文件 AI／推論最佳化
Two-Stage Table Compression Uses 64 Visual Tokens to Find Tables, Then Reasons at Full Resolution
University of Edinburgh research shows that although low-resolution tables are unsuitable for directly reading values, they still provide enough information for a VLM to assess relevance. In long-document tests, the training-free, two-stage pipeline reduced decoder token usage by 41% and achieved an accuracy 7.1 percentage points higher than single-pass inference on all tables at full resolution.
A University of Edinburgh team has proposed a “identify first, reason later” pipeline for question answering over documents containing multiple tables. In the first stage, each table is compressed to as few as 64 visual tokens, and the model is asked to return only the IDs of the relevant tables. In the second stage, the selected tables are resubmitted at their original resolution, allowing the model to read values and calculate the answer. The method requires no training and does not modify the VLM architecture.
The study covers five configurations, including Gemma 4 E4B, Gemma 4 26B-A4B, Qwen 3 VL-8B, and Qwen 3.5-9B, and evaluates them on MultiHiertt and FinLongDocQA. A typical FinLongDocQA document contains about 100 pages and 80 tables, with HTML inputs reaching up to 127,816 tokens. Experiments found that directly downscaling tables makes cells difficult to recognize, causing models to produce longer but ineffective reasoning traces. Table-relevance assessment, however, is less sensitive to resolution. Compared with a control that retained only table positions, 64-token images improved table-retrieval F1 by 16 and 10 points, respectively, for the two main models.
On FinLongDocQA, the two-stage method used an average of 40.6% fewer tokens than single-pass question answering at full resolution while improving accuracy by 7.1 points. It also used 14.9% fewer tokens than the best compressed single-pass configuration with comparable accuracy. This suggests a practical design for document RAG: use low-cost representations for routing and reserve high-fidelity content for final reasoning. However, the measurements cover only the language decoder’s input and output tokens, excluding visual encoding, the latency of two requests, and API pricing. FinLongDocQA also sampled only 400 questions. The authors state that the code and data will not be released until formal publication, so the results cannot yet be independently reproduced.