Machine Learning Expedition

Agents communicating through latent states instead of natural language arxiv.org/abs/2604.25917

Paper: arxiv.org/pdf/2601.03192

Explicitly separates stable LLM reasoning from a plastic, evolving memory, effectively addressing the stability-plasticity dilemma to allow for continuous runtime improvement and avoid catastrophic forgetting.

MEMRL: SELF-EVOLVING AGENTS VIA RUNTIME REINFORCEMENT LEARNING ON EPISODIC MEMORY 1. Introduces a two-phase retrieval system that first filters memory candidates by semantic relevance and then selects the final ones based on learned utility (Q-values), moving beyond purely semantic matching.

#science-mathematics" target="_blank" rel="nofollow noopener">blog.google/technology/ai/…

3. Modular and Multi-Path-Aware Offline Benchmarking for Mobile GUI Agents arxiv.org/pdf/2512.12634

2. WEBOPERATOR: ACTION-AWARE TREE SEARCH FOR AUTONOMOUS AGENTS IN WEB ENVIRONMENT arxiv.org/pdf/2512.12692

Recent LLM Agent Papers 1. MAC: A Multi-Agent Framework for Interactive User Clarification in Multi-turn Conversations

Paper URL: arxiv.org/pdf/2511.21689

ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

Paper Link: arxiv.org/pdf/2512.08296

• Capability Saturation Once a single agent achieves ~45% task performance baseline, additional agents often degrade results because coordination overhead dominates.

Towards a Science of Scaling Agent Systems- This paper aims to move agentic AI systems (LLM-powered systems that reason, plan, and act) from heuristic practice toward a principled, quantitative science—especially for scaling those systems effectively

This paper got best paper award from Neurips 2025

Authors introduce infinity-chat dataset with 26k real world open-ended questions. The also provide human annotations including absolute rating and pairwise preferences. arxiv.org/abs/2510.22954