Sep‑Latent is an adaptive implicit reasoning framework that treats separator tokens as structural semantic anchors. An Attention‑Decay mechanism forces the model to aggregate and compress reasoning logic into these anchors during training. Multi‑level distillation across all layers aligns the explicit CoT pathway with the implicit one to preserve logical consistency. The framework also adapts reasoning depth by terminating the latent thinking phase based on query complexity, enabling dynamic depth. On the GSM8k‑Aug benchmark Sep‑Latent reaches 56.8% accuracy while using only 5.4 tokens on average, markedly improving efficiency over fixed‑length implicit models and narrowing the gap to explicit CoT. PCA and symbolic variant analyses confirm semantic grounding and superior robustness to perturbations.
Review