Research map / Large language models / Language, safety and interpretability
WET: Mitigating World-Conditioned Knowledge Conflicts via World Entropy Tethering
Zixuan Wang, Yifei He, Zihan Wang, Kun Wang, Chaomeng Chen
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 |
|---|---|---|---|
| replace geometry-selected heads with random heads (Random-K) (ablation) | WET · TimeChara | TWCC: 55.8 -> 48.5 | Hurt |
| use norm-based sensitivity instead of signed projection (ablation) | WET · TimeChara | CSC: 62.0 -> 69.4 | Improved |
| patch head to within-world mean instead of zero when scoring heads (ablation) | WET · TimeChara | TWCC: 49.3 -> 55.8 | Improved |
| train score model on true world labels instead of shuffled labels (ablation) | WET · TimeChara | TWCC: 48.7 -> 55.8 | Improved |
| use geometry-selected heads instead of random heads (ablation) | WET · TriviaQA | F1: 84.7 -> 86.8 | No clear effect |
| add entropy-gated head tethering to Llama-3.1-8B | Llama-3.1-8B · TimeChara | TWCC: 35.0 -> 55.8 | Improved |
| add entropy-gated head tethering to Llama-3.1-8B | Llama-3.1-8B · NeoQA | Accuracy: 55.3 -> 91.1 | Improved |
| add entropy-gated head tethering to Llama-3.1-8B | Llama-3.1-8B · TriviaQA | EM: 60.9 -> 86.2 | Improved |
| add a fixed world-aware Static Vector without gating | Llama-3.1-8B · TriviaQA | Oracle EM: 72.2 -> 68.8 | Hurt |
| vary gate threshold tau and scaling lambda (ablation) | WET · TriviaQA | F1: 86.3-86.8 | No clear effect |
| reorder structured prompt blocks (ablation) | WET · TimeChara | TWCC: 55.8 -> 55.5 | No clear effect |
| remove explicit anchor markers, use end of prompt as anchor (ablation) | WET · NeoQA | Accuracy: 92.5 -> 92.6 | No clear effect |
Closest papers by meaning
- Evaluating the World Model Implicit in a Generative Model (NeurIPS 2024 · similarity 0.86)
- Monitoring Latent World States in Language Models with Propositional Probes (ICLR 2025 · similarity 0.86)
- Entropy-Aware On-Policy Distillation of Language Models (ICML 2026 · similarity 0.85)
- BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (NeurIPS 2026 · similarity 0.85)
- Integrating Local and Global Entropy for Uncertainty Quantification in LLMs (NeurIPS 2026 · similarity 0.85)
- Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 2024 · 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.
