Research map / Large language models / Instruction and fine-tuning
Latent Paraphrasing: Perturbation on Layers Improves Knowledge Injection in Language Models
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 LaPael, which trains input-dependent learned noise (latent paraphrasers) inserted in early LLM layers to mimic paraphrase-induced latent distributions, enabling paraphrase-like knowledge injection without repeated LLM paraphrasing.
Problem
Fine-tuning-based knowledge injection benefits from paraphrased documents, but paraphrasing with LLMs is computationally expensive for each knowledge update and yields limited diversity; existing latent noise methods ignore text semantics.
Approach
A small learned latent paraphraser inserted before the MLP in early transformer layers produces input-dependent multiplicative Gaussian-based noise with a learnable token mask; it is trained by matching Gaussian approximations of last-token hidden-state distributions with and without paraphrased inputs (symmetric KL), plus a mask loss that spares named entities.
Key results
- On SQuAD-syn with Vicuna-7b-v1.5, LaPael (1k training sentences) reaches 72.47 EM / 84.50 F1 versus 66.30 EM / 76.11 F1 for standard fine-tuning and 70.70 / 79.67 for FreeLB.
- On ArchivalQA-syn, LaPael (1k) achieves 68.37 EM and 84.75 F1, exceeding fine-tuning with 10x paraphrases (64.90 EM, 81.24 F1).
- On raw documents, LaPael obtains the best scores on all four datasets, e.g., Events 2024-raw F1 46.45 vs 39.62 for fine-tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase latent-paraphraser training data from 50 to 1k sentences (ablation) | LaPael · SQuAD-syn | F1 | Helped |
| Increase latent-paraphraser training data from 50 to 1k sentences (ablation) | LaPael · StreamingQA-syn | F1 | Mixed |
| Insert three latent paraphrasers in early LLM layers (ablation) | LaPael · StreamingQA-syn | QA accuracy | Helped |
| Place latent paraphrasers before the MLP layer (ablation) | LaPael · StreamingQA-syn | QA accuracy | Helped |
| Add a trainable token-dependent mask (ablation) | LaPael · StreamingQA-syn | QA performance | Helped |
| Use a concrete-distribution mask instead of sigmoid-only mask (ablation) | LaPael · StreamingQA-syn | QA performance | Helped |
| Use stochastic Gaussian noise instead of deterministic noise (ablation) | LaPael · StreamingQA-syn | QA performance | Helped |
| Use KL loss instead of mean-squared error between distribution means (ablation) | LaPael · StreamingQA-syn | QA performance | Helped |
| Use learnable multiplicative noise instead of alternative noise distributions (ablation) | LaPael · StreamingQA-syn | QA performance | Helped |
| Add a learnable mask to fixed Gaussian or Uniform noise (ablation) | Fixed-noise latent perturbation · StreamingQA-syn | QA performance | No clear effect |
| Combine LaPael with document paraphrases | LaPael · SQuAD-syn | F1 | Helped |
| Use LaPael instead of standard fine-tuning | Vicuna-7b-v1.5 · SQuAD-raw | F1: 29.73 -> 34.65 | Helped |
Surprising findings
- LaPael trained on only 50 sentences already performs competitively, and on ArchivalQA-syn and StreamingQA-syn it outperforms fine-tuning with 10x more paraphrase data on some metrics.
- A learnable mask is reported to be ineffective for fixed (Gaussian or uniform) noise, but important for input-dependent learned noise.
Limitations
- Cannot address the reversal curse; requires reverse-direction information in documents.
- Adds training cost and still underperforms RAG on Events 2024.
- Evaluated only on 7B-scale models (and Phi3-mini), not 13B or 70B; not applied to continual pretraining or instruction tuning.
Experimental setup
Datasets: SQuAD StreamingQA ArchivalQA Films 2024 Events 2024 NovelQA MedMCQA EntityQuestions
Benchmarks: SQuAD-syn StreamingQA-syn ArchivalQA-syn SQuAD-raw StreamingQA-raw Films 2024-raw Events 2024-raw NovelQA-syn MedMCQA-syn EntityQuestions
Metrics: Exact Match Recall F1
Baselines: Fine-Tuning Fine-Tuning (seq.) Fine-Tuning (+ para.) FreeLB NEFTune LoRA fine-tuning RAG
Builds on: NEFTune FreeLB concrete distribution (Gumbel-style relaxation) LoRA CaMeLS
Closest papers by meaning
- Seeking Neural Nuggets: Knowledge Transfer in Large Language Models from a Parametric Perspective (ICLR 2024 · similarity 0.88)
- Training with “Paraphrasing the Original Text” Teaches LLM to Better Retrieve in Long-Context Tasks (AAAI 2025 · similarity 0.88)
- Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis (NeurIPS 2026 · similarity 0.88)
- BLADE: Enhancing Black-Box Large Language Models with Small Domain-Specific Models (AAAI 2025 · similarity 0.87)
- WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · similarity 0.87)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · 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.