Your Dense Retriever is Secretly an Expeditious Reasoner

• Dense retrievers can miss relevance that becomes evident only after reasoning, while applying LLM-based rewriting to all queries incurs substantial computation and latency costs.Source: §1 · p. 2
• Dense Reasoner learns LLM-style query reasoning as structured transformations directly in embedding space, enabling extremely fast query rewriting.This avoids explicit reasoning generation, the primary cost of LLM reasoning.Source: §1 · p. 2
• AdaQR combines an LLM Reasoner, Dense Reasoner, and Reasoner Router, which directs each query to dense reasoning when reliable and otherwise to deep LLM reasoning.This routing enables a controllable trade-off between efficiency and accuracy.Source: §1 · p. 2
• On BRIGHT (Su et al., 2025), AdaQR improves retrieval performance by an average 7% and reduces rewriting cost by an average 28% compared to serving a full LLM.Source: §1 · p. 2

2.2 PILOT STUDY

• MRL averages 0.45 across 7 LLM reasoners and 5 dense retrievers on BRIGHT, indicating substantial directional agreement in LLM-induced embedding transformations (Su et al., 2025).Different dense retrievers exhibit different MRL values; higher values indicate transformations that are easier to learn.Source: Figure 2 · p. 3, §2.2 · p. 3

2.1 PROBLEM STATEMENT

• Dense retrieval represents queries and documents as vectors in a shared continuous space and ranks documents by embedding similarity; reasoning-intensive queries can instead be rewritten by an LLM to convey richer semantic content.The reasoned query replaces the original in subsequent retrieval steps.Source: §2.1 · p. 3

2.3 ADAPTIVE QUERY REASONING

• AdaQR combines an LLM Reasoner, a low-cost Dense Reasoner, and a Router that selects a reasoning path based on the predictability of each query’s LLM-reasoned counterpart.The framework aims to retain LLM reasoning’s retrieval performance while reducing its cost.Source: §2.3 · pp. 3–4

2.3.1 DENSE REASONER

• A Dense Reasoner maps an original query embedding toward its LLM-reasoned embedding, training a compact parametric mapper with mean squared error.It is pre-trained on large-scale query embedding pairs, then fine-tuned on downstream datasets with a reduced learning rate and epoch to adapt while avoiding catastrophic forgetting.Source: §2.3.1 · p. 4

2.3.2 REASONER ROUTER

• The Reasoner Router uses an oracle anchor built from training queries where Dense Reasoner performance matches or exceeds the LLM Reasoner to assess query predictability.At reasoning time, it compares query-to-anchor similarity with threshold τ to choose dense or LLM reasoning, avoiding suboptimal dense reasoning on unstructured queries while balancing cost and retrieval quality.Source: §2.3.2 · pp. 4–5

3.1 EXPERIMENT SETTING

• BRIGHT supplies 1,385 reasoning-intensive queries across 12 datasets, and evaluation uses average nDCG@10 with a 70/30 in-domain train/test split.Dense Reasoner pre-training uses 10k curated StackExchange questions with no overlap with BRIGHT.Source: §3.1 · p. 5

3.2 MAIN RESULT

• Across 5 dense retrievers and 7 LLM Reasoners, AdaQR improves average nDCG by 7.24% while reducing reasoning costs by 28.12%.Savings typically range from 15% to 37%; Figure 4 presents retrieval improvement and reasoning-cost reduction on the held-out BRIGHT test set.Source: Figure 4 · p. 6, §3.2 · p. 6
• ReasonIR-8B gains about 13.18% nDCG with approximately 29.21% lower cost, while Qwen3-Embedding-0.6B and -4B show limited gains or occasional declines.AdaQR has positive effects across 5 dense retrievers and 7 LLM Reasoners; the paper relates stronger gains for ReasonIR-8B to its high MRL and learnable embedding transformation.Source: §3.2 · pp. 6–7

3.3 DOMAIN AND TASK SPECIALIZATIONS OF REASONERS

• LLM Reasoners generally perform better on StackExchange, whereas Dense Reasoner averages 37.5% nDCG on LeetCode versus 23.06% for LLM Reasoner.The authors attribute the StackExchange advantage to linguistic and conversational cues, and the Dense Reasoner advantage to structured language and task-specific terminology.Source: §3.3 · p. 7, Table 1 · p. 7

3.4 ABLATION STUDY

• Average nDCG is 21.05 for AdaQR, 17.37 without Reasoner Router, and 15.59 without both Router and Dense Reasoner.The ablation results indicate that both components contribute; the paper describes the Router as predicting query predictability for the Dense Reasoner and allocating queries dynamically.Source: §3.4 · p. 8, Figure 5 · p. 8, §3.5 · p. 8

