Intel / article

TEFM: Token-Efficient Faithful Modeling for Structured Data

Source: arxiv.org
Author: Zhichao Hou, Lingdao Sha, Xueyu Mao, Yang Liu, Peijie Qiu, Rui Song
Date: Published Sep 10, 2026

Context

TEFM shortens the input a must read when the data is structured records, such as a patient's hospital history or a log of network traffic. An encoder turns each record into a short sequence of entries from a learned codebook, and those codes are added to the language model's vocabulary so the model reads them in place of the lengthy text version.

The model is first trained to match the codes with natural-language descriptions, then fine-tuned to classify records from the codes. The authors report retaining approximately 1% of the original count on clinical data (predicting whether a patient died during a hospital stay) and 2% on security data (detecting attack traffic) while keeping competitive classification accuracy. These are results from their own experiments on the specific datasets and models they tested.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Key quotes

“TEFM achieves token efficiency by compressing lengthy structured observations into compact Behavioral Code tokens, dramatically reducing token consumption with minimal information loss.”

“Comprehensive experiments across various domain datasets and model backbones (Qwen3, Gemma-2, Phi-4) show that TEFM achieves competitive classification accuracy with dramatic token reduction (approximately 1% token retention in clinical and 2% in security domains) while producing faithful rationales.”

Read the source arxiv.org