When Context Becomes a Systems Problem: Recursive LM

Description

Current language models exhibit severe quality degradation as the size and complexity of input information increase—a phenomenon known as context rot. This limitation stems not only from the context window size but from a deeper constraint: LLMs are forced to manipulate long sequences exclusively within their token space, where attention and memory do not scale. In this talk, I introduce Recursive Language Models (RLMs), a general inference paradigm proposed by MIT CSAIL that reframes the context problem as a systems problem. Rather than feeding the full prompt into the model’s context window, RLMs place it in an external persistent environment accessible via a REPL, and allow the LLM itself to generate code to inspect, filter, decompose, and process that context through recursive calls to other models. On benchmarks such as CodeQA and BrowseComp+, RLMs achieve up to 91% accuracy where traditional approaches fail completely. RLMs scale effective context by orders of magnitude without relying on long-context architectures, and demonstrate how training natively recursive models (e.g., RLM-Qwen3-8B) can become a new axis of scaling. Throughout the talk, I will explain how RLMs work, why they outperform classical approaches such as truncation or summarization, and how to implement them in practice using rlm-runtime—an open-source runtime that materializes the RLM paradigm and reproduces key results from the paper. Agenda --- Speaker Antonio Peña Batista - Caixabank Tech (Technical Lead) I'm a Computer Engineering graduate with over 9 years of experience in software development. I specialize in technologies such as React, Django, and Python, and I've played a pivotal role in crafting innovative solutions for various industries. My passion for coding is complemented by a strong background in designing and implementing design systems, allowing me to create intuitive and visually… Hosted By Laura Morillo-Velarde, GDG & WTM Organizer Ruben Aguilera, GDG Organizer Jesus Cuesta, GDG Organizer Jesús Cuesta is Senior Specialist Lead / Senior Manager at Deloitte, where he leads web architecture projects, innovation initiatives, and serves as recruiting manager. He also collaborates with various tech communities as administrator and/or mentor, notably GDG Madrid, Adalab, Codemotion Ambassador, DASA Ambassador, Openexpo Europe Ambassador, etc. He is currently pursuing a PhD in AI at Rey Juan Carlos University, and holds multiple master’s degrees in AI and Innovation, Web Solutions, specialized university training in Human Resources, and technology. Jaime Conde Segovia, GDG Organizer Xavier Portilla Edo, GDG Organizer Xavier was born in Valencia, but currently lives in Madrid. He has over 7 years of experience. His focus areas include Java, DevOps, QA, microservices, and voice-based applications. He is the organizer of DevOpsDays Madrid and the Alexa community in Spanish, and GDG Madrid. In his spare time, he enjoys music, beer, and friends. Author of the book "The Revolution of Alexa". Maritere Nieto Galán, WTM Organizer Maria Teresa (Maritere) is a computer engineer and currently works at Telefónica as Senior Product Manager. Since 2017, she has combined her main job with teaching at Spanish universities. She is a Women Techmakers ambassador—a Google-led initiative that provides visibility, community, and resources for women in technology—and co-organizes the WTM Madrid community. In 2022, she was recognized by Mujer Hoy in the Next Generation category as a leading woman in technology. Rei Aguilera, Business Development --- Partner Sirviendo Codigo (https://sirviendocodigo.com/) We are a space for sharing knowledge. Here you’ll find a place to share your technical knowledge and access knowledge created by others—freely, simply, and within a community. --- Complete your event RSVP here: https://gdg.community.dev/events/details/google-gdg-madrid-presents-cuando-el-contexto-se-convierte-en-un-problema-de-sistemas-recursive-lm/.