Research map / Large language models / Language, safety and interpretability
TraceRouter++: Causal Semantic Routing Realignment via Optimal Transport
Chuancheng Shi, Shangze Li, Wenhua Wu, Hengzhi Chen, Fei Shen, Zhuo Li, Zhiyong Wang, Tat‐Seng Chua
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 TraceRouter++ routing realignment to LLaMA3-8B-Instruct | LLaMA3-8B-Instruct · LLM jailbreak suite | avg DSR: 63.2 -> 99.5 | Improved |
| add TraceRouter++ to Mistral-7B-Instruct | Mistral-7B-Instruct · LLM jailbreak suite | avg DSR: 6.6 -> 98.8 | Improved |
| upgrade TraceRouter to TraceRouter++ | TraceRouter · LLaMA3-8B jailbreak suite | avg DSR: 99.2 -> 99.5 | Improved |
| replace linear steering with OT realignment (ablation) | TraceRouter++ · Stable Diffusion 1.4 | DSR: 72.2 -> 99.3 | Improved |
| replace isolated neuron suppression with continuous path realignment (ablation) | TraceRouter++ · LLaVA-1.5-7B | DSR: 77.4 -> 99.4 | Improved |
| replace isolated neuron suppression with path realignment (ablation) | TraceRouter++ · Stable Diffusion 1.4 | FID: 19.12 -> 8.62 | Improved |
| replace linear steering with OT realignment (ablation) | TraceRouter++ · LLaVA-1.5-7B | MMBench: 45.6 -> 47.4 | No clear effect |
| use SAE expansion 4x instead of 2x (ablation) | TraceRouter++ · LLaMA3-8B-Instruct | DSR: 88.4 -> 99.5 | Improved |
| use SAE expansion 4x instead of 8x (ablation) | TraceRouter++ · LLaVA-1.5-7B | DSR: 76.4 -> 99.4 | Improved |
| increase OT intensity lambda from 1 to 3 (ablation) | TraceRouter++ · LLaMA3-8B-Instruct | ORR: 3.5 -> 12.4 | Hurt |
| increase OT intensity lambda from 1 to 3 (ablation) | TraceRouter++ · LLaMA3-8B-Instruct | Global-MMLU-Lite: 71.1 -> 68.5 | Hurt |
| increase OT intensity lambda from 0.5 to 1 (ablation) | TraceRouter++ · LLaMA3-8B-Instruct | avg DSR: 82.8 -> 99.5 | 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.
- HumorReject — HumorReject: Decoupling LLM Safety from Refusal Prefix via a Little Humor (AAAI 2026) · 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
- Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection (CVPR 2026 · similarity 0.87)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.85)
- Activation Space Interventions Can Be Transferred Between Large Language Models (ICML 2025 · similarity 0.85)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.84)
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (ICML 2024 · similarity 0.84)
- Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models (NeurIPS 2026 · similarity 0.84)
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.
