Research map / Large language models / Language, safety and interpretability
InstEmb: Instruction-Following Embeddings through Glimpses of the Future
Tianhao Gao, Jun Fang, Xiaohui Zhang, Zhiyuan Liu, Chao Liu, Pengzhang Liu, Qixia Jiang
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 |
|---|---|---|---|
| apply teacher-guided MSE or KL self-distillation on look-ahead tokens instead of CE (ablation) | InstEmb (no contrastive loss) · Instruction-following mean | mean score: 60.13 -> 64.16 (MSE), 63.41 (KL) | Improved |
| add look-ahead tokens with CE loss instead of plain SFT (ablation) | Llama-3-8B-instruct · FollowIR score | score: 5.0 -> 27.3 | Improved |
| use KL instead of MSE distillation (ablation) | InstEmb · Generic sentence embedding tasks | mean score: 61.81 -> 63.01 | Improved |
| use MSE instead of KL distillation (ablation) | InstEmb · Instruction-following tasks | mean score: 63.41 -> 64.16 | Improved |
| increase look-ahead length from 0 to 1 token (ablation) | InstEmb · most tasks (not IntSTS, IntEmo) | performance | Mixed |
| vary look-ahead length among 1, 4, 8 (ablation) | InstEmb · all tasks | performance | No clear effect |
| add the dropout-augmented student view in contrastive loss (ablation) | InstEmb · Instruction-following mean | mean score: 61.30 -> 67.08 | Improved |
| add teacher input view in contrastive loss (ablation) | InstEmb · Instruction-following mean | mean score: 65.05 -> 67.08 | Improved |
| use DAAP pooling instead of InputLast pooling (ablation) | InstEmb · Instruction-following mean | mean score: 66.72 -> 67.08 | Improved |
| use DAAP pooling instead of mean over all tokens (ablation) | InstEmb · Instruction-following mean | mean score: 63.48 -> 67.08 | Improved |
| use a larger 14B teacher with MLP alignment (ablation) | InstEmb on Qwen2.5-7B-Instruct · Instruction-following mean | mean score: 67.70 -> 68.75 | Improved |
| train on extractive QA or MS-MARCO instead of abstractive QA (ablation) | InstEmb-MSE · Inst. and Gen. means | mean score: 67.08 -> 66.51 (extQA), 66.56 (MS-MARCO) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Promptriever — Promptriever: Instruction-Trained Retrievers Can Be Prompted Like Language Models (ICLR 2025) · better on 2 of 2 shared results
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
- Instruction Embedding: Latent Representations of Instructions Towards Task Identification (NeurIPS 2024 · similarity 0.89)
- SPaR: Self-Play with Tree-Search Refinement to Improve Instruction-Following in Large Language Models (ICLR 2025 · similarity 0.87)
- Instructive Decoding: Instruction-Tuned Large Language Models are Self-Refiner from Noisy Instructions (ICLR 2024 · similarity 0.87)
- InstructRetro: Instruction Tuning post Retrieval-Augmented Pretraining (ICML 2024 · similarity 0.86)
- Steering Information Utility in Key-Value Memory for Language Model Post-Training (NeurIPS 2025 · similarity 0.86)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.86)
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.
