs L ml p (X),(29) e v L t = e u L t + e m L t .(30) The modified block output is then passed to the re- maining layers and the language-model head. The coefficientsÎąandβindependently regulate the attention- side and MLP-side interventions. Setting both coef- ficients to zero recovers the unsteered forward com- putation for the same generation prompt; setting either coefficient to zero yields the corresponding single-site intervention. The order of these operations is essential. The attention-side perturbation first changes the residual entering the MLP, so the MLP is evaluated on e u L t , not on the originalu L t . The MLP-side direction is then added to this recomputed output before resid- ual addition. The resulting operation is therefore not, in general, equivalent to addingÎąs L attn +βs L ml p to the unmodified block output. Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot13 In the cached dual-site configuration, these inter- ventions are applied during both prompt processing and autoregressive decoding. This inference-time scope is distinct from the response-token ranges used to estimate the directions. Matching extraction and injection interfaces establishes where each vec- tor acts; it does not require identical token ranges between preparation and inference. 5 Experiment 5.1 Setup 5.1.1 Datasets To evaluate FleetPilot under different scenarios, we utilize datasets including short personalized con- tent generation (headline generation 2) ), long person- alized content generation (abstract writing 3) ), and personal web function calling (PersonalWAB 4) ) for evaluation. Each of them contains user id, corre- sponding profile, and interaction histories. Table 1 Statistics of Used Datasets. Datasets#Users Avg. Profiles Avg. Text Length (per user) Avg. Text Length (per sample) Headline Generation 20068.262,25733 Abstract Writing 200117.9118,369156 PersonalWAB (product) 20060.7220,348335 PersonalWAB (interaction) 20012.265654 â˘News-Headline Generation. This dataset Misra [2022] evaluates the ability of a large language model to capture the stylistic patterns of an author, querying the LLM to generate a headline given the input news article. 2) https://lamp-benchmark.github.io/ 3) https://longlamp-benchmark.github.io/download 4) https://github.com/HongruCai/PersonalWAB/tree/main/PersonalWAB â˘Abstract Writing. This dataset evaluates LLMsâ ability to distill complex ideas and generate accu- rate, concise, and coherent output over the span of multiple paragraphs on domain-specific tasks. The input of the model is the title of the paper along with some keywords to guide the content. The expected output is an abstract conditioned on the title and keywords in the userâs styleKumar et al. [2024]. ⢠PersonalWAB. This dataset integrates personal- ized instruction comprehension and action execu- tion, where LLMs must infer personalized user requirements and preferences to determine which Web function to call and formulate the correspond- ing function parameters. Subsequently, the results of these function calls are returned to usersCai et al. [2025]. We sample 200 users from the original dataset, based on the time-split setting, earlier question- answer (QA) pairs form the train set and the latter pairs form the test set. The statistics of used datasets are summarized in Table 1. 5.1.2 Baselines We evaluate FleetPilot against: 1) Prompt-based method (In Context Learning) and 2) Parametric adaptation methods including One-Peft-All-UserLin et al. [2024] and One-Peft-Per-User Tan et al. [2024a,b]. 3) Personalized reward modelsRyan et al. [2025], 3) Steering-based baseline Chen et al. [2025b], Zhang et al. [2025a], Zou et al. [2023]. 5.1.3 Metric We use Rouge Values Lin [2004] and METEOR Banerjee and Lavie [2005] to measure the quality of content generation and use Overall Accuracy (ACC) to measure the quality of web function calling. 