Renmin University Builds a Local Harness That Teaches AI Agents Your Habits
Researchers propose a bandit-based local harness that decouples preference learning from LLM reasoning, beating memory-augmented agents on skill selection accuracy and regret.
- Researchers from Renmin University of China propose a Local Harness architecture for personal AI agents that separates preference learning from LLM reasoning.
- A lightweight bandit algorithm runs locally to track implicit user skill preferences, replacing expensive memory-augmented LLM calls.
- The approach achieves the lowest cumulative regret and highest accuracy vs. all baselines including in-context memory and profile-based memory agents.
- Two modes: Bandit-as-Context (soft hints to LLM) and Bandit-as-Override (bandit decides, LLM handles explicit exceptions only).
- Key insight: conflating statistical preference tracking with semantic intent parsing inside one LLM call hurts both tasks.
- No code released; the architecture uses standard bandit algorithms (UCB/Thompson Sampling) and is straightforward to implement.
Personal AI agents that run locally but call out to remote LLMs for reasoning are becoming a real deployment pattern. You install the agent, connect it to a cloud model via API, and let it orchestrate a growing library of skills -- calendar tools, messaging apps, music services, and more. The problem nobody has cleanly solved yet: how does the agent learn which skill you actually prefer when you never explicitly say so?
Researchers Zeyu Gan, Huayi Tang, and Yong Liu from the Gaoling School of Artificial Intelligence at Renmin University of China published a paper proposing a new architecture they call the Local Harness -- a lightweight module that sits between the user and the remote LLM, learning implicit preferences locally using statistical methods rather than asking the LLM to figure it out.
The problem nobody talks about
As LLM capabilities advance, locally deployed personal agents relying on API-based remote models and external skills have emerged as a novel paradigm -- and with the rapid expansion of available skills, enabling personal agents to learn and adapt to implicit user preferences becomes a critical challenge. Think of it this way: if you have three coffee-ordering apps installed and you always end up using one specific service, the agent should learn that pattern without you having to say "I prefer this one."
Local deployment constraints preclude complex centralized selection algorithms, creating an urgent need for a lightweight local preference harness. The naive fix -- stuffing your preference history into the LLM's context window as memory -- runs into real problems at scale. Memory-augmented agents grow expensive and slow as history accumulates, and they conflate two fundamentally different jobs: understanding what you
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
