acceptodds
Under review as a conference paper at ICLR 2027

GateReranker: Listwise Reranking with Joint Ranking and Adaptive Selection

Abstract

Reranking is a crucial component in modern multi-stage retrieval and retrieval-augmented generation systems, where it is widely used to refine the ordering of candidate documents before downstream processing. In practical pipelines, however, the reranked list is usually followed by a selection step, such as fixed top- truncation or global score thresholding, to decide which documents should be retained. Such post-hoc selection strategies are simple and effective, but they rely on static decision rules and may be brittle across queries, domains, and candidate sets. In this work, we propose GateReranker, a listwise LLM reranker that unifies document ranking and adaptive selection within a single model. GateReranker introduces a dedicated gate token to learn a list-conditioned cutoff, allowing the model to retain documents whose relevance scores exceed the gate score while simultaneously producing their relative ordering. Experiments on BEIR show that GateReranker achieves the best retained-set F1 and the highest nDCG compared with strong open-source rerankers paired with fixed top- truncation and global score thresholding. These results suggest that adaptive selection can be naturally integrated into reranking, providing an effective in-model alternative to static post-hoc filtering.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Listwise Reranking, Adaptive Selection, Retrieval-Augmented Generation, Information Retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_xhoiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a method for simultaneous listwise re-ranking and filtering of documents. Filtering uses a new token to set the threshold and is jointly fine-tuned with the re-ranking objective.

Critical strengths
  • The problem of relevance filtering (or defining a cutoff threshold for ranking) is mostly underexplored, and it's nice to see work that contributes to this space.
  • The results are really strong compared to the provided baselines (though I have concerns about the soundness of the baselines; see below). Table 5 is critical in showing the direct effects (the w/o Gate Loss rows).
Critical weaknesses
  • The system is compared against really weak baselines. Most importantly, they're missing experimental and prose comparisons with a conceptually similar paper. Beyond this, less naive truncation methods (like the ones covered in the related work) should have been considered. In short, too many of the conclusions hinge on comparisons against weak top-k and fixed thresholding baselines. As a related issue, the baselines also do not share the backbone model or the exact training procedure, but the controlled ablations in Table 5 partially make up for this, though they still would have benefited from stronger truncation baselines there too.

EDIT: I realized that I forgot to mention the paper, https://arxiv.org/abs/2402.02764, at WWW 2024

Questions for authors
  • For the baselines that use a globally tuned score threshold, what was the average number of documents returned per query?
2 comments

Response to Reviewer xhoi -- Part 1/2

We thank the reviewer for recognizing the value of jointly studying reranking and relevance filtering, and for identifying the related WWW 2024 paper. We have added evidence addressing the three main concerns: comparison with prior work, baseline strength, and the number of documents actually retained.

1. Clarifying the relationship to GenRT.

GenRT is directly relevant prior work. We will add an explicit citation and methodological comparison and revise the original characterization of the literature.

GenRT uses an encoder--decoder architecture to generate the reranked list sequentially while making truncation decisions. GateReranker instead reads document identifier logits and a gate logit at the first prediction step and obtains the retained set by comparing them, without sequentially generating document identifiers (Xu et al., List-aware Reranking-Truncation Joint Model for Search and Retrieval-augmented Generation, WWW 2024; arXiv:2402.02764).

The contribution should therefore be positioned as a specific first-step-logit formulation that learns a list-conditioned boundary in a shared output space, rather than as the first method to jointly model reranking and truncation. We have not completed a direct experimental comparison with GenRT and do not claim empirical superiority over it.

2. Completed matched training and more extensive tuning of truncation baselines.

We have completed training and evaluation of FIRST-style, Rank-only, and Gate on all 13 tasks. The three models share an initialization checkpoint, the same seven-task training mixture, and a fixed budget of 2,000 optimizer steps. We also verified that the evaluation queries, candidate documents, and relevance labels match across models.

FIRST-style removes the gate token and gate losses, while Rank-only retains the gate-aware input and removes only the gate losses. FIRST-style denotes a controlled first-step scoring implementation sharing our ranking objective, rather than a complete reproduction of the original FIRST recipe. The budget is matched in data and optimizer steps, not GPU-hours; actual training implementation details will be documented.

