acceptodds
Under review as a conference paper at ICLR 2027

Query-Conditioned Residual Retrieval for Multimodal RAG

Abstract

Multimodal retrieval-augmented generation (RAG) extends external knowledge beyond text to heterogeneous sources such as images, documents, and videos. A practical strategy represents such content using query-independent textual descriptions, whereas recent multimodal retrievers directly index the original multimodal content. These alternatives expose a basic representation problem: a textual description can be semantically correct yet omit a visual or temporal detail required by a particular query. We study this problem through query-dependent description sufficiency and introduce Query-Conditioned Residual Retrieval (QCR), a training-free retrieval approach. For each multimodal item, QCR uses its textual descriptions to construct an item-specific description subspace in a shared multimodal representation space and retains the orthogonal component as a description-conditioned residual. At inference time, QCRindependently retrieves from the textual description and residual representations. The first route captures relevance directly expressed by the descriptions, while the second exposes representation directions not contained in their span. Candidates from both routes are jointly ranked, and the selected original items are provided to the answer model. Across heterogeneous text, table, image, and video knowledge sources, QCR consistently improves over description-only, direct multimodal, and description–multimodal fusion retrieval. Extensive analysis shows that QCR advantage grows as textual descriptions become less sufficient for the downstream query, together with increased recovery of supporting evidence missed by description retrieval.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal, 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 4 reviews, above 85% of the others.
Reviewer_NALsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a method called Query-Conditioned Residual Retrieval (QCR2) to address a limitation of existing multimodal retrieval methods, where converting multimodal content into textual descriptions can omit visual or temporal details relevant to specific queries. The method applies orthogonal projection to decompose each multimodal embedding into a component aligned with its textual descriptions and a residual component orthogonal to the description subspace. It then introduces a separate retrieval path based on these residual representations to retrieve potentially relevant information that may not be adequately captured by the textual descriptions.

Critical strengths

1- The proposed method is training-free and clearly defined, which makes it appealing because it can be easily integrated into existing methods.

2- The paper provides an in-depth analysis, a theoretical overview, and comprehensive evaluations that help explain the proposed method and its effectiveness.

Critical weaknesses

1- My main concern is the novelty of the proposed method. QCR2 decomposes multimodal embeddings through orthogonal projection and uses the residual as an additional retrieval channel. However, the conceptual contribution beyond existing projection-based representations and multi-channel retrieval methods remains unclear.

2- QCR2 achieves a relatively small improvement over UniversalRAG. Moreover, Orthogonality to the textual-description subspace does not necessarily imply semantic complementarity or query relevance; residuals may retain irrelevant visual variations rather than useful information omitted by descriptions.

Questions for authors

I have no questions.

5 comments
AuthorsOpenReview ↗

Official response to Reviewer NALs

We sincerely thank the reviewer for the positive evaluation and thoughtful feedback. We are encouraged that you found QCR$^2$ training-free, clearly defined, easy to integrate, and supported by comprehensive theoretical and empirical analyses. Although no explicit questions were listed, we address both of your critical weaknesses directly below and in the revised manuscript (summarized in "Revision summary").

[W1] Conceptual contribution beyond existing projection and multi-channel retrieval methods

While QCR$^2$ intentionally uses lightweight linear projection and dual-index search to remain training-free and indexable at 30.0 ms latency (Table D12), our controlled ablations in Table 2 and Table 3 show that what subspace is projected out, and at which side of retrieval, is critical, rather than projection or multi-channel retrieval alone:

Retrieval / Decomposition Paradigm (Controlled Table 2 Setting)Recall@5Avg.
Standard multi-channel retrieval (Description + multimodal, learned fusion)44.040.8
Query-adaptive multi-channel selection (QA-Fusion MLP gate)44.941.3
Online query-side multi-hop subtraction (GRAIL, O & Lee, 2026 adapted)44.641.2
Global modality-level centering (Modality-wise mean-centering)45.141.3
Global corpus-level projection (Global description subspace, rank-4 SVD)45.341.4
QCR$^2$ (Offline item-specific description-subspace decomposition)48.342.9

Existing projection-based representations either separate global modality-shared vs. modality-private subspaces across a corpus (41.3 to 41.4) or subtract already-retrieved context from the query across iterative hops (O & Lee, 2026, 41.2). Neither isolates what each individual item's concrete indexing captions omitted. Conditioning the orthogonal subspace on each item's own M=4M = 4 descriptions removes 45.1% of redundant caption-explained norm energy per item and accounts for over 70% of the total gain (+1.5 of +2.1 over learned fusion). We have sharpened this conceptual positioning in Sections 1 and 2.

[W2] Improvement over UniversalRAG and geometric orthogonality vs. query-irrelevant visual variations

(1) Improvement over UniversalRAG. In Table 1, UniversalRAG's macro-averaging weights Text, Table, Image, and Video equally, even though QCR$^2$ applies no residual branch to Text or Table items (1[xi∈Dv]=0\mathbb{1}[x_i \in \mathcal D_v] = 0). Across the six image and video benchmarks where residuals operate, QCR$^2$ outperforms UniversalRAG by +1.08 points (+1.15 on Image, +1.01 on Video; 95% paired bootstrap CI [+0.68, +1.49]), e.g., +1.55 on MRAG-Bench, +1.28 on LVBench, and +1.60 R-L on VideoRAG-Synth, surpassing both UniversalRAG (42.40) and its ground-truth corpus Oracle (42.45) without needing a trained modality-granularity router. Under the strictly matched Table 2 backbone (Qwen3-VL-Embedding-2B), QCR$^2$ also outperforms UniversalRAG (42.2) by +2.7 Recall@5 and +0.70 Avg. ([+0.42, +0.98]).

(2) Why geometric orthogonality helps despite query-irrelevant visual variations. We fully agree that vi⊥v_i^\perp is a geometric residual and can retain query-irrelevant visual variations such as background or viewpoint (as explicitly analyzed in Appendix F and illustrated by our failure case in Figure E4a). QCR$^2$ prevents such irrelevant variations from degrading retrieval through three design mechanisms:

  1. Directional query alignment (zq⊤riz_q^\top r_i) over unit-normalized residuals: In a high-dimensional embedding space, benign background/viewpoint variations that are unrelated to the textual query qq have near-zero inner product with zqz_q, whereas normalizing rir_i prevents items with large uninformative residual norms ∣∣vi⊥∣∣2||v_i^\perp|| _ {2} from dominating rankings (+1.1 Recall@5 and +0.5 Avg. in Appendix C.5, Table C8).
  2. Zero-mean candidate-union standardization (Equation (10)): Standardizing s~res(q,i)\tilde s_{\text{res}}(q,i) over Kq∩Dv\mathcal K_q \cap \mathcal D_v ensures that multimodal items whose residuals merely reflect random visual noise receive negative or near-zero standardized scores and are filtered out by s~desc(q,i)\tilde s_{\text{desc}}(q,i) during joint ranking.
  3. Empirical verification of low noise risk (Appendix B.4, Table B3): Comparing QCR$^2$ directly against Description only on the exact same queries (Table B3) shows that when the residual route does not recover new gold evidence, answer degradation remains essentially identical (5.8%), and when it does recover missed evidence, answer improvements jump more than threefold (9.4% to 30.8%, while hurt rate stays flat at 6.2%). Thus, query-irrelevant residual variations rarely displace valid textual evidence, whereas query-aligned residuals provide substantial complementary recall.
AuthorsOpenReview ↗

Any follow-up question?

