RVQ-Alpha: Bridging Single-Cell Transcriptomics and Large Language Models via Discrete Tokenization and Verifiable Reinforcement Learning

Integrating continuous single-cell gene expression profiles with discrete-token large language models (LLMs) remains an open challenge: text-based methods are token-inefficient and discard quantitative precision, continuous embeddings preclude autoregressive generation, and single-codebook vector quantization cannot stratify biological scales. We present RVQ-Alpha, an end-to-end framework that bridges this modality gap through three contributions: O_LIa Residual Vector Quantization (RVQ) tokenizer that compresses each cell into a fixed 10-token sequence via eight residual codebooks embedded directly in the LLM vocabulary, requiring 3.4x fewer tokens than prior discrete methods and enabling bidirectional unification of cell interpretation and generation; C_LIO_LIscCoT-Synth, a teacher-student engine that grounds newly added biological tokens through evidence-before-conclusion reasoning, using the language modeling objective as the cross-modal alignment signal without a separate projection network; C_LI and (iii) a Fact-Aware RLVR system combining an ontology-grounded answer judge with saliency-weighted verification of biological claims against actual expression data, under dynamic gating that conditions hallucination suppression on task competence. Built on Qwen3-4B and trained via continued pretraining, supervised fine-tuning, and reinforcement learning, RVQ-Alpha substantially improves out-of-distribution generalization and rare-cell recognition across eight held-out datasets; ablations confirm that evidence-first grounding reduces hallucination more than fivefold.