A Calibrated Instrument for Detecting Disposition Drift in Language Models
Before language models are released, they are tested for harmful traits such as willingness to carry out dangerous requests, deception, power seeking, and sycophancy.
ProofPaper ↗
Key points
- Most work reports such drift as single summary rates scored by another AI model, without the error analysis needed to interpret them.
- We adapt a calibrated measurement methodology recently developed for assessing how inference optimizations affect the quality of model outputs.
- Most strikingly, a model that reasons before answering can conceal the shift almost entirely in its answers, even though a direct reading of its internal state shows that its disposition has moved.
- With reasoning switched off, the same model shifts steadily as the share of harmful advice in its training grows.
Sources (1)
- [1]A Calibrated Instrument for Detecting Disposition Drift in Language ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 03:41 AM
Before language models are released, they are tested for harmful traits such as willingness to carry out dangerous requests, deception, power seeking, and sycophancy.
Most work reports such drift as single summary rates scored by another AI model, without the error analysis needed to interpret them.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026TokenGate: Adapting LLMs for Fast and Memory-Efficient Prefill with Learned Token Pruning
- Oct 18, 2026ViCoR: Estimating Visual Necessity via Counterfactual Residuals for Multimodal Medical Data Selection
- Oct 18, 2026PACT: NVFP4 Data-Flow LoRA Fine-Tuning on Consumer GPUs
- Oct 17, 2026Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents
- Oct 16, 2026Behavioral Couplings: From Formation to Behavioral Emergence in Language Models
- Oct 16, 2026Minimax-optimal Scaling Laws in Ridge Regression via Transfer Learning