acceptodds
Under review as a conference paper at ICLR 2027

Rethinking RAG in Long Videos: What to Retrieve and How to Use It?

Abstract

Retrieval-augmented generation is extending beyond text to long videos, where query-relevant chunks can be represented across multiple modalities and temporal granularities. Progress in this setting, VideoRAG, is limited by two gaps: existing benchmarks allow queries to be answered without the video, obscuring retrieval errors, and prior methods apply a single modality-granularity configuration per query, ignoring chunk-level variability. We address both by introducing V-RAGBench, a benchmark of ⟨query, evidence chunk, answer⟩ triplets over hour-scale videos, in which each answer depends on a unique evidence chunk, enabling decoupled evaluation of retrieval and generation, and CARVE, a training-free method that runs parallel retrievers across configurations and uses chunk-adaptive reranking to select a winning configuration for each chunk, which is then carried into generation. On V-RAGBench, CARVE outperforms eight recent VideoRAG baselines on both stages, with the evidence it supplies to the generator interleaving multiple configurations rather than sharing one. Its gains hold across egocentric and third-person long videos and extend to an expanded configuration space.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Retrieval, Long Video Understanding, Retrieval-Augmented Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_khq8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies VideoRAG for long videos and focuses on how video chunks should be represented for retrieval and generation. It introduces V-RAGBench, a benchmark of 2,100 <query, evidence chunk, answer> triplets over 216 hour-scale videos. It also proposes CARVE, a training-free method that retrieves candidates from multiple modality–granularity configurations, reranks them at the chunk level, and passes each retrieved chunk to the generator using its selected representation.

Critical strengths
  1. V-RAGBench is proposed so that retrieval and generation can be evaluated separately. The benchmark uses multiple filtering stages and human verification to reduce shortcut answerability and spurious evidence.
  2. CARVE allows different retrieved chunks to use different modality–granularity configurations. The method is training-free, easy to integrate into existing retrieval pipelines, and shows consistent gains across multiple baselines and generator backbones.
Critical weaknesses
  1. Although V-RAGBench assumes that each question can only be answered using its annotated evidence chunk, only 97.1% of non-evidence windows were judged unanswerable. More importantly, the paper does not clarify whether the problematic samples identified during human verification were removed or corrected. These issues should be addressed, particularly for the test set, to ensure reliable evaluation.

  2. The empirical answerability filter removes questions that Gemini-2.5-Flash fails to answer even when given the ground-truth evidence. While this helps ensure answerability, it may also exclude difficult but valid questions. As a result, the benchmark's question distribution may be influenced by the capabilities of the filtering model.

Questions for authors
  1. Datasets such as NExT-GQA already provide questions, answers, and temporal evidence annotations. Could the authors clarify what V-RAGBench adds beyond the unique-evidence constraints and longer videos?
  2. CARVE uses the configuration selected during reranking directly for generation. However, a representation that is effective for retrieving a chunk may not be the best one for answering the question. Have the authors considered selecting the generation representation separately?

My rating may change depending on how the authors address my concerns in their rebuttal.

5 comments

We sincerely thank the reviewer for the positive evaluation of our work and for providing important feedback on our benchmark and method design. Your comments motivated us to run additional experiments for W1 and Q2 to deepen our analysis, and we have done our best to address each of your points as thoroughly as possible below.

W1.

W1. Although V-RAGBench assumes that each question can only be answered using its annotated evidence chunk, only 97.1% of non-evidence windows were judged unanswerable. More importantly, the paper does not clarify whether the problematic samples identified during human verification were removed or corrected. These issues should be addressed, particularly for the test set, to ensure reliable evaluation.

We thank the reviewer for highlighting this point. We fully share your view that maintaining strict evidence uniqueness is critical for reliable evaluation. While there is a remaining gap to 100%, we respectfully note that the 97.1% figure was measured exhaustively across all 10,206 non-evidence window–query pairs in the test videos, which we believe provides a solid empirical basis for evidence localization in hour-scale recordings.

Having said that, to address your concern directly and ensure complete test-set reliability, we re-evaluated all 10 methods after strictly excluding the test queries associated with the flagged windows from the human audit.

