TLM: Trajectory-Aware LLM Agents for Temporal Decision-Making
A recent study available on arXiv (2607.21625) presents TLM (Trajectory Language Model), an agentic framework designed for decision-making from extensive, temporally organized text utilizing large language model (LLM) agents. Traditional retrieval-augmented generation (RAG) methods disrupt chronological context, neglecting the temporal structure essential for accurate decision-making. TLM enhances the evidence set through SHAP-guided feedback in an iterative manner. Its significant technical innovation is a latent growth curve model (LGCM) applied to retrieved chunk embeddings, which identifies trajectory trends, turning points, and gaps in information. Assuming scorer calibration, the iterative refinement process consistently increases the probability assigned to the correct label. TLM has been tested on three datasets with temporal grounding.
Key facts
- arXiv paper 2607.21625 introduces TLM (Trajectory Language Model).
- TLM is a closed-loop agentic framework for decision-making from temporally structured text.
- Standard RAG pipelines fragment chronological context, discarding temporal structure.
- TLM iteratively refines evidence using SHAP-guided feedback.
- Key contribution: latent growth curve model (LGCM) over retrieved chunk embeddings.
- LGCM detects trajectory trends, turning points, and information gaps.
- Under scorer-calibration assumption, refinement is monotonically non-decreasing in correct label probability.
- TLM evaluated on three temporally grounded datasets.
Entities
Institutions
- arXiv