14Front. Comput. Sci., 2026, 0(0): 1â22 Table 2 Main results on content generation and personal web function calling. All baselines are done with n-shot demon- strations, consistent with the ICL setting. (The symbol * indicates statistically significant improvement with p-valueâ¤0.001. Hyperparameters sensitivity is discussed in Appendix 5.4.4.) DatasetsHeadline GenerationAbstract WritingPersonalWAB MethodsRouge-1Rouge-LMETEORRouge-1Rouge-LMETEORACC DirectZeroShot0.14320.12970.10020.35030.20240.19990.6017 In Context Learning k=10.13890.12870.09940.36480.20930.22980.6767 k=30.16430.15150.12570.37010.21040.24430.8415 k=50.16650.15060.12670.37710.21530.25170.8330 k=100.16300.14880.12710.37680.21270.25590.7966 Parametric Adaptation Rella0.16350.14630.12630.38180.21410.25620.8330 OPPU0.17790.16130.13480.38480.21460.25800.8394 PER-PCS0.17630.15770.12550.36020.21690.24460.8402 Reward ModelsSynthesizeMe0.16720.14900.12880.39060.21560.25650.8275 Steering REPE0.16850.15310.13250.37830.21340.24700.8501 PersonaVector (Response)0.17140.15400.13390.39540.22910.25670.8116 PersonaVector (All)0.17000.15360.13060.38250.21690.25120.8308 PersonaVector (Prompt)0.16400.14810.12670.38340.21480.24940.8458 StyleVector (Mean)0.16940.15400.13450.37900.21780.25470.8137 StyleVector (Logistic)0.17120.15390.13420.38310.21360.25050.8436 FleetPilot0.1816*0.1666*0.1473*0.3990*0.2306*0.2609*0.8566* Table 3 Impact of hooking strategy. (The symbol * indicates statistically significant improvement with p-value⤠0.001.) DatasetsHeadline GenerationAbstract WritingPersonalWAB MethodsRouge-1Rouge-LMETEORRouge-1Rouge-LMETEORACC FleetPilot Fine-Grained Hooking Post-Attention0.17520.15900.13840.39380.22710.25560.8588 Post-MLP0.17040.15440.13010.38260.21330.25030.8501 Post-Block0.17370.15750.13690.39460.22770.25560.8458 Hybrid Strategy0.1816*0.1666*0.1473*0.3990*0.2306*0.2609*0.8566 5.2 Research Questions â˘(RQ1) Does FleetPilot outperform the competing baselines? â˘(RQ2) Does fine-grained hooking bring perfor- mance gain? ⢠(RQ3) How is the robustness of FleetPilot against heterogeneous/fast changing distribution compared to other methods? â˘(RQ4) How is the data efficiency of FleetPilot compared to other methods? Particularly, under high data sparsity? â˘(RQ5) How is the overhead brought by FleetPilot? Is FleetPilot real-world feasible? ⢠(RQ6) How is the hyperparameters sensitivity of FleetPilot? 5.3 Main Results (RQ1-RQ2) The main results of the experiments are included in Table 2. (RQ1). From the results, we can see that Fleet- Pilot outperforms all other baselines. Concretely, parametric methods show better performances than the prompt-based methods, showcasing the bene- fits of personalization in parameters to better digest user patterns. Furthermore, FleetPilot outperforms traditional personalized methods, justifying the su- periority of the steering-based methodology. (RQ2). To evaluate the efficacy of steering vec- tors beyond layer-wise granularity, we investigate different hooking strategies: â˘Post-Attention. The direction is extracted from Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot15 Table 4 Results of different methods against heterogeneous distribution among temporal dimension. (Headline Generation)Time-Range 1Time-Range 2Time-Range 3 MethodsRouge-1Rouge-LMETEORRouge-1Rouge-LMETEORRouge-1Rouge-LMETEOR In Context Learning0.16200.13620.11570.12910.11090.08230.14910.13680.1216 Parametric adaptationOPPU0.17510.15440.12450.14960.13390.11210.15600.13970.1425 Personalized Reward ModelsSynthesizeMe0.15830.13800.11090.14970.13550.09740.15510.13360.1295 Steering REPE0.16590.15170.12190.14350.12760.09290.16570.14740.1325 PersonaVector0.17090.15360.12200.14750.13630.09280.16420.14030.1341 