Research map / Multi-agent systems and games / Human-AI and goal-conditioned
The Crowded Embedding Space: A Mean-Field Mechanism for Emergent Marginalization in Retrieval-Augmented Agents
Shwan Ashrafi, Dan Roth
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase majority document population Nmaj beyond critical threshold (ablation) | dense retrieval (exact NN) · Synthetic Gaussian (d=2, L=20) | minority retrieval success: collapse once Nmaj exceeds Nc = 490 | Hurt |
| increase embedding dimensionality (ablation) | synthetic retrieval · Synthetic Gaussian | minority retrieval success | Mixed |
| increase majority interferers (ablation) | CLIP image retrieval · CIFAR-100 (otter minority) | minority Recall@1: near-perfect -> below 0.3 | Hurt |
| increase majority interferers (ablation) | text retrieval · Wikipedia Movie Plots (Film Noir) | minority Recall@1: below 0.2 at Nmaj about 2000 | Hurt |
| increase majority interferers (ablation) | text retrieval · Quora Question Pairs | minority Recall@1: near-perfect -> below 0.6 | Hurt |
| increase majority interferers (ablation) | text retrieval · 20 Newsgroups | minority Recall@1: near-perfect -> roughly 0.2 | Hurt |
| raise retrieval budget k from 1 to 10 (ablation) | retrieval under crowding · CIFAR-100, Movie Plots, 20 Newsgroups, QQP | minority Recall@k | No clear effect |
| add cross-encoder reranking to the two-stage pipeline (ablation) | SBERT retriever · 20 Newsgroups, varying shortlist L | minority success rate @1 | No clear effect |
| enlarge shortlist size L above 100 (ablation) | retriever plus cross-encoder · 20 Newsgroups | minority success rate @1 | Hurt |
| enlarge shortlist size L with an oracle reranker (ablation) | retriever plus oracle reranker · 20 Newsgroups | minority success rate @1 | Improved |
| apply accuracy-driven feedback updates to document embeddings | document embeddings · Wikipedia Movie Plots (98% majority queries) | minority Recall@1: > 0.8 for t < 1500 -> below 0.1 within 500 steps after collapse | Hurt |
Closest papers by meaning
- C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation (ICML 2025 · similarity 0.84)
- Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation (NeurIPS 2026 · similarity 0.84)
- DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.84)
- Demystifying and Enhancing the Efficiency of Large Language Model Based Search Agents (ICLR 2026 · similarity 0.84)
- Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations (NeurIPS 2025 · similarity 0.84)
- REGENT: A Retrieval-Augmented Generalist Agent That Can Act In-Context in New Environments (ICLR 2025 · similarity 0.84)
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.