Overall standard nDCG@10 is 0.4247, 0.4630, and 0.4763 for FIRST-style, Rank-only, and Gate, respectively. For retained-set selection, we additionally allowed the post-hoc baselines to use test labels to select the best global threshold:

Model and selection ruleOverall F1
FIRST-style + oracle global threshold0.2584
Rank-only + oracle global threshold0.2892
Gate scores + oracle global threshold0.3013
Gate + learned boundary0.3594

This oracle diagnostic exhaustively covers the distinct retained sets obtainable by globally thresholding the saved scores. Thus, within the current scoring implementation, the observed difference cannot simply be attributed to an insufficient global-threshold sweep.

The diagnostic applies to the saved sigmoid scores. Because the current implementation applies sigmoid in BF16 before conversion to floating-point values, saturation and ties may occur. We therefore do not interpret these results as upper bounds for arbitrary raw-logit representations or alternative calibration methods.

Similarly, sweeping fixed K over {1,2,3,5,10,20,100,150} gives best F1 scores of 0.2772 for FIRST-style and 0.3039 for Rank-only, compared with 0.3594 for Gate. In particular, the comparison between 0.3013 and 0.3594 holds the Gate model and its document scores fixed, isolating the inference-time selection rule.

These experiments strengthen the controlled evidence in the paper. We agree, however, that they serve a different purpose from direct comparisons with learned truncation methods such as GenRT, and will clearly distinguish the two.

Throughout these selection analyses, F1 is computed per query and macro-averaged across tasks, with recall conditioned on relevant documents in the BM25 top-150 candidate pool.

Response to Reviewer xhoi -- Part 2/2

3. Average number of documents returned by global thresholds.

We have computed the requested statistics from the formal per-query outputs of the original paper models. Each value below is obtained by first averaging the retained count over queries within each task and then macro-averaging across the 13 tasks.

Original paper model and selection ruleMean documents retained
BGE-v2-m3 + training-calibrated threshold5.66
Qwen3-0.6B + training-calibrated threshold4.60
Qwen3-4B + training-calibrated threshold2.01
Qwen3-8B + training-calibrated threshold2.31
Jina-v3 + training-calibrated threshold3.12
Gate scores + training-calibrated threshold0.55
Gate + learned boundary12.97

For completeness, in the additional matched-training experiments, FIRST-style and Rank-only retain averages of 1.72 and 1.50 documents under their respective overall test-oracle global thresholds. The newly trained Gate model retains an average of 9.60 documents using its native boundary. These results are reported separately from those of the original paper models.

We will report retained counts together with precision, recall, and F1, rather than treating a larger retained set as an improvement by itself. Under the unified offline recomputation for the original paper models, Gate achieves an average F1 of 0.4174, compared with 0.3778 for Jina, the strongest global-threshold baseline. The difference is 0.0396, with a paired-bootstrap 95% confidence interval of [0.0327, 0.0461]. This supports an aggregate selection benefit, not a claim of superiority on every task.

We have also separated standard full-list nDCG from post-truncation selection metrics and will correct the original ranking-performance statements accordingly.

We hope that the completed matched-training comparisons, exhaustive global-threshold diagnostic, and retained-set statistics address the main concerns about baseline reliability and the sources of improvement, and provide useful evidence for reconsidering the paper.

Reviewer_UijLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Standard retrieval pipelines typically apply separate post-hoc selection steps after reranking which can be brittle. To address this issue, the authors introduce GateReranker, a listwise LLM reranker built on Llama-3.2-3B. The model uses first-step logits of document identifiers for relevance scoring and adds a dedicated gate token to learn a list-conditioned cutoff. Both ranking and gating objectives are trained jointly. On the BEIR benchmark, GateReranker outperforms baselines like BGE and Qwen3-Reranker, reaching an average retained-set F1 of 0.42 and an nDCG of 0.65