Additional results after excluding the flagged test queries

MethodRecall@5nDCG@5Qwen3-VL-8BQwen3-VL-32BGemma-4-26B
VideoRAG-A0.5070.3280.2390.3140.306
GQR0.5000.3360.1710.1130.154
FreeRet0.2600.1840.2330.1990.161
GME0.4070.2460.2570.2810.213
RRF0.4590.2940.1750.1300.188
DAT0.4560.3100.2190.1470.205
VideoRAG-B0.4830.3230.3110.3010.228
UniversalRAG0.4450.2960.2590.2590.227
UniversalRAG-LoRA0.4660.3060.2360.2740.219
CARVE0.6030.4310.3520.3700.323

All metrics change by at most ±0.008\pm 0.008 compared to Table 2, and all method rankings remain completely unchanged. We are deeply grateful to you for highlighting this point, as it truly helped us address a critical detail and make our evaluation much more rigorous. We will include this sensitivity analysis and provide the cleaned test split in the revision.

W2.

W2. The empirical answerability filter removes questions that Gemini-2.5-Flash fails to answer even when given the ground-truth evidence. While this helps ensure answerability, it may also exclude difficult but valid questions. As a result, the benchmark's question distribution may be influenced by the capabilities of the filtering model.

We appreciate the reviewer’s important observation. We agree that filtering with a specific model could exclude difficult yet valid questions. However, we would like to respecfully emphasize two key aspects of our setup that prevent this filter from making the benchmark overly easy or model-dependent.

Firstly, in the Empirical Answerability filter, Gemini-2.5-flash inspects only the isolated, short ground-truth event clip (median 8.6s). Because localized verification on a single short clip is far simpler than long-video RAG, the filter primarily screens out visually ambiguous or ill-posed candidates rather than reasoning-heavy questions.

Secondly, there is still a high difficulty ceiling under full RAG evaluation. Passing this localized check leaves substantial challenge at evaluation time, where models must retrieve evidence from 1–9 hour videos and reason amidst top-kk distractors. In Table 2, end-to-end pass rates remain at 0.318–0.368 even for CARVE. Moreover, in our new GT-insertion experiment (see our response to Reviewer M7Wp, W1), even when the GT chunk is guaranteed in the top-5 context, Qwen3-VL-8B achieves 0.322 and Gemini-3.8-Flash reaches 0.630, demonstrating a large gap yet to be closed.

We will revise the manuscript to clarify this distinction between localized verification and full RAG evaluation, and note how future extensions can audit rejected candidates to incorporate harder perceptual questions.

Q1.

Q1. Datasets such as NExT-GQA already provide questions, answers, and temporal evidence annotations. Could the authors clarify what V-RAGBench adds beyond the unique-evidence constraints and longer videos?

We thank the reviewer for the opportunity to clarify how V-RAGBench complements grounded video QA benchmarks such as NExT-GQA. While scaling to hour-long videos and enforcing unique evidence are foundational for moving beyond single-pass in-context processing, V-RAGBench introduces two additional benchmark design properties specifically tailored for VideoRAG evaluation.

First, whereas NExT-GQA is formulated for in-context temporal grounding where a short clip is fed entirely into a model to predict a continuous time interval [ts,te][t_s, t_e], V-RAGBench is structured around a discrete retrieve-then-generate pipeline. This allows researchers to separately measure first-stage candidate retrieval and reranking (via Recall@kk and nDCG@kk) and downstream generation under retrieved top-kk distractors.

Second, V-RAGBench is constructed to prevent retrieval and generation shortcuts. During query generation, each question is required to embed contextual situation anchors rather than explicit temporal cues (Section 4), forcing retrievers to match visual and semantic content. Moreover, unlike multiple-choice setups, V-RAGBench pairs open-ended reference answers with blind shortcut filtering and empirical answerability checks so that downstream accuracy reflects grounded generation rather than option guessing or language priors.

We will cite NExT-GQA and clearly articulate these complementary evaluation goals in Section 2 (Related Work).

Reviewer_M7WpOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies Retrieval-Augmented Generation for long-video question answering (VideoRAG) and identifying two key challenges: existing benchmarks obscure retrieval failures due to shortcut-based answering, and existing methods overlook chunk-level differences in optimal video representations. To address these challenges, it introduces V-RAGBench, a benchmark with uniquely grounded evidence that enables independent evaluation of retrieval and generation. It further proposes CARVE, a training-free, chunk-adaptive reranking method that dynamically selects the most suitable modality and temporal granularity for each video chunk, improving retrieval and generation performance on V-RAGBench, LongVideoBench, and Video-MMMU.

