RLearner-LLM Achieves 6x NLI Improvement via Hybrid-DPO

RLearner-LLM addresses the verbosity bias in DPO by fusing DeBERTa-v3 NLI signals with verifier LLM scores in a hybrid preference pipeline. On Gemma 4 E4B-it, the method yields up to a 6x improvement in NLI entailment over SFT, mitigating the trade-off between fluency and logical grounding.