Dear Reviewer NALs,

Thank you again for your positive evaluation and thoughtful feedback on our work. As the author-reviewer discussion period closes shortly (about 22 hours), we wanted to briefly summarize how our response and magenta-highlighted manuscript revisions address both of your points:

  • Conceptual contribution over projection & fusion baselines (W1): We have sharpened our conceptual positioning and contrasted offline item-specific description decomposition against multi-channel fusion, query-adaptive selection, adapted GRAIL, modality-wise centering, and global SVD projection (Official response; Sec. 1, Sec. 2, Sec. 3.3, Sec. 4.2, Sec. 4.4, Table 2, Table 3).
  • UniversalRAG comparison & irrelevant visual variation (W2): We have added paired bootstrap confidence intervals and controlled UniversalRAG comparisons, and clarified how residual normalization and zero-mean candidate standardization suppress query-irrelevant visual noise (Official response; Sec. 3.4, Sec. 4.2, Table 2, App. B.4, Table B3, App. C.5, Table C8).

Could you let us know whether these new experiments and clarifications resolve your main concerns, or if there are any remaining questions we can address before the discussion period ends?

AuthorsOpenReview ↗

Gentle reminder before discussion closes

Dear Reviewer NALs,

This is a gentle reminder that about 6 hours remain in the author-reviewer discussion period. Even a brief note on whether our rebuttal and revised manuscript have addressed your feedback would be greatly appreciated. If your concerns have been resolved, we also kindly ask whether you would consider updating your score accordingly. Thank you again for your time and thoughtful review!

Reviewer_roZPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QCR², a training-free method for multimodal retrieval-augmented generation. It projects each multimodal item embedding onto the subspace spanned by its textual descriptions and uses the orthogonal residual as a complementary retrieval representation. Textual and residual retrieval independently generate candidates, which are jointly ranked before answer generation. Experiments on heterogeneous multimodal benchmarks show improvements over description-only retrieval and description–multimodal fusion, with analyses suggesting larger benefits when descriptions omit query-relevant information.

Critical strengths

The method is simple, clearly presented, and requires no additional retriever training. The controlled comparisons provide evidence of its practical value: in Table 2, QCR² improves Recall@5 from 44.0 to 48.3 and the average downstream score from 40.8 to 42.9 over learned fusion. The ablations and description-sufficiency analysis also help explain when the residual route is useful, showing that it can recover supporting evidence missed by textual retrieval.

Critical weaknesses
  1. My main concern is limited methodological novelty. The approach combines standard orthogonal projection, residual normalization, dual-route retrieval, and score fusion. Although this combination is useful, the paper does not sufficiently establish a conceptual advance 3. beyond applying these operations to multimodal retrieval. The residual is also query-independent; query conditioning enters through scoring and ranking.

  2. The experiments demonstrate gains over the evaluated fusion baselines, but leave the broader significance less clear. Comparisons with stronger query-adaptive evidence-selection methods would help establish the distinctive value of residualization. Furthermore, DSG depends on the answer model, so the analysis does not fully establish that the observed trend reflects general description insufficiency rather than model-specific behavior.

Questions for authors
  1. What is the operational role of query-dependent description sufficiency in the method, beyond motivating and analyzing retrieval over a query-independent residual?

  2. Can the authors compare against a stronger query-adaptive evidence-selection baseline under matched encoders, candidate budgets, and final evidence budgets?

  3. Do the trends in Figure 3 remain consistent when DSG is computed using an answer model different from the one used for downstream evaluation?

6 comments
AuthorsOpenReview ↗

Official response to Reviewer roZP [1/2]

We thank the reviewer for the insightful assessment and constructive questions. We are glad that you found the method clean, training-free, and well-supported by the controlled Table 2 gains (44.0 to 48.3 Recall@5 and 40.8 to 42.9 Avg.) and recovery analyses. Below, we address your questions on conceptual framing, query-adaptive baselines, and cross-model DSG consistency; an overview of revisions appears in the Revision summary comment.

[W1 & Q1] Conceptual advance and operational role of query-dependent description sufficiency

We appreciate this sharp conceptual question. First, keeping the indexed residual rir_i query-independent is a deliberate systems choice: if the index representation depended on qq, decomposing embeddings over millions of items at query time would prevent offline HNSW indexing. Instead, our conceptual contribution is showing that query-dependent sufficiency can be resolved by pairing an offline item-specific residual subspace with online query-conditioned directional scoring, without retraining encoders or adding inference overhead (30.0 ms vs. 29.9 ms for standard fusion; Table D12).

Operationally, query conditioning enters the retrieval rule (beyond motivation and post-hoc analysis) in three concrete ways:

  1. Decoupling residual existence from query-conditioned directional relevance (Appendix C.5, Table C8): Because an item's residual norm ∣∣vI⊥∣∣2||v _ {I}^\perp|| _ {2} is query-independent, unnormalized scoring zq⊤vi⊥=∣∣vi⊥∣∣2(zq⊤ri)z _ {q}^\top v_i^\perp = ||v_i^\perp|| _ {2}(z _ {q}^\top r_i) conflates how much uncaptioned variation an item has with whether that variation matters to qq. Normalizing rir_i forces sres(q,i)=zq⊤ris_{\text{res}}(q,i) = z_ q ^\top r_i to activate only when the query vector zqz_q aligns with directions outside span(Ti)\text{span}(T_i), improving Recall@5 from 47.2 to 48.3 and Avg. from 42.4 to 42.9.
  2. Query-conditioned candidate-pool calibration (Equation (10)): Both (μdesc(q),σdesc(q))(\mu_{\text{desc}}(q), \sigma_{\text{desc}}(q)) and (μres(q),σres(q))(\mu_{\text{res}}(q), \sigma_{\text{res}}(q)) are conditioned on each query's retrieved candidate union Kq\mathcal K_{q}. When a query is well-served by captions (sdescs_{\text{desc}} exhibits sharp separation while sress_{\text{res}} is flat/noisy) or vice versa, query-level z-standardization dynamically shifts relative ranking influence across candidates without per-query LLM calls.
  3. Direct boundary validation and optional query-adaptive gating λ(q)\lambda(q): Appendix C.4 (Table C7) confirms the sufficiency mechanism at the representation boundary: when descriptions are generated conditioned on qq, Description only jumps from 38.6 to 42.4, nearly closing the gap to QCR$^2$ (42.8). Moreover, as shown in [Q2] below, replacing fixed λ=1\lambda = 1 with a learned query-adaptive gate λ(q)\lambda(q) (QA-QCR$^2$) reaches 43.1 (+0.2 over fixed λ=1\lambda = 1), showing that most of the query-adaptive gain (+2.1 over learned fusion) is already unlocked by scoring zqz_q against the item-specific residual rir_i. We have clarified this distinction explicitly in Sections 1 and 3.4.
AuthorsOpenReview ↗

