The Intrusion Pathway: How Non-Evidential Context Degrades Language Model Performance
cs.AI
Submitted: 2026-11-05
Updated: 2026-11-05
Code: https://github.com/dk4248/the-intrusion-pathway
Project page: https://dk4248.github.io/the-intrusion-pathway
Terminology
Sources
- Lost in the Middle at Birth: An Exact Theory of Transformer Position Bias
- Localizing Model Behavior with Path Patching
- The Llama 3 Herd of Models
- Mistral 7B
- Copy Suppression: Comprehensively Understanding an Attention Head
- Qwen2.5 Technical Report
Related papers
- Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Embeddings for Preferences, Not Semantics
- Can We Trust Item Response Theory for AI Evaluation?
- Diagnosing Task Insensitivity in Language Agents