RTriever-4B: A Reasoning-Intensive Retriever

• RTriever-4B is a 4B parameter dense retriever designed for reasoning-intensive tasks, achieving competitive performance with streamlined computational overhead.
• It employs LoRA fine-tuning on Qwen3-Embedding-4B using a specialized synthetic corpus to boost multi-aspect reasoning and evidence retrieval.
• Evaluated with static and agentic benchmarks, it demonstrates efficient aspect coverage and high-quality, iterative retrieval in complex search workflows.

RTriever-4B is a 4 billion-parameter dense retriever designed for reasoning-intensive retrieval tasks, with a particular focus on surfacing complementary evidence to support downstream reasoning within agentic search systems. Developed by LoRA fine-tuning Qwen3-Embedding-4B, RTriever-4B integrates a specialized synthetic training corpus and is evaluated using multi-aspect, agentic benchmarks to probe its reasoning capacity, aspect coverage, and efficiency. The model demonstrates superior performance relative to larger general-purpose embedders, and approaches the best results among reasoning-focused retrievers of greater scale (Zhao et al., 5 May 2026).

1. Model Architecture and Fine-Tuning

RTriever-4B is initialized from Qwen3-Embedding-4B, a general-purpose text embedding model that implements a standard (query, document) dual-encoder interface, facilitating cosine-similarity-based retrieval (Section 5.2). To adapt this backbone for reasoning-intensive tasks, RTriever-4B employs Low-Rank Adaptation (LoRA) by injecting adapters into every linear projection layer of the base transformer. Each LoRA module consists of learnable matrices $A \in \mathbb{R}^{n \times r}$ and $B \in \mathbb{R}^{r \times d}$, with rank $r=16$ and scaling factor $\alpha=32$. During fine-tuning, the weights of the base model remain frozen; only the LoRA parameters (approximately 0.4% of total parameters) are updated, minimizing computational overhead and promoting transfer efficiency (Figure 1, right).

2. Synthetic Training Corpus Construction

The RTriever-Synth corpus underpins RTriever-4B’s training, composed of 140,000 synthetic query bundles, each engineered to foster aspect-level reasoning. The corpus generation pipeline begins by sampling from one million MS MARCO queries, pairing each with human-curated personas from PersonaHub. An LLM is prompted to rewrite each seed into a DeepResearch-style long-form query plus background text. A secondary classifier determines whether the query is “factual” (single positive) or “analytical” (multi-aspect decomposition). Analytical queries are decomposed by LLM into self-contained answers, which are further partitioned into 2–3 non-overlapping reasoning aspects, each realized as a full passage and accompanied by a TL;DR and metadata blueprint (Section 5.1).

For negative sampling, each positive set is paired with aspect-conditioned hard negatives generated by LLMs that share topical cues but explicitly omit one of the reasoning aspects, producing challenging negatives that remain contextually plausible (Section 5.1, Figure 1). This design compels the retriever to discern complementary evidence rather than simple topical matches.

3. Training Objectives and Optimization Protocols

RTriever-4B is trained using a contrastive InfoNCE loss over (query, one-positive, one-negative) triplets with additional in-batch negatives (Section 5.2). Using dot-product similarity and a temperature $T=0.02$, the objective encourages high similarity between query and positive passage embeddings while contrasting against sampled negatives. The loss for a single sample is

$$L_{(q, d^+)} = -\log \frac{\exp(\mathrm{sim}(q, d^+)/T)}{\exp(\mathrm{sim}(q, d^+)/T) + \sum_{i=1}^N \exp(\mathrm{sim}(q, d_i^-)/T)}$$

Training hyperparameters are: LoRA rank $r=16$, scaling $\alpha=32$, peak learning rate $1\times10^{-5}$ with 5% linear warm-up, five epochs over 140K bundles, batch size 384 per GPU (two NVIDIA B200 GPUs yielding 768 effective batch), bf16 mixed precision, DeepSpeed ZeRO-2 optimization, and a sequence length cap of 2,048 tokens.

