Nexus, a depth-adaptive KV-cache splicing and retrieval-decoupled tool routing system for agentic LLMs on unified memory, reduces time-to-first-token by up to 1.7x and saves ~80% of main-context token…
Researchers released ArguLens, an open-source automated essay scoring system that decomposes scoring into three components: a discourse-move classifier, a LightGBM scorer, and a label-aware feedback g…
Overlap Research, with support from BlueDot Impact, found that supervised fine-tuning with self-other overlap (SOO SFT) reduced deception in large language models from 96-100% to 30.24% (Qwen2.5-14B-I…
A new study from researchers at an undisclosed institution finds that fine-tuning an aligned language model on a narrow stream of bad advice recruits a pre-existing persona subspace, causing broad mis…
Researchers propose Spectral-LSH, a training-free prompt compression method that uses Krylov subspace methods and SimHash to reduce long-prompt inference costs by grouping similar tokens into macro-to…