Critical strengths
  1. The paper identifies two important limitations in current VideoRAG research: the difficulty of independently evaluating retrieval and generation, and the lack of chunk-level adaptation across modalities and temporal granularities. I think both problems are reasonable, and highly relevant to the community. Further, the V-RAGBench provides a valuable step toward enabling more systematic research on these challenges.

  2. CARVE successfully improve the retrieval performance from V-RAGBench through chunk-adaptive representation selection in VideoRAG. The methodology is reasonable and the experiment results support that different video chunks may require different modalities and temporal granularities to effectively capture relevant evidence.

Critical weaknesses
  1. The gap between retrieval improvement and downstream QA performance remains insufficiently addressed. The paper correctly identifies the importance of evaluating retrieval and generation separately. However, although CARVE improves retrieval performance, these improvements do not consistently translate into successful QA. For example, in Table 3, the text-only frame-and-clip configuration achieves higher Recall@5 than the visual-only counterpart (0.567 vs. 0.497), but substantially lower QA performance (0.217 vs. 0.327).

  2. There is limited coverage of diverse retrieval scenarios in V-RAGBench. In real-world long-video understanding, many questions require retrieving and integrating multiple evidence chunks, potentially across different modalities and temporal granularities. Some questions even require causal reasoning capabilities. Therefore, single-evidence retrieval already seems like an easy setting for the long-video understanding field.

Questions for authors
  1. Could you explain why substantial improvements in retrieval performance not consistently translate into comparable improvements in downstream QA performance, as illustrated in Table 3?

  2. Given that many QA failures occur despite successful retrieval, what are the practical benefits and limitations of further optimizing retrieval performance?

  3. Why does V-RAGBench focus exclusively on single-evidence retrieval, given that real-world long-video understanding often requires integrating evidence from multiple chunks?

3 comments

We sincerely thank the reviewer for the positive evaluation of our paper and for raising thoughtful questions about the relationship between retrieval and downstream QA. Your feedback motivated us to conduct new controlled experiments to deepen our analysis. Below, we address your comments in detail, grouping related Weaknesses and Questions together for clarity.

W1, Q1, & Q2.

W1. The gap between retrieval improvement and downstream QA performance remains insufficiently addressed. The paper correctly identifies the importance of evaluating retrieval and generation separately. However, although CARVE improves retrieval performance, these improvements do not consistently translate into successful QA. For example, in Table 3, the text-only frame-and-clip configuration achieves higher Recall@5 than the visual-only counterpart (0.567 vs. 0.497), but substantially lower QA performance (0.217 vs. 0.327).

Q1. Could you explain why substantial improvements in retrieval performance not consistently translate into comparable improvements in downstream QA performance, as illustrated in Table 3?

Q2. Given that many QA failures occur despite successful retrieval, what are the practical benefits and limitations of further optimizing retrieval performance?

We thank the reviewer for encouraging a deeper examination of how retrieved evidence is used during generation. Exposing and quantifying this stage-wise discrepancy is one of the core contributions of V-RAGBench. Two factors explain this gap:

Firstly, there is a generator bottleneck under distracting context (new controlled experiment below). Prompted by your question, we conducted a new experiment where the ground-truth evidence chunk was inserted at a random position among the retrieved chunks for k∈{3,5,10}k \in \{3, 5, 10\}. Every setting thus has 100% retrieval recall, isolating how well each generator utilizes the true evidence alongside distractors.

Additional generation results with ground-truth evidence inserted at a random position (pass rate):

Generatork=3k = 3k=5k = 5k=10k = 10
Qwen3-VL-4B0.34670.28890.2600
Qwen3-VL-8B0.36850.32220.2778
Gemini-3.8-Flash0.64360.63000.6000

At k=5k = 5, Gemini-3.8-Flash answers 63.0% of questions correctly compared with 32.2% for Qwen3-VL-8B, even though both receive the exact same annotated evidence. Moreover, accuracy decreases monotonically as kk increases from 3 to 10 across all models due to distractor interference. This aligns with our error analysis in Appendix D.4 (Tables 19–20), where 105 (Qwen3-VL-8B) and 104 (Qwen3-VL-32B) failed QA cases occurred despite successful retrieval (e.g., failing to preserve temporal order in the tortilla example).

