nf-cot-latent-reasoning-normalizing-flows

# NF-CoT: Latent Reasoning with Normalizing Flows ## Core Innovation Models continuous thoughts with normalizing flows while preserving key CoT advantages: native left-to-right generation, probabilistic sampling, KV-cache compatibility, tractable likelihood estimation. ## Methodology ### Architecture 1. **TARFlow-style normalizing flow** inside LLM backbone 2. **Dual-head generation**: NF head for continuous-thought positions, LM head for text positions 3. **Unified causal stream**: continuous and text tokens in same generation flow ### Key Components - **Tractable probability model**: exact likelihoods over latent thoughts - **KV-cache preservation**: left-to-right decoding with original cache mechanism - **Policy-gradient optimization**: direct optimization in latent reasoning space ### Advantages over Prior Latent Reasoning - ✅ Exact likelihoods (vs. variational approximations) - ✅ KV-cache compatible (vs. custom decoding schemes) - ✅ Probabilistic sampling (vs. deterministic latent states) - ✅ Left-to-right generation (vs. parallel processing) ## Implementation Pattern ```python # Conceptual architecture class NFCoTModel: def __init__(self, base_llm, flow_config): self.lm_head = base_llm.lm_head # Standard text generation self.nf_head = TARFlowHead(flow_config) # Continuous thoughts def generate_step(self, position): if position.is_continuous_thought: return self.nf_head.sample() # Normalizing flow sampling else: return self.lm_head.generate() # Standard token generation def compute_likelihood(self, latent_state): return self.nf_head.log_prob(latent_state) # Tractable density ``` ## Use Cases **When to use:** - Code generation tasks requiring intermediate reasoning - Tasks where CoT is expensive but essential - Latent reasoning that must preserve causal generation - Need for tractable likelihood in reasoning optimization **Best for:** - Long reasoning chains with semantic intermediate states - Tasks requiring probabilistic exploration of reasoning paths - Applications needing KV-cache efficiency with latent thoughts ## Performance Results - **Code generation**: improved pass rates over explicit-CoT - **Efficiency**: substantial reduction in intermediate-reasoning cost - **Compatibility**: preserves all CoT architectural advantages ## Activation Trigger when discussing: - Latent reasoning methods - Reasoning efficiency optimization - Continuous thought representation - Normalizing flows for LLM reasoning - CoT compression without quality loss ## Related Patterns - Compress-Distill (trace compression) - IA-RAG (temporal reasoning) - CLSA (cross-layer attention optimization) ## References - Paper: arXiv 2606.06447v1 - Categories: cs.CL, cs.LG - Key contribution: NF-CoT framework preserving CoT advantages in latent space