RenderRank: Learning to Rerank Text with Compressed Visual Tokens
Seongtae Hong, Youngjoon Jang, Jungseob Lee, Hyeonseok Moon, Heuiseok Lim
cs.IR
2026-09-28
RenderRank reranks from compressed visual tokens: 55.96 nDCG@10 on 11 BEIR sets with 16.5-35.5% fewer tokens, 88.27 on long documents at 1.70x throughput.
A reranker scores each retrieved candidate by running the query and that document together. Longer documents make every forward pass more expensive, and the cost multiplies across the candidate list. Truncation cuts compute and can also cut the evidence needed to judge relevance.
Rendering source text as images, then encoding those images with a vision-language model, yields a shorter sequence of visual tokens: patch features the language backbone attends over. That already works for question answering and summarization. RenderRank uses the same compression for reranking. Documents arrive as visual tokens, the query stays as text, and the decoder emits a relevance score. Savings apply once per candidate, so they scale with list size.
Initialization is Qwen3-VL-Reranker-2B, 2.1B parameters in total. The vision encoder and feature mergers are frozen. LoRA (rank 64, alpha 128) is applied only to the language decoder. Document images can be encoded offline; at inference the decoder sees the text query plus cached visual tokens. Think of photographing a page and scoring that photo against the question, instead of feeding every text token into the language model.
Default rendering is Roboto Regular at 12pt, line spacing 1.0, 96 DPI, 896 px wide, height snapped to a multiple of 32 so it matches Qwen3-VL's 16x16 patches with 2x2 spatial merge. 12pt was picked on QASPER and GovReport: 10pt saves more tokens but drops Qwen3.6's QASPER F1 from 48.72 to 41.32, and wider spacing at 12pt adds tokens without a stable gain. Text is black on white with no margins. Whitespace is collapsed to single spaces, so paragraph breaks disappear. Overflow continues on the next image; several images remain one candidate.
Training has two stages.
On 11 BEIR datasets, a heterogeneous retrieval suite, the system reranks BM25's top 100. The metric is nDCG@10, which weights relevant documents more when they sit higher in the top 10. Throughput is GPU forward time on one NVIDIA A6000, reported as pairs per second.
| Method | Params | nDCG@10 | Input tokens | PPS |
| RenderRank | 2.1B | 55.96 | 290 | 76.83 |
| Qwen3-Reranker-4B | 4B | 57.99 | 439 | 22.41 |
| bge-reranker-v2-gemma | 2.5B | 55.82 | 399 | 35.51 |
| LAMAR-600m | 600M | 55.19 | 409 | 67.55 |
| zerank-2-reranker | 4B | 51.98 | 381 | 27.61 |
Input length is 16.5-35.5% below the text rerankers in the comparison. The score is 0.14 above 2.5B bge-reranker-v2-gemma; every other text baseline under 4B is lower, as are zerank-2 at 51.98 and LightOn-rerank-PW-4B at 52.17. The 4B teacher remains ahead at 57.99. Estimated cost is 0.63 TFLOPs per pair, excluding offline vision encoding, versus 1.10 for bge-v2-gemma. 76.83 pairs/s is about twice bge-v2-gemma and above LAMAR-600m at 67.55. If vision encoding is included in the forward pass, throughput falls to 37.31 pairs/s.
On four long-document datasets (MLDR English, plus LongEmbed's 2WikiMQA, QMSum, and SummScreenFD) average nDCG@10 is 88.27, with about half the tokens of the text baselines. Average throughput is 4.51 pairs/s versus 2.66 for the fastest baseline, a 1.70x gap. On MLDR with a 2K length cap RenderRank scores 97.9, above gte-reranker-modernbert at 4K (95.8) and LightOn-PW-4B at 4K (97.3). At a 4K cap it reaches 99.5, close to text models allowed 8K.
Both stages help. Untuned image input scores 50.20, 54.86 after stage 1, 55.96 after stage 2. The same stage-2 weights on text input score 57.21, so the visual path still trails by 1.25 nDCG. Raising font size from 10pt to 14pt moves nDCG@10 from 55.04 to 56.43, tokens from about 236 to 371, and throughput from about 94 to 63 pairs/s. Default 12pt is about 0.83% below 14pt on score, with fewer tokens and higher throughput.
The setting that fits is a mostly static corpus, long documents, and many candidates per query: render once, cache visual tokens, run a shorter decoder, and pack more source text under the same length cap. This is an efficiency increment. It does not set a new quality ceiling. If nDCG is the only target, Qwen3-Reranker-4B is the simpler choice. If documents are rewritten often, vision encoding returns to the hot path and the throughput edge drops by about half.
The paper has no limitations section.
Training overlaps the BEIR eval. Stage 1 includes FiQA, HotpotQA, and FEVER; stage 2 adds ArguAna and SciDocsRR. That is 5 of 11 sets. ArguAna jumps from 60.56 to 69.11 after stage 2. On Touche-2020 RenderRank scores 34.27, below BM25 at 44.22 and mxbai-rerank-large-v2 at 46.94.
Throughput and TFLOPs treat vision encoding as free. Storage for cached visual embeddings is not reported. Rendering wipes paragraph structure. Font and spacing were tuned on generation, not ranking. Results come from a single English run. The vision encoder is frozen, so the compression ratio is a rendering knob rather than a ranking-trained compressor.