Seq2Seq2Seq: lossless data compression via discrete latent transformers and reinforcement learning
Abstract Efficient lossless compression is essential for reducing storage and transmission requirements while preserving data exactly. Traditional dictionary-based and statistical compressors can be limited in their ability to exploit complex, long-range dependencies. In this paper, we propose a lossless compression framework that combines a T5-small architecture with Advantage Actor-Critic reinforcement learning to generate a variable-length sequence of discrete compressed tokens. Rather than relying on a continuous autoencoder bottleneck, the proposed framework directly optimizes the length of this discrete representation while preserving the information required for exact reconstruction. The method operates without external grammatical rules or world knowledge. On the enwik8 benchmark, it achieves a compression ratio of 4.14, improving upon XZ (4.05) by approximately 2.3% and GZIP (2.74) by 51.0%. Although this result remains below NNCP v3.2 (6.70), it demonstrates that reinforcement learning can learn a compact, lossless discrete representation using two T5-small networks evaluated on a single 12 GB GPU device.