Are decision models useful for reranking? This new paper compares decision models with fine-tuned rerankers in a controlled experiment. Before you replace your reranker with a general decision model, run your current reranker and the decision model on your own queries and compare their rankings. The paper uses the same backbone model and the same candidate passages for every version it compares. The authors tested six Qwen backbones from 0.6B to 27B. Every version reranked the same BM25 top-100 candidates on ten collections. Five come from TREC DL 2019-2023, and five come from BEIR. They compared three versions of each backbone. The first is the untuned model asked zero-shot. The second is a decision fine-tune. The third is a plain reranker fine-tune. From 0.6B to 9B, the reranker fine-tune ranks better than the other two on every backbone by 5 to 29 points in nDCG@10. The gap is statistically significant on 48 of 60 comparisons. The reranker fine-tune also wins on collections that none of the fine-tunes were trained on. At 27B, the reranker fine-tune and the decision fine-tune score about the same. No reranker fine-tune had been released at that size, so the authors trained one. Up to 9B, decision fine-tuning stays within 2.5 points below and 0.7 points above the untuned model asked the same question. Reranker fine-tuning adds 6 to 30 points. Decision models are mainly promoted for their calibrated probabilities. The paper shows that the decision models' calibration depends on how the question is worded. When asked whether a passage answers the query, decision models from 4B up are better calibrated than published reranker fine-tunes. With a slightly different question, taken from prior work, the decision models are no longer better calibrated. Better calibration also does not fix the ranking. Up to 9B, a reranker fine-tune keeps better passages at a fixed threshold. arxiv.org/abs/2610.32318 Chat with Paper: academy.dair.ai/papers/syste…

RAG is so back! Another paper showing that plain RAG keeps winning the agent memory comparison. They show that many dedicated memory systems score below naive retrieval over the raw history on LongMemEval, PersonaMem, and MINT-Eval, while costing much more. They split a memory system's value into two parts. Memory benefit counts questions it answers that RAG misses. Memory tax counts questions RAG answers that the memory system gets wrong. Why does the tax appear? Building memory rewrites the history. Some facts get dropped or distorted during extraction, and some survive but become harder to retrieve. Memory systems also vary more than RAG when you change the backbone model. A simple fix helped. Retrieving from the raw history alongside memory and reranking both together reduced the tax. The practical takeaway here is that before adding a memory layer to your agent, compare it with plain retrieval over the transcript on your own data. arxiv.org/abs/2610.34073 Chat with Paper: academy.dair.ai/papers/memor…