Yeco Flash Mini-IT v1
118M-parameter Italian micro-LLM, trained for under €15, ~200 tok/s on CPU. Beats 5× larger models on Italian logic benchmarks.
Paper
Anti-Loop Layer
Mini-LLM cognitive layer that detects and stops infinite agent loops (~12ms).
Abstract
As LLMs move from chat to agentic loops (Thought → Action → Observation), stochastic failure modes — thank-you loops, semantic collapse — become catastrophic. Anti-Loop Layer is a deterministic exogenous monitor that watches the agent from outside the model and intervenes early.
Design
A second frontier LLM monitoring the first is cost-prohibitive and slow. Anti-Loop Layer is built on Mini-LLMs: highly specialized, quantized models that run on CPU with minimal latency (~12ms typical, ~38.85 MB peak).
Results
- Loop detection: catches repetitive and semantically collapsed trajectories
- Reported F1 up to 0.90 / 0.71 / 0.63 across detection regimes
- CPU-only, no GPU required, drop-in around existing agent loops
Availability
Open source under Apache 2.0: pip install yecoai-cognitive-layer. Technical whitepaper and reference implementation published for research and production use.