Yifan Zhang has introduced the Recurrent Looped Transformer (RLT), an architecture that re-runs a 48-layer decoder stack to create an extensible computation path. The design aims to maintain flat per-token costs for long-horizon AI agent tasks, though the technical report currently lacks performance benchmarks.
The introduction of this architecture signals a shift toward models designed specifically for long-horizon, stateful AI agent tasks where traditional transformer memory resets create bottlenecks. By keeping per-token compute fixed at 48 blocks while extending the total path, the design attempts to solve the cost-scaling issues inherent in deep reasoning workflows. For the streaming and broader tech ecosystem, this highlights a growing dependency on memory bandwidth rather than raw FLOPs for next-generation model deployment. Industry observers should watch for the release of first-party benchmarks to verify if this extensible path translates into measurable reasoning gains compared to standard transformer stacks.
The agentic AI landscape for network operations is consolidating rapidly around a handful of vendor frameworks, each targeting different layers of the telco stack. In June 2026, Ericsson launched its AI in RAN commercial software subscription claiming up to 20% higher downlink throughput and up to 10% better spectral efficiency across more than 15 live deployments using existing baseband silicon. Verizon simultaneously disclosed that its 60,000-site vRAN network is now applying agentic AI to planned configuration changes, service assurance, and network optimization, while publicly calling for industry-wide interoperability standards for agentic systems. The absence of a standardized protocol for agentic command and control across multi-vendor networks mirrors the early RAN interoperability challenges that Open RAN later addressed.
Nokia has taken a markedly different strategic path, building its autonomous network ambitions around cloud and data partnerships rather than incremental RAN intelligence. At DTW Ignite in June 2026, Nokia announced work with AWS and Databricks to build the data, cloud, and control layers for its Autonomous Network Fabric, positioning the fabric as an operating system spanning radio, core, transport, and service domains. The Databricks collaboration addresses fragmented telco data silos through a unified lakehouse architecture, while the AWS integration brings cloud scalability and access to Amazon Bedrock and SageMaker tools. Nokia reported that operators using its autonomous networks portfolio are achieving automation rates above 90 percent, service delivery times under four hours, and up to 85 percent reduction in slice rollout time.
The technical divergence between Ericsson and Nokia on AI-RAN is now stark, with implications for how operators plan long-horizon AI agent deployments across their infrastructure. Nokia's entire RAN strategy is built on its close partnership with Nvidia, cemented by the chipmaker's $1 billion investment, with an entire Layer 1 RAN designed to run on Nvidia's CUDA platform. T-Mobile US is testing both approaches, highlighting the split between incremental RAN intelligence and broader shared AI compute. Meanwhile, Nokia is pushing agentic AI deeper into the mobile core itself. Nokia's mobile core team reported that AI-driven paging reduces call setup times from roughly 10 seconds to one or two seconds in certain use cases, with the company offering a Mobile Core Early Access program so operators can trial AI-native features before full deployment. These developments underscore that the architectural choices being made now around enterprise AI agents will determine which vendors control the automation layer as networks move toward level-four autonomy. As these systems evolve, Sony and Truist executives warn of AI agent governance gaps that could impact future deployment safety.
Yifan Zhang has introduced the Recurrent Looped Transformer, an architecture that re-runs a 48-layer decoder stack to create an extensible computation path. By separating memory from computation, the design maintains flat per-token costs for long-horizon AI tasks, potentially solving scaling bottlenecks inherent in traditional transformer models that require memory resets.
It is a new AI architecture introduced by Yifan Zhang that utilizes a 48-layer decoder stack which re-runs recurrently to create an extensible computation path for long-horizon tasks.
The design maintains flat per-token costs by separating memory from computation, allowing the system to keep compute fixed at 48 blocks while extending the total reasoning path.
No, the current technical report, zhang2026recurrentlooped, does not include performance benchmarks, and reasoning improvements remain research goals.
The system uses compatible weights across its 48 encoder and 48 decoder layers to carry the complete state forward without resetting at prompt boundaries.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source