LLM-as-RNN Enables Recurrent Inference With Memory
Researchers propose LLM-as-RNN, an inference-only framework (submitted Jan 19, 2026) that turns frozen LLMs into recurrent predictors by representing hidden state as structured natural-language prompt memory. The state is updated each timestep via feedback-driven text rewrites, enabling online learning without parameter updates; evaluated on healthcare, meteorology, and finance across Llama, Gemma, and GPT families, it improves predictive accuracy by 6.5% on average.
Key Points
- 1Introduces LLM-as-RNN representing hidden state as structured natural-language system-prompt summaries
- 2Demonstrates online learning without parameter updates, correcting errors and retaining task-relevant patterns
- 3Enables practitioners to improve sequential predictions across domains with interpretable, human-readable learning traces
Scoring Rationale
High novelty and broad, cross-domain applicability; limited credibility because results come from a single arXiv preprint without peer review.
Sources
Public references used for this report.
Practice with real Logistics & Shipping data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Logistics & Shipping problems
