Kevork Sulahian

Kevork Sulahian

Staff Machine Learning Engineer

Grid Dynamics

Yerevan, Armenia

Talk

Recursive Language Models: Architectural Fragilities, and the Path Toward Robust Agentic Intelligence
Track: Data Science Duration: 25 minutes View on Schedule
AI Agents Computer Vision Testing Data Science

The recent introduction of Recursive Language Models (RLMs) represents a paradigm shift from linear token processing to inference-time computing. By offloading context to a REPL environment and enabling symbolic recursion, RLMs have demonstrated the ability to handle effectively infinite context windows.
However, the original RLM paper focuses almost exclusively on capability. It does not address the reliability of the recursive mechanism itself.
Current literature on recursive scalable oversight suggests high failure rates in hierarchical supervision, and spiral of hallucination effects in long-horizon agents. We propose to rigorously test whether RLMs are susceptible to these failure modes when subjected to adversarial or high-complexity inputs.
The questions will be specifically on:
1. The Decay Hypothesis.
2. The Halting Problem.
3. State Persistence Vulnerability

About the Speaker

AI researcher and Staff Machine Learning Engineer building end-to-end AI systems across healthcare, finance, safety, and education. Led development of large-scale LLM pipelines, evaluation frameworks, and time-series forecasting tools, driving client adoption and helping scale SuperAnnotate from pre-seed through Series B. Co-authored an AAAI-accepted paper on LLM safety and contributed national-level economic analytics for Armenia’s Ministry of Economy. Combines deep technical expertise with cross-functional leadership, partnering with product, engineering, and research teams to translate advanced research into scalable, production-ready AI solutions that deliver measurable business impact, improve decision-making, and create durable long-term value for organizations and users worldwide.
Big fan of pubs

Recording

Video will be available after the conference.