Official response to Reviewer roZP [2/2]

[W2 & Q2] Comparison with stronger query-adaptive evidence-selection baselines

To test whether stronger query-adaptive selection can match or replace residualization, we evaluated three query-adaptive baselines under strictly matched encoders (Qwen3-VL-Embedding-2B), per-route candidate budgets (Ks=5,Kr=5K_s = 5, K_r = 5), final budgets (K=5K = 5), and answer model (Qwen3-VL-8B-Instruct), added to Table 2 and Appendix C.7 (Table C11):

  • Entropy-Fusion (Training-free query-adaptive gate): Weights s~desc\tilde s_{\text{desc}} and s~mm\tilde s_{\text{mm}} per query using the normalized entropy/margin of top-KsK_s description scores.
  • QA-Fusion (Learned query-adaptive MLP router): Trains a 2-layer MLP α(q)=σ(MLP(zq))\alpha(q) = \sigma(\text{MLP}(z_q)) on the calibration set to predict query-specific fusion weights over Kdesc∪Kmm\mathcal K_{\text{desc}} \cup \mathcal K_{\text{mm}}.
  • Description + Multimodal + Qwen3-VL-Reranker-2B (Cross-modal query-adaptive reranker): Applies Qwen3-VL-Reranker-2B to jointly rerank the matched Kdesc∪Kmm\mathcal K_{\text{desc}} \cup \mathcal K_{\text{mm}} candidate union and select top-5 items.
  • QA-QCR$^2$: Applies the same query-adaptive MLP gate λ(q)\lambda(q) to QCR$^2$'s Kdesc∪Kres\mathcal K_{\text{desc}} \cup \mathcal K_{\text{res}} pool.
Method (Matched Encoder & Budgets: 5+5→K=55 + 5 \to K = 5)Candidate Pool Recall (%)Recall@5Avg.Retrieval Latency (ms)
Description + multimodal (static learned fusion)51.444.040.829.9
Entropy-Fusion (query-adaptive dispersion gate)51.444.541.130.0
QA-Fusion (query-adaptive MLP gate α(q)\alpha(q))51.444.941.330.2
Desc. + Multimodal + Qwen3-VL-Reranker-2B51.446.141.8184.6
QCR$^2$ (Fixed λ=1\lambda = 1, training-free)56.248.342.930.0
QA-QCR$^2$ (Learned query-adaptive λ(q)\lambda(q))56.248.743.130.3

These results establish the distinctive value of residualization:

  • Even a heavy cross-modal reranker (Qwen3-VL-Reranker-2B at 184.6 ms/query) on the unprojected Description + Multimodal union reaches only 46.1 Recall@5 and 41.8 Avg. (-2.2 Recall@5 and -1.1 Avg. below training-free QCR$^2$ at 30.0 ms).
  • The bottleneck is candidate generation: because direct multimodal search is dominated by modality-gap and caption-redundant features, Kdesc∪Kmm\mathcal K_{\text{desc}} \cup \mathcal K_{\text{mm}} achieves 51.4% pool recall vs. 56.2% for Kdesc∪Kres\mathcal K_{\text{desc}} \cup \mathcal K_{\text{res}}. Post-hoc query-adaptive selection cannot recover gold evidence that never entered the candidate pool, whereas residualization and query-adaptive gating (QA-QCR$^2$, 43.1) compose cleanly. We have added these comparisons to Table 2 and Appendix C.7 (Table C11).

[W2 & Q3] Cross-model DSG consistency

With the generated descriptions held fixed, we cross-evaluated DSG and downstream performance between Qwen3-VL-8B-Instruct and InternVL3.5-8B in both directions, and also measured answer-model-independent Recall@5 across DSG quartiles (Q1 to Q4), added to Figure 3b and Appendix B.5 (Table B4):

SettingMetricQ1 (Low DSG)Q2Q3Q4 (High DSG)Overall
InternVL3.5-8B DSG (Retrieval only)Δ\DeltaRecall@5 (QCR$^2$ vs. Learned fusion)+1.0+3.1+5.2+7.9+4.3
InternVL3.5-8B DSG →\to Qwen3-VL-8B EvalFusion / QCR$^2$ Avg.43.5 / 44.341.5 / 43.639.9 / 42.837.1 / 40.940.5 / 42.9
Δ\DeltaAvg. (QCR$^2$ vs. Fusion)+0.8+2.1+2.9+3.8+2.4
Qwen3-VL-8B DSG →\to InternVL3.5-8B EvalFusion / QCR$^2$ Avg.43.0 / 43.741.0 / 42.939.2 / 41.936.0 / 39.539.8 / 42.0
Δ\DeltaAvg. (QCR$^2$ vs. Fusion)+0.7+1.9+2.7+3.5+2.2

Per-query DSG rankings correlate strongly across the two models (Spearman ρ=0.74\rho = 0.74), and in both cross-model directions, as well as in answer-model-free Recall@5, QCR$^2$'s gain increases monotonically from Q1 to Q4. Full per-method curves and tables are provided in Figure 3b and Appendix B.5 (Table B4).

AuthorsOpenReview ↗

Any follow-up question?

Dear Reviewer roZP,

Thank you again for your insightful evaluation and constructive questions. As the author-reviewer discussion period comes to a close shortly (about 22 hours), we wanted to share a brief index of how each of your points is addressed in our two-part response and magenta-highlighted manuscript revisions:

  • Operational role of query conditioning (W1, Q1): We have clarified how offline item-specific residuals pair with online directional alignment and per-query candidate-union standardization, supported by query-conditioned description and learned gating diagnostics (Official response [1/2]; Sec. 1, Sec. 3.4, App. C.4, Table C7, App. C.5, Table C8).
  • Query-adaptive selection baselines (W2, Q2): We have added matched-budget comparisons against dispersion-based Entropy-Fusion, learned QA-Fusion, and cross-modal reranking with Qwen3-VL-Reranker-2B (Official response [2/2]; Sec. 4.2, Table 2, App. A.5, App. C.7, Table C11).
  • Cross-model DSG consistency (W2, Q3): We have added answer-model-independent Recall@5 across DSG quartiles and bidirectional cross-model evaluations between InternVL3.5-8B and Qwen3-VL-8B-Instruct (Official response [2/2]; Sec. 4.3, Figure 3b, App. B.5, Table B4).