StyleVector0.15560.14440.11180.15080.13500.09610.16020.13590.1272 FleetPilot0.17870.15540.12510.15310.14110.12190.17820.14240.1448 and injected into the post-attention residualu L , after the attention skip addition and before the MLP normalization. â˘Post-MLP. The direction is extracted from and injected into the raw MLP sub-layer outputm L , before its addition to the residual stream. ⢠Post-Block. The direction is extracted from and injected into the post-block residualv L =u L +m L . â˘Hybrid Strategy. Position-matched directions are injected sequentially atu L andm L , following Equations (28)â(30). Table 3 reports the performance of the hooking variants. The Hybrid Strategy improves content- generation metrics by 0.36%â1.64% over the strongest single-site or post-block alternative ( p⤠0.001). 5.4 Ablation Studies 5.4.1 (RQ3) Robustness to Drifting Distribution In this section, we study the robustness of different methods against drifting distributions. Particularly, we use the realistic temporal drift setting, where samples are split based on the timestamps and only earlier logs are available. We then test modelsâ per- formances on different time ranges. The results of different methods are summarized in Table 4. From the results, we can see that FleetPilot out- performs all methods in each time-range, showing FleetPilotâ strong superiority in realistic temporal drift setting. Since the parametric methods cannot quickly realign to rapidly changing user preferences Sample fr om Cluster 1 -I nput: 'Generate a headline for the following article: One thing is certain, I still stumble over the question when people ask, ?How many children do you have?? I dread the decision' -Output: 'I Still Stumble Over The Question, How Many Children Do You Have?' -Categor y: Personal Essay / Parenting & Family -Output Style: First-person Narrative / Reflective Sample fr om Cluster 2 -I nput: 'Generate a headline for the following article: Filmmaker Ava DuVernay, best known for the critically-acclaimed film ?Selma? about the equal voting rights movement led by' -Output: 'Filmmaker Ava DuVernay?s Oscar-Nominated Documentary '13th' Is A Wake-up Call' -Categor y: Entertainment News / Documentary Film & Social Commentary -Output Style: Informative & Advocacy-Driven Sample fr om Cluster 3 -I nput: 'Generate a headline for the following article: Eduardo Feldman?s photo project celebrates the work of the people who took to the streets to help.' -Output: 'After The Earthquake In Mexico, Everyone Was A Hero' -Categor y: Humanitarian News / Community Resilience -Output Style: Inspirational & Inclusive Histor ical Behavior s fr om User 10000051 Fig. 3 Illustration of the heterogeneous data. due to their lock to sluggish and offline training loops. While FleetPilot offers inference-time and instance-tailored adaptation, and therefore it offers flexibility and robustness to heterogeneous distribu- tions. This feature makes FleetPilot distinctive in scenario where user patterns change fast, which is a commonly-seen and important problem in LLMs personalization. 16Front. Comput. Sci., 2026, 0(0): 1â22 Table 5 Data efficiency analysis of different methods (Abstract Writing). #Used Data5% (Avg. 5.82)15% (Avg. 17.46)25% (Avg. 29.10)50% (Avg. 58.72) MethodsRouge-1Rouge-LMETEORRouge-1Rouge-LMETEORRouge-1Rouge-LMETEORRouge-1Rouge-LMETEOR Parametric adaptation ReLLa0.37010.20360.23820.37050.21340.24410.37300.21370.24370.37580.21850.2508 OPPU0.37640.20000.24880.37880.20330.25150.37940.21320.25230.38340.22000.2571 PER-PCS0.36000.21540.24450.36440.21580.24910.35970.21500.24410.35940.21520.2442 Personalized Reward Models SynthesizeMe0.38200.20910.25640.38530.21070.25840.38620.21090.25850.38750.21230.2575 Steering StyleVector 0.38270.21510.24990.38510.21580.25180.38190.21470.24880.38290.21300.2507 