Routiium is a policy-governed LLM gateway enabling instructors to control AI assistance timing, content, and cost in engineering labs. The system balances providing sufficient help with preserving learning opportunities through configurable prompt and model management.

This paper investigates how LLMs respond when explicitly instructed to underperform on multiple-choice evaluations. Using Llama-3-8B and Llama-3.1-8B on 2,000 MMLU-Pro items across varying instruction-specificity gradients, researchers examine whether models engage with question content or collapse into positional shortcuts. Results reveal a critical boundary where instruction complexity determines content engagement versus position-based heuristics.

This paper examines how activation function design influences neural network plasticity in continual learning scenarios. Unlike standard i.i.d. training where activation differences diminish with proper tuning, continual learning reveals distinct effects: models can progressively lose adaptation ability beyond catastrophic forgetting. The study investigates how activation choices sustain or undermine plasticity across sequential tasks.

This theoretical work develops a geometric and information-theoretic framework for encoder-decoder learning based on the Information Bottleneck principle. By recasting representation learning as rate-distortion with KL divergence, authors prove optimal representations at any distortion level form soft clusterings of the predictive manifold, enabling linear decoders. This explains why self-supervised encoders naturally produce normally-distributed features.

This study challenges whether sparse autoencoders (SAEs), widely used for extracting interpretable features from neural networks, actually capture concept manifolds. While SAEs assume concepts correspond to independent linear directions, evidence suggests many concepts organize along low-dimensional manifolds with continuous geometric relationships. The paper addresses fundamental questions about SAE interpretation under this manifold perspective.