Secondly, there is a representation mismatch between locating vs. answering (See Table 3 in the manuscript). Textual summaries (m={text}) match query keywords well for retrieval (0.567 Recall@5) but compress away fine-grained visual details needed for answering (0.217 pass rate), whereas visual representations preserve those details (0.327). This is precisely why CARVE selects representations per chunk and passes the winning modality-interleaved forms to generation (reaching 0.603 Recall@5 and 0.352 pass rate).

Regarding Q2 specifically, which is about practical benefits and limitations of optimizing retrieval, we appreciate you raising this practical question. Optimizing retrieval provides three concrete benefits.

  1. Mitigating distractor noise via higher precision at small kk: As our k=3,5,10k=3, 5, 10 experiment shows, generators degrade when fed more distractor chunks. Improving retrieval ranking (higher nDCG@kk) allows systems to operate at a smaller retrieval depth kk, directly improving generator utilization.
  2. Consistent downstream gains across backbones: CARVE’s retrieval gains consistently translate into the best QA pass rate across all three backbones (0.352, 0.368, and 0.318), outperforming all baselines with statistical significance (see Table 2 in the manuscript.)
  3. Scaling with stronger generators: As shown by Gemini-3.8-Flash reaching 63.0% in our new experiment, stronger generators convert a much higher fraction of retrieved evidence into correct answers. We will incorporate this new experiment and discussion into the revision.

W2 & Q3.

W2. There is limited coverage of diverse retrieval scenarios in V-RAGBench. In real-world long-video understanding, many questions require retrieving and integrating multiple evidence chunks, potentially across different modalities and temporal granularities. Some questions even require causal reasoning capabilities. Therefore, single-evidence retrieval already seems like an easy setting for the long-video understanding field.

Q3. Why does V-RAGBench focus exclusively on single-evidence retrieval, given that real-world long-video understanding often requires integrating evidence from multiple chunks?

We thank the reviewer for highlighting multi-evidence reasoning. Focusing V-RAGBench on uniquely sufficient single-chunk evidence was a deliberate diagnostic design to establish clean stage-wise evaluation. The benefits of this design can be summarized into the following three points.

The first is diagnostic identifiability. When a query requires multiple chunks, an incorrect answer (or partial retrieval) makes it impossible to disentangle whether the failure came from missing one required event or from the generator failing to synthesize multiple events across time. A uniquely sufficient chunk makes the retrieval target unambiguous and allows Recall@kk and nDCG@kk to be interpreted independently of multi-hop reasoning.

Second, this design enables controlled VideoRAG analyses. It directly allowed us to conduct our controlled experiments on representation switching (Table 3), distractor/position scaling (the k=3,5,10k=3, 5, 10 experiment above), and stage-wise failure attribution (Table 19).

Third, locating a single event across 1-to-9-hour egocentric videos filled with recurring daily routines remains highly challenging. Existing baselines achieve only 0.263–0.510 Recall@5 (Table 2), and CARVE reaches 0.603. This shows that while integrating multiple evidence chunks is an important goal, our current benchmark already poses sufficient and targeted challenges to current long-video QA systems.

Moving forward, we completely agree that extending VideoRAG evaluation to the multi-evidence and causal reasoning scenarios you suggested is an essential next step for the field. Building on our current pipeline, this can be achieved by pairing independently verified unique events (e.g., an early action and a later outcome) and verifying that neither chunk suffices alone. We will clarify the rationale for our current diagnostic setting and explicitly highlight this multi-evidence extension as a key future direction in the revised manuscript.

Thank you once again for your time and constructive comments, which have greatly improved our paper. We hope our responses and new analyses address your questions well, and we would be glad to discuss any further points.

Reviewer_M7WpOpenReview ↗

Reply

Thanks for addressing my questions, and the clarifications did help. I would retain my positive rating.

Reviewer_ejX3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies RAG over hour-long videos. It argues that the representation of each chunk, in both modality and temporal granularity, should be chosen per chunk rather than per query. The authors introduce V-RAGBench, a benchmark built from long egocentric videos where each answer depends on a single unique evidence chunk, so retrieval and generation can be evaluated separately. They also propose CARVE, a training-free method. It retrieves with several representations in parallel, reranks each candidate under the representation that retrieved it, and passes each chunk to the generator in its winning representation. CARVE outperforms prior VideoRAG methods on both retrieval and generation.

