Paper Radar

ELI5A benchmark that tracks how detectable AI writing becomes as humans and AI gradually edit a document together over multiple revisions—like watching a document transform from human-written to AI-written in slow motion, across word, sentence, and paragraph levels.

Problem solvedAI detection tools fail in real-world editing where humans and AI collaborate incrementally. This benchmark reveals that mixed-authorship documents (partially edited by AI) are actually harder to detect than purely AI-written ones, exposing why existing detectors fail on real co-authored work.

ELI5Discrete diffusion language models generate text by repeatedly guessing all words at once and keeping confident guesses. This paper uses the low-confidence guesses as hints to retrieve relevant information mid-generation, improving answers before finalizing output.

Problem solvedDiffusion language models are fast at generation but struggle with retrieval-augmented tasks because they commit to outputs without dynamic information lookup. This enables real-time evidence retrieval during generation without retraining, boosting factual accuracy on QA tasks.

ELI5A new way to catch when LLMs make mistakes on complex questions: ask the model to break down a question into steps, then check if its direct answer matches what it gets by actually following those steps. If they disagree, the model is probably wrong.

Problem solvedDetecting when LLMs fail at multi-step reasoning without needing correct answers as reference. Current methods like chain-of-thought sampling are inconsistent across question types; this approach works reliably everywhere and catches errors better than existing confidence scores.