PersonaVector0.39080.22560.25160.39080.22770.25430.39530.22680.25440.39500.22880.2580 REPE0.37600.21000.24460.37790.21030.24400.38060.21330.24660.37880.21430.2495 FleetPilot0.39200.22360.25420.39260.22470.25610.39560.22570.25630.39740.22700.2594 To further illustrate the discrepancy existing in the drifting distributions, we select different samples of the same user 10000051 as illustrated in Figure 3. These illustrations demonstrate the heterogeneous content themes, audience engagement strategies, and emotional tones existing in Question-Answer pairs of the same user. 5.4.2 (RQ4) Robustness to Data Scarcity and Data Efficiency Discussion Despite the robustness of heterogeneous distribu- tion, data sparsity is also an important problem in LLMs personalization, where only a small amount of data is available to conduct effective personal- ization. Therefore, we study the data efficiency of different methods in this section. Concretely, we use different amounts of personalized data to train models/serve as steering materials. The results are summarized in Table 5 and visualized in Figure 4. The results show that FleetPilot remains effective with limited personalization histories, achieving the highest ROUGE-1 across all four data fractions. Particularly, under extreme data sparsity (<5 shots), steering presents strong superiority over the para- metric methods. The increasing trend indicates that FleetPilot preserves much of its performance under substantial history reduction while still benefiting from additional evidence. A plausible explanation for this effectiveness un- der sparse histories is that activation steering reuses 5152550 Percentage of Data Used (%) 0.36 0.37 0.38 0.39 0.40 Rouge-1 Score 0.3701 0.3764 0.3600 0.3820 0.3827 0.3908 0.3760 0.3920 0.3705 0.3788 0.3644 0.3853 0.3851 0.3908 0.3779 0.3926 0.3730 0.3794 0.3597 0.3862 0.3819 0.3953 0.3806 0.3956 0.3758 0.3826 0.3594 0.3875 0.3829 0.3950 0.3788 0.3974 Data Efficiency Analysis - Rouge-1 Score (All Methods) ReLLa OPPU PER-PCS SynthesizeMe StyleVector PersonaVector REPE FleetPilot Fig. 4 Data efficiency analysis of different methods. behavioral structure already encoded in the pre- trained model through a constrained intervention, rather than estimating a parameter update from lim- ited user-specific data. For FleetPilot, matched queryâresponse contrasts may reduce variation un- related to contextual conditioning, while response- token averaging may stabilize the extracted signals. Query-conditioned retrieval further selects poten- tially relevant historical directions instead of merg- ing all available records into a single user-level rep- resentation. Together, these mechanisms offer a plausible account of its effective use of limited his- tories. 5.4.3(RQ5) Overhead Discussion and Deploy- ment Analysis We sample 200 datapoints from each dataset and conduct the evaluation. The results are displayed in Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot17 Table 6. Table 6 Latency Analysis. Datasets (200 samples) Headline Generation Direct Inference+ FleetPilot Latency (s)63.2372.49 Datasets (200 samples) Abstract Writing Direct Inference+ FleetPilot Latency (s)670.34712.36 Datasets (200 samples) PersonalWAB Direct Inference+ FleetPilot Latency (s)1161.881343.89 The inference overhead for FleetPilot is primarily attributed to retrieving the relevant steering vectors from the set, which is the most time-consuming step, and subsequently inserting the aggregated steering signal into the LLMâs forward pass. While latency varies with the steering vector size per user and the queryâs token count, Table 6 displays the results of time latency analysis brought by FleetPilot, which is conducted on a single NVIDIA H100 device. Re- sults demonstrate