AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research

• Mainstream rerankers select individually relevant documents, although downstream answering needs a jointly supportive, comprehensive set without redundancy or conflict.Source: §1 · p. 1
• For the Blockchain example, two documents repeat the same point and all three cover benefits, leaving limitations uncovered.The figure contrasts that relevance-based selection with AdaTutoRank's complementary, comprehensive set.Source: §1 · p. 2, Figure 1 · p. 2
• Rubric rewards spread one set-level scalar across documents, leaving redundant free riders unpunished and decisive documents indistinguishable from their set's other members.The paper also identifies reward hacking through selecting a single document and rubric coverage that omits density, completeness, and complementarity.Source: §1 · p. 2
• Existing on-policy distillation varies information content across samples but uses one fixed guidance type, which cannot adapt to rollouts with different failure modes.The paper motivates on-policy, dense, adaptive tutoring as requirements for supervision.Source: §1 · p. 2
• AdaTutoRank carries three-level, nine-dimension rubrics through training and matches hints to rollout quality, combining token-level distillation with group-relative outcome advantage.The rubrics provide silver labels, reinforcement rewards, and distillation hints; high-, medium-, and low-reward rollouts receive rubrics, a reflection, or a sibling-set correction, respectively.Source: §1 · pp. 2–3

• In deep research, rerankers serve each self-issued sub-query, and their selected evidence becomes the observation conditioning the next reasoning step.A defect in that set can propagate through subsequent steps.Source: §2 · p. 3
• Ad hoc rerankers order documents by relevance, while setwise methods predict subsets using answer-generation quality or a strong teacher.Neither setwise approach described here addresses open-ended queries with unverifiable answers.Source: §2 · p. 3
• RubricRanker trains on an aggregate set score that cannot distinguish document contributors from free riders, whereas AdaTutoRank turns rubrics into rollout-matched hints and token-level shaping.Source: §2 · p. 3

3.1 Preliminary

• The setwise formulation predicts a subset with non-additive utility and query-dependent size, rather than truncating a relevance ranking at a fixed cutoff.In RAG the selected set supplies the generator’s evidence once; in deep research it supplies an observation after each search action.Source: §3.1 · p. 4

3.2 Rubrics Construction

• The three-level rubric hierarchy scores documents, inter-document synergy, and the set as LLM context across nine dimensions.The dimensions are relevance, authenticity, quality; complementarity, redundancy, conflict; and completeness, density, reachability.Source: §3.2 · p. 4

3.3 Setwise Supervised Fine-Tuning

• Silver labels from a frontier LLM cold-start the policy, which is fine-tuned to select documents from the query and candidate pool without inference-unavailable rubrics.The silver labels are generated using the query, candidate pool, and query-specific rubrics; the resulting checkpoint initializes the next stage and serves as its frozen distillation teacher.Source: §3.3 · pp. 4–5

3.4 Adaptive Tutoring Optimization

• The selector and reflector derive hints from the frozen policy’s rollouts, and the trained policy uses query and candidates without hints at inference.The sibling-set is gated against the supervised rollout so that a lower-utility sibling-set is not delivered as the reference.Source: §3.4 · pp. 5–7
• A weighted hierarchy of document-, set-, and global-level rubric scores supplies the set utility used as the reinforcement reward.Set and global rubrics are rated for the whole set, while document-level rubric scores are averaged across documents.Source: §3.4 · pp. 5–6
• ATO matches hint specificity to rollout quality and distills the hint’s effect into token-level supervision alongside outcome feedback.Strong rollouts receive rubric criteria, mid-quality rollouts receive reflections, and weak rollouts receive a sibling-set hint; the hint-conditioned teacher is frozen.Source: §3.4 · p. 6

4.1 Settings

• Evaluation covers five RAG and four deep-research answer benchmarks, plus a setwise benchmark scoring document sets on nine dimensions beyond relevance.Baselines include pointwise, reasoning-enhanced, and setwise rerankers, alongside unreranked Initial Retrieval.Source: §4.1 · p. 8

4.2 Analysis of Experimental Results

• AdaTutoRank raises scenario averages by 0.80 on RAG (39.05 vs. 38.25) and 2.05 on deep research (53.06 vs. 51.01), and leads all three rubric-level averages.It ranks first on seven of nine fine-grained dimensions and lifts the overall setwise score by 2.17 points over RubricRanker.Source: §4.2 · p. 9, Table 2 · p. 9
• Ablations show both training stages and both advantages contribute: removing ATO or SFT yields 46.33 or 47.04, while combined advantages reach 49.19.Performance also declines away from λ = 0.7; fixed hints score 47.48 (rubrics), 46.88 (sibling sets), and 47.62 (reflections).Source: §4.2 · p. 9, Table 3 · p. 9
• Across candidate pools from 10 to 50, AdaTutoRank stays strongest; on two deep-research benchmarks it uses the fewest search calls while selecting fewer documents per round.On WebWalkerQA, all three rerankers peak at 30 candidates and decline thereafter.Source: Figure 4 · p. 9, Figure 5 · p. 9, §4.2 · p. 10

