Harvard's Gary King and Steven Worthington compared AI-written article reviews with human reviews that editors had rated unusually good. An AI judge preferred the AI review in 132 of the 132 decided pairs, and human journal editors preferred it 87% of the time. AI reviews cited no fabricated sources, and 5.5% of their claims about the manuscript were judged incorrect, versus 15.9% for the human reviews. The authors also discuss downsides for journal policy.
Standard softmax and cross-entropy code, the functions that turn a model's scores into probabilities and a training loss, have two rounding bugs, the authors show. Softmax Collapse zeroes the gradient of a very confident correct answer, so gradients no longer sum to zero; Softmax Cancellation leaves non-zero gradients when a student model already matches its teacher. Their drop-in rewrite removes the Slingshot loss spikes and stops students drifting from teachers in distillation.
Raw-RAG keeps the raw conversation, finds relevant past messages by meaning, reranks them, and hands them to the answering model, with no rewriting of memory. Across LoCoMo, LongMemEval, MEME and MemoryAgentBench, it ranked first in six of seven benchmark-model settings, ahead of the strongest agentic memory system by 0.13-5.53 points. The authors say keeping raw evidence, not rewritten memories, is what matters.