AD-Rank & Shape-CFD: Physics-Inspired Reranking via Conservative Advection-Diffusion on Document Point Clouds for Dense Retrieval
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Every dense retrieval system makes the same bet: one vector per document is enough. A 4096-dimensional embedding is supposed to capture everything—topic, argument structure, terminology, nuance —in a single point. Then cosine similarity ranks these points, and the top results go to the user. This paper starts from a different premise. A document is not a point. It is a cloud of tokens in embedding space, each carrying a fragment of meaning. Collapsing that cloud into its centroid throws away structure—the spread of topics within a paragraph, the cluster of domain-specific terms, the outlier sentence that makes a document uniquely relevant to a query. We do not train a new model to fix this. Instead, we take the hidden states that any LLM already produces—one vector per token —and treat them as a point cloud. On this cloud we perform three geometric operations: (1) PQ-Chamfer distance: split each 4096d vector into 64 subspaces of 64 dimensions, compute cosine distance independently in each subspace, then aggregate via Chamfer matching. This is Product Quantization repurposed—not for compression, but as a distance metric. Each subspace acts as a different "lens" on the same token pair, and subspaces can match different nearest neighbors independently. (2) Graph Laplacian smoothing: build a k-nearest-neighbor graph over the candidate documents using point cloud distances, then propagate scores through the neighborhood. Documents that are geometrically close to high-scoring documents get a boost; isolated high scores get dampened. (3) Score fusion: combine token-level and graph-level signals with a simple linear weight. The results surprised us. On FiQA, a financial QA benchmark where every document is a single sentence, this pipeline improves NDCG@10 by 136% over cosine similarity. On SCIDOCS, a citation prediction task, the gain is 94%. More unexpectedly, the full pipeline surpasses BGE-large-en-v1.5—a model trained specifically for retrieval—on both benchmarks, without any training whatsoever. These gains are not specific to one embedding model. We verified them on three architectures (Qwen3-8B, BGE-M3, BGE-large), finding statistically significant improvements on all three. The weaker the baseline, the larger the gain—from +2.8% on BGE-large to +32.1% on Qwen3. Graph smoothing alone improves all six BEIR datasets we tested, spanning 2,500 to 523,000 documents. Not everything worked. This project began as an attempt to apply computational fluid dynamics to retrieval—convection-diffusion equations on document similarity graphs. That approach failed. So did 20 other ideas we tried: Allen-Cahn reaction terms, Helmholtz decomposition, density-weighted matching, multi-probe retrieval, and more. We report these failures in detail, because they shaped the method that survived: the simplest geometric operations turned out to be the most effective. Our contributions: - PQ-Chamfer distance, a new dissimilarity measure that uses product quantization for distance computation rather than compression. Prior work has not explored this use of PQ subspace decomposition in retrieval. - A training-free pipeline that takes any LLM's hidden states and improves retrieval quality by +2.7% to +136% across six BEIR benchmarks. - Evidence that the pipeline is model-agnostic, improving three embedding models of different sizes and training objectives. - The finding that PQ subspace decomposition benefits even single-vector documents—what we call "implicit point clouds" —demonstrated by FiQA's +136% gain on single-sentence documents. - A record of 21 falsified approaches, establishing where geometric post-processing helps and where it does not.
Fields of Study & Subjects
Information Retrieval and Search Behavior
Topic Modeling
Advanced Image and Video Retrieval Techniques
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 26 March 2026
- Language
- EN
- Document type
- Preprint