CogSpace is a personal research notebook at the meeting point of computational neuroscience and artificial intelligence. One question runs through every project: how a system, brain or model, keeps track of what it knows and what it does not. Each project is a small, reproducible experiment on that question. The code of some projects is on GitHub.
Results and write-ups are published here. Code goes on GitHub once a project is stable.
Each experiment is meant to be rerun: versioned code, documented pipelines, fixed random seeds.
Neuroscience theories become computational models, and those models test cognitive hypotheses in return.
Predictive Representation for Introspective Spatial Monitoring
PRISM is an agent that learns a map of a maze and, at the same time, a map of what it does not yet know about it. The idea comes from neuroscience: the hippocampus, the brain's spatial mapping system, may also monitor its own knowledge (the meta-map thesis). PRISM is the first computational test of that idea. The uncertainty map lets the agent doubt where it should, explore where it is ignorant, and notice when the maze changes.
Instead of storing where it is, the agent stores where it will probably go next (the successor representation, inspired by hippocampal place cells). One map serves both navigation and self-monitoring.
A second map, derived from the first, records how reliable each part of the first map is: what the agent knows, and what it does not.
The agent's doubt is checked against its real errors with the tools used for human judgement: reliability diagrams, calibration error, and a continuous "I don't know" signal.
Wager-based Assessment of Grounded Epistemic Reasoning
WAGER is a metacognition benchmark for LLMs. The model answers a question and wagers 0-100 points on its own correctness. Correct = +wager, wrong = -wager. Total profit is the single metacognitive score. 292 items across 5 categories test factual knowledge, reasoning, inference, awareness of ignorance, and self-knowledge.
A simple, interpretable protocol: the model risks points on its own answer. Profit separates models that know what they know from those that don't.
Factual knowledge, multi-step reasoning, post-cutoff inference, true unknowables (March 2026), and self-knowledge about own capabilities.
Claude, GPT, Gemini, and Mistral compared head-to-head. Accuracy and metacognition turn out to be two independent abilities.
Exploring LLM Uncertainty, Calibration & Internal Directions
LUCID is an exploratory line of work on how large language models represent, express and regulate their own uncertainty. Each experiment starts from one concrete question (can a model be trained to hedge better? what does fine-tuning actually change inside it?) and looks at it from both sides: behaviour and mechanism.
A supervised fine-tuning method that doubles the hedging-accuracy correlation of Qwen2.5-7B without collapse, and transfers unexpectedly to numerical wager calibration.
Linear probes on base and fine-tuned activations show the base model already has strong latent metacognition (AUC 0.83). TEMPER does not amplify it. It re-routes it.
A three-line prompt forcing an assertive style erases 100% of TEMPER's gain on Qwen, but only 37% of the refusals RLHF installed in Llama-3.1. Routing depth grows with post-training compute. The latent signal does not move.