Critical strengths
  • Clean, intuitive idea, and inexpensive during inference. Using the no-relevant-passage token’s logit as the list-conditioned decision boundary and in the same output space as document-identifier logits is a good idea.
  • Sufficient baselines and strong results reported consistently
  • Good isolation study (Table 4) and ablations (Table 5) holding the trained model fixed and varying only the selection rule
  • Authors state plainly that the prompt elicits a single-choice distribution and that the losses reshape the full logit distribution so multiple documents can pass the gate.
Critical weaknesses
  • Related work mentions techniques like FIRST and SCaLR which are omitted as baselines. Reading first-step identifier logits as relevance scores is similar to FIRST and adapting listwise scores for retained-set selection is the idea in SCaLR. The only critical addition is the learned gate token acting as an in-model cutoff. The paper only compares the gate against fixed top-K truncation on GateReranker's own logits in Table 4. This serves only as a self-ablation rather than a benchmark against prior methods
  • Related to above, the conceptual novelty is slightly incremental only
  • GateReranker is trained on 7 BEIR tasks and tested on the same tasks plus another 6. Every baseline is run zero-shot with no task-specific finetuning thus inflating the overall average F1and the seen average. The unseen set is not a clean control for the above reason and tasks like Climate-FEVER and SCIDOCS closely mirror the seen FEVER and SciFact training sets.
  • The paper focuses completely on English retrieval tasks from the BEIR benchmark. Appendix C figure 3 evaluates score separation using CMedQAv2-reranking which is a Chinese medical dataset. This dataset is never mentioned in the experimental setup, the method section, or the main tables. The authors do not explain why a Chinese dataset appears in an English study. This inconsistency raises questions about what data the model was actually trained on
  • The gate over-selects documents on unseen tasks with few relevant answers, going against the paper's core motivation. On NQ and ArguAna, where queries typically have only one target document, the model retains an average of about 12 and 5 passages. This hurts precision, causing F1 to drop to 0.18 on NQ (baselines reach 0.44) and 0.21 on ArguAna (baselines reach 0.65). So the learned boundary fails exactly where adaptive depth should help most which is out-of-domain queries with little relevant evidence
Questions for authors

Can you provide evaluation results where at least one baseline reranker is fine-tuned on the exact same seven-task BEIR training mixture and compute budget before applying post-hoc truncation rules? This is a clarification request and not another future ablation study

2 comments

Response to Reviewer UijL -- Part 1/2

We thank the reviewer for the positive assessment of the gate design, inference efficiency, and isolation analyses. In response to your concerns about training fairness, contribution positioning, and data provenance, we have completed additional training and evaluation and added more direct evidence for the proposed mechanism.

1. Completed comparisons with matched training data and optimization budgets.

We trained three additional models: FIRST-style, Rank-only, and the full Gate model. They share the same initialization checkpoint, seven-task BEIR training mixture, and main optimization settings, and each is trained for a predetermined 2,000 optimizer steps. The final test metrics were not used for checkpoint selection in these runs.

FIRST-style removes the gate token and gate losses. Rank-only retains the gate-aware input but removes the gate losses. These two controls examine the complete gate design and the contribution of gate supervision, respectively.

All three models have now been evaluated on all 13 tasks:

ModelSeen nDCG@10Unseen nDCG@10Overall nDCG@10
FIRST-style0.52730.30490.4247
Rank-only0.57000.33820.4630
Gate0.59470.33810.4763

For retained-set selection, which is the central objective of the paper, the results are:

Selection methodOverall F1
FIRST-style + best fixed K0.2772
Rank-only + best fixed K0.3039
FIRST-style + oracle global threshold0.2584
Rank-only + oracle global threshold0.2892
Gate + learned boundary0.3594

Here, fixed K is selected from {1,2,3,5,10,20,100,150}, and the global threshold is swept over every distinct operating point in the saved scores. Both are selected to maximize test F1 and therefore constitute optimistic diagnostic references, not validation-tuned deployment results. Gate directly uses its predicted boundary. Selection metrics use relevant documents within the candidate pool as the recall denominator and are macro-averaged across tasks.

These completed comparisons provide evidence beyond comparisons with external rerankers that did not receive the same BEIR fine-tuning. The matched budget refers to data and optimizer steps, rather than identical GPU-hours. We will document implementation differences, including the training attention backends. FIRST-style is a controlled implementation sharing our ranking objective, rather than a complete reproduction of the original FIRST recipe.

