Chronological feed of everything captured from Bernhard Schölkopf.
PaperMentor is a human-centered writing assistant designed for AI research papers in Overleaf. It leverages a curated expert skill library and 12 specialized agents to provide actionable inline comments, enhancing the drafting process. User studies indicate superior performance over baseline models in terms of comment actionability and validity, with the system released as open-source.
This work introduces Loop-OWM, an object-centric world-modeling architecture designed to learn compositional rules for the Abstract Reasoning Corpus (ARC) by treating rules as transitions over visual-symbolic states. The architecture integrates color-prototype slots, task summaries, and a looped transition model with sophisticated correction mechanisms. Loop-OWM demonstrates superior performance on ARC-1 and ARC-2 compared to baseline models, suggesting an alternative paradigm for rule induction beyond language or symbolic programming.
This work introduces "Echoes of the Prior," an interactive installation that employs a Feed-Forward 3D Reconstruction model to artistically visualize the subjective experience of forgetting. By simulating synaptic decay within the neural network, the project explores the erosion of predictive priors and positions AI as a cognitive proxy to represent the loss of cognitive faculties. The ultimate aim is to catalyze exploration into neuromorphic aesthetics for depicting the fragility of intelligence.
As MoE models trend toward many small ("granular") experts for better performance, the cost of routing tokens to the correct experts becomes a dominant computational bottleneck. AIR-MoE addresses this with a two-stage approach: a coarse vector-quantization shortlisting step followed by exact fine-grained scoring only over the candidate set, approximating true top-k routing without imposing constraints on expert parameters or requiring architecture/loss changes. The method is backed by a theoretical lower bound on mass recall, providing analytical grounding for its empirical gains over existing routers in granular MoE settings.
This paper formally proves, using incomplete contract theory, that mechanism design — the standard tool for aligning individual and collective incentives — cannot eliminate welfare loss in multi-agent settings where contracts fail to anticipate all future contingencies. The gap is strictly positive and irreducible by any realistic mechanism. The authors demonstrate both theoretically and experimentally that LLM-powered agents with intrinsic prosocial preferences (weighting others' welfare alongside their own) can close this gap. The implication for AI safety is that cooperative AI at scale requires agents built with prosocial values, not just better-designed incentive structures.