ShadowStream: A Trainable Auxiliary State Stream for Frozen LLMs: Improving Dialogue-state Retention and Persona Persistence in a Japanese 3B Model

Description This paper introduces ShadowStream, a trainable auxiliary state stream added to a frozen Transformer-based language model. The goal is not to increase the model’s factual knowledge, but to improve retention of temporary dialogue state such as corrections, role assignments, local names, answer-style constraints, and persona settings. Experiments were conducted on sbintuitions/sarashina2.2-3b-instruct-v0.1, a Japanese-capable 3B Llama-style model. In the main noinit setting, ShadowStream is not initialized from token embeddings, reducing direct token-copy shortcuts and forcing the auxiliary stream to rely on layer-wise writes and reads. On a 100-case held-out dialogue-state probe, active ShadowStream achieved 81.7% ± 2.9% across three seeds, compared with 56.0% for the frozen base model, 56.0% ± 0.0% under read ablation, and 58.3% ± 0.5% under write ablation. On a long-context character/persona probe, active ShadowStream achieved 7/8, compared with 4/8 for base and ablated conditions. The results suggest that frozen LLMs can be augmented with a small trainable internal pathway that exhibits working-memory-like behavior for dialogue-state retention. The paper also discusses limitations, including synthetic evaluation scope, limited model coverage, non-exhaustive LoRA baselines, and the need for future work on natural dialogue, larger models, forgetting, and safety.