2. The contribution is boundary learning, not first-token scoring itself.

We agree that the contribution should be defined more precisely. FIRST's first-step scoring and SCaLR's relevance-score calibration are important foundations. SCaLR focuses on obtaining comparable relevance scores through self-calibration; our focus is to explicitly learn a query- and list-conditioned selection boundary in the same output space as the document identifier logits (Ren et al., Self-Calibrated Listwise Reranking with Large Language Models, arXiv:2411.04602).

The additional experiments further isolate this distinction. For the same Gate model, the test-oracle global threshold yields an F1 of 0.3013, whereas its native gate yields 0.3594. Because the model and document scores are fixed, this comparison directly supports the value of the learned selection boundary.

The threshold diagnostic is specific to the saved sigmoid scores. The current implementation applies sigmoid in BF16 before conversion to floating-point values, so saturation and ties may occur; the diagnostic is not an upper bound for arbitrary raw-logit representations or calibration methods.

We will accordingly refine the contribution statements: neither first-token scoring itself nor the general idea of jointly modeling ranking and selection is claimed as new. We have not completed a direct experimental comparison with SCaLR and do not present these controlled experiments as a substitute for that comparison.

Response to Reviewer UijL -- Part 2/2

3. CMedQAv2 and training-data provenance.

The mismatch between the appendix visualization and the stated experimental scope requires clarification. CMedQAv2 is not included in the seven-task BEIR fine-tuning files used in these runs, nor is it one of the 13 formal BEIR evaluation tasks. We have generated a replacement score-distribution analysis from the formal BEIR outputs so that the appendix visualization matches the scope of the main experiments.

In this replacement analysis, positive--negative score AUROC is 0.9713 on seen tasks and 0.8682 on unseen tasks. These results illustrate both the model's relevance discrimination and the change under transfer. The figure depicts document-score separation; we do not describe it as a gate-logit margin analysis.

We also verified that the models are initialized from the paper's existing SFT checkpoint, rather than directly from an original Llama checkpoint without task-specific fine-tuning. The "seven-task training" description refers specifically to the subsequent BEIR fine-tuning stage. We will distinguish initialization provenance from this training stage so that the paper does not imply that the model's entire training history consists only of those seven tasks. All three additional variants share this initialization, controlling for it within the new comparison.

4. Over-selection on NQ and ArguAna, and the meaning of unseen tasks.

We confirmed the over-selection identified in the review. The original paper model retains an average of 11.95 documents on NQ and 4.92 on ArguAna, indicating that its learned boundary can be too permissive when few documents are relevant. We will state this limitation explicitly: "adaptive" does not mean that the model always finds the optimal truncation depth for every task.

At the same time, the additional experiments provide evidence for the aggregate value of adaptive selection. Across the six unseen tasks, Gate achieves an F1 of 0.2024. Even when FIRST-style and Rank-only are each allowed to choose their best single global threshold using the unseen test labels, their corresponding F1 scores are 0.1293 and 0.1487. This supports the overall benefit of cross-task adaptive selection without resolving the calibration limitations on individual tasks.

We also accept the distinction concerning task similarity. "Unseen" means that a task was not included in the BEIR fine-tuning mixture; it does not establish complete semantic-domain isolation.

We appreciate these comments and hope that the completed matched-training comparisons, direct selection analyses, and clearer account of data provenance address the central concerns about experimental fairness and the credibility of the contribution.

Reviewer_Gp7xOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper fine tunes a listwise reranker that can also group documents into relevant and irrelevant, adoptively selecting the optimal k (the final number of returned candidates). The listwise reranker follows the FIRST training recipe using the logits of document identifiers as their relevance score, with a new gate token and gate loss functions added. At inference time document identifiers scored below the gate are considered irrelevant. The training is done with a subset of BEIR, eval is only on BEIR too with BM25 as the first stage.