that FleetPilot incurs only a light overhead during inference. This efficiency offers a crucial advantage by circumventing the huge up- load burden typically associated with transferring distinct personalized weights. We further compare the per-user storage overhead of FleetPilot with that of personalized parametric adaptation methods. Letddenote the hidden dimen- sion andNdenote the number of steering vectors retained for each user. A steering vector hasdel- ements, resulting in a storage complexity ofO(d). In our implementation,d= 4096 and each element is stored inbfloat16format, which occupies two bytes. Thus, each steering vector requires 4096Ă 2 = 8192 bytes,(31) or approximately 10 KB on disk after accounting for serialization overhead. Since fewer than 100 steering vectors are retained for each user, the total per-user storage remains below 1 MB. The resulting per-user storage complexity of FleetPilot is there- fore O(Nd).(32) By contrast, personalized parametric adaptation methods must store a separate set of LoRA weights for each user. Letrdenote the LoRA rank andL denote the number of adapted Transformer layers. The storage complexity of a personalized LoRA checkpoint is O(rdL),(33) up to constant factors determined by the number and dimensions of the adapted projection matrices. In our experiments, wherer= 8 andL= 32, each user-specific LoRA weight dictionary occupies ap- proximately 18 MB on disk. BecauseN ⪠rLunder our configuration, Fleet- Pilot requires substantially less per-user storage than personalized LoRA methods: less than 1 MB compared with approximately 18 MB per user. This difference becomes increasingly consequential as the number of users grows. Moreover, updating a user profile in FleetPilot only involves construct- ing and storing lightweight steering vectors, with- out gradient computation or optimizer-state storage. Parametric adaptation methods, in contrast, require iterative training and the associated computational and memory overhead for every personalized model update. Together with the high flexibility and ro- bustness validated above, FleetPilot is a real-world feasible method that offers superior performance gains under scarce and shifting environment. 5.4.4 (RQ6) Hyperparameters Analysis We perform grid search on (Îą,β) in a range of (0.6,1.4)Ă(0.6,1.4). The results are visualized in Figure 5. Concisely, we compute the variances 18Front. Comput. Sci., 2026, 0(0): 1â22 0.70.80.91.01.11.21.3 Beta 0.7 0.8 0.9 1.0 1.1 1.2 1.3 Alpha 0.3920.3900.3980.3940.3960.3940.395 0.3910.3950.3880.3940.3880.3980.388 0.3900.3870.3890.3950.3940.3950.391 0.3950.3960.3920.3930.3910.3890.389 0.3990.3970.3920.3880.3890.3830.391 0.3850.3940.3910.3860.3830.3840.384 0.3880.3910.3850.3910.3820.3840.391 Abstract_Writing Dataset - ROUGE-1 Metric 0.70.80.91.01.11.21.3 Beta 0.7 0.8 0.9 1.0 1.1 1.2 1.3 Alpha 0.1810.1810.1780.1790.1790.1740.175 0.1740.1810.1790.1730.1760.1770.178 0.1820.1770.1770.1780.1780.1750.178 0.1760.1770.1720.1770.1780.1750.178 0.1790.1740.1790.1810.1790.1770.179 0.1760.1790.1760.1740.1770.1780.181 0.1790.1780.1800.1810.1810.1810.172 Headline_Generation Dataset - ROUGE-1 Metric 0.70.80.91.01.11.21.3 Beta 0.7 0.8 0.9 1.0 1.1 1.2 1.3 Alpha 0.8460.8540.8540.8480.8350.8440.846 0.8440.8500.8500.8460.8390.8420.848 0.8350.8480.8540.8440.8570.8460.846 0.8460.8500.8420.8500.8480.8370.846 0.8330.8390.8310.8420.8460.8390.839 0.8240.8270.8350.8440.8220.8220.827 0.8390.8330.8240.8370.8120.8350.835 PersonalWAB Dataset - ACC Metric 0.382 0.384 0.386 0.388 0.390 0.392 0.394 0.396 0.398 ROUGE-1 Score 0.172 0.174 0.176 0.178 0.180 ROUGE-1 Score 0.815 0.820 0.825 0.830 0.835 0.840 0.845 0.850 0.855 ACC Score Fig. 5 Hyperparameter sensitivity analysis. of the values and find that resulting performances show low variance: 0.000019 on abstract writing, 0.000006 on headline generation, and 0.000092 on PersonalWAB. We further conduct one-way/two-way analysis and visualize the main effects analysis in Figure 6, Figure 7, and Figure 8.From the anova analysis, we can find thatÎąis the only systematically impactful parameter, with ι⤠1.1 as a general guideline. 