Agentic AI

This paper addresses the difficulty of adapting LLM agent behaviors to individual users without expensive fine-tuning. It proposes FABLE, a small, factorized policy model that sits outside the main agent to learn execution decisions from scalar feedback. FABLE separately optimizes memory retrieval, tool usage, and response style to tailor interactions.

• FABLE enables personalization of frozen, proprietary LLM agents without retraining the base model.
• It uses online learning from scalar feedback to adapt execution policies per user.
• The approach factorizes decisions into memory, information acquisition, and style layers.
• This method avoids the high cost of fine-tuning while improving user-specific relevance.

CrystalMem addresses memory hysteresis in self-evolving LLM agents, where capability drops after quota squeezes due to irreversible deletion or one-way compression. The paper proves that policies retaining only entries have a residual-deficit floor, then proposes a sidecar that demotes entries across four fidelity states using a crystallization-energy schedule. This approach orders demotions by advantage weighting to preserve essential knowledge while managing cloud resource constraints.

• Memory hysteresis occurs when agent capability fails to recover after budget cuts due to structural data loss.
• Standard keep-or-drop policies create a permanent performance deficit floor.
• CrystalMem uses a four-state fidelity system to manage memory elasticity without losing core knowledge.
• Demotion decisions are driven by an advantage-weighted crystallization-energy schedule.
• Acts as a sidecar, allowing existing LLM agents to adopt elastic memory management.

Self-improving LLM agents that rely on external memory for learning face a critical failure mode called the Echo Gap. Because ground-truth labels are unavailable in deployment, agents use LLM assessments as proxy rewards for stored episodes. This substitution leads to incorrect episodes receiving inflated scores, causing the system to preferentially reuse flawed experiences.

• Non-parametric self-improving agents are vulnerable to reward inflation in memory systems.
• Proxy rewards from LLM assessments can mislead policy improvement steps.
• The Echo Gap causes agents to reinforce incorrect behaviors over time.
• Score production mechanisms must be audited to prevent bias in retrieval.

Researchers introduce Bazaar, a dynamic sealed-bid benchmark designed to evaluate how well LLM agents price in real-world market conditions with hidden preferences and shifting demand. Unlike static tests, this framework uses closed-form customer utilities to allow for exact performance evaluation despite the environment's complexity. The study assesses 11 frontier models from four major providers, finding that top-tier agents like Gemini 3.1 Pro show competence in customer acquisition.

• Bazaar benchmarks LLMs in dynamic, multi-attribute auctions rather than static scenarios.
• Closed-form utilities enable exact evaluation of agent pricing performance.
• Tests reveal which frontier models can handle real-time market adaptation.
• Gemini 3.1 Pro leads in customer acquisition among the 11 evaluated agents.