5 Conclusion and Discussion

• AdaTutoRank combines hierarchical query-specific rubrics, rubric-conditioned silver-label cold starting, and adaptive tutoring optimization; inference selects a subset without rubrics or hints.Its reported experiments show consistent gains at answer and setwise levels, with fewer search calls in deep research agents.Source: §5 · p. 10

A What Adaptive Tutoring Optimization Optimizes

• The derivation sums token-level divergences over rollout prefixes sampled from the old policy, and its stated gradient result is exact at the snapshot for the first inner step.The clip is inactive at that snapshot because the importance ratio is one and lies inside the clipping interval.Source: A What Adaptive Tutoring Optimization… · p. 15
• ATO’s first inner update follows a reinforcement-learning ascent direction with a reverse-KL pull toward the hint-conditioned teacher.The appendix identifies this objective and notes that its derivation uses advantages fixed within the update.Source: A What Adaptive Tutoring Optimization… · p. 15

AGRPO(y)

• The distillation branch combines token-level hint shaping with a reverse-KL anchor to the cold-start policy, while the objective trades this branch against the group-relative outcome objective.The hint lift depends on the frozen teacher and is independent of the current policy parameters.Source: AGRPO(y) · pp. 15–16
• The cold-start anchor weakens as malformed rollouts are excluded, and neither it nor an explicit KL term bounds cumulative displacement from the cold start.The explicit penalty changes the coefficient of the same type of regularizer rather than adding a different constraint.Source: AGRPO(y) · p. 16
• A moving self-teacher makes the anchor’s gradient vanish at the snapshot, leaving no divergence to a parameter setting held fixed across updates.A separate explicit KL term would be needed to repair that design.Source: AGRPO(y) · p. 16
• Hint-conditioned policy behavior receives no direct gradient in either training stage and evolves only as a side effect of shared-parameter updates.The paper describes hint use as an in-context ability inherited from pre-training and says hints are not invoked at inference.Source: AGRPO(y) · p. 16
• Freezing the teacher requires one extra forward pass per update rather than two because the old-policy probability is already computed for the importance ratio.This is the appendix’s stated computational comparison.Source: AGRPO(y) · p. 16
• The frozen teacher keeps the hint lift fixed, but the lift is measured under that teacher rather than the current policy, making it an off-policy estimate.The appendix identifies this as the price of a stationary target.Source: AGRPO(y) · p. 16

B.1 Details of RAG, Deep Research and Setwise Benchmarks

• The ten-benchmark suite comprises five short-form RAG datasets, four deep-research datasets, and one benchmark that scores retrieved sets.The setwise benchmark reports per-level and overall scores; its results are treated as diagnostic because its criteria and judge differ from those used in training.Source: §B.1 · pp. 16–17
• All five RAG datasets use short, closed-ended answers scored by exact match, covering real queries, trivia, entity-centric, multihop, and curated two-hop questions.The benchmarks include Natural Questions, TriviaQA, PopQA, 2WikiMultihopQA, and Bamboogle.Source: §B.1 · pp. 16–17
• The four deep-research datasets use long-form open-ended responses scored by LLM judges; evaluation samples 100 queries per benchmark, randomly when official sets are larger.The suite includes WebWalkerQA, HealthBench, DeepResearchBench, and ResearchQA.Source: §B.1 · p. 17
• SetwiseEvalKit scores sets across nine dimensions: document relevance, authenticity, and quality; set complementarity, redundancy, and conflict; and global completeness, density, and reachability.The main results use its short-form scenario, where each query has a candidate pool from which a set is selected.Source: §B.1 · p. 17

B.2 Details of Reranker

• Ad hoc baselines rank candidates by pointwise relevance and take the top five documents as the selected set.The listed methods are BGE-Reranker-Large, MonoT5, RankT5, RankLLaMA, RankVicuna, and RankZephyr.Source: §B.2 · pp. 17–18
• Reasoning-enhanced baselines add an explicit reasoning process before ranking, trading inference cost for accuracy on harder queries.Rearank optimizes reasoning-then-ranking with reinforcement learning, while ReasonRank warms up on synthesized reasoning traces before ranking-reward refinement.Source: §B.2 · p. 18
• Setwise baselines predict a retained subset directly, using answer-generation quality, teacher imitation, or aggregated rubric scores as supervision.RubricRanker’s aggregate score is a single set-level scalar shared by all documents in the set.Source: §B.2 · p. 18

