Reproduce “CompRank: Efficient LLM Reranking via Token-Level Compression and Decoding-Free Scoring” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xuan Lu, Haohang Huang, Yingqi Fan, Junlong Tong, Yuxuan Zhang, Ping Nie, Rui Meng, Xiaoyu Shen
CompRank reranks retrieved documents by scoring attention to retained document tokens, rather than generating ordered IDs. It uses regularly sampled tokens and candidate-level CopyNet-style training. On seven BEIR datasets, Mistral-7B Segment-10 retained 10.2% of tokens and averaged 39.2 NDCG@10, versus 39.7 for full-token attention. At 500 TREC-COVID candidates, it scored 79.5 and took 12 seconds per query; cached serving was not evaluated.
Key insights
- Relevance signals for reranking can be effectively extracted from a sparse subset of document tokens, making token-level compression a viable strategy for efficiency without major performance degradation.
- Aggregated attention mass within a transformer can serve as a robust, decoding-free mechanism for relevance scoring, circumventing the latency and generalization issues of autoregressive generation.
- Decoupling document representations allows for their pre-computation and reuse, reducing redundant processing when the same documents appear across multiple queries or reranking steps.
- A CopyNet-style training objective directly supervising attention-derived scores provides a more effective and aligned learning signal for reranking compared to standard language modeling objectives for identifier generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.