Research map / Large language models / Retrieval and knowledge
Information-Theoretic Generative Clustering of Documents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes generative clustering (GC) that clusters documents by KL divergence between LLM-generated text distributions, estimated via regularized importance sampling with a doc2query T5 model, and applies it to hierarchical indexing for generative document retrieval.
Problem
Document clustering typically embeds documents as fixed vectors and runs k-means, which fails to capture implicit knowledge in sparse texts and relies on Gaussian assumptions; the paper asks whether LLM-generated texts and their probabilities give a rigorous information-theoretic clustering basis.
Approach
Represent each document by its generation distribution p(Y|x) from a generative LM, define dissimilarity as KL divergence between these distributions, and optimize clusters with a Bregman-style hard clustering whose KL terms are estimated by regularized importance sampling from a shared proposal (the second-moment-minimizing proposal), with log-probability clipping.
Key results
- On R2, GC reaches 96.1% ACC, 77.8 NMI and 84.9 ARI, versus 92.0% ACC for the best SBERT k-means baseline.
- On MS Marco Lite generative retrieval, GCI improves Recall@1 from 23.78 (BMI) to 32.41, a 36% relative gain.
- Setting alpha=1 (unbiased KL estimate) drops R2 NMI from 77.8 to 25.8, showing that variance reduction via alpha=0.25 matters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace embedding-based clustering with generative clustering | document clustering · R2 | NMI: 65.6 -> 77.8 | Helped |
| Replace the BMI document index with GCI | generative document retrieval · MS Marco Lite | Rec@1: +8.63 (36%) | Helped |
| Set α=0.25 rather than α=1 (ablation) | GC · R2 | NMI: 25.8% to 77.8% | Helped |
| Use the naive p(Y) estimator instead of the proposed estimator (ablation) | GC · R2 | NMI: 77.8 -> 77.1 | Hurt |
| Remove probability clipping (ablation) | GC · R2 | NMI: 77.8 -> 73.2 | Hurt |
| Use α=0.1 instead of α=0.25 (ablation) | GCI · MS Marco Lite | Rec@1: 32.41 -> 31.66 | Hurt |
| Use α=0.5 instead of α=0.25 (ablation) | GCI · MS Marco Lite | Rec@1: 32.41 -> 29.80 | Hurt |
| Use α=1 instead of α=0.25 (ablation) | GCI · MS Marco Lite | Rec@1: 32.41 -> 24.43 | Hurt |
| Reduce J from 4096 to 1024 (ablation) | GCI · MS Marco Lite | Rec@1: 32.41 -> 32.16 | No clear effect |
| Reduce J from 4096 to 300 (ablation) | GCI · MS Marco Lite | Rec@1: 32.41 -> 29.08 | Hurt |
| Remove localized proposal distribution φ (ablation) | GCI · MS Marco Lite | Rec@1: 32.41 -> 30.37 | Hurt |
| Use k-means++ instead of random initialization (ablation) | GC · R5 | ARI: 69.1 -> 69.6 | No clear effect |
Surprising findings
- Unbiased importance-sampling estimation (alpha=1) performs much worse than the biased alpha=0.25 variant, because clustering is invariant to scaling and tolerant of bias.
Limitations
- Convergence only to a local minimum; global optimality not guaranteed.
- Uses doc2query models not designed for clustering; fine-tuning for clustering left to future work.
- Requires computing p(y|x) for every document-sample pair, which scales with dataset size times J.
Experimental setup
Datasets: R2 R5 AG News Yahoo! Answers NQ320K MS Marco Lite
Benchmarks: R2 R5 AG News Yahoo! Answers MS Marco Lite NQ320K
Metrics: ACC NMI ARI Recall@1 MRR@100
Baselines: k-means on BoW/tf-idf LDA BERT k-means SBERT k-means GSDMM DEC IDEC STC DFTC NCI BMI
Builds on: k-means Bregman hard clustering doc2query regularized importance sampling k-means++
Closest papers by meaning
- Generative Retrieval Meets Multi-Graded Relevance (NeurIPS 2024 · similarity 0.85)
- Bottleneck-Minimal Indexing for Generative Document Retrieval (ICML 2024 · similarity 0.85)
- DOGR: Leveraging Document-Oriented Contrastive Learning in Generative Retrieval (AAAI 2025 · similarity 0.85)
- Descriptive and Discriminative Document Identifiers for Generative Retrieval (AAAI 2025 · similarity 0.85)
- Retrieval is Accurate Generation (ICLR 2024 · similarity 0.83)
- Learning to Rank in Generative Retrieval (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.