RAG Generator

• In RAG, Llama-3.1-8B-Instruct generates short answers from each reranker’s top-k passages.The mid-scale generator is chosen to exploit good evidence without compensating for a poor set through parametric knowledge.Source: §B.3 · p. 18

Deep Research Agent

• In deep research, DR.Tulu-8B alternates planning, tool use, and answer synthesis, with web search, browsing, and paper retrieval available.The agent is initialized from Qwen3-8B and trained end-to-end with reinforcement learning against evolving rubrics.Source: §B.3 · p. 18
• The deep-research agent is held fixed while the reranker changes, so differences in final reports can be traced to evidence composition at search steps.Source: §B.3 · p. 18

C.1 Details of Meta Rubrics

• The nine rubric dimensions form three levels, with templates instantiated into query-specific, weighted criteria that drive silver labels and the reward.Each criterion must name specific entities, facts, or figures from the query and reference answer.Source: §C.1 · pp. 18–19
• Document-level ratings assess relevance, authenticity, and extractability for each document before averaging scores across the set.Authenticity alone requires the reference answer, so it is available during training but not inference.Source: §C.1 · pp. 18–19
• Set-level ratings capture distinct contributions, repeated information, and contradictory claims—properties that belong to document interactions, not individual documents.A set can be non-redundant yet still lack complementarity if its distinct documents are jointly uninformative.Source: §C.1 · p. 19
• Global-level ratings assess whether the set is complete, dense, and sufficient for a model without external knowledge to reach the answer.Completeness is measured against the information need, while complementarity compares contributions among set members.Source: §C.1 · p. 19
• Inapplicable criteria are omitted from both reward numerator and denominator, renormalizing weights over applicable criteria without rewarding trivial non-conflict.Relevance is never null, and completeness and reachability score zero when the relevant subset covers nothing, keeping the utility defined.Source: §C.1 · p. 19

C.2 Details of Training Query

• RAG training uses short questions from four multi-hop datasets, while deep-research training reuses agent-issued sub-queries from RubricRanker rather than tying data generation to one agent.The combined sources vary in form, length, and difficulty and empirically generalize better across both scenarios.Source: §C.2 · pp. 19–20

C.3 Cold-Start SFT Stage

• DeepSeek-V4-Pro produces document-selection silver labels from each query, candidate pool, and query-specific rubrics; Qwen3-8B is fine-tuned without those rubrics because they require unavailable inference-time reference answers.Source: §C.3 · p. 20

13: AGRPO

• The group-relative outcome advantage is standardized within each rollout group and complements the distillation advantage in the update.The implementation excludes malformed identifier outputs from distillation and trains them on outcome reward alone.Source: §C.4 · pp. 20–21, §13 · p. 21

35: end

• Training uses a fixed cold-start teacher and batches rubric-judge criteria; the reported checkpoint is step 30, selected for best validation reward after 1,920 distinct queries.The stated training setup uses 32 NVIDIA H20 GPUs and selects λ = 0.7 by ablation.Source: §C.4 · p. 20, §35 · pp. 21–22

C.4 Adaptive Tutoring Optimization Stage

• Rollouts scoring at least 6 receive rubrics, scores in (3, 6) receive corrective reflections, and scores at most 3 receive sibling sets, subject to a sibling-set quality gate.If the sibling set does not outscore the rollout, the hint falls back to rubrics.Source: §C.4 · p. 20

28: AATD

• The distillation advantage is the hint-conditioned teacher’s log-probability minus the rollout snapshot’s log-probability on sampled tokens.The teacher distribution is truncated to its top 128 logits.Source: §C.4 · p. 20, §28 · p. 21

C.5 Algorithm and Hint Examples

• Representative examples show rubric-only hints, corrective reflections naming documents to add or drop, and sibling sets as concrete references across three reward bands.The table includes rollouts for one RAG query and one deep-research query.Source: Table 5 · p. 22, §C.5 · pp. 22–23

25: end

• A frozen policy snapshot generates rollouts, selects a sibling set, and reflects on differences; its hint-conditioned teacher and hint-free scores yield a token-level advantage combined with GRPO.The teacher remains at the cold-start checkpoint, and the combined advantage is optimized in one clipped policy-gradient step.Source: §C.4 · p. 20, §C.5 · p. 22

D.1

• Appendix D.1 extends setwise-level evaluation to long-form scenarios to check whether evidence-quality gains also hold beyond short, closed-ended queries.Source: D More Experimental Results · p. 23