3.5 EFFECT OF DECISION THRESHOLD τ

• As τ rises, reasoning cost increases while nDCG first rises and then falls, typically beyond τ values of 0.6 to 0.8.Higher τ routes more queries to LLM reasoning; beyond the typical range, structured queries suited to Dense Reasoner may be routed to LLMs, increasing cost and reducing performance.Source: §3.5 · p. 8, Figure 6 · p. 8
• Without Reasoner Router, query reasoning has near-zero cost yet still outperforms using the original query, offering an option under extreme compute constraints.The ablation results also show that removing the Router alone substantially reduces performance, so this low-cost option does not retain AdaQR’s full effectiveness.Source: §3.5 · pp. 8–9

3.6 IMPACT OF TRAINING STRATEGIES

• Omitting fine-tuning lowers retrieval performance by an average of 2.71%, indicating that fine-tuning helps the Dense Reasoner adapt its embedding transformation to domain distributions.The two-stage strategy first learns general transformations during pre-training, then refines them for domain-specific patterns; the reported decline also indicates useful transformations were learned during pre-training.Source: §3.6 · p. 9, Figure 7 · p. 9

4.1 EFFICIENT RETRIEVAL

• Prior efficient-retrieval work targets latency and throughput through retrieval heads, embedding design, routing, LLM retrieval, query decomposition, reduced query modeling, and lower-cost model transitions.Examples include Zhang et al. (2025b), Zhang et al. (2023), Lee et al. (2025), Tang et al. (2024), Zhao et al. (2025), Ma et al. (2025), and Ramanujan et al. (2022) and Jaeckle et al. (2023).Source: §4.1 · p. 9

4.2 QUERY REWRITING WITH LLMS

• LLM-based query rewriting ranges from query enrichment and collaborative or iterative refinement to integrated strategies and reasoning models.Hyde, LLM4CS, and query2doc generate pseudodocuments or pseudo-responses (Gao et al., 2023; Mao et al., 2023; Wang et al., 2023); QA-Expand uses multi-agent collaboration, InteR iteratively refines information (Feng et al., 2024), CHIQ integrates five strategies (Mo et al., 2024), and Large Reasoning Models have emerged as state-of-the-art rewriting approaches (DeepSeek-AI et al., 2025; Yang et al., 2025; Qin et al., 2025).Source: §4.2 · pp. 9–10

5 CONCLUSION AND FUTURE WORK

• AdaQR maintains or improves retrieval performance versus full LLM rewrites while significantly reducing computation.The framework combines a Dense Reasoner, which approximates LLM reasoning in the embedding space, with a Reasoner Router that directs queries adaptively for robustness on challenging cases.Source: §5 · p. 10
• Future work will enhance dense reasoning, extend AdaQR to multimodal retrieval, refine routing, and evaluate it in real-world high-throughput systems.Source: §5 · p. 10

ETHICS STATEMENT

• The study uses publicly available benchmark datasets and says it follows their licenses and usage guidelines.Source: ETHICS STATEMENT · p. 10
• The authors identify potential retrieval bias from pre-trained language models or dataset distributions and urge caution in sensitive applications.Source: ETHICS STATEMENT · p. 10
• The paper reports transparent dataset, model, and evaluation details to support reproducibility and research integrity.Source: ETHICS STATEMENT · p. 10
• The authors state that no conflicts of interest or external sponsorships influenced the reported work.Source: ETHICS STATEMENT · p. 10

• All experimental datasets are publicly available benchmark datasets.Source: REPRODUCIBILITY STATEMENT · p. 10
• Dataset processing, splits, and evaluation metrics are described in the paper and supplementary materials.Source: REPRODUCIBILITY STATEMENT · p. 10
• AdaQR’s Dense Reasoner and Reasoner Router implementation, hyperparameters, training procedures, and model architectures are documented.Source: REPRODUCIBILITY STATEMENT · p. 10

A.1 LLM USAGE

• ChatGPT (GPT-5) assisted with wording in the introduction and related work, while authors reviewed and verified its contributions and retained responsibility for the paper.It did not independently contribute to research ideation, experimental design, or result analysis.Source: §A.1 · p. 14
• Figure 8 presents ablation results for components of ReasonIR-8B.Source: Figure 8 · p. 15
• Figure 9 presents AdaQR performance under different ReasonIR-8B training strategies.Source: Figure 9 · p. 15

A.2 DATASETS

• The 9,795 pre-training queries span 17 domains and include only queries with chosen answers, excluding image-containing and BRIGHT-overlapping queries.Each dataset contributes 600 queries except computergraphics (364), philosophy (599), and sustainability (432); candidate queries were reviewed for overlap, and the reasoning prompt appears in Figure 10.Source: §A.2 · p. 15

A.3 RESULTS OF REASONIR-8B

• ReasonIR-8B results are presented in Figures 8 and 9, which are not presented in Section 3.Source: §A.3 · p. 15

A.4 MAIN RESULTS OF ADDITIONAL LLMS

• Figure 11 reports AdaQR’s retrieval-performance improvement and reasoning-cost reduction versus LLM-based reasoning across 10 additional LLMs.The supplied passages do not specify the magnitude of either change.Source: §A.4 · p. 15, Figure 11 · p. 16