TL;DR
This paper introduces a learnable adapter that dynamically estimates false negatives in dense retrieval, improving training and inference by reweighting negatives and reranking documents, leading to better retrieval performance.
Contribution
The paper proposes a novel adapter module that models false negatives contextually, enhancing dense retrieval through improved negative sampling and reranking.
Findings
Outperforms strong Bi-Encoder baselines on benchmarks.
Effectively identifies false negatives during training.
Improves retrieval accuracy through adaptive reweighting and reranking.
Abstract
In dense retrieval, effective training hinges on selecting high quality hard negatives while avoiding false negatives. Recent methods apply heuristics based on positive document scores to identify hard negatives, improving both performance and interpretability. However, these global, example agnostic strategies often miss instance specific false negatives. To address this, we propose a learnable adapter module that monitors Bi-Encoder representations to estimate the likelihood that a hard negative is actually a false negative. This probability is modeled dynamically and contextually, enabling fine-grained, query specific judgments. The predicted scores are used in two downstream components: (1) resampling, where negatives are reweighted during training, and (2) reranking, where top-k retrieved documents are reordered at inference. Empirical results on standard benchmarks show that our…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The motivation of this paper is clear. 2. The use of the learnable adapter in both training and inference-stage increases its practical value. 3. The reported improvements across standard dense retrieval benchmarks demonstrate the effectiveness of the proposed method.
1. The adapter is trained in Stage 2 on a frozen Bi-Encoder, meaning both embeddings and hard negatives remain static. This may limit its ability to capture evolving training dynamics. When transitioning to Stage 3, the encoder is updated, which could misalign the adapter's decision boundary with the new representation space. The authors should examine whether this decoupled training causes suboptimal performance. 2. While the method is claimed to be lightweight, the paper does not provide a de
(1) This paper addresses an important problem in retrieval training: mining-based hard negatives can sometimes be false negatives, which may harm the retriever’s performance. (2) The effectiveness of RRRA is demonstrated across multiple datasets, including NQ, TQ, and MS MARCO.
(1) Inconsistent comparison results. According to Tables 1 and 2, RRRA outperforms all baselines, including SimANS and TriSampler. However, in Table 4, RRRA performs worse than both baselines across all reported metrics. I assume that the results in Table 4 are taken directly from the original baseline papers, but it remains unclear what configuration differences lead to this discrepancy. (2) Several parts of the paper lack sufficient detail, which makes it difficult to fully understand the pro
* The idea of using one learned signal both to filter negatives during training and re-rank results at test time is **conceptually** interesting. The method avoids cross-encoders yet recovers part of their query-aware behavior via a residual correction and a simple re-ranking rule. * In principle, the adapter sits on top of existing embeddings without adding cross-encoder/LLM reranker cost; it is known that distillation from these models is usually effective but incurs a costly (but one-time, of
My main overall concern is the clarity and readability of the paper. Although I am quite familiar with the IR field, I found the paper difficult to follow—from the motivation to the proposed approach. Many of the presented solutions seemed to lack clear motivation or intuitive explanation. Below, I provide a more detailed discussion of specific concerns. * Clarity: I read carefully the paper twice, and still struggled to understand the proposed solution (and especially, how it answers the probl
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