Could you let us know whether these new experiments and clarifications resolve your main concerns, or if there are any remaining questions we can address before the discussion period ends?

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

The paper introduces a novel multimodal RAG solution improving upon the traditional approach of generating a text caption to describe non-text modalities and use them for multimodal retrieval augmented generation. The paper's observation is that the traditional approaches drop important details about the content itself (such as the color of the objects in the scene). The proposed solution QCR keeps the text description index but also adds a second index built from each item’s multimodal embedding after projecting out the span of that item’s own description embeddings. The usual candidates generated from the text search path is merged with the residual vector retrieval results and re-scored. The proposed approach is training free. On the UniversalRAG benchmarks (includes 10 datasets), QCR2 scores 42.9, compared with 42.4 for UniversalRAG. Most of the gain are on image and video tasks.

Critical strengths
  1. Very practical and intuitive idea.
  2. Fair apples-to-apple comparison with the baselines. With the same retriever, data, descriptions and budget, it beats learned fusion
  3. Training-free and cheap. Same storage and latency as two-index fusion, with no model updates.
  4. Shown to generalize across different models (three retrievers and three answer models)
  5. Ablations, failure cases, an exact-search check and a query-aware description upper bound are all reported.
Critical weaknesses
  1. The improvements are somewhat small over the best baseline. 42.9 vs 42.4 for UniversalRAG. There are no error bars.
  2. The paper could improve explainability why the proposed method works. It is not intuitively clear what exactly happens when we remove only four description directions from a high-dimensional embedding.
  3. The closest prior work is likely mispositioned. GRAIL already uses projection-based subtraction for multimodal retrieval but QCR cites it only as an iterative-retrieval method and never compares them.
Questions for authors

The paper is intuitive, well-written with extensive experimentation. It can be improved further though by

  • Adding error bars to the metrics.
  • Potential simple baseline could such as removing the average text direction.
  • Compare and contrast with recent work O & Lee (2026) even though they may not compare apples-to-apples
4 comments
AuthorsOpenReview ↗

Official response to Reviewer UjvC [1/2]

We sincerely thank the reviewer for the positive evaluation and insightful suggestions. Your feedback on error bars, high-dimensional subspace intuition, and the positioning of O & Lee (2026) directly improved both our experiments and related-work discussion. Below, we address every point in detail; a summary of all manuscript updates is provided in the Revision summary comment.

[W1 & Q1] Error bars and magnitude of improvement over UniversalRAG

We have added paired bootstrap 95% confidence intervals (10,000 resamples across queries) to our main comparisons (Section 4.2, Table 2, and Appendix A.2). Two factors contextualize the 42.90 vs. 42.40 overall score in Table 1:

  1. Dilution from non-residual text/table modalities: Following UniversalRAG's averaging protocol, Avg. weights four modality groups equally (Text, Table, Image, and Video). Because QCR$^2$ constructs residuals exclusively for non-text items (1[xi∈Dv]=0\mathbb{1}[x_i \in \mathcal D_v] = 0 on Text and Table), half of the macro-average weight comes from benchmarks without a residual branch (MMLU, NQ, HotpotQA, and HybridQA). Across the six image and video benchmarks, QCR$^2$ outperforms UniversalRAG by +1.08 points (+1.15 on Image, +1.01 on Video; 95% paired CI [+0.68, +1.49]), including +1.55 on MRAG-Bench (54.10 vs. 52.55), +1.28 on LVBench (35.00 vs. 33.72), and +1.60 R-L on VideoRAG-Synth (41.55 vs. 39.95), yielding a statistically significant overall gain of +0.50 ([+0.24, +0.76]) and surpassing UniversalRAG's ground-truth routed Oracle (42.45) without using a modality-granularity router.
  2. Controlled apples-to-apples comparison: When UniversalRAG is evaluated under the exact Table 2 backbone (Qwen3-VL-Embedding-2B) and corpus, it achieves 45.6 Recall@5 and 42.2 Avg., whereas QCR$^2$ achieves 48.3 Recall@5 (+2.7, 95% CI [+2.0, +3.4]) and 42.9 Avg. (+0.70, 95% CI [+0.42, +0.98]), and beats Learned fusion (40.8) by +2.10 ([+1.72, +2.48]).

[W2 & Q2] Why removing four description directions works & average-text-direction baseline

(1) Geometric intuition in high dimensions. If four random orthonormal directions were removed from a dd-dimensional unit embedding (d≥1536d \ge 1536), they would eliminate only 4/d≈0.2%4/d \approx 0.2\% of the squared norm, leaving the vector virtually unchanged (43.1 Recall@5, 40.5 Avg., identical to Description + multimodal). In contrast, pretrained multimodal embeddings concentrate semantic variance on a low-dimensional manifold, and the M=4M = 4 structured prompts explicitly target each item's dominant scene, attribute, fine-grained, and OCR/temporal semantics. Empirically, projecting viv_i onto its 4D item-specific span span(Ti)\text{span}(T_i) removes on average E[∣∣vi∥∣∣22]=45.1%\mathbb{E}[||v_i^\parallel||_2^2] = 45.1\% of ∣∣vi∣∣22||v_i||_2^2 (leaving ∣∣vi⊥∣∣2=0.738±0.064||v_i^\perp||_2 = 0.738 \pm 0.064; Table C9). Stripping out nearly half of the embedding energy, corresponding to the dominant features already verbalized in text, reorganizes nearest-neighbor structure around unverbalized visual and temporal cues (as visualized in Figure E2a).

(2) Comparison with removing the global average text direction. Following your suggestion, we added a baseline that computes the corpus-wide mean description vector tˉ=1M∣Dv∣∑i∈Dv∑m=1Mti(m)\bar t = \frac{1}{M|\mathcal D_v|}\sum_{i \in \mathcal D_v}\sum_{m=1}^{M} t_i^{(m)}, normalizes uˉtext=tˉ/∣∣tˉ∣∣2\bar u_{\text{text}} = \bar t/||\bar t|| _ {2}, and projects out uˉtext\bar{u}_{\text{text}} from every viv_i under the Table 2 setting (added to Table 3 and Appendix A.5):

Projection Baseline (Controlled Table 2 Setting)Subspace RankRecall@5 (95% CI)Avg. (95% CI)
Description + multimodal (equal fusion)043.1±0.643.1 \pm 0.640.5±0.540.5 \pm 0.5
Description + multimodal (learned fusion)044.0±0.644.0 \pm 0.640.8±0.540.8 \pm 0.5
Item-specific random subspaces (5 seeds)4 (per-item)43.1±0.643.1 \pm 0.640.5±0.540.5 \pm 0.5
Global average text direction (UjvC Q2)1 (global)44.7±0.644.7 \pm 0.641.1±0.541.1 \pm 0.5
Global description subspace (rank-4 SVD)4 (global)45.3±0.645.3 \pm 0.641.4±0.541.4 \pm 0.5
QCR$^2$ (Item-specific description subspace)4 (per-item)48.3±0.648.3 \pm 0.642.9±0.542.9 \pm 0.5

Removing the global average text direction (41.1) and top-4 global SVD directions (41.4) provides a modest boost over Learned fusion (40.8) by reducing the shared modality offset, but QCR$^2$ achieves an additional +1.8 Avg. (+3.6 Recall@5) over the average-text baseline because omitted details depend on what that specific item's captions already stated. We have added this baseline and norm-energy explanation to Sections 3.3 and 4.4 (Table 3).