Critical strengths
  • The core idea is interesting and well motivated in RAG and agentic search applications.
  • Additional experiments such as the robustness study, effect of different loss function removals, the retained set size (effective k), and score separation provide further insight on how the finetuned reranker behaves.
Critical weaknesses
  • The main concern is about the evaluation experiments:
    • GateReranker is a listwise reranker that looks at 150 documents at once, yet is not compared with any other listwise rerankers. Without this comparison it is not clear how much of the reranking superiority (table 3) comes from this fact rather than the gate-based training. For an apple to apple comparison, I suggest training the same backbone with the FIRST recipe without the gate token, and measure its ndcg based on a properly tuned score threshold (See my comments below about the current tuning limitations).
    • As evident by tables 7 and 8, the optimal K varies significantly across different datasets, yet only one sweep is done per model with the resulting value used for all datasets. In real world different tasks will have separate tunings. I strongly recommend to use dataset specific sweeps for the seen set, and for the unseen average over the optimal values for seen set.
    • The optimization sweep must happen for the same metric that is being measured. The main ranking quality results with the ndcg metric (table 3) are reported on the sweep that is done for F1 score.
    • The evaluation is only done with one benchmark, BEIR, that is also used for training, and with one first stage retriever BM25, I recommend expanding the evaluations to newer more challenging benchmarks like BRIGHT, OBLIQUE, or agentic retrieval benchmarks like BCP, TRQA, as well as dense retrievers as the first stage.
Questions for authors
  • both gate and rank loss have group level and pairwise parts. Did you consider the min score of positive docs - the max of negatives for broadening the separation? This way only one function for ranking and one for loss would be required.
  • I appreciate that the hyper parameters are listed in a table, but I am curios how they are decided? For example the weights of different loss functions?
  • How is 1000 selected for the number queries used for global score threshold sweeping? how sensitive are th
4 comments

Response to Reviewer Gp7x -- Part 1/2

We thank the reviewer for recognizing the motivation of our work and the value of the supplementary analyses. Your comments helped us identify the key questions that require stronger evidence: whether the improvements arise from listwise modeling, unequal training conditions, or the choice of truncation parameters. We have now completed three additional training runs and evaluated all three models on all 13 tasks.

1. Listwise comparisons with matched initialization and training budgets.

We trained three variants:

  • FIRST-style: first-step identifier scoring, with the gate token and gate losses removed.
  • Rank-only: the gate-aware input and gate token are retained, but the gate losses are removed.
  • Gate: the complete joint ranking and gating objective.

All three variants share the initialization checkpoint used in the paper, the same seven-task BEIR training mixture, and the main optimization settings. Each run stops at a predetermined budget of 2,000 optimizer steps. The final test metrics were not used for checkpoint selection in these additional runs. We also verified that the 33,234 evaluation queries, candidate documents, and relevance labels match across all three models.

The standard full-list nDCG@10 results are:

ModelSeenUnseenOverall
FIRST-style0.52730.30490.4247
Rank-only0.57000.33820.4630
Gate0.59470.33810.4763

Gate improves overall nDCG@10 by 0.0516 over FIRST-style and by 0.0133 over Rank-only, while its unseen ranking performance is essentially unchanged relative to Rank-only. These results support the overall value of joint training, without implying a ranking improvement on every task.

Here, FIRST-style denotes a controlled no-gate implementation that uses first-step identifier scoring and shares our ranking objective; it is not a complete reproduction of the original FIRST training recipe. We will make this distinction explicit and document implementation details, including the attention backends used in training. The matched budget refers to training data and optimizer steps, rather than identical GPU-hours.

2. Selection comparisons with more extensively tuned baselines.

To examine whether the results could be explained by insufficient threshold tuning, we exhaustively swept global thresholds over the saved test scores and report the highest task-macro F1 achievable by a single threshold. This is an oracle diagnostic that gives the baselines access to test labels, rather than a deployable validation-tuned setting.

Model and selection ruleOverall F1
FIRST-style + best fixed K0.2772
FIRST-style + oracle global threshold0.2584
Rank-only + best fixed K0.3039
Rank-only + oracle global threshold0.2892
Gate scores + oracle global threshold0.3013
Gate + learned boundary0.3594

Fixed K was swept over {1,2,3,5,10,20,100,150}, and the threshold sweep covered every distinct retained set obtainable from the saved scores using a global threshold. Selection F1 is computed per query and then macro-averaged across tasks; its recall denominator is the number of relevant documents in the BM25 top-150 candidate pool.

