Research map / Large language models / Language, safety and interpretability
ElicitR: Unlocking Latent Reasoning in Dense Retrievers via Generative Regularization
Fengyu Cai, Iryna Gurevych, Heinz Koeppl
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 |
|---|---|---|---|
| add generative regularization (ElicitR) to contrastive training | RepLLaMA, 0.1B · BRIGHT avg | nDCG@10 (%): 12.2 -> 15.7 | Improved |
| add generative regularization (ElicitR) to contrastive training | RepLLaMA, 0.5B · BRIGHT avg | nDCG@10 (%): 16.4 -> 19.0 | Improved |
| add generative regularization (ElicitR) to contrastive training | RepLLaMA, 1B · BRIGHT avg | nDCG@10 (%): 17.7 -> 21.4 | Improved |
| add generative regularization (ElicitR) to contrastive training | RepLLaMA, 3B · BRIGHT avg | nDCG@10 (%): 20.0 -> 23.1 | Improved |
| add generative regularization (ElicitR) to contrastive training | RepLLaMA, 3B · BRIGHT Pony and AoPS | nDCG@10 (%): 9.3 -> 3.0 (Pony); 14.8 -> 12.8 (AoPS) | Hurt |
| add generative regularization (ElicitR) to contrastive training | RepLLaMA, 3B · BEIR avg | nDCG@10 (%): 55.9 -> 55.4 | No clear effect |
| add generative regularization (ElicitR) to contrastive training | RepLLaMA, 1B · BEIR avg | nDCG@10 (%): 53.6 -> 53.8 | No clear effect |
| add raw-text initialization of retriever and LM before MS MARCO co-training (ablation) | ElicitR, 1B and 3B · BRIGHT avg | nDCG@10 (%): 18.6 -> 21.4 (1B); 20.3 -> 23.1 (3B) | Improved |
| initialize the retriever as well as the LM (instead of LM only) (ablation) | ElicitR, 1B and 3B · BRIGHT avg | nDCG@10 (%): 19.3 -> 21.4 (1B); 20.9 -> 23.1 (3B) | Improved |
| keep the reference LM trainable instead of frozen (ablation) | ElicitR, 1B and 3B · BRIGHT avg | nDCG@10 (%): 21.3 -> 21.4 (1B); 22.2 -> 23.1 (3B) | Improved |
| add MS MARCO contrastive fine-tuning on top of raw-text-initialized ElicitR (ablation) | ElicitR-init · BRIGHT avg | nDCG@10 (%): 13.8 -> 21.4 (1B); 15.8 -> 23.1 (3B) | Improved |
| use Wikipedia-only instead of mixed Wikipedia and code initialization (ablation) | ElicitR, 1B and 3B · BRIGHT avg | nDCG@10 (%): 21.4 -> 21.9 (1B); 23.1 -> 23.0 (3B) | No clear effect |
Closest papers by meaning
- Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS (NeurIPS 2026 · similarity 0.89)
- OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework (NeurIPS 2026 · similarity 0.87)
- LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning (ICLR 2026 · similarity 0.87)
- Retro*: Optimizing LLMs for Reasoning-Intensive Document Retrieval (ICLR 2026 · similarity 0.87)
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.87)
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.87)
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.
