LATTE: Forecasting Peer Anchored Preference Trajectories for Personalized LLM Generation

Technical Summary: LATTE – Forecasting Peer Anchored Preference Trajectories for Personalized LLM Generation

Problem Statement
Personalized generation with frozen Large Language Models (LLMs) requires a conditioning signal that is both compact and current. Existing approaches typically rely on retrieving or summarizing user histories in text, or compressing them into static latent profiles and soft prompts. While efficient, these methods treat user history as an aggregate profile, conflating stable identity, recent behavioral drift, and item-specific content into a single representation. This static aggregation fails to capture temporal shifts in user preferences (e.g., a reviewer becoming more technical over time) and often results in stale or content-dominated profiles that do not match the user's current state. The core challenge is how to condition a frozen LLM on a user's current preference state without fine-tuning the model or adding user-specific parameters.

Methodology: LATTE Framework
The authors propose LAtent Trajectory Tracking and Extrapolation (LATTE), a framework that reframes personalization as forecasting a "peer-anchored relative preference state." The method operates in three distinct, decoupled stages:

1. Construction of Peer-Anchored Relative States:
   For each historical session, LATTE constructs a state that isolates the user's specific preference from shared item variations.

   • Peer Baseline: For a target user's interaction with an item $i_t$ at time $\tau_t$, the system identifies a set of comparable "peer" users who reviewed the same item prior to $\tau_t$.
   • Residual Calculation: The system computes a weighted average of the peers' response embeddings (the baseline) and subtracts this from the target user's response embedding.
   • Normalization: The resulting residual is normalized to form a peer-anchored relative state vector. This process mathematically removes the shared item component (under an additive embedding model assumption), leaving a signal that represents how the user differs from similar peers under the same context.

2. Trajectory Forecasting:
   Instead of aggregating these states into a static average, LATTE treats them as a time-ordered sequence.

   • A lightweight sequence predictor (e.g., a GRU, Transformer, or learned attention mechanism) is trained offline to forecast the next state in the trajectory based on the observed history of peer-anchored states.
   • The training objective is a direct regression loss (cosine similarity and MSE) to predict the next constructed state, decoupling the state prediction from the language modeling loss. This prevents the conditioning vector from collapsing into a low-rank shortcut that relies solely on item metadata.

3. State-to-Token Bridge (STB):
   The forecasted state is injected into the frozen LLM.

   • A lightweight bridge module maps the predicted latent vector into the token embedding space of the LLM.
   • This vector replaces a single placeholder token ([PREF_TOKEN]) in the prompt.
   • A natural language anchor precedes the token to instruct the LLM on how to interpret the injected state.
   • Crucially, the LLM weights remain frozen; only the bridge and the predictor are trained.

Key Contributions

• Formulation: The paper formulates frozen LLM personalization as forecasting a peer-anchored relative preference state, transforming user history from a static profile into a time-ordered latent trajectory.
• Framework Design: It introduces a modular framework (LATTE) that constructs same-item peer residual states, predicts the next state using a lightweight sequence model, and injects the forecast via a single anchored soft token.
• Theoretical and Empirical Validation: The authors provide analytical proofs (Propositions 1 and 2) demonstrating that peer anchoring reduces shared item variation and that trajectory forecasting offers a bias-variance trade-off superior to static averaging when user states drift. Empirically, they show that gains are driven by forecasting user-specific trajectory information rather than merely adding a soft prompt interface.

Experimental Results
The method was evaluated on Amazon Reviews 2023 (Books, Movies, CDs) and MemoryCD against a comprehensive set of baselines, including retrieval-based methods, summary memory, static latent profiles, difference-aware static profiles (DEP style), and soft prompt compression (PERSOMA style).

• Performance: LATTE consistently outperformed all baselines. On Amazon Reviews 2023, LATTE (using a GRU predictor) achieved an average ROUGE-L of 0.259, improving upon the strongest static latent baseline (0.245) and the strongest soft prompt compression baseline (0.245).
• Preference Fidelity: Beyond lexical overlap, LATTE showed significant improvements in history-aware pairwise win rates (64.0% vs. 58.6% for PERSOMA) and preference fidelity metrics, particularly in "Recency" alignment, which directly ties to the trajectory formulation.
• Ablation and Diagnostics:
  • Removing the peer anchor nearly eliminated the performance gain over static compression, confirming the necessity of peer anchoring.
  • End-to-end training (without decoupled regression) led to representation collapse and poor generation, validating the decoupled training strategy.
  • Diagnostics confirmed that the model does not leak peer consensus or item content, and that the learned states retain user variation without collapsing to a constant vector.

Significance and Claims
The paper claims that modeling users as evolving trajectories rather than fixed aggregates is critical for personalized generation, especially in long-history settings where recent behavior matters. The significance of LATTE lies in its ability to:

1. Decouple Representation from Injection: By separating the state prediction from the LLM conditioning, it avoids the "shortcut" problem where the model ignores the user signal.
2. Handle Temporal Drift: By forecasting the next state rather than averaging the past, it adapts to changing user preferences more effectively than static profiles.
3. Control for Content: By using peer anchoring, it isolates the user's specific preference signal from the item's inherent characteristics, providing a cleaner conditioning signal for the LLM.

The authors maintain a modest scope, noting that the method requires sufficient user history for trajectory estimation and assumes the existence of comparable peers for historical items. For cold-start users, the framework falls back to existing text or static latent baselines. The work focuses on improving the conditioning mechanism for frozen LLMs without requiring model fine-tuning.