D.2

• Appendix D.2 reports per-benchmark breakdowns for all four ablations, with overall scores averaged across nine answer-level benchmarks.Source: D More Experimental Results · p. 23

D.3

• Appendix D.3 completes Section 4.2 analyses on omitted benchmarks, covering candidate-pool sizes on NQ and PopQA and search behavior on HealthBench and DRB.Source: D More Experimental Results · p. 23

D.4

• Appendix D.4 tracks hint-type distribution, admission-gate frequency, and sibling-set reference strength as the policy improves to assess whether adaptive assignment is exercised.Source: D More Experimental Results · p. 23

D.1 Setwise-Level Evaluation

• 43.51 overall score puts AdaTutoRank 2.30 points ahead of the strongest baseline, RubricRanker, in the long-form evaluation.The comparison is reported for Table 6.Source: §D.1 · pp. 23–24
• 34.10 versus 29.70 on document-level average and 63.27 versus 60.26 on set-level average show gains over RubricRanker at both levels.The passage attributes the advantage to document choice and combination rather than retaining more documents.Source: §D.1 · p. 24
• AdaTutoRank scores 37.71 on density but trails RubricRanker on completeness (32.00 vs. 35.13), reachability (29.79 vs. 32.27), and global-level average (33.17 vs. 33.66).The passage characterizes the long-form advantage as coming from document- and set-level composition rather than global coverage.Source: §D.1 · p. 24

D.2 Complete Ablation Experiment Results

• Removing ATO costs 1.65 points and removing the cold start costs 2.19, indicating that neither training stage is dispensable.Without the cold start, Bamboogle falls from 20.80 to 14.40; the cold-start removal cost is 1.85 over the remaining eight benchmarks.Source: §D.2 · p. 24
• The combined outcome and distillation advantages score best across both scenario groups, while every fixed hint form underperforms adaptive assignment.Outcome-only scores 42.25 overall; distillation-only reaches 44.70, and the interpolation coefficient performs worse on either side of λ = 0.7.Source: §D.2 · p. 24

D.3 Varying Number of Candidate Documents and Search Call Rounds

• On NQ and PopQA, AdaTutoRank remains strongest as candidate pools grow from 10 to 50; on HealthBench and DRB, it uses 3.03 and 3.39 search calls per query, respectively.It also returns the fewest documents per search call on HealthBench and DRB.Source: §D.3 · p. 24

D.4 Training Dynamics of Adaptive Tutoring Hints

• Rubric hints rise from roughly two fifths to three fifths of rollouts, sibling-set hints shrink from under a fifth to about a tenth, and reflections ease from about two fifths to three tenths.The sibling-set gate passes every rollout in its band, while the reflection band passes 85% to 90%; sibling sets score 7.2–7.6 versus 5.0–5.6 for policy rollouts.Source: §D.4 · pp. 24–25, Figure 7 · p. 25, Figure 8 · p. 25

E.1 Case Study of Query-Specific Rubrics

• The query-specific rubric examples use the same three tiers and nine dimensions, but tailor dimension content and weights to the query.Relevance and completeness dominate the factoid case, while document complementarity carries the most weight in the deep-research case.Source: §E.1 · pp. 25–26, Table 9 · p. 26

E.2 Case Study of RAG

• Across three RAG queries, AdaTutoRank selects the supporting passages without conflicting documents and is the only method reported to answer all three correctly.The baselines sometimes include answer-bearing passages but pair them with conflicting evidence.Source: §E.2 · p. 26

Reranker Selected documents Answer Reranker Selected documents Answer

• For the Natural Questions example, AdaTutoRank selects exactly the two passages that date the 13-inch model, while eleven of twelve baselines forward at least one passage dating another generation.The table marks answer-supporting selections and reports the generator's answer for each method.Source: Table 10 · p. 27

E.3 Case Study of Set Selection

• 3.22 against 2.67 overall scores distinguish the two selected sets, although both receive a global-level score of 3.33.Both sets retain the answer-establishing passages and; RubricRanker additionally forwards and, which add no distinct information and lower relevance and authenticity from 4.0 to 2.5 and complementarity from 2.0 to 1.0.Source: §E.3 · pp. 27–28, Table 13 · p. 30

F.1 ADATUTORANKER Prompt: RAG Setting

• The RAG prompt asks the selector to choose and rank a nonredundant subset of at most 10 passages by their joint usefulness under nine rubrics.It frames selection as answering the question with the best subset, rather than ranking passages solely by individual usefulness.Source: §F.1 · p. 31

