MobileLLM's JAMEL Teaches AI Agents to Explore Unfamiliar Apps Without Human Labels
JAMEL trains GUI agents to explore and remember simultaneously, using JavaScript code coverage as a free training signal — no human labels needed.
Takeaways
- JAMEL is a new open-source framework that jointly trains an LLM agent's memory and exploration policy together, solving a core chicken-and-egg problem in agentic AI.
- It uses JavaScript code coverage as a free, annotation-free reward signal — new UI paths executed = novel behavior discovered.
- A lightweight compressor model (Qwen3-VL-2B) distills raw interaction history into compact memory, keeping token costs low over long trajectories.
- On the ScaleWoB benchmark, JAMEL outperforms all open-weight baselines and matches the closed-source MAI-UI model while using fewer tokens.
- The model generalizes to unseen apps at test time, the critical requirement for real-world deployment on unfamiliar software.
- Code, model weights (JAMEL-9B on HuggingFace), and the full repo are publicly available.
Getting a language model agent to genuinely explore an unfamiliar app is harder than it sounds. Ask it to poke around a new e-commerce site and it will often loop through the same three buttons, convinced it has seen everything. The root cause is a chicken-and-egg problem: effective exploration requires memory, and without remembering what has already been tried, an agent can easily repeat exhausted behaviors instead of discovering new states. A new paper from the MobileLLM group tackles this directly with JAMEL (Joint Agent Memory and Exploration Learning), an open-source framework that trains memory and exploration together rather than treating them as separate problems.
The vicious cycle agents are stuck in
Modern LLM agents operating over long sessions face a brutal trade-off. Keeping the full interaction history in context is accurate but burns tokens fast. Compressing that history into a learned latent memory (a compact vector summary the model can query) is efficient, but training that compressor has always been the hard part: what does a "good" memory even look like, and who labels it?
In open-ended environments, agents often struggle to explore effectively, especially when external task rewards are sparse or unavailable. Agent performance degrades when instructions are underspecified and the true entry point is visually inconspicuous, and the key bottleneck is weak proactive exploration after an initial strategy fails. This is not a niche edge case , it is the default condition for any agent deployed on a real app it has never seen before.
Memory and exploration as a feedback loop
JAMEL's main insight is that memory and exploration form a mutually reinforcing loop: memory helps the agent avoid repeated behaviors and identify unexplored actions, while novelty-driven exploration provides supervision for what the memory should encode. This reframes the labeling problem entirely. Instead of asking humans to annotate what the agent should remember, the system generates its own training signal by measuring how much new ground it covers.