0.70.80.91.01.11.21.3 alpha 0.175 0.176 0.177 0.178 0.179 0.180 Rouge Main effect of alpha 0.70.80.91.01.11.21.3 beta 0.175 0.176 0.177 0.178 0.179 0.180 Rouge Main effect of beta Fig. 6 Main effects analysis on Headline Generation. 0.70.80.91.01.11.21.3 alpha 0.384 0.386 0.388 0.390 0.392 0.394 0.396 Rouge Main effect of alpha 0.70.80.91.01.11.21.3 beta 0.386 0.388 0.390 0.392 0.394 Rouge Main effect of beta Fig. 7 Main effects analysis on Abstract Writing. Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot19 0.70.80.91.01.11.21.3 alpha 0.825 0.830 0.835 0.840 0.845 0.850 ACC Main effect of alpha 0.70.80.91.01.11.21.3 beta 0.825 0.830 0.835 0.840 0.845 0.850 ACC Main effect of beta Fig. 8 Main effects analysis on PersonalWAB. 6 Conclusion In this paper, we present FleetPilot, a novel inference- time personalization framework that addresses the critical challenges of data efficiency and adaptabil- ity in LLM personalization. Unlike traditional para- metric methods that require extensive fine-tuning, FleetPilot operates through activation steering at inference time, making it particularly suitable for scenarios with fast-changing user preferences and limited training data. The core innovations of Fleet- Pilot include: (1) matched-contrast that produces user-relevant signals, (2) fine-grained hooking that separately extracts steering signals from attention and MLP layers, capturing nuanced user prefer- ences that whole-layer approaches miss; and (3) query-conditioned aggregation that dynamically re- trieves relevant historical steering vectors based on semantic similarity to the current query, enabling instance-level adaptation without retraining. Exten- sive evaluation across diverse tasks demonstrates FleetPilotâs superior performance over both prompt- based and parameter-efficient fine-tuning baselines. Notably, FleetPilot maintains strong performance even under extreme data scarcity (as few as 5-6 examples per user) and rapidly shifting user distri- butions, scenarios where traditional methods exhibit significant performance degradation. AcknowledgementsWe thank the anonymous reviewers for their constructive comments and suggestions, which helped us refine the methodological presentation, strengthen the experi- mental evaluation, and clarify the contributions and limitations of this work. References Satanjeev Banerjee and Alon Lavie. Meteor: An au- tomatic metric for mt evaluation with improved correlation with human judgments. In Proceed- ings of the acl workshop on intrinsic and extrin- sic evaluation measures for machine translation and/or summarization, pages 65â72, 2005. Hongru Cai, Yongqi Li, Wenjie Wang, Fengbin Zhu, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. Large language models empowered personalized web agents. In Proceedings of the ACM on Web Conference 2025, pages 198â215, 2025. Daiwei Chen, Yi Chen, Aniket Rege, Zhi Wang, and Ramya Korlakai Vinayak. PAL: Sample- efficient personalized reward modeling for plu- ralistic alignment.In The Thirteenth Inter- national Conference on Learning Representa- tions, 2025a. URLhttps://openreview.net/ forum?id=1kFDrYCuSu. Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. Persona vectors: Mon- itoring and controlling character traits in lan- guage models. arXiv preprint arXiv:2507.21509, 2025b. 20Front. Comput. Sci., 2026, 0(0): 1â22 Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A sur- vey on in-context learning.arXiv preprint arXiv:2301.00234, 2022. