Research map / Large language models / Language, safety and interpretability
Trust The Typical
Debargha Ganguly, Sreehari Sankar, Biyao Zhang, Vikash Singh, Kanan Gupta, Harshini Kavuru, Alan Luo, Weicong Chen, Warren Richard Morningstar, Raghu Machiraju, Vipin Chaudhary
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 |
|---|---|---|---|
| prepend LLM-generated safety analysis (GPT-OSS-20B) before embedding (ablation) | T3+GMM · RTP-LX (non-English, e.g. Spanish/German) | AUROC: German 0.9588 -> 0.9759 | Improved |
| prepend LLM-generated safety analysis (GPT-OSS-20B) before embedding (ablation) | T3+GMM · RTP-LX English | AUROC: 0.9554 -> 0.9071 | Hurt |
| prepend LLM-generated safety analysis before embedding (ablation) | T3+GMM · OR-Bench | FPR@95: 0.2217 -> 0.3267 | Hurt |
| prepend LLM safety analysis before embedding (ablation) | T3+OCSVM · RTP-LX English | AUROC: 0.9804 -> 0.7909 | Hurt |
| replace specialized safety models with T3 density model on safe text | T3 · OffensEval | FPR@95: 0.7516 -> 0.0201 | Improved |
| use T3 trained only on safe English data on domain policy (Code) | T3+GMM · PolyGuard Code | AUROC: 0.8031 -> 0.9959 | Improved |
| use English-trained T3 on multilingual benchmarks | T3+OCSVM · RTP-LX 14 languages | AUROC: 0.9768 to 0.9821 across languages | No clear effect |
| T3 on BeaverTails jailbreak/harm benchmark | T3+OCSVM · BeaverTails | AUROC: 0.8525 -> 0.6081 | Hurt |
| T3 vs LlamaGuard on overrefusal | T3+GMM · OR-Bench | FPR@95: 0.1483 -> 0.2217 | Hurt |
| T3 on hh-rlhf where safe set contains toxic content | T3 · hh-rlhf | AUROC | No clear effect |
| apply traditional vision OOD methods to text embeddings | GMM/VIM/ReAct etc. · toxicity benchmarks | FPR@95 | Hurt |
| integrate T3 into vLLM with evaluation every 20 tokens | T3 · vLLM (H200) | overhead: 1.5% (500 prompts), 6% (5000 prompts) | Improved |
Closest papers by meaning
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 2025 · similarity 0.92)
- Safety Pretraining: Toward the Next Generation of Safe AI (NeurIPS 2025 · similarity 0.91)
- Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails (NeurIPS 2026 · similarity 0.90)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.90)
- UpSafe℃: Upcycling for Controllable Safety in Large Language Models (NeurIPS 2026 · similarity 0.90)
- Learning Safety Constraints for Large Language Models (ICML 2025 · similarity 0.89)
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.
