Research map / Large language models / Language, safety and interpretability
Let LLMs Speak Embedding Languages: Generative Text Embeddings via Iterative Contrastive Refinement
Yu-Che Tsai, Kuan-Yu Chen, Yuan-Chi Li, Yuan-Hao Chen, Ching-Yu Tsai, Shou-De Lin
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 soft-token generation (without stepwise loss or refinement) to Causal-EOS embedding (ablation) | Mistral-7B, 50K samples · MTEB (English, v2) | average score: 63.84 -> 65.21 | Improved |
| add stepwise contrastive loss to generative embedding (ablation) | Mistral-7B, 50K samples · MTEB (English, v2) | average score: 65.21 -> 65.69 | Improved |
| add iterative refinement regularization on top of stepwise loss (ablation) | Mistral-7B, 50K samples · MTEB (English, v2) | average score: 65.69 -> 66.27 | Improved |
| add soft-token generation to Causal-EOS embedding (ablation) | Mistral-7B, 50K samples · Instruction following (IntentEmotion, NYT) | average score: 47.05 -> 56.47 | Improved |
| add stepwise loss to generative embedding (ablation) | Mistral-7B, 50K samples · Instruction following | average score: 56.47 -> 60.13 | Improved |
| add iterative refinement regularization (ablation) | Mistral-7B, 50K samples · Instruction following | average score: 60.13 -> 62.97 | Improved |
| add iterative refinement regularization on top of stepwise loss (ablation) | Mistral-7B, 50K samples · MTEB Summarization | score: 35.28 -> 34.19 | Hurt |
| increase generated tokens at inference (K up to 20) though trained with K=5 (ablation) | GIRCSE · MTEB STS, clustering, classification | relative improvement over K=1 | Improved |
| generate more tokens at inference for non-generative Causal-EOS baseline (ablation) | Causal-EOS · MTEB tasks | score | No clear effect |
| use GIRCSE instead of Causal-EOS with 50K samples on Qwen2.5-0.5B | Qwen2.5-0.5B · MTEB average | average score: 55.5 -> 61.2 | Improved |
| use GIRCSE instead of Causal-EOS with 50K samples on Llama-3B | Llama-3B · MTEB average | average score: 62.7 -> 65.5 | Improved |
| enable KV caching for GIRCSE inference (ablation) | GIRCSE · FLOPs at sequence length 512, k=5 | FLOPs (T): 44.17 -> 7.40 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- NV-Embed-v1 — NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GRACE: Generative Representation Learning via Contrastive Policy Optimization (ICLR 2026 · similarity 0.87)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.87)
- NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025 · similarity 0.86)
- Generative Representational Instruction Tuning (ICLR 2025 · similarity 0.85)
- Scaling Language-centric Omnimodal Representation Learning (NeurIPS 2025 · similarity 0.85)
- Repetition Improves Language Model Embeddings (ICLR 2025 · similarity 0.85)
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.