Critical strengths

S1. The benchmark is the main contribution. In existing long-video QA data, models can often answer from priors, so final accuracy says little about retrieval. V-RAGBench addresses this with careful filtering and an exhaustive human audit, which makes stage-wise evaluation of VideoRAG possible. I expect it to be useful regardless of the proposed method.

S2. The generation result is well controlled. The trained query-level routers receive the same retrieved chunks as CARVE, so the gain can only come from how each chunk is represented. This directly supports the main claim that the best representation varies across chunks within a query.

Critical weaknesses

W1. The retrieval comparison does not separate the effect of the larger candidate pool from the effect of chunk-level selection. CARVE reranks candidates from multiple retrievers, while the reranked baselines seem to reorder only their own top results. The paper’s own ablation suggests that part of the gain comes from pooling alone. This does not invalidate the claim, since tag-aware reranking still beats fixed-configuration reranking on the same pool. But the contribution of chunk-level selection is likely overstated.

W2. The test set is small, and significance is measured only across generation runs, not across queries. Some margins over the strongest baselines are narrow, so their robustness is unclear.

Questions for authors

Q1. In the reranked-baseline experiment, was the reranker applied only to each baseline's own top-k? If so, please report baselines that rerank a pool of the same size as CARVE's, to separate the effect of pool size from that of chunk-level selection.

3 comments

We sincerely thank the reviewer for the positive assessment of our work and for the time and effort invested in reviewing our paper. Below, we have carefully addressed each of your points, grouping W1 and Q1 together for clarity.

W1 & Q1.

W1. The retrieval comparison does not separate the effect of the larger candidate pool from the effect of chunk-level selection. CARVE reranks candidates from multiple retrievers, while the reranked baselines seem to reorder only their own top results. The paper’s own ablation suggests that part of the gain comes from pooling alone. This does not invalidate the claim, since tag-aware reranking still beats fixed-configuration reranking on the same pool. But the contribution of chunk-level selection is likely overstated.

Q1. In the reranked-baseline experiment, was the reranker applied only to each baseline's own top-k? If so, please report baselines that rerank a pool of the same size as CARVE's, to separate the effect of pool size from that of chunk-level selection.

We appreciate the reviewer raising this important point. To directly separate the effect of candidate pool size from that of chunk-level selection under a strictly matched budget, we provide two complementary controlled comparisons.

1. New experiment: Expanding baseline retrieval budgets to match CARVE (N=20N=20):

Following your suggestion in Q1, we evaluated all baselines under an expanded retrieval budget of 20 candidates prior to reranking, matching the nominal maximum pool size of CARVE's four parallel top-5 retrievals.

MethodRecall@5nDCG@5
VideoRAG-A0.4870.341
GQR0.4900.343
FreeRet0.4370.316
GME0.5130.374
RRF0.4900.344
DAT0.4930.341
VideoRAG-B0.4730.336
UniversalRAG0.5230.359
UniversalRAG-LoRA0.5670.387
CARVE0.6030.433

Under the exact same 20-candidate budget, CARVE consistently outperforms the strongest reranked baseline (UniversalRAG-LoRA) by +0.036 Recall@5 and +0.046 nDCG@5.

2. Controlled comparison on the exact same candidate pool: This result complements our controlled ablation in Section 6.2 (Table 5), where all strategies rerank the exact same candidate pool (P(q)\mathcal{P}(q)) assembled by CARVE:

Reranking strategyRecall@5nDCG@5
Text-frame / Text-clip0.497 / 0.4900.343 / 0.339
Visual-frame / Visual-clip0.547 / 0.5600.391 / 0.396
Random / Concatenation0.513 / 0.5130.312 / 0.339
CARVE (Chunk-adaptive)0.6030.433

Together, these two comparisons isolate both components. We are grateful for this suggestion and will revise the manuscript to include the matched-budget table, explicitly attributing broader initial evidence coverage to parallel pooling and the further ranking improvement (+4.3%p Recall@5 on the shared pool) to chunk-adaptive selection.

W2.

