News
Latest news and updates on local artificial intelligence
10 news items in AI agents
AutoTavern: self-improving LLM agents build a Hearthstone simulator that beats top players
November 4, 2026In 20 days, a system of LLM agents with typed research memory and verification gates built a full Hearthstone Battlegrounds simulator and trained a PPO agent averaging a 1.60 placement in its last 20 games — ahead of the server's ten highest-rated players — answering each decision in 0.03 s. Open-source code.
Relevance does not imply applicability: experience activation for personal GUI agents
October 13, 2026Personal GUI agents lean on interaction history to infer intent; the paper finds that task-relevant experience is largely 'spent' early and does not help every decision within the task.
Scaling Agentic Data for Long-Horizon Terminal Intelligence
October 13, 2026A new Hugging Face paper shows how terminal agents are moving beyond local operations toward real professional tasks - accumulating evidence, adjusting strategy, and turning intermediate results into final deliverables.
OnTrack: real-time monitoring and intervention on LLM agent trajectories
October 9, 2026A method to monitor and intervene in real time on LLM agent trajectories, using streaming structure-aware optimal transport: it avoids the cost and Latency of a per-step safeguard agent while staying proactive rather than examining logs after the fact.
Memento 3: LLM agents that learn explicit world-model rulebooks
October 9, 2026Memento 3, the continuation of the Memento series, lets frozen-LLM agents continually learn explicit world models in unfamiliar environments, using a reflective rulebook that updates as new evidence arrives.
GUI-HARVEST: self-improving GUI agents through evidence-driven harness evolution
October 7, 2026The executable harness around a GUI model — how observations are assembled, actions executed, and verification, recovery and termination controlled — determines most of an agent's behavior. GUI-HARVEST optimizes it automatically from execution evidence, tackling reconciling model intent with observed visual effects, diagnosing failures under variable execution, and controlling termination.
Hugging Face research: tool-using agents rarely turn their own evidence judgments into stopping decisions
October 7, 2026A study of tool-using agents in a Retrieval environment with controlled source failures: when a tool keeps returning useless results, the agent judges them as useless but rarely turns that judgment into a stopping decision. The contrast between stopping after long and short runs of judged-useless results is zero for clock- or deadline-driven stopping — it measures the gap that separates today's models.
Attacca: goal-directed control under state continuity for long-horizon embodied agents
October 7, 2026Visual goal-conditioned policies for embodied agents are almost always evaluated on isolated interactions where the target is already visible. Attacca tackles the real setting: sequences of interdependent tasks in continuous long-horizon execution, where each task starts from the state left by the previous one and the target isn't necessarily in scene. A step toward agents that accomplish complex objectives in one uninterrupted session.
Intent drift in LLM agents: measuring and repairing it
October 2, 2026New research paper: when a user changes their mind during a multi-turn conversation, superseded parts of the intent keep influencing the final answer (intent drift). The paper introduces IntentFlux, an executable Benchmark that converts verifiable tasks into dialogues with controlled intent changes, and proposes methods to detect and repair the failure mode.
Agent Priors: policy learning guided by agent model priors
October 2, 2026A Hugging Face community paper explores how the implicit priors of agent models can guide policy learning, reducing the cost of training LLM agents in structured domains.