Trajectory-Aware Retrieval Agents for Temporal Decision- Making
| Source: arXiv AI
Tags: RAG, LLM agents, temporal reasoning, SHAP, medical QA, financial prediction
TLM (Trajectory Language Model) improves RAG for temporally ordered decisions by iteratively refining evidence with SHAP-guided feedback and latent growth curves over embeddings — substantially outperforming standard RAG on medical QA and yielding economically meaningful gains on financial tasks.
Details
Standard RAG pipelines fragment documents into isolated chunks, discarding temporal order. For decisions that depend on trend, turning point, or trajectory — medical history, earnings calls, stock context — this destroys critical signal. TLM addresses this with a closed-loop agentic framework. The core novelty is the Latent Growth Curve Model (LGCM) applied over retrieved chunk embeddings, providing an interpretable mechanism for detecting trends, turning points, and information gaps in temporal sequences. Retrieval is iteratively refined using SHAP values to identify which chunks most influence the current decision. Improvements are demonstrated on three tasks: medical question answering (substantially beats zero-shot LLM and standard RAG), earnings call surprise prediction (consistent gains), and overnight stock gap prediction (economically meaningful returns). A convergence guarantee shows iterative refinement is monotonically non-decreasing in accuracy under a scorer-calibration assumption that holds approximately in practice. The SHAP-guided feedback loop adds computational overhead versus single-pass RAG, and the scorer-calibration assumption requires validation in production settings. The approach is most suited for high-stakes decisions where latency is not critical.