F.2 ADATUTORANKER Prompt: Deep-Research Setting

• The deep-research prompt likewise selects and ranks a nonredundant subset of at most 10 documents, using query intent only to clarify the query.The selector is instructed to stay anchored on the search query and use only relevant parts of the agent’s reasoning as intent.Source: §F.2 · pp. 31–32

F.3 Rubric Generation Prompt: Short-form Scenario

• Rubric generation requires binary, grounded, query-specific questions with weights from 1 to 5, correctly scoped across document-, set-, and global-level dimensions.The nine dimensions cover relevance, authenticity, and quality; complementarity, redundancy, and conflict; and completeness, density, and reachability.Source: §F.3 · p. 32
• The short-form example turns the question of which magazine was founded earlier into weighted rubrics covering factual accuracy, evidence for both founding dates, redundancy, contradictions, and whether the set supports the comparison.The example’s answer identifies Arthur’s Magazine as earlier; its rubrics specify 1844 for Arthur’s Magazine and 1989 for First for Women.Source: §F.3 · p. 33

F.4 Rubric Generation Prompt: Long-form Scenario

• Rubric generation assigns doc-level, set-level, and global-level dimensions to individual documents, relationships between documents, and the whole set, respectively.Rubrics must be query-specific, binary, grounded in the query-answer pair, salient, correctly scoped, and weighted; weights range from 1 to 5 according to how severely failure damages the answer.Source: §F.4 · p. 34
• The long-form prompt distinguishes complementarity, redundancy, conflict, completeness, density, and reachability as separate judgments of coverage, duplication, consistency, answer sufficiency, useful text proportion, and self-contained reasoning.It also limits rubrics to three per dimension and calls for coverage without redundant questions.Source: §F.4 · pp. 34–35

F.5 Rubric-based Judge Scoring Prompt: Short-form Scenario

• The judge scores relevance first, excludes documents scoring 0 from other dimensions, and returns those dimensions as null when no relevant documents remain.When relevant documents remain, the other dimensions are evaluated only over that subset and are null if their applicability prerequisites are unmet; relevance is always scored.Source: §F.5 · p. 36
• The short-form rubric separates cross-document division of labor (Complementarity) from whether any required information is missing (Completeness).A single document may cover everything yet yield low Complementarity; missing points affect Completeness, not Complementarity. The Rams example illustrates a complete set with low Complementarity because one document supplies both required points.Source: §F.5 · pp. 37–39
• The short-form prompt scores Relevance, Authenticity, and Quality per document, and Complementarity, Redundancy, Conflict, Completeness, Density, and Reachability for the set.Off-topic documents are excluded from other dimensions; Complementarity, Redundancy, Conflict, and Density are null when their applicability prerequisites are unmet, while Completeness and Reachability measure gaps and derivability.Source: §F.5 · p. 39

F.6 Rubric-based Judge Scoring Prompt: Long-form Scenario

• The long-form prompt requires judging all nine dimensions from document text, scoring Relevance for every document first, and excluding documents scored 0 from all other dimensions.If every document is irrelevant, the other eight dimensions are null; the prompt also directs judges not to use external knowledge to fill gaps in the documents.Source: §F.6 · p. 40
• The judge scores document-level Relevance, Authenticity, and Quality; set-level Complementarity, Redundancy, and Conflict; and global Completeness, Density, and Reachability.Document-level dimensions are scored per document, while set- and global-level dimensions receive one score for the whole set.Source: §F.6 · pp. 40–41
• Complementarity measures division of labour across relevant documents, whereas Completeness measures missing required information, and Redundancy penalizes repeated points that add no new coverage.A set can be incomplete yet complementary, or complete through one document while having low complementarity.Source: §F.6 · pp. 40, 41, 43
• Complementarity, Redundancy, Conflict, and Density are null when their applicability prerequisites fail; Completeness and Reachability instead score 0 when relevant documents supply no required content or reasoning links.For example, Conflict requires at least two relevant documents making comparable statements about the same fact.Source: §F.6 · pp. 41, 43
• In the example, Doc A alone covers both required facts, so Completeness and Reachability are 10 while Complementarity is 3; Conflict is null because only Doc A states the relevant fact.The example distinguishes complete coverage from cross-document division of labour and treats the Titans’ record as non-comparable to the Rams’ record.Source: §F.6 · p. 42
• If every document has Relevance = 0, the other eight dimensions are null; otherwise, assess them only on relevant documents and required information points.Off-topic documents do not count as coverage, distinct points, useful length, or comparable statements.Source: §F.6 · p. 43
• The prompt requires output as JSON containing only the “score” object.Source: §F.6 · p. 43