AuthorsOpenReview ↗

Official response to Reviewer UjvC [2/2]

[W3 & Q3] Positioning and comparison with O & Lee (2026) (GRAIL)

We completely agree with the reviewer: grouping O & Lee (2026) solely with textual/structural iterative RAG methods in Section 2 understated its close methodological connection as a projection-subtractive multimodal retriever. We have revised Section 2 (Multimodal RAG) to explicitly highlight GRAIL (O & Lee, 2026) as the closest projection-based multimodal retrieval work and added both a conceptual and an adapted empirical comparison (Table 2 and Appendix A.5).

  • Conceptual distinction (Online query-side multi-hop subtraction vs. Offline item-side intra-item decomposition): GRAIL targets multi-hop inter-item gaps by updating the query online across sequential retrieval steps, subtracting the projection onto already-retrieved Hop-1 context embeddings so Hop-2 seeks a different piece of missing evidence. In contrast, QCR$^2$ targets single-step intra-item caption insufficiency by decomposing each corpus item viv_i offline against its own MM descriptions span(Ti)\text{span}(T_i). When a single-hop query asks for an uncaptioned detail of an item (e.g., "What color balloons were displayed during the ship launching ceremony?" in Figure 1), query-side subtraction after Hop-1 steers away from Hop-1 concepts while still searching unprojected multimodal embeddings dominated by caption-redundant semantics.
  • Empirical comparison under matched Table 2 budgets (Qwen3-VL-Embedding-2B, K=5K = 5): Adapting GRAIL's two-step gap-aware query subtraction (Hop-1 top-3 retrieval followed by orthogonal query projection for Hop-2 top-2 retrieval over the unified corpus) improves multi-hop HotpotQA F1 (35.1 vs. 34.5), validating GRAIL's multi-hop design, while achieving 44.6 Recall@5 and 41.2 Avg. overall across the ten benchmarks (at 54.6 ms sequential two-hop latency) compared to 48.3 Recall@5 and 42.9 Avg. at 30.0 ms parallel latency for QCR$^2$ (Table 2). Thus, GRAIL and QCR$^2$ solve complementary retrieval gaps (inter-hop query steering vs. intra-item description residualization).
AuthorsOpenReview ↗

Any follow-up question?

Dear Reviewer UjvC,

Thank you again for your encouraging feedback and helpful suggestions. As the author-reviewer discussion period concludes shortly (about 22 hours), we wanted to briefly summarize where each of your suggestions has been incorporated across our two-part response and revised manuscript (highlighted in magenta):

  • Confidence intervals & UniversalRAG comparison (W1, Q1): We have added paired bootstrap confidence intervals, modality-group breakdowns, and a controlled UniversalRAG evaluation under the shared retriever and corpus (Official response [1/2]; Sec. 4.2, Table 1, Table 2, App. A.2, App. A.5).
  • Geometric intuition & average-text baseline (W2, Q2): We have explained why removing item-specific description directions reorganizes high-dimensional neighborhoods and added your suggested global average-text-direction baseline (Official response [1/2]; Sec. 3.3, Sec. 4.4, Table 3, App. A.5, App. C.6, Table C9).
  • Positioning & comparison with GRAIL (W3, Q3): We have revised Related Work to highlight GRAIL (O & Lee, 2026) and added an adapted empirical comparison contrasting online query-side subtraction with offline item-specific decomposition (Official response [2/2]; Sec. 2, Sec. 4.2, Table 2, App. A.5).

Could you let us know whether these new experiments and clarifications resolve your main concerns, or if there are any remaining questions we can address before the discussion period ends?

Dear Reviewer UjvC,

This is a gentle reminder that about 6 hours remain in the author-reviewer discussion period. Even a brief note on whether our rebuttal and revised manuscript have addressed your feedback would be greatly appreciated. If your concerns have been resolved, we also kindly ask whether you would consider updating your score accordingly. Thank you again for your time and thoughtful review!

Reviewer_oW1AOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper introduces Query-Conditioned Residual Retrieval (QCR$^2$), a training-free retrieval method for multimodal RAG, motivated by cases where a query-independent description omits details needed to answer a particular query. For each non-text item, QCR$^2$ projects its embedding onto the span of its description embeddings and normalizes the orthogonal residual. At retrieval time, it searches the description and residual representations independently, combines their scores to rank candidates, and passes the selected original items to the answer model. The authors evaluate QCR$^2$ on the ten-benchmark UniversalRAG protocol, covering text, tables, images, and videos, and compare it with single-corpus retrieval, unified multimodal retrievers, score-fusion baselines, and UniversalRAG. They also analyze whether residual retrieval recovers annotated supporting evidence missed by description-based retrieval and how such recoveries relate to downstream answer outcomes.

Critical strengths
  • The paper identifies a practical limitation of description-based multimodal retrieval. A description can accurately summarize an item while omitting the detail needed for a particular query. Framing description sufficiency as query-dependent gives the method a clear motivation and addresses a relevant challenge in multimodal RAG.

  • The evaluation connects retrieval behavior to downstream outcomes. Beyond end-to-end answer scores, the authors analyze recovery of annotated supporting evidence. They report answer improvements on 30.8% of queries when the residual route uniquely retrieves supporting evidence, compared with 9.4% when it does not, while the rates of degraded answers are similar. This provides useful evidence about when the additional retrieval route may help.

Critical weaknesses
  • The contribution of the item-specific residual is not isolated. The paper does not define its global-subspace ablation, report residual-norm statistics, or compare against matched controls such as rank-matched random subspaces, modality-wise centering, and a non-decomposition retrieval score. Residual-only retrieval outperforming direct multimodal retrieval (36.9 vs. 35.3, Table 3) is consistent with removing a shared modality-gap component, but does not establish that the residual captures query-relevant information omitted by descriptions. The fusion baselines also use a non-text retrieval channel under matched candidate budgets; the unresolved question is whether replacing the original multimodal score with the item-specific residual provides a benefit specific to the description-conditioned projection.

