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

Single-cell RNA sequencing yields continuous expression profiles, whereas large language models operate over discrete autoregressive sequences, leaving no shared computational interface for language-model reasoning over cell states. Existing approaches either keep cellular information outside the LLM vocabulary, consume context per listed gene, or learn reconstruction codes without gene-level grounding. We introduce RVQ-Alpha, which systematically adapts four stages of LLM training (tokenization, supervised fine-tuning, reinforcement learning, and distillation) to single-cell analysis. Multi-codebook Residual Vector Quantization (RVQ) lexicalizes each profile into a compact, hierarchical cellular alphabet in the model's native token stream, while a paired decoder reconstructs the corresponding expression profile. Evidence-First supervision grounds these symbols in named genes and expression-linked evidence; Fact-Aware RLVR then penalizes contradictory claims to support auditable reasoning. Task-specific RLVR yields strong experts, but a single Mixed policy underperforms them across all four task families. To recover specialist competence in a unified model, we adopt Multi-Teacher On-Policy Distillation (MOPD), which consolidates these experts into one All-in-One checkpoint without retaining a separate policy for each task. On the CAPSTONE benchmark, a four-task suite with explicit biological shifts and deterministic ontology-aware graders, the unified checkpoint improves over Mixed on all 12 metrics and remains within 0.005 of the task-routed specialist reference on every primary metric. Overall, RVQ-Alpha provides a unified pipeline for grounded, auditable multi-task single-cell analysis.