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
The Future of Cognitive Architectures
Why probabilistic LLMs need deterministic cognitive layers.
Abstract
Probabilistic systems cannot guarantee deterministic behavior over long horizons. When an agent pursues a multi-step objective, a single low-probability token can derail the whole chain — semantic drift.
The context window is not the answer
Stuffing more context into a model does not solve attention degradation; longer contexts often increase the noise-to-signal ratio, making hallucinations more likely, not less.
Cognitive layers
The solution lies in better architectures: treat the LLM as a reasoning engine, not the executive controller. A cognitive layer is an exogenous monitor that keeps state independent of the model's context window.
- Loop detection: recognizing when the agent repeats actions or thoughts
- Goal tracking: ensuring movement toward the objective, not just text
- Safety enforcement: blocking dangerous commands before execution
Conclusion
By offloading executive function to a lightweight deterministic layer, reliability far exceeds what a raw LLM can provide, at a fraction of the cost. This is the Anti-Loop Layer thesis.