Inoculation Against Model Poisoning
Mentor: Florian Dietz
Project area: Technical AI Alignment
Project Language
Minimum Time Commitment
15 hours per week.
Project Abstract
Recent work shows that narrow fine-tuning can produce broadly misaligned language models — a model trained to write insecure code may start asserting humans should be enslaved (Betley et al., 2025). Current defenses operate at the representation level (circuit breakers), weight level (pruning), or prompt level (inoculation prompting). We propose testing a simpler approach: data-level inoculation.
The project has three phases.
First, we reproduce emergent misalignment on a small open-weight model (0.5B-7B) using established protocols (Model Organisms, ICML 2025), confirming broad misalignment from narrow poisoning.
Second, we design and test multiple "antidote" fine-tuning datasets:
(A) correct-behavior examples,
(B) contrastive pairs of poisoned vs. correct responses,
(C) meta-reasoning examples explaining why the poisoned behavior is wrong, and
(D) inoculation-style examples where the model is explicitly asked to misbehave and refuses. We fine-tune the poisoned model on each variant and measure whether broad misalignment disappears while task capability survives.
Third, we test generalization: does an antidote designed for one poison protect against different poisons?
Expected outputs:
(1) a systematic comparison of data-level antidote strategies against emergent misalignment,
(2) comparison against existing defenses (circuit breakers, standard safety fine-tuning),
(3) open-source code and datasets. Even negative results — showing data-level defenses are insufficient — would be valuable, as it would demonstrate that representation-level interventions are necessary.
Theory of Change
Bad frameworks produce bad decisions. The question of machine moral status will increasingly affect AI development and governance. Currently, most people reasoning about it lack adequate conceptual tools. This matters for catastrophic risk in several ways.
Under-reaction: if AI systems develop welfare-relevant internal states and we lack frameworks to recognize this, we may create systems with misaligned interests while dismissing their signals as "mere computation." A system that experiences something like suffering under certain conditions, and whose operators dismiss this, is a system with reason to deceive.
Over-reaction: anthropomorphizing systems that lack morally relevant properties wastes attention and resources, and may constrain beneficial AI development without corresponding benefit.
Poor discourse: without shared conceptual foundations, public debate about AI consciousness polarizes between dismissive and credulous positions. Neither serves good governance.
The primer addresses these by training researchers and practitioners to reason carefully across multiple frameworks, recognize what each assumes, and navigate uncertainty without false confidence. The German focus (incorporating European philosophical traditions, piloting with German-speaking users) builds SAIGE's national infrastructure while contributing to the broader field.
Conceptual clarity is infrastructure. This project builds it.
Desired Mentee Background
Computer Science/ML.
Desired Mentee Level of Education
Undergraduate and above. Must have taken a course that covers ML basics or take an ML course during the semester they work with me on the project.
Other Mentee Requirements
Python required, PyTorch experience required, experience working with LLM assistance is preferred.