LFM2.5-350M-RLCD
⚠️ Why not a model
Parallel constrained scoring on unchanged LiquidAI weights — name 'RLCD' ≠ TypeSafe RLCD training method.
Technical specs
- Base LLMs
- LiquidAI/LFM2.5-350M
- Decision types
- choice
- Features
- parallel-scoring, liquid-backbone
- License
- Apache-2.0
Links
LFM2.5-350M-RLCD is a research method for parallel constrained scoring on frozen LiquidAI weights.
Important clarification
The name “RLCD” in this project is not the same as TypeSafe AI’s RLCD (Reinforcement Learning for Calibrated Decisions) training method. The author explicitly disclaims any connection:
“No FT, no proprietary Jev method.”
How it works
The method performs parallel constrained scoring on the unchanged LiquidAI/LFM2.5-350M weights:
- Load frozen LiquidAI model
- Score options in parallel using constrained inference
- Return log-likelihood scores (not calibrated confidence)
Why it’s not a model
- Uses unchanged LiquidAI weights
- No fine-tuning or training
- Outputs are log-likelihood scores, not calibrated probabilities
- Speedup claims vs autoregressive generation are UnverifiedClaim
Same class as
This project is in the same category as SemIf, simple-jev, and litjev — inference-time techniques on frozen LLMs.