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv e-prints, pages arXivâ2407, 2024. Abraham Frandsen, Rong Ge, and Vatsal Sharan. Extracting latent state representations with linear dynamics from language models. In Proc. of ICML, 2022. Ramanathan Guha, Vineet Gupta, Vivek Raghu- nathan, and Ramakrishnan Srikant. User model- ing for a personal assistant. In Proceedings of the Eighth ACM International Conference on Web Search and Data Mining, pages 275â284, 2015. Lin Gui, Cristina Gârbacea, and Victor Veitch. Bon- bon alignment for large language models and the sweetness of best-of-n sampling. Advances in Neural Information Processing Systems, 37: 2851â2885, 2024. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. Muddesar Iqbal, Sohail Sarwar, Muhammad Safyan, and Moustafa Nasralla. Personalized and adap- tive e-learning systems for semantic web: a sys- tematic review and roadmap. International Jour- nal of Web Information Systems, 21(4):327â352, 2025. Yifan Jiang et al.Importance weighting can help large language models self-improve. arXiv preprint arXiv:2408.09849, 2024. Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A Rossi, Franck Dernon- court, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, et al. Longlamp: A benchmark for personalized long-form text gen- eration. arXiv preprint arXiv:2407.11016, 2024. Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74â81, 2004. Jianghao Lin, Rong Shan, Chenxu Zhu, Kounian- hua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. Rella: Retrieval- enhanced large language models for lifelong se- quential behavior comprehension in recommen- dation. In Proceedings of the ACM Web Confer- ence 2024, pages 3497â3508, 2024. Jiahong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Jieming Zhu, Minda Hu, Menglin Yang, and Irwin King. A survey of personalized large lan- guage models: Progress and future directions. arXiv preprint arXiv:2502.11528, 2025a. Jianghao Liu, Rong Shan, Chenxu Zhu, Kounan- hua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. Relia: Retrieval- enhanced large language model for lifelong se- quential behavior comprehension in recommen- dation. In Proceedings of the ACM Web Confer- ence 2024, pages 3497â3508, 2024. Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J Liu, and Jialu Liu.Statistical rejection sampling im- Kounianhua Du 1 , Jianxing Liu 1 , Kangning Zhang 1 , Wenxiang Jiao 2 , Yuan Lu 2 , Jiarui Jin 2 , Jianghao Lin 1 , Weiwen Liu 1 , Yong Yu 1 , Weinan Zhang 1 . FleetPilot21 proves preference optimization. arXiv preprint arXiv:2309.06657, 2023. Xingyu Bruce Liu, Shitao Fang, Weiyan Shi, Chien- Sheng Wu, Takeo Igarashi, and XiangâAnthonyâ Chen. Proactive conversational agents with inner thoughts. In Proceedings of the 2025 CHI Con- ference on Human Factors in Computing Systems, pages 1â19, 2025b. Aman Madaan, Niket Tandon, Peter Clark, and Yim- ing Yang. Memory-assisted prompt editing to improve gpt-3 after deployment. arXiv preprint arXiv:2201.06009, 2022. Rishabh Misra. News category dataset. arXiv preprint arXiv:2209.11429, 2022. Hongjin Qian, Peitian Zhang, Zheng Liu, Kelong Mao, and Zhicheng Dou. Memorag: Moving to- wards next-gen rag via memory-inspired knowl- edge discovery. arXiv preprint arXiv:2409.05591, 1, 2024. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. Steering Llama 2 via contrastive activation addi- tion. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15504â15522. Association for Computational Linguistics, 2024. doi: 10.18653/v1/2024.acl-long.828. Michael J Ryan, Omar Shaikh, Aditri Bhagirath, Daniel Frees, William Barr Held, and Diyi Yang. Synthesizeme! inducing persona-guided prompts for personalized reward models in llms. In Pro- ceedings of the 63rd Annual Meeting of the Asso- ciation for Computational Linguistics (Volume 1: Long Papers), pages 8045â8078, 2025. Sadaf Saleem, Muhammad Umar Aziz, Muham- mad Jawed Iqbal, and Shahid Abbas. Ai in edu- cation: Personalized learning systems and their impact on student performance and engagement. The Critical Review of Social Sciences Studies, 3 (1):2445â2459, 2025. Alireza Salemi, Sheshera Mysore, Michael Bender- sky, and Hamed Zamani. