Multi-View Guided Passage Reranking (MVP)

• MVP is a multi-view guided passage reranking framework that leverages large language models to compute passage relevance efficiently while mitigating biases.
• It employs fixed soft view tokens for independent encoding and a non-generative, anchor-guided decoding process to ensure diverse semantic representations.
• Empirical evaluations on benchmarks like TREC-DL19 demonstrate that MVP delivers competitive nDCG@10 scores with significantly lower latency and FLOPs compared to autoregressive models.

Multi-View-guided Passage Reranking (MVP) is a contemporary framework in information retrieval that leverages LLMs to compute passage relevance by integrating multiple, complementary semantic views for robust, efficient, and unbiased reranking. MVP departs from autoregressive generation-based methods by introducing a non-generative, multi-view encoding and anchor-guided decoding process. The paradigm combines query-aware passage embeddings into several distinct "views"—with explicit architectural and loss function regularization to encourage diversity across views—yielding performance and efficiency gains validated on standard in-domain and out-of-domain benchmarks.

1. Multi-View-Driven Architecture

MVP is architected atop the Fusion-in-Decoder (FiD) backbone, which allows the independent encoding of query–passage pairs into multiple view-specific representations. The MVP input for each candidate passage $c_i$ given query $q$ is constructed by prepending $m$ learnable soft "view" tokens. Each such input $x_i$ is structured as:

$$x_i = \langle v_1 \rangle, \ldots, \langle v_m \rangle~\Vert~ \text{“Query: }q\text{”}~\Vert~\text{“Context: }c_i\text{”}$$

Each view token $\langle v_k \rangle$ is assigned a unique positional embedding, and the encoder produces hidden states $H_i \in \mathbb{R}^{L \times d}$. Extracting $\langle v_k \rangle$ positions yields $m$ distinct passage relevance vectors $e_{ik} = H_i[\langle v_k \rangle]$ for $k=1\ldots m$. This explicit separation enforces the modeling of multiple semantic relevance dimensions, simultaneously mitigating risks of position and selection bias.

2. Anchor-Guided Decoding and Score Computation

Instead of sequential, token-wise autoregressive ranking, MVP introduces a single-step anchor-guided decoding phase. For each view $k$, the encoder outputs across all $n$ candidates are aggregated as $$E_k = [e_{1k}; e_{2k}; \ldots; e_{nk}] \in \mathbb{R}^{n \times d}$$. The decoder employs a [BOS] token to compute the view-specific anchor vector:

$$a_k = \text{FiD}_\text{decoder}(\text{[BOS]}, E_k) \in \mathbb{R}^{1 \times d}$$

The relevance score for each candidate $c_i$ is thereby computed as the average dot product between its $m$ view embeddings and the corresponding anchor vectors:

$$s_i = \frac{1}{m} \sum_{k=1}^{m} \langle a_k, e_{ik} \rangle$$

This direct scoring avoids the latency and resource overhead of autoregressive permutation generation and enables unbiased comparison across candidates regardless of input order.

Orthogonality among anchor vectors is enforced via an explicit regularization term:

$$\mathcal{L}_\text{Orthogonal} = \sum_{k=1}^{m} \sum_{l \neq k} [\text{cos\_sim}(a_k, a_l)]^2$$

where $\text{cos\_sim}(\cdot, \cdot)$ is the normalized dot product. This encourages the views to be maximally diverse and prevents collapse into redundant representations.

3. Performance Metrics and Benchmark Evaluation

MVP is quantitatively assessed using Normalized Discounted Cumulative Gain at rank 10 (nDCG@10) on retrieval benchmarks including TREC-DL19, TREC-DL20, and BEIR. Efficiency is measured via inference latency and floating-point operation counts (FLOPs).

At the T5-base scale (220M parameters), MVP achieves nDCG@10 of 74.3 on DL19 and 69.2 on DL20, surpassing models such as MonoT5, RankT5, and ListT5. The 3B variant achieves further improvements, establishing new state-of-the-art. Significantly, MVP matches the performance of fine-tuned 7B models with only 220M parameters and demonstrates up to 100× lower inference latency and an 82% reduction in FLOPs relative to listwise generative rerankers (see Figure 1 (Na et al., 9 Sep 2025)).

Robustness to input order and identifier permutation is established through extensive ablations (Table 2 (Na et al., 9 Sep 2025)), with multi-view encoding eliminating sensitivity observed in prior architectures.

4. Bias Mitigation

Traditional generative rerankers incur position and selection biases due to sequential decoding and windowed input processing. MVP counters these biases by:

• Using fixed-position soft view tokens across all candidates;
• Independently encoding each query–passage pair;
• Scoring all candidates jointly in a single decoding step, decoupled from input ordering.

Empirical robustness tests confirm invariance to candidate shuffling and identifier changes.

5. Efficiency and Scalability

MVP's design affords practical advantages for deployment in high-throughput retrieval environments. Relevance computation across all candidates is performed in parallel, and the non-generative scoring obviates the bottlenecks of sequence generation. This permits deployment at scale for tasks requiring real-time or large-list reranking without tradeoffs in ranking fidelity.

The model remains effective with compact parameter budgets (e.g., 220M), making it suitable for environments with tight compute constraints. The architecture is directly applicable to other domains that benefit from multi-view listwise scoring, including recommendation systems and multimodal ranking scenarios.

6. Comparative Analysis and Applicability

MVP's multi-view guided reranking strategy generalizes well across in-domain and out-of-domain contexts. It demonstrates state-of-the-art performance without reliance on large parameter counts or specific fine-tuning regimens. The released implementation (https://github.com/bulbna/MVP) offers a practical resource for researchers and practitioners seeking to integrate efficient, robust reranking modules into information retrieval pipelines.

Key advances over prior work include:

• Elimination of input-order sensitivity;
• Efficient single-stage inference versus multi-step generative pipelines;
• Enhanced interpretability and control via explicit multi-view anchor vectors and regularization.

7. Future Directions

The MVP framework is extensible. Future research may consider:

• Increasing the number or functional diversity of view tokens;
• Integrating adaptive or query-conditioned view construction;
• Applying the methodology to further domains requiring listwise, unbiased scoring.

This suggests that multi-view-guided reranking represents a promising direction for modern IR systems, balancing the competing demands of effectiveness, efficiency, and robustness to bias.

---

The MVP architecture advances the state of passage reranking by combining multi-view semantic modeling and anchor-guided score aggregation in a non-generative, bias-resistant framework, and it establishes strong empirical performance across retrieval tasks (Na et al., 9 Sep 2025).