Continuous Next-Vector LLMs
🔗 Source: arXiv
CONTINUOUS AUTOREGRESSIVE LANGUAGE MODELS
🚀 Technical Novelty
- Mechanism: Compresses chunks of K discrete tokens into a single dense continuous vector via a lightweight autoencoder, enabling next-vector prediction instead of next-token prediction.
- Nuance: Abandons the softmax vocabulary bottleneck entirely by adopting a likelihood-free energy-based framework and BrierLM metric, avoiding iterative diffusion/flow sampling while maintaining controllable temperature sampling.
💡 Yield
- Achieves >99.9% token reconstruction with K=4 using only 10 latent dimensions; establishes BrierLM as a strictly proper likelihood-free evaluation metric; demonstrates superior performance-compute trade-offs compared to discrete baselines at equivalent quality.
⚠️ Limitations
- Current autoencoder lacks semantic grounding in the latent space (focuses only on reconstruction); performance gap remains between CALM (K=1) and standard Transformers; context-aware autoencoders and semantically rich latent spaces are left for future work.