Technical Summary: Statistical Priors for Implicit Preferences

Problem Definition

The paper addresses a critical bottleneck in the deployment of locally executed personal agents that rely on API-based remote Large Language Models (LLMs). As the ecosystem of external skills (tools) expands, agents must dynamically select the most appropriate skill for a user. However, user preferences are often implicit, manifesting as stochastic, high-frequency habits rather than explicit instructions.

Current paradigms conflate two distinct cognitive tasks:

1. Semantic Intent Parsing: Understanding explicit user instructions (e.g., "Order from HouseBrew").
2. Statistical Preference Learning: Inferring latent habits from historical interactions (e.g., a user always prefers "VibeCof'ing" for coffee).

Existing solutions typically rely on memory-augmented LLMs, where historical logs are injected into the prompt to guide a single remote model. The authors argue this approach suffers from:

• Systemic Failures: LLMs often lose logical consistency in multi-turn conversations or fail to track fine-grained statistics.
• Resource Constraints: High API latency and context window overflow make complex centralized algorithms impractical for local, privacy-preserving agents.
• Lack of Interpretability: Memory-based approaches lack mathematical rigor in modeling preference distributions.

Methodology: The LOCAL HARNESS

To resolve these issues, the paper proposes LOCAL HARNESS, a novel architecture that strictly decouples statistical preference learning from semantic intent parsing.

Core Architecture

The system operates on a "Local Statistical Estimator + Remote Exception Handler" model:

1. Local Statistical Prior (The Harness): A lightweight, on-device module that manages the user's preference distribution. It handles the high-frequency decision-making and the exploration-exploitation trade-off. The paper implements two specific priors:
   • Frequency Prior: A simple empirical success-rate table tracking $(user, domain, skill)$ counts.
   • Bandit Prior: A contextual bandit approach (specifically LINUCB) that maintains a parameter vector and a regularized Gram matrix. It uses feature hashing to create context vectors from the query, skill name, and domain, allowing for principled exploration via Upper Confidence Bounds (UCB).
2. Remote LLM (The Override Handler): The remote LLM is removed from the critical execution path for standard queries. It is invoked only to check for explicit lexical overrides.
   • Step 1 (Domain Classification): A single LLM call infers the target domain to restrict the candidate skill set.
   • Step 2 (Local Default): The local harness selects the default skill based on the statistical prior.
   • Step 3 (Semantic Override Probe): A binary probe asks the LLM: "Does this query explicitly name a specific tool?" If yes, the named tool overrides the local default; if no, the local default is executed.

Theoretical Foundation

The authors provide a theoretical analysis (Theorem 4.1) proving that under mild assumptions, this decoupled architecture strictly reduces the expected per-round regret compared to an LLM-only baseline. The regret converges to a value strictly lower than the LLM's inherent error rate, provided the local statistical prior is consistent. The proof relies on the fact that the final policy is a convex combination of the local prior and the LLM's policy, a structural property lost in memory-augmented approaches where signals are conflated.

Experimental Evaluation

Since no existing benchmark addresses implicit preference-driven skill selection, the authors constructed TOOLBENCH-60, a sandbox with 60 skills across 10 domains and synthetic users with controllable preference distributions (ranging from deterministic "one-hot" to stochastic Dirichlet distributions).

Evaluation Setup:

• Backbones: Evaluated across three diverse LLMs: GPT-5.2, DeepSeek-V4-Flash, and Qwen3-30B-Instruct.
• Baselines: Nine agents were compared, categorized into:
  • No Learning: Random, ZeroShot-LLM.
  • Statistical Only: Freq-Greedy, Pure-Bandit.
  • LLM with Memory: InContext-Memory (recent history), Profile-Memory (full serialized logs).
  • LLM with Statistical Prior: Bandit-as-Context (LLM sees bandit probs), Freq-as-Override, Bandit-as-Override (the proposed method).

Key Results:

1. Superior Performance: The BANDIT-AS-OVERRIDE agent consistently achieved the lowest cumulative regret and highest test accuracy across all backbones and preference regimes.
2. Decoupling vs. Memory: Memory-augmented baselines (InContext/Profile) significantly underperformed, confirming that injecting history into prompts does not substitute for a dedicated statistical estimator.
3. Handling Overrides: Purely statistical agents failed on explicit queries (zero-shot semantic instructions). The LOCAL HARNESS achieved near-perfect accuracy on explicit queries by reserving the LLM strictly for override detection.
4. Exploration Value: The Bandit-based prior outperformed simple frequency counting, particularly in stochastic (soft) preference regimes, demonstrating the necessity of principled exploration.

Significance and Claims

The paper claims to identify a fundamental architectural flaw in current personal agent designs: the conflation of statistical learning and semantic reasoning.

• Architectural Shift: It proposes a paradigm shift from "LLM-as-Brain" (handling everything) to "LLM-as-Exception-Handler" (handling only explicit overrides), delegating habit modeling to lightweight local statistics.
• Efficiency and Privacy: By keeping statistical learning local, the approach reduces API latency, avoids context overflow, and adheres to strict privacy constraints without requiring complex centralized infrastructure.
• Robustness: The decoupled design ensures that the agent can adapt to user habits while remaining responsive to explicit user commands, a balance that memory-augmented approaches fail to maintain.

The authors conclude that this decoupled framework establishes a robust, mathematically grounded paradigm for personalized agent design, validated by both theoretical regret bounds and extensive empirical evaluation.