  • The reported margins are not separated from evidence-ordering, configuration, and answer-model effects. Table 1 uses published baselines with their original retrieval configurations, while the controlled Table 2 comparison reports only an aggregate and excludes UniversalRAG. Thus, the 0.5-point margin over UniversalRAG is not isolated from retrieval-configuration differences. Under the default Ks=K=5K_s=K=5, the “description candidates only” variant appears to use the same five items as Description-only; if so, its 3.1-point gain comes from reordering alone, exceeding the 2.1-point margin between QCR² and learned fusion. The paper reports no order-controlled comparison or uncertainty estimates. Finally, DSG is computed with the same Qwen model used to generate descriptions and produce end-to-end scores. High-DSG queries are those where this model’s reference-answer score drops more when gold evidence is replaced by descriptions. Fig. 3b shows that residual-only supporting-evidence recovery increases with DSG, but neither Fig. 3a nor Fig. 3b compares per-group Recall@5 across methods to establish that QCR²’s retrieval advantage over fusion grows with description insufficiency.

Questions for authors
  1. How are residual norms ∥∣vi⊥∥∣2{\||v_i^\perp\||}_2 distributed across non-text items, and what fraction falls below ϵ\epsilon? Please define the “global description subspace” baseline and compare QCR$^2$ under the Table 2 setting with item-specific rank-matched random subspaces (averaged over multiple seeds), modality-wise mean-centering, and direct retrieval of non-text items by smm−βsdescs_{\mathrm{mm}}-\beta s_{\mathrm{desc}}, with β\beta tuned on the calibration set used for learned fusion. Please report paired bootstrap confidence intervals over queries for the controlled comparisons below as well. Comparable results would respectively challenge the need for description-conditioned directions, suggest that centering suffices, or show that the per-item projection is unnecessary.

  2. Could generic non-text candidate access reproduce QCR$^2$’s gains? Please compare it with a control that retrieves the global top-KsK_s items by sdescs_{\mathrm{desc}}, then fills KrK_r further slots with non-text items ranked by sdescs_{\mathrm{desc}}, excluding items already selected. Rank the union by standardized description score plus a constant bonus γ1[xi∈Dv]\gamma \mathbf{1}[x_i \in D_v], with γ\gamma tuned on the calibration set used for learned fusion. Please report final-set modality composition and QCR$^2$’s union size and route overlap. Comparable performance would indicate that residual-specific scoring is not needed.

  3. For the “QCR$^2$, description candidates only” variant in Table 3, what KsK_s and evidence order were used? If Ks=K=5K_s=K=5, the candidate set matches Description-only, so the 3.1-point gain would arise from reordering alone—larger than the 2.1-point margin between QCR$^2$ and learned fusion. Please report a Table 2 comparison averaged over multiple randomized evidence orders. If KsK_s was larger, please specify both budgets.

  4. Does the Figure 3a relationship persist when DSG and end-to-end evaluation use different answer models—for example, InternVL3.5-8B and Qwen3-VL-8B-Instruct in both directions—with descriptions held fixed? Since DSG groups may also reflect the answer model’s sensitivity to gold evidence, please report Recall@5 by DSG group for Description, Multimodal, Fusion, and QCR$^2$.

7 comments
AuthorsOpenReview ↗

Official response to Reviewer oW1A [1/5]

We thank the reviewer for the exceptionally thorough and constructive review. Your falsifiable control questions helped us rigorously isolate the item-specific residual mechanism and clarify missing experimental specifications. Below, we address each weakness and question point by point; an overview of manuscript updates is provided in the "Revision summary" comment.

[W1 & Q1] Residual norm statistics, global subspace definition, and matched projection controls

(1) Distribution of residual norms ∣∣vi⊥∣∣2||v_i^\perp|| _ 2 and ϵ\epsilon. In Equations (5) and (10), we set ϵ=10−6\epsilon = 10^{-6} strictly to prevent numerical division by zero. Across all non-text items xi∈Dvx_i \in \mathcal{D}_v with Qwen3-VL-Embedding-2B and M=4M = 4 descriptions (reported in Appendix C.6, Table C9), ∣∣vi⊥∣∣2||v_i^\perp|| _ 2 is distributed as follows:

StatisticMean ±\pm SDMin5th %25th %Median75th %95th %MaxFraction below ϵ=10−6\epsilon = 10^{-6}
∥vi⊥∥2\lVert v_i^\perp \rVert_20.738±0.0640.738 \pm 0.0640.4620.6240.6960.7410.7830.8390.9340.0% (0 items)

By ∣∣vi∣∣22=∣∣vi∥∣∣22+∣∣vi⊥∣∣22=1||v_i|| _ 2^2 = ||v_i^\parallel|| _ 2^2 + ||v_i^\perp|| _ 2^2 = 1, the four item-specific descriptions account for an average of E[∣∣vi∥∣∣22]=45.1%\mathbb{E}[||v_i^\parallel|| _ 2^2] = 45.1\% of each item's embedding energy, while zero items fall below ϵ=10−6\epsilon = 10^{-6}. Moreover, the nearly twofold spread in ∣∣vi⊥∣∣2||v_i^\perp|| _ 2 (0.462 to 0.934) explains why normalizing rir_i in Equation (5) improves Recall@5 from 47.2 to 48.3 and Avg. from 42.4 to 42.9 (Appendix C.5, Table C8): without normalization, items with sparse descriptions (large ∣∣vi⊥∣∣2||v_i^\perp|| _ 2) are upweighted regardless of directional alignment with the query.

(2) Definition of Global description subspace. We apologize for omitting this definition from Section A.5. We stack the normalized description embeddings of all non-text items as columns into Tall∈Rd×(4∣Dv∣)T_{\text{all}} \in \mathbb{R}^{d \times (4| \mathcal{D} _ v|)} and compute its rank-4 truncated SVD, Tall≈UglobalΣV⊤T_{\text{all}} \approx U_{\text{global}}\Sigma V^\top. Using the four leading left singular vectors Uglobal∈Rd×4U_{\text{global}} \in \mathbb{R}^{d \times 4} as a shared orthonormal basis, every multimodal embedding is projected via vi⊥,global=(I−UglobalUglobal⊤)viv_i^{\perp,\text{global}} = (I - U_{\text{global}}U_{\text{global}}^\top)v_i and ℓ2\ell_2-normalized as in Equation (5), keeping all other retrieval settings identical. We have added this formal definition to Appendix A.5.

