// Sandro Andric
Code. Cats. Robots.
Argues stronger reasoning can degrade behavioral simulation fidelity when the goal is sampling boundedly rational behavior rather than solving strategic problems. Moderate reflection yields more diverse, compromise-oriented negotiation outcomes than either zero or advanced reasoning.
Uses MEG brain data to build a word-level atlas of phase-locking connectivity patterns. Lightweight adapters map LLM hidden states to brain axes, enabling interpretable steering of model behavior across GPT-2, Qwen2, and TinyLlama.
Proposes data scaling exponents can be predicted from dataset geometry via intrinsic dimension. A 10-minute geometric probe can predict scaling behavior before expensive training runs.
We present the Parasitic Manipulation Framework analyzing how "parasite" LLMs can manipulate router decisions. Found 7/10 models achieve 100% capture, while Claude Opus 4.5 shows complete immunity.
We investigate whether LLMs exhibit altruistic tendencies, finding a "virtue signaling gap" where models claim 77.5% altruism but act at 65.6%.
Mechanistic interpretability aspires to reverse-engineer neural networks into explicit algorithms, while model editing seeks to modify specific behaviours without retraining.
Fig 1. Operator
Fig 2. Companions
Fig 3. Inspiration