Search for a command to run...
Open research notes on AI safety, alignment, infrastructure, and agents.
A. Mitchell, R. Patel, S. Kim, J. Hernandez, L. Chen
Scaling-law analysis for inference efficiency in transformer language models. We map model size, batch settings, and latency, and show configurations that raise throughput about 3.2x without measured quality loss.
Read on arXivK. Osei, T. Nakamura, P. Svensson
We introduce a novel alignment methodology that uses structured debate between AI agents to surface and resolve value conflicts. This approach reduces reward hacking by 47% compared to standard RLHF while maintaining task performance.
D. Williams, F. Zhang, M. Kowalski
We propose a dynamic routing mechanism for mixture-of-experts models that reduces compute requirements by 60% during inference while maintaining 98.7% of the dense model's accuracy across 15 benchmark tasks.
Read on arXivL. Chen, A. Mitchell, B. Okonkwo
We develop techniques for understanding how multimodal transformers attend across visual and textual inputs. Our methods reveal emergent cross-modal reasoning pathways and enable targeted interventions for bias mitigation.
R. Patel, K. Osei, J. Hernandez, T. Nakamura
We study how language models develop the ability to select, compose, and adapt external tools during multi-step reasoning. Our analysis identifies critical training signals and architectural features that enable reliable tool use at scale.
Read on arXiv