AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research
Kailin Jiang, Lei Liu, Jian Xi, Yangqi Chen, Hui Xu, Hongwei Zhao, Bin Li, Yu Lu and 1 others
University of Science and Technology of China · Tencent
Published September 26, 2026 · on the feed September 29, 2026 · ▲ 18 · arXiv 2609.32472
The abstract
Document rerankers determine what evidence reaches the downstream model in RAG and deep research, yet mainstream rerankers select by relevance matching, and individually relevant documents rarely constitute the complete, complementary, non-redundant set a complex information need demands. Prior work rewards a set by its aggregate rubric score, shifting the objective from ranking documents to composing sets. Yet that score is one scalar shared by every document in the set, so the supervision is sparse: a redundant document is rewarded with the rest whenever the set scores well, and a decisive one penalized with the rest whenever it does not; credit assignment leaves contributors indistinguishable from free riders. On-policy distillation could densify this supervision, but existing methods give every rollout the same fixed guidance, too prescriptive for strong rollouts and too abstract for weak ones. We therefore propose AdaTutoRank, a setwise reranker trained with Adaptive Tutoring Optimization (ATO) under a three-level hierarchy of nine rubric dimensions, which supplies silver labels for the cold start, rewards for reinforcement learning, and hints for distillation. ATO draws three hint forms of increasing specificity from the policy's own frozen snapshot: the rubrics alone, a self-selector's sibling-set chosen under rubrics, and a self-reflector's reflection contrasting the rollout with that sibling-set; each rollout receives the form matched to its quality. Re-scoring that rollout under the hint-conditioned frozen teacher and the hint-free snapshot distills the hint's effect into a token-level advantage that complements the group-relative outcome advantage. Across ten benchmarks spanning RAG, deep research, and setwise evaluation, AdaTutoRank attains the best overall performance while issuing fewer retrieval calls.
The Brief
3:10A short podcast summary of the paper. Written and voiced by AI.
Transcript
AdaTutoRank. Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research. By Kailin Jiang and eight others, from the University of Science and Technology of China and Tencent. Published on the twenty-sixth of September. Its code repository is a placeholder: a readme, and nothing to run yet.
In retrieval-augmented generation and in deep research, the reranker decides which evidence reaches the model downstream. Mainstream rerankers choose by relevance matching, one document at a time. But a complex information need wants a set that is complete, complementary and free of redundancy, and a pile of individually relevant documents rarely adds up to that. Earlier work moved toward sets by rewarding a whole set with its aggregate rubric score. The trouble is that the score is one number, shared by every document in the set. A redundant document gets rewarded along with the rest whenever the set does well. A decisive document gets penalized with the rest whenever it does not. Contributors and free riders look the same.
Put another way, the supervision is sparse, and credit assignment breaks down. On-policy distillation could make that supervision denser, by giving feedback along the way rather than one verdict at the end. Existing versions of it, however, hand every rollout the same fixed guidance. For a strong rollout that guidance is too prescriptive, telling it things it already got right. For a weak one it is too abstract to act on. One level of help does not fit every attempt.
AdaTutoRank is a setwise reranker trained with Adaptive Tutoring Optimization, or ATO. Its core move is to tutor each rollout according to how well it did. Underneath sits a three-level hierarchy of nine rubric dimensions, and that hierarchy does triple duty. It supplies silver labels for the cold start, rewards for reinforcement learning, and hints for distillation. The same rubric therefore shapes the model at every phase of training, from its first labelled examples to its last round of tutoring, and the reranker it produces judges documents as a set rather than one by one.
The hints come in three forms, each more specific than the last, and all are drawn from a frozen snapshot of the policy itself. The mildest is the rubrics alone. The next is a sibling set chosen by a self-selector under those rubrics. The most specific is a self-reflector's reflection that contrasts the rollout with that sibling set. Each rollout gets the form that matches its quality. The rollout is then scored twice, once by the frozen teacher conditioned on the hint and once by the snapshot with no hint, and comparing the two distills the hint's effect into a token-level advantage. That sits alongside the usual group-relative outcome advantage rather than replacing it.
Across ten benchmarks spanning retrieval-augmented generation, deep research and setwise evaluation, AdaTutoRank achieves the best overall performance, and it does so while issuing fewer retrieval calls. The abstract gives no per-benchmark figures, but the pairing of those two outcomes is the notable part: better sets, reached with less searching.
Picture a teacher grading group projects with one mark for the whole team. The paper replaces that with a tutor who looks at each attempt, decides how much help it needs, and hands back feedback far finer than a single mark for the team.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is one score shared by a whole set of documents.
You are building a reranker: it picks which retrieved documents an AI reads before answering, in search-augmented answering and in deep research. Picking each document for relevance alone tends to give piles that repeat themselves and miss half the question.
The road ahead
- 1 · The problem ~6 min Choosing documents one by one for relevance gives poor evidence sets, and scoring the whole set gives one number that cannot tell helpful documents from passengers.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Re-score each attempt with and without a hint matched to its quality, and add that per-token signal to the usual whole-set reward.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Warm up the reranker on rubric silver labels, then train it with rubric rewards plus hint-based per-token credit from its own frozen snapshot.
- 6 · The experiment ~6 min You’ll guess: Did AdaTutoRank's setwise reranking buy its results with more retrieval calls from the agent, or fewer?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.