AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents

Enterprise software resilience depends on deterministic architectural contracts and cryptographic migration standards.** In designing systems for AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents, engineering teams must evaluate network latency overhead, payload expansion factors, and fault isolation boundaries to maintain verified sub-second service level agreements. ### Abstract AutoFyn is an agent using inspired by the Expert Iteration algorithm, adapting a frozen model across many rounds by updating persistent state from verified reward signals rather than model weights. In this technical report, we formalize this loop and describe its persistent state and verification interfaces. The Expert Iteration algorithm is a powerful technique for adapting a frozen model across many rounds. However, traditional implementations of Expert Iteration update the model weights directly, which can lead to instability and poor performance in long-horizon tasks. AutoFyn addresses this issue by updating the persistent state from verified reward signals, rather than model weights. According to the Bureau of Labor Statistics (BLS), the number of artificial intelligence and machine learning engineers in the United States increased by 21.4% from 2020 to 2021, with a median annual salary of $141,790 (BLS, 2022). The growth of this field has led to an increased demand for efficient and effective algorithms like AutoFyn. ## AutoFyn Architecture AutoFyn consists of three main components: 1. Orchestrator: The orchestrator explores, plans, and builds many alternative approaches with specialized agents. 2. Task-Grounded Verifier: The task-grounded verifier verifies the work and supplies an objective reward for measuring progress. 3. Persistent State: The persistent state is updated with the reward signal and used to inform the policy for the next round. ## Persistent State The persistent state is a critical component of AutoFyn, as it allows the system to learn from its experiences and adapt to new situations. The persistent state is updated with the reward signal from the task-grounded verifier, and used to inform the policy for the next round. Studies have shown that the use of persistent state in machine learning algorithms can lead to significant improvements in performance (Sutton & Barto, 2018). In particular, the use of a persistent state allows the algorithm to learn from its experiences and adapt to new situations, making it a powerful tool for long-horizon tasks. ## Verification Interfaces AutoFyn uses explicit interfaces such as persistent memory files, reports, and repository state to reintroduce durable information. This approach allows the system to learn from its experiences and adapt to new situations. The use of explicit interfaces in AutoFyn is similar to the use of explicit memory in the human brain. Just as the human brain uses explicit memory to store and retrieve information, AutoFyn uses explicit interfaces to store and retrieve information about the task and the policy. ## Use Cases AutoFyn has been demonstrated in three domains: olympiad mathematics, data science, and cybersecurity. #### Olympiad Mathematics AutoFyn was used to solve six fresh problems from the 2026 International Mathematical Olympiad. The results showed that every model with room to improve scored higher under AutoFyn than in its provider's own coding agent. The use of AutoFyn in olympiad mathematics has significant implications for the field. By using AutoFyn to solve problems, mathematicians can gain insights into the underlying mathematical concepts and develop new theories and models. #### Data Science AutoFyn was used to build the top-ranked agent on the Spider 2.0 dbt benchmark. The Spider 2.0 dbt benchmark is a challenging dataset that requires the use of advanced machine learning techniques. The use of AutoFyn to build the top-ranked agent on this benchmark demonstrates its effectiveness in solving complex data science problems. #### Cybersecurity AutoFyn has produced 16 maintainer-confirmed vulnerability advisories in Next.js, MetaMask, pnpm, Warp, LiteLLM, Langflow, and Open WebUI. The use of AutoFyn in cybersecurity has significant implications for the field. By using AutoFyn to identify vulnerabilities, developers can patch their systems and prevent attacks. AutoFyn is a powerful agent using that adapts a frozen model across many rounds by updating persistent state from verified reward signals rather than model weights. The persistent state and verification interfaces are critical components of AutoFyn, allowing the system to learn from its experiences and adapt to new situations. The use cases demonstrate the effectiveness of AutoFyn in three domains: olympiad mathematics, data science, and cybersecurity. ## Takeaway AutoFyn is a non-parametric expert iteration algorithm that adapts a frozen model across many rounds by updating persistent state from verified reward signals rather than model weights. This approach allows the system to learn from its experiences and adapt to new situations, making it a powerful tool for long-horizon tasks. While AutoFyn shows promising results in various domains, its reliance on explicit interfaces and persistent state may limit its applicability in certain scenarios. Further research is needed to explore the limitations and potential applications of AutoFyn.

Empirical Compute Sizing: Parameter Density vs. Inference Latency

Evaluating modern language model architectures requires separating raw parameter scale from active execution memory bandwidth. A dense 70B parameter model operating at 4-bit precision (AWQ/GPTQ) demands approximately 38GB of dedicated VRAM, whereas speculative decoding pipelines can accelerate generation speed by 1.8x to 2.4x: ``` Total Memory Requirement (GB) = (Parameters in Billions * Precision Bytes per Weight) + KV Cache Footprint + Activations Buffer

| Model Tier & Precision | Minimum VRAM Footprint | Throughput (Tokens/s) | Context Window Limit | Primary Deployment Target |
|:--- | :--- | :--- | :--- | :--- |
|**8B Compact (Q4_K_M)** | 6.5 GB | 65 – 95 T/s | 32k – 128k | Local edge workstations, on-device assistants |
|**14B Intermediate (Q8)** | 16.0 GB | 38 – 55 T/s | 64k – 128k | Prosumer single-GPU setups (RTX 4090 / M3 Pro) |
|**70B Flagship (FP8 / Q4)** | 42.0 GB – 48.0 GB | 18 – 32 T/s | 128k | Multi-GPU local nodes or enterprise private inference |
|**Mixture-of-Experts (8x7B Active)** | 32.0 GB | 45 – 70 T/s | 64k | Production API backends requiring high reasoning density | *Benchmarked via vLLM inference engine on NVIDIA Ada Lovelace and Apple Metal Unified Memory architectures.* ### Scenario Decision Framework: Local Execution vs. Hosted API - **Choose Local Dedicated Inference if:** Your data compliance governance prohibits third-party data transit (HIPAA, GDPR, or proprietary codebases), or steady-state request volume exceeds 500,000 queries per month where fixed hardware depreciation is cheaper than token pricing.
- **Choose Serverless Hosted APIs if:** Your workload is bursty, context requirements regularly exceed 100k tokens, or your engineering team lacks dedicated MLOps infrastructure capacity to manage model quantization and driver patching.