W2. The test set is small, and significance is measured only across generation runs, not across queries. Some margins over the strongest baselines are narrow, so their robustness is unclear.

We thank the reviewer for this valuable feedback. To address this concern, we provide two complementary analyses: (1) whether the results hold on a larger evaluation set, and (2) whether the margins are significant across queries rather than only across generation runs.

(1) Results on a larger evaluation set. The official test split is fixed to preserve its balance across the three query categories and the held-out Outdoor domain (Appendix A.2). We therefore extended it with 200 randomly drawn queries that were not used to train or test any method, and evaluated all methods on the resulting 500 queries. CARVE retains first place even on this expanded set. The table below reports the scores on the 500-query set for CARVE and the five strongest baselines (Qwen3-VL-8B).

MethodRecall@5nDCG@5Pass rate
VideoRAG-A0.4800.3100.266
GQR0.4640.3110.140
RRF0.4220.2700.188
VideoRAG-B0.4580.3050.330
UniversalRAG-LoRA0.4360.2880.240
CARVE0.5560.4060.422

CARVE outperforms every baseline on all three metrics. This shows that the improvements are preserved when the test set is extended.

(2) Query-level significance. We additionally ran a query-level paired bootstrap on the 500-query set (10,000 resamples), computing difference with CARVE and baseline on the same resampled queries. Positive differences indicate improvements achieved by CARVE over the corresponding baseline.

BaselineΔ Pass [95% CI]pΔ R@5 [95% CI]pΔ nDCG@5 [95% CI]p
VideoRAG-A+0.156 [+0.116, +0.198]<0.001+0.076 [+0.042, +0.112]<0.001+0.096 [+0.069, +0.124]<0.001
GQR+0.282 [+0.238, +0.328]<0.001+0.092 [+0.058, +0.128]<0.001+0.094 [+0.066, +0.123]<0.001
RRF+0.234 [+0.190, +0.276]<0.001+0.134 [+0.096, +0.172]<0.001+0.136 [+0.108, +0.165]<0.001
VideoRAG-B+0.092 [+0.052, +0.132]<0.001+0.098 [+0.064, +0.134]<0.001+0.101 [+0.075, +0.128]<0.001
UniversalRAG-LoRA+0.182 [+0.140, +0.224]<0.001+0.120 [+0.084, +0.156]<0.001+0.118 [+0.091, +0.146]<0.001

All comparisons are significant at p < 0.001, and none of the confidence intervals include zero. The gains are therefore consistent at the query level. We will add both analyses to the revised manuscript.

We sincerely thank you again for your time and thoughtful feedback. We hope our additional experiments and responses have fully addressed your concerns, and we remain happy to provide any further clarification during the discussion period.

I thank the authors for their detailed response and additional experiments. I have carefully read the rebuttal, and it addresses most of my concerns. I continue to view V-RAGBench and CARVE as valuable contributions and remain positive about the paper overall.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of responses

We thank all three reviewers for time and effort in evaluating our work. We feel very fortunate to have received such thoughtful and constructive feedback, and we are glad that all three reviewers maintained their initial positive ratings of 3. Below is a brief summary of the comment and discussion.

Reviewer ejX3: The reviewer highlighted V-RAGBench's stage-wise evaluation and our well-controlled generation comparison. During the discussion, we added a matched-budget reranking comparison to separate candidate pool size from chunk-level selection, and verified robustness on an expanded query set with query-level significance testing.

Reviewer M7Wp: The reviewer recognized the high relevance of the two VideoRAG limitations we identify, V-RAGBench as a valuable step for systematic research, and CARVE's retrieval improvements through chunk-adaptive representation selection. During the discussion, we conducted a controlled ground-truth insertion experiment to demonstrate generator bottlenecks under distractors and explained the representation mismatch between retrieval and answering. We also clarifed the practical benefits of retrieval optimization and our single-evidence design.

Reviewer khq8: The reviewer acknowledged V-RAGBench's design for separate and reliable evaluation, as well as CARVE's practical, training-free design with consistent gains across baselines and backbones. During the discussion, we conducted a re-evaluation on a cleaned test split, clarified the localized scope of our answerability filter and distinctions from NExT-GQA, and added comparisons against alternative generation-stage routers.

All new experiments and clarifications will be incorporated into the revised manuscript. Thank you again for your time and careful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.