Latest Research: Cognitive bias as a prompt-level behavioural setting for LLMs
Building from Philosophical hyperparameters, I supervised the research of Vladyslav Fridmo, in his MSc dissertation, Cognitive bias as a prompt-level behavioural setting for LLMs looks at how explicitly injecting biased instructions into the prompt can enact and control certain behavioural characteristics
The project investigates a new approach to steering Large Language Model (LLM) behaviour by using cognitive bias as an explicit prompt-level behavioural setting. The aim is to determine whether deliberately induced cognitive biases can reliably influence LLM responses. A central challenge lies in isolating and quantifying these effects, as model randomness and prompt sensitivity can obscure or mimic the intended steering.
The project was carried out as an experimental study using models from two major commercial LLM providers. Ten cognitive biases were organised into opposing pairs to test whether behaviour could be steered both with and against existing model tendencies. Repeated sampling and controlled test variation were used to reduce the influence of unrelated variability. The analysis examined the direction, strength and consistency of the resulting behavioural shifts across the tested conditions.
Latest Paper: Philosophical Hyperparameters, Grounding LLMs via Configurable Ethical and Philosophical Frameworks
Today’s AI systems behave—but they don’t believe.
Their alignment is reactive, shaped by post-hoc constraints and surface-level safety tuning rather than stable, principled grounding.
Philosophical Hyperparameters introduces a new approach to AI alignment, treating ethical and philosophical frameworks not as abstract guidance, but as configurable control parameters for behaviour. Instead of training models toward a single, risk-averse norm, the paper demonstrates how large language models can be grounded in explicit ethical worldviews that act as a moral operating system.
The paper presents a proof-of-concept system in which LLMs are configured with 55 distinct philosophical frameworks, spanning classical ethics, religious traditions, and modern meta-ethical positions. Each configuration is evaluated across 20 complex dilemmas and quantitatively measured along 11 behavioural axes, including harm avoidance, risk appetite, autonomy, rule adherence, and collective good orientation. This produces a multidimensional behavioural signature for each framework, allowing alignment to be measured, compared, and audited rather than assumed.
The results show that LLMs can exhibit deep, coherent, and repeatable behavioural alignment when grounded in explicit philosophical convictions. These behaviours cluster into a small number of emergent moral regimes—such as harm-minimisation absolutism, risk-on activation, and rule-bound legalism—revealing a new design space for controllable, interpretable AI behaviour. The work reframes alignment as configuration rather than constraint, offering a path toward AI systems whose decisions are transparent, intentional, and anchored in known ethical postures rather than opaque optimisation pressures.
Cognitive Geometry for Agentic Memory Systems: Bridging Neural Correlation and Symbolic Reasoning
Today’s AI systems remember—but they don’t think.
Their memories function as static repositories, not dynamic structures of understanding.
Cognitive Geometry in Memory Systems introduces a new way to design memory for AI—one that transforms information into reasoning. The paper presents a four-layer cognitive memory model that maps different types of knowledge to unique geometries:
Facts → Knowledge Webs: semantic graphs that enable deep associative recall.
Experiences → Semantic Temporal Knowledge Webs: time-indexed graphs that preserve sequence, context, and meaning.
Procedures → Causal Trees: hierarchical chains of cause, effect, and subgoal.
Personalisation → Forests of Knowledge Trees: adaptive, individualised reasoning spaces.
This geometry creates a bridge between neural memory (pattern-based learning) and symbolic memory (logic-based reasoning)—allowing systems to store, transform, and reason over knowledge dynamically.
The result is a neuro-symbolic substrate that tackles long-standing LLM challenges: catastrophic forgetting, weak interpretability, and the inability to learn continuously. It reframes memory as the cognitive engine of agentic AI.
How can we help you scale intelligence?
Whether you have a question, an idea, or just want to say hello, feel free to reach out—we’re here to help.