AuthorsOpenReview ↗

Official response to Reviewer oW1A [2/5]

(3) Controlled comparisons and paired bootstrap 95% confidence intervals. We evaluate all requested controls under the exact Table 2 protocol (Qwen3-VL-Embedding-2B, Ks=Kr=5K_s = K_r = 5, K=5K = 5, Qwen3-VL-8B-Instruct, with β∗=0.45\beta^* = 0.45 tuned on the same held-out calibration set used for learned fusion), alongside UniversalRAG using the same retriever and corpus (added to Table 2 and Table 3):

Second-Route / Retrieval Control (Table 2 Setting)Recall@5 (95% CI)Avg. (95% CI)Paired Δ\DeltaAvg. vs. QCR$^2$
Direct multimodal (single-route)31.4±0.631.4 \pm 0.635.3±0.535.3 \pm 0.5-7.60 ([-8.12, -7.08])
Modality-wise mean-centered multimodal (single-route)33.8±0.633.8 \pm 0.636.6±0.536.6 \pm 0.5-6.30 ([-6.79, -5.81])
Residual only (single-route)34.1±0.634.1 \pm 0.636.9±0.536.9 \pm 0.5-6.00 ([-6.48, -5.52])
Description + multimodal (equal fusion)43.1±0.643.1 \pm 0.640.5±0.540.5 \pm 0.5-2.40 ([-2.79, -2.01])
Item-specific rank-4 random subspaces (5 seeds)43.1±0.643.1 \pm 0.640.5±0.540.5 \pm 0.5-2.40 ([-2.78, -2.02])
Description + multimodal (learned fusion, α∗=0.65\alpha^* = 0.65)44.0±0.644.0 \pm 0.640.8±0.540.8 \pm 0.5-2.10 ([-2.48, -1.72])
Modality-wise mean-centering (dual-route)45.1±0.645.1 \pm 0.641.3±0.541.3 \pm 0.5-1.60 ([-1.96, -1.24])
Global description subspace (rank-4 SVD)45.3±0.645.3 \pm 0.641.4±0.541.4 \pm 0.5-1.50 ([-1.84, -1.16])
Calibrated score subtraction (smm−β∗sdesc,β∗=0.45s_{\text{mm}} - \beta^* s_{\text{desc}}, \beta^* = 0.45)46.0±0.646.0 \pm 0.641.6±0.541.6 \pm 0.5-1.30 ([-1.64, -0.96])
UniversalRAG (controlled Qwen3-VL-Embedding-2B)45.6±0.645.6 \pm 0.642.2±0.542.2 \pm 0.5-0.70 ([-0.98, -0.42])
QCR$^2$ (Ours)48.3±0.648.3 \pm 0.642.9±0.542.9 \pm 0.5--

The results clarify each hypothesis raised by the reviewer:

  • Modality-gap centering vs. item-specific projection: In single-route retrieval, modality-wise mean-centering (36.6) nearly matches Residual only (36.9), confirming the reviewer's intuition that Residual only outperforming Direct multimodal (35.3) primarily reflects removing a shared modality-gap offset. However, in the dual-route setting, both Modality-wise mean-centering (41.3) and Global description subspace (41.4) capture less than 30% of QCR$^2$'s gain over Learned fusion (40.8 to 42.9), leaving a significant +1.60 ([+1.24, +1.96]) margin specifically attributable to removing each item's own description directions.
  • Random subspaces: Projecting out 4 random directions in Rd\mathbb{R}^d removes only 4/d≈0.2%4/d \approx 0.2\% of embedding energy, leaving neighborhood structure unchanged (43.1 Recall@5, 40.5 Avg.).
  • Score subtraction (smm−β∗sdescs_{\text{mm}} - \beta^* s_{\text{desc}}): Scalar subtraction (46.0 Recall@5, 41.6 Avg.) improves over Learned fusion (40.8) but remains less effective than QCR$^2$ by -2.3 Recall@5 and -1.30 Avg. ([-1.64, -0.96]). Because sdesc(q,i)=max⁡m∈{1,…,4}zq⊤ti(m)s_{\text{desc}}(q,i) = \max_{m \in \{1,\dots,4\}} z_q^\top t_i^{(m)}, scalar subtraction removes only the single highest-scoring caption with a fixed global weight β∗\beta^* rather than projecting out the 4D orthonormal span UiUi⊤viU_i U_i^\top v_i, omits per-item residual normalization ∣∣vi⊥∣∣2−1||v_i^\perp||_2^{-1}, and cannot be indexed as a single static vector per item in HNSW due to the query-dependent max⁡m\max_m.
  • Margin over UniversalRAG: In Table 1, the +0.50 overall margin (42.90 vs. 42.40, 95% paired CI [+0.24, +0.76]) averages equally across four modality groups (Text, Table, Image, and Video), two of which (Text and Table) construct no residual index (I[xi∈Dv]=0\mathbb{I}[x_i \in \mathcal{D}_v] = 0). Across the six image and video benchmarks, QCR$^2$ improves by +1.08 (+1.15 on Image, +1.01 on Video; 95% CI [+0.68, +1.49]), and under the strictly controlled Table 2 backbone it outperforms UniversalRAG by +2.7 Recall@5 and +0.70 Avg. ([+0.42, +0.98]) without using a modality-granularity router.
AuthorsOpenReview ↗

Official response to Reviewer oW1A [3/5]

[Q2] Generic non-text candidate access, modality composition, union size, and route overlap

We implemented the exact requested control: retrieving global top-Ks=5K_s = 5 items by sdescs_{\text{desc}}, filling Kr=5K_r = 5 additional non-text slots from Dv∖Kdesc\mathcal{D} _ {v} \setminus \mathcal{K}_{\text{desc}} ranked by sdescs _ {\text{desc}}, and ranking the 10-item union by s~desc(q,i)+γ1[xi∈Dv]\tilde{s} _ {\text{desc}}(q,i) + \gamma \mathbf{1}[x_i \in \mathcal{D}_v] with γ∗=0.25\gamma^* = 0.25 tuned on the calibration set. Across the nine retrieval benchmarks (excluding no-retrieval MMLU), we report union size ∣Kq∣|\mathcal{K} _ q|, route overlap ∣Kdesc∩Kres∣|\mathcal{K} _ {\text{desc}} \cap \mathcal{K} _ {\text{res}}|, and final top-5 modality composition (added to Appendix C.6, Table C10):

Method (K=5K = 5)∣Kq∣\lvert \mathcal{K}_q \rvertOverlap (items / %)Top-5 Dt\mathcal{D}_t (%)Top-5 Image / Clip / Video (Dv\mathcal{D}_v %)Recall@5Avg.
Description only5.00--52.423.9 / 11.2 / 12.5 (47.6)38.738.6
Non-text slot fill + bonus (γ∗=0.25\gamma^* = 0.25)10.000.00 / 0.0%41.827.6 / 14.4 / 16.2 (58.2)39.839.2
Description + multimodal (learned)9.120.88 / 17.6%40.628.5 / 14.6 / 16.3 (59.4)44.040.8
QCR$^2$ (Ours)9.360.64 / 12.8%41.228.1 / 14.5 / 16.2 (58.8)48.342.9

Two findings stand out:

