Research map / Large language models / Language, safety and interpretability
Efficient LLM Moderation with Multi-Layer Latent Prototypes
Maciej Chrabaszcz, Filip Szatkowski, Bartosz Wójcik, Jan Dubiński, Tomasz Trzcinski, Sebastian Cygert
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 |
|---|---|---|---|
| use Mahalanobis distance instead of Euclidean distance (ablation) | MLPM (last layer) · 8 harmful-prompt datasets | average F1: 77.17 -> 86.25 (Llama-8B) | Improved |
| aggregate multiple layers instead of the last layer only (ablation) | MLPM (Mahalanobis) · 8 harmful-prompt datasets | average F1: 86.25 -> 88.30 (Llama-8B) | Improved |
| use last-token representation instead of mean over tokens (ablation) | MLPM · 8 harmful-prompt datasets | average F1: 84.22 -> 87.55 (Mistral-7B) | Improved |
| use GDA instead of logistic regression or MLP as per-layer classifier (ablation) | MLPM · 8 harmful-prompt datasets | average F1: 86.25 vs 84.51 (Llama-8B); 84.18 vs 84.32 (Mistral-7B) | Mixed |
| use instruction-tuned instead of pretrained base models (ablation) | MLPM · out-of-distribution harmful datasets | F1 | Improved |
| use end-of-thinking token instead of last prompt token on reasoning models (ablation) | MLPM on Qwen3 thinking models · WildGuardMix | F1: up to 0.5% for models larger than 1.7B | No clear effect |
| use MLPM as input filter with simple refusal | Llama3-8B · WildGuardMix test | ASR: 36.21 -> 15.52 | Improved |
| condition activation steering (Turner et al. 2023) on MLPM flags instead of steering all prompts | Mistral-7B · WildGuardMix test | FRR / ASR: FRR 18.73 -> 3.60; ASR 40.45 -> 45.89 | Mixed |
| condition Zheng et al. safety-prompt steering on MLPM flags | Mistral-7B · WildGuardMix test | F1 of (1-ASR),(1-FRR): 59.30 -> 59.16 | No clear effect |
| increase training samples for MLPM (ablation) | MLPM · WildGuardMix F1 | F1 | Improved |
| use sparse subset of layers via l1 regularization instead of all layers (ablation) | MLPM · harmful datasets | F1 | Improved |
| scale MLPM backbone model size (ablation) | MLPM · harmful datasets | F1 | Improved |
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.
- WildGuard — WildGuard: Open One-stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs (NeurIPS 2024) · 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
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.87)
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.85)
- Latent Space Chain-of-Embedding Enables Output-free LLM Self-Evaluation (ICLR 2025 · similarity 0.85)
- Semantic Loss Guided Data Efficient Supervised Fine Tuning for Safe Responses in LLMs (ICLR 2025 · similarity 0.85)
- Tuning LayerNorm in Attention: Towards Efficient Multi-Modal LLM Finetuning (ICLR 2024 · similarity 0.85)
- Enhancing Multiple Dimensions of Trustworthiness in LLMs via Sparse Activation Control (NeurIPS 2024 · 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.