Notably, even for the same Gate model and document scores, the test-oracle global threshold achieves an F1 of 0.3013, compared with 0.3594 for the learned boundary. This provides direct evidence for list-conditioned selection without changing the backbone.

These oracle results are specific to the score representation used by the current inference implementation. The saved scores are obtained by applying sigmoid in BF16 before conversion to floating-point values, which can introduce saturation and ties. Accordingly, these results are not upper bounds for arbitrary raw-logit representations or other calibration methods.

Response to Reviewer Gp7x -- Part 2/2

3. Dataset-specific tuning.

We agree that dataset-specific tuning is a stronger reference point. We therefore also computed per-task oracle thresholds. These yield average F1 scores of 0.3438 for FIRST-style and 0.3765 for Rank-only, compared with 0.3594 for Gate without task-specific threshold tuning on test labels.

We therefore do not claim that Gate surpasses an oracle that separately optimizes the threshold using each task's test labels. The more precise conclusion is that the learned gate substantially improves over a single global selection rule and provides competitive cross-task selection without requiring task-specific truncation parameters.

These oracle analyses do not replace independent validation-based tuning, and we will clearly distinguish the two settings.

4. Separating ranking evaluation from F1-based selection.

We accept this concern and have separated ranking and selection evaluation. The nDCG@10 results above use the full candidate ranking and an IDCG computed from the complete qrels; they do not depend on the K or threshold selected for F1.

Our audit also identified that the normalization used for the original post-truncation nDCG was not consistent with standard full-list nDCG. We will correct the corresponding tables and the claim of overall ranking superiority. The revised conclusions will instead be supported by standard ranking metrics, retained-set F1, and the controlled comparisons above.

5. Loss design, hyperparameters, and the 1,000-query calibration set.

The proposed min-positive/max-negative margin is a valuable alternative. It directly constrains the hardest positive--negative pair, whereas our group-level and pairwise terms cover aggregate separation and multiple local relationships, respectively. The design motivation is to avoid concentrating supervision exclusively on extreme examples. Without a direct ablation, however, we do not claim that our formulation is empirically superior to this alternative.

The additional runs inherit a fixed configuration: ranking weights of 0.4 and 0.3, gate weights of 0.2 and 0.1, and a margin of 0.25. We did not perform separate hyperparameter searches for the different variants. These are fixed experimental settings, not theoretically optimal weights; we will distinguish the configuration choices from the evidence provided by the ablations.

For calibration-set size, we conducted a stratified bootstrap analysis using the saved scores for the 1,000 training queries, with 30 repetitions at each sample size. For post-hoc global thresholding of the original paper model, the resulting test F1 scores were:

Calibration sample sizeF1, mean ± standard deviation
5000.3004 ± 0.0096
7500.2989 ± 0.0084
1,0000.3001 ± 0.0090

Within the available query pool, we do not observe substantial sensitivity over the range of 500--1,000 queries. This supports 1,000 queries as a practical calibration-set size. This analysis concerns post-hoc global thresholding, not the learned gate itself.

We also agree that additional benchmarks and dense first-stage retrievers would broaden the evidence. In this revision, we prioritized matched comparisons and evaluation corrections. Our empirical conclusions remain scoped to BEIR with BM25 candidates and do not imply demonstrated performance in agentic retrieval settings.

We hope that these completed experiments and clarifications address the main concerns about the sources of improvement and the reliability of the evaluation, and provide a stronger basis for reconsidering the paper.

thank you for your response. This addresses most of my concerns [with the exception of evaluation on BEIR only with only BM25 as the first retriever]. I will increase the soundness score and the overall score.

Thank you for your thoughtful follow-up and for reconsidering your assessment. We are glad that our response addressed most of your concerns. We acknowledge your remaining concern regarding the BEIR evaluation, where BM25 was used as the only first-stage retriever. We appreciate you highlighting this limitation and will clarify the scope of this evaluation in the paper. Thank you again for your constructive feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.