4. Multi-Aspect and Agentic Evaluation Methods

RTriever-4B is assessed on BRIGHT-PRO, an expert-annotated benchmark expanding on BRIGHT with multi-aspect gold passages and both static and agentic search protocols (Sections 3–4, Figure 2). Human annotators decompose each query into reasoning aspects, weighted on a normalized 1–5 Likert scale. The primary static metric is aspect-novelty-aware nDCG (a-nDCG@$k$) with novelty penalty $B \in \mathbb{R}^{r \times d}$0:

• Gain at rank $B \in \mathbb{R}^{r \times d}$1: $B \in \mathbb{R}^{r \times d}$2
• $B \in \mathbb{R}^{r \times d}$3-nDCG@$B \in \mathbb{R}^{r \times d}$4 normalizes traditional DCG by ideal DCG given multi-aspect coverage.

Weighted Aspect Recall (A-Recall@$B \in \mathbb{R}^{r \times d}$5) counts each aspect once if any retrieved evidence satisfies it, penalizing redundant coverage (Appendix C).

In agentic protocols, retrievers serve as a search tool within an LLM-powered agent that iteratively plans, retrieves, reads, and synthesizes answers, with variants for fixed rounds (R ∈ {1,2,3}) and adaptive rounds (agent halts upon sufficient coverage for ≤ 100 rounds). Efficiency-Quality Reward (AER) is reported for adaptive runs:

$B \in \mathbb{R}^{r \times d}$6

Answer quality and aspect completeness are LLM-judged (Appendix E).

5. Quantitative Results and Performance Analysis

RTriever-4B is benchmarked against 12 baselines spanning lexical (BM25), general-purpose embedders (OpenAI-3L, Qwen3-8B, GTE-7B, etc.), and reasoning-intensive retrievers (ReasonIR-8B, DIVER-4B, BGE-Reasoner-8B, INF-Pro):

Model	Static a-nDCG@25	Adaptive AER	Fixed R=3 Quality (k=15)
BGE-Reasoner-8B	33.8	3.65	4.31
DIVER-4B-1020	30.6	—	—
DIVER-4B	28.9	3.29	4.29
RTriever-4B	27.7	3.51	4.25
INF-Pro-7B	26.3	—	—
Qwen3-8B	23.7	—	—
BM25	14.5	3.31	4.12

(Condensed from Tables 2–4, Section 6.2)

RTriever-4B surpasses all tested 7–8B parameter general-purpose embedders and closely approaches the performance of larger, specialized reasoning retrievers. It achieves the second-best adaptive efficiency-quality trade-off (AER) and maintains strong answer completeness and quality in fixed-round agentic evaluation.

6. Empirical Insights and Observed Behaviors

Multi-aspect and agentic evaluation illuminate deployment-relevant dynamics that traditional metrics obscure (Section 6.2–6.3, Figures 7–11). Standard NDCG@10 fails to penalize aspect imbalance, allowing models to score well when concentrating on a single aspect. In contrast, a-nDCG and A-Recall explicitly surface unbalanced retrieval.

Key findings include:

• Early-round efficiency: RTriever-4B often saturates aspect coverage within one or two agentic rounds (Fig. 7).
• Evidence deprivation: If all gold aspects are missed, LLM agents frequently hallucinate answers using unrelated information (Fig. 8).
• Repetition and “aspect tunnel vision”: Retrieval may stall on a subtopic cluster, producing near-duplicate passages and neglecting diverse aspects (Figs. 9–10).
• Hypothesis hopping: After achieving aspect completeness, agents may pursue unnecessary exploratory queries, raising iteration cost with diminishing quality returns (Fig. 11).

A plausible implication is that multi-aspect training and evaluation are essential for robust agentic retrieval in high-reasoning settings.

7. Deployment Recommendations and Best Practices

Effective deployment of reasoning-intensive retrievers such as RTriever-4B requires synthetic corpora structured with aspect-level positives and hard negatives. Agentic, aspect-aware evaluation protocols are recommended to detect and address evidence imbalance, inefficiencies, and failure cases hidden by classic static metrics. Practical guidance includes:

• Monitoring per-round aspect coverage and halting search agents when saturation is achieved, minimizing hypothesis hopping and redundant retrieval.
• Prioritizing novelty-penalized ranking (a-nDCG) during ranking and evaluation to reflect evidence portfolio balance.
• Combining RTriever-4B with adaptive agentic querying and carefully constructed training corpora maximizes both efficiency and completeness for iterative reasoning workflows.

In summary, RTriever-4B validates the efficacy of LoRA-based fine-tuning on aspect-aware synthetic corpora, achieving competitive or superior performance in reasoning-intensive retrieval relative to much larger general-purpose models when assessed under rigorous static and agentic evaluation protocols (Zhao et al., 5 May 2026).