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406, 2023. Amazon Science and Research Teams.Con- text length alone hurts llm performance despite perfect retrieval.Amazon Sci- ence Publications, 2025.URLhttps: //w.amazon.science/publications/ context-length-alone-hurts-llm-performance-despite-perfect-retrieval . Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Han Li. Retrieval aug- mented generation with collaborative filtering for personalized text generation. In Proceed- ings of the 48th International ACM SIGIR Con- ference on Research and Development in Infor- mation Retrieval, SIGIR â25, page 1294â1304, New York, NY, USA, 2025. Association for Computing Machinery. ISBN 9798400715921. doi: 10.1145/3726302.3730075. URLhttps: //doi.org/10.1145/3726302.3730075. Zhaoxuan Tan, Zheyuan Liu, and Meng Jiang. Per- sonalized pieces: Efficient personalized large lan- guage models through collaborative efforts. arXiv preprint arXiv:2406.10471, 2024a. Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang.Democra- tizing large language models via personalized parameter-efficient fine-tuning. In Proceedings 22Front. Comput. Sci., 2026, 0(0): 1â22 of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6476â6491, 2024b. Alexander M Turner, Laura Thiergart, Gabriel Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Steering language mod- els with activation engineering. arXiv preprint arXiv:2308.10248, 2023. Tung Vuong, Salvatore Andolina, Giulio Jacucci, and Tuukka Ruotsalo. Does more context help? effects of context window and application source on retrieval performance. ACM Transactions on Information Systems (TOIS), 40(2):1â40, 2021. Haoming Wang, Boyuan Yang, Xiangyu Yin, and Wei Gao.Never start from scratch: Expe- diting on-device llm personalization via ex- plainable model selection.arXiv preprint arXiv:2504.13938, 2025. Shaofeng Wang and Zhuo Sun. Roles of artificial intelligence experience, information redundancy, and familiarity in shaping active learning: In- sights from intelligent personal assistants. Educa- tion and Information Technologies, 30(2):2525â 2546, 2025. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. Habeeb Yusuf, Arthur Money, and Damon Daylamani-Zad. Pedagogical ai conversational agents in higher education: a conceptual frame- work and survey of the state of the art. Educa- tional technology research and development, 73 (2):815â874, 2025. Jinghao Zhang, Yuting Liu, Wenjie Wang, Qiang Liu, Shu Wu, Liang Wang, and Tat-Seng Chua. Personalized text generation with contrastive acti- vation steering. arXiv preprint arXiv:2503.05213, 2025a. Weizhi Zhang, Yuanchen Bei, Liangwei Yang, Henry Peng Zou, Peilin Zhou, Aiwei Liu, Yinghui Li, Hao Chen, Jianling Wang, Yu Wang, et al. Cold-start recommendation towards the era of large language models (llms): A com- prehensive survey and roadmap. arXiv preprint arXiv:2501.01945, 2025b. You Zhang, Jin Wang, Liang-Chih Yu, Dan Xu, and Xuejie Zhang. Personalized lora for human- centered text understanding, 2024. URLhttps: //arxiv.org/abs/2403.06208. Yuchen Zhuang, Haotian Sun, Yue Yu, Rushi Qiang, Qifan Wang, Chao Zhang, and Bo Dai. Hydra: Model factorization framework for black-box llm personalization. Advances in Neural Information Processing Systems, 37:100783â100815, 2024. Andy Zou, Long Phan, Sarah Chen, James Camp- bell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405, 2023.