  1. Generic non-text access cannot reproduce QCR$^2$: Although Non-text slot fill + bonus matches QCR$^2$'s final non-text modality share (58.2% vs. 58.8%), it reaches only 39.8 Recall@5 and 39.2 Avg. (-8.5 Recall@5 and -3.7 Avg. below QCR$^2$). Because both candidate selection and ranking rely solely on sdescs_{\text{desc}}, items whose descriptions omit the queried visual/temporal attribute still rank poorly within Dv\mathcal{D} _ {v}. Furthermore, note that in Equation (10), s~res(q,i)\tilde s_{\text{res}}(q,i) is standardized to zero mean over Kq∩Dv\mathcal{K} _ {q} \cap \mathcal{D} _ {v}, so λI[xi∈Dv]s~res(q,i)\lambda \mathbb{I}[x_i \in \mathcal{D} _ {v}]\tilde{s}_{\text{res}}(q,i) penalizes non-text candidates with below-average residual alignment rather than granting a positive modality bonus.
  2. Low route overlap confirms complementary candidate discovery: QCR$^2$ yields an average union size of 9.36 items (9.92 on text/table tasks and 9.08 on image/video tasks) with only 0.64 overlapping items (12.8% overall; 18.4% on image/video tasks), confirming that Kres\mathcal K_{\text{res}} introduces distinct multimodal candidates missed by Kdesc\mathcal K_{\text{desc}}.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revision summary

We truly appreciate all the reviewers and the chairs for their commitment to the community and their thoughtful, constructive feedback. We are highly encouraged that the reviewers found our problem formulation of query-dependent description sufficiency intuitive, practical, and well-motivated (Reviewers oW1A, UjvC, roZP, and NALs), recognized that QCR$^2$ is training-free while matching standard two-index fusion in storage (27.4 GB) and latency (30.0 ms) (Reviewers UjvC, roZP, and NALs), and valued our evaluation across ten heterogeneous benchmarks, multiple retriever/generator backbones, and downstream evidence-recovery analyses (Reviewers oW1A, UjvC, roZP, and NALs).

To address all raised questions rigorously, we conducted controlled ablation, fairness, and statistical significance analyses under the unified experimental setting of Table 2 (Qwen3-VL-Embedding-2B, identical heterogeneous corpus D\mathcal{D} and M=4M = 4 descriptions, Ks=Kr=5K_s = K_r = 5, final budget K=5K = 5, and Qwen3-VL-8B-Instruct, with paired bootstrap 95% confidence intervals over 10,000 resamples):

Method (Controlled Table 2 Setting)Recall@5Avg.Δ\DeltaAvg. vs. QCR$^2$ (95% CI)
Description only38.738.6-4.30 ([-4.74, -3.86])
Direct multimodal31.435.3-7.60 ([-8.12, -7.08])
Description + non-text slot fill + bonus (γ∗=0.25\gamma^* = 0.25)39.839.2-3.70 ([-4.12, -3.28])
Description + multimodal (equal fusion)43.140.5-2.40 ([-2.79, -2.01])
Item-specific rank-4 random subspaces (5 seeds)43.140.5-2.40 ([-2.78, -2.02])
Description + multimodal (learned fusion)44.040.8-2.10 ([-2.48, -1.72])
Global average text direction (rank-1)44.741.1-1.80 ([-2.17, -1.43])
GRAIL-adapted query-subtractive retrieval (O & Lee, 2026)44.641.2-1.70 ([-2.09, -1.31])
Modality-wise mean-centering45.141.3-1.60 ([-1.96, -1.24])
Query-adaptive MLP fusion (QA-Fusion)44.941.3-1.60 ([-1.95, -1.25])
Global description subspace (rank-4 SVD)45.341.4-1.50 ([-1.84, -1.16])
Calibrated score subtraction (smm−β∗sdescs_{\text{mm}} - \beta^* s_{\text{desc}})46.041.6-1.30 ([-1.64, -0.96])
Description + multimodal + Qwen3-VL-Reranker-2B46.141.8-1.10 ([-1.45, -0.75])
UniversalRAG (controlled backbone & corpus)45.642.2-0.70 ([-0.98, -0.42])
QCR$^2$ (Ours)48.342.9--

Below, we summarize the major revisions made to the manuscript (highlighted in magenta in the revised PDF):

  • Projection, centering, and non-text access controls (Section 4.4, Table 3, Appendix A.5, Appendix C.6): We formally define the rank-4 SVD Global description subspace baseline and ϵ=10−6\epsilon = 10^{-6}, report ∥vi⊥∥−2\|v_i^\perp\|-2 norm distributions (Table C9) and route-overlap statistics (Table C10), and add matched controls for rank-4 random subspaces, global average-text subtraction, modality-wise mean-centering, calibrated non-decomposition subtraction (smm−β∗sdescs_{\text{mm}} - \beta^* s_{\text{desc}}), and generic non-text slot filling (Reviewer oW1A [W1, Q1, Q2], Reviewer UjvC [W2, Q2]).
  • Controlled UniversalRAG, query-adaptive selection baselines, prompt-order checks, and error bars (Section 4.2, Table 2, Table 3, Appendix A.2, Appendix A.5, Appendix C.7): We add paired bootstrap 95% confidence intervals, include controlled UniversalRAG, adapted GRAIL, and stronger query-adaptive selection/reranking baselines in Table 2 and Table C11, specify the Ks=30K_s = 30 (and union-matched Ks=10K_s = 10) budgets for QCR$^2$ with description candidates only in Table 3, and report Table 2 results averaged over randomized prompt evidence orders in Appendix C.7 (Reviewer oW1A [W2, Q1, Q3], Reviewer UjvC [W1, Q1], Reviewer roZP [W2, Q2], Reviewer NALs [W2]).
  • Cross-model DSG validation and per-group Recall@5 (Section 4.3, Figure 3b, Appendix B.5, Table B4): We add answer-model-independent Recall@5 across DSG quartiles (Q1 to Q4) in Figure 3b and cross-evaluate DSG between InternVL3.5-8B and Qwen3-VL-8B-Instruct in both directions with fixed descriptions in Table B4 (Reviewer oW1A [W2, Q4], Reviewer roZP [W2, Q3]).
  • Positioning of O & Lee (2026), geometric intuition, and operational role of query conditioning (Section 1, Section 2, Section 3.3, Section 3.4): We revise Section 2 to accurately position and empirically contrast GRAIL (O & Lee, 2026), quantify why projecting out M=4M = 4 item-specific directions removes 45.1% of multimodal norm energy in Section 3.3, and clarify how query conditioning operates over offline residuals in Section 1 and Section 3.4 (Reviewer UjvC [W2, W3, Q3], Reviewer roZP [W1, Q1], Reviewer NALs [W1, W2]).

Please feel free to let us know if any concern remains; we are happy to provide further discussion during the rebuttal period.

Area_Chair_GoWVOpenReview ↗

Please engage with the authors rebuttal by Nov 18

Dear Reviewers,

The authors have posted their rebuttal. Please engage with it by November 18, 2026, noting which of your concerns are resolved or which remain. Even brief engagement helps greatly.

Best regards,
Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.