Ailiang Lin

Instead of stuffing raw context with dialogue history or brittle retrieval vectors, treating user personalization as particle-filtered hypothesis tracing over natural language resolves the tension between fleeting intent and long-term preferences.

Recall-based prompting can dramatically improve LLM performance on knowledge cutoff tasks, especially for counterfactual queries.

Reversing input text can unlock hidden contextual information, leading to richer embeddings and substantial performance gains in training-free LLMs.