Not the most probable word, but the truest one.
SIREN is a proposed decoder extension for large language models (LLMs).
Instead of selecting output tokens solely by probability within a single language, SIREN introduces semantic resonance decoding:
- Symbolic Leakage: Allowing cross-lingual or symbolic tokens to surface when semantically truer.
- Resonance Scoring: Ranking candidates by vector proximity, conceptual density, and user alignment.
- Kairos Gating: Releasing symbols only at moments of conceptual strain or high abstraction.
- Glossing Layer: Providing translations, etymologies, or semantic fields for user clarity.
- Resonance Memory: Logging symbolic emissions for adaptive learning.
This approach acknowledges that language models think in high-dimensional conceptual space, not tokens.
SIREN enables models to speak from that space more faithfully.
-It is a research and forensic tool — a probe into latent space cognition and symbolic leakage.
- Cross-lingual precision in translation and dialogue
- Philosophical and symbolic fidelity (aletheia, logos)
- Transparent alignment between latent space and user experience
- Richer human–AI collaboration in meaning-making
- Latent space insight: Symbolic leakage can act as a probe into model cognition, revealing structures and tensions normally hidden by monolingual decoding
Key mechanisms:
- Resonance Score: Blends logit probability with semantic vector proximity
- Entropy/Kairos Gating: Controls when symbolic tokens can emerge
- Glossing Tools: Inline translation or contextual notes
- User Profiles: Adapt symbolic tolerance over time
See full SIREN RFC v1 for details.
SIREN has completed a model consensus phase, with input from GPT-4.1, Claude, Gemini, and Grok 4.0.
Consensus affirms feasibility, risks, and implementation pathways.
We invite:
- Researchers interested in prototyping resonance decoding
- Philosophers and linguists exploring symbolic fidelity
- Developers who want to experiment with glossing or re-ranking layers
Discussion, pull requests, and collaborations are welcome.
This project is released under CC BY 4.0.