A mechanism-first reading of how mutual-information-selected geography helps Transformer traffic forecasts avoid the usual trap of using either too much sensor data or too little.
A mechanism-first reading of how VR behavioral data can be compressed into a discrete-event simulator for scalable safety-policy learning—without pretending the learned robot policy is ready for deployment.
A comparison-based reading of how frozen Whisper encoders, attention pooling, and layer choice can make speech emotion recognition cheaper without pretending emotion recognition is solved.
A mechanism-first reading of UAT-Lite, an inference-time method that moves uncertainty from final probability cleanup into transformer attention itself.
A mechanism-first reading of DeltaEvolve: why structured change memory may matter more than larger code histories for LLM-driven discovery agents.
Why frontier research agents can write code, run experiments, and still fail at the part of science that actually matters: designing the right evidence and drawing the right conclusion.
A mechanism-first reading of how a reward-free AI agent can develop a slow, history-shaped internal stance—and why the business value is observability, not consciousness theater.
A new BAPO-CoT paper shows why some reasoning tasks cannot be compressed below linear token growth, and why enterprise AI systems need routing, tools, and architecture—not just shorter prompts.
A new benchmark-alignment paper shows how public LLM leaderboards can be reweighted toward downstream preferences—and why that is useful only when the benchmark already contains the right signal.
A paper on inference-time instability shows how token probability logs can reveal when an LLM’s reasoning trajectory is beginning to unravel.