acceptodds
Under review as a conference paper at ICLR 2027

Min-Cost Flow Routing for Evidence Assembly in Long Multimodal Documents

Abstract

Answering questions about long multimodal documents requires distributing a fixed evidence budget across relevant facets in text, tables, figures, and slides while avoiding near-duplicates. We present \flowreader, which formulates evidence selection as a single minimum-cost flow problem with capacity limits over a multimodal content graph. Spectral decomposition identifies latent aspects of query-relevant content and allocates the budget among them in proportion to their spectral energy. These capacity limits enforce aspect coverage during routing without requiring a language-model planning call. Query-conditioned costs prioritize chains of relevant, mutually consistent evidence. Decomposing the optimal flow produces short evidence chains, which a vision-language model reads in parallel and a reasoner reconciles. On VisDoMBench with Qwen3-VL-32B, \flowreader achieves the highest macro accuracy (), surpassing the strongest prior system by points, leading on three of five subsets and attaining the highest worst-subset accuracy. It uses a measured content nodes per query and maintains its lead at . Ablation studies with a fixed graph, scorer, reader, and judge show that cost design drives accuracy, capacity limits preserve it while using about three-quarters of the reader tokens required by shortest-path routing without these limits on the same network, and spectral aspects align with LLM-generated sub-questions without a planning call.

Primary area:
foundation or frontier models, including LLMs
Keywords:
retrieval-augmented generation, multimodal document understanding, question answering, vision-language models, graph-based retrieval

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_4HH3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FLOWREADER for evidence selection and assembly in long multimodal document question answering. The method computes query relevance over an offline content graph, applies spectral decomposition to relevance-weighted node embeddings, and assigns flow quotas to latent aspects in proportion to their spectral energy. A single capacitated minimum-cost flow optimization produces candidate evidence chains, which are subsequently selected, read in parallel by a vision-language model, and reconciled by a reasoner. The theoretical result guarantees that the complete routed flow satisfies the prescribed aspect allocation, but does not guarantee semantic coverage or preservation of that allocation after path selection. On VisDoMBench, the system achieves a macro accuracy of 68.9, exceeding the strongest reported baseline by 2.7 percentage points. Controlled ablations indicate that cost design contributes substantially to accuracy, while capacity constraints reduce prompt-token consumption by approximately 27% relative to the uncapacitated shortest-path configuration at similar accuracy.

Critical strengths

1.The paper formulates multimodal evidence assembly as a minimum-cost flow problem with aspect quotas, providing an alternative to online LLM-based sub-question planning. On VisDoMBench, the system achieves a macro accuracy of 68.9, improving over the strongest reported baseline by 2.7 percentage points, leading on three of five subsets, and achieving the highest worst-subset accuracy. Accuracy remains at 68.0 when the average measured number of distinct nodes read is reduced to 12.9. These results demonstrate practical utility in the evaluated setting, with improvements extending beyond a single subset 2.The paper compares multiple selection strategies, routing methods, and cost components while holding the graph, node scores, reader, reasoner, and judge fixed. The shortest-path comparison is particularly informative: it indicates that the main benefit of capacity constraints is reduced token consumption at similar accuracy, rather than an additional accuracy improvement. The authors also explicitly disclose the scope of the coverage guarantee, its limitations after path selection, and confounding factors in some ablations and evaluation procedures. These analyses and disclosures help identify which parts of the method are effective and where its conclusions apply

Critical weaknesses

1.The theoretical guarantee constrains only the numerical allocation of the complete flow. Final path selection need not preserve that allocation, and coverage of the actual semantic facets required by the question is not directly evaluated. The single-sink ablation substantially changes the termination scale κ\kappa when removing aspect nodes, so its performance degradation cannot be attributed specifically to the removal of spectral aspects. The 0.45-percentage-point difference between spectral aspects and LLM-generated sub-questions, with p=0.35p=0.35, establishes only that the experiment did not detect a significant difference; it does not establish semantic alignment or statistical equivalence.Meanwhile, the shortest-path configuration achieves similar accuracy, but the paper does not provide a comparison with matched token consumption or constrained chain length, nor does it isolate the respective effects of content-arc and aspect capacities. Because spectral allocation and capacity constraints are central contributions, these gaps limit assessment of their independent value and necessity

2.Figure 2 compares G²-Reader’s configured node upper bounds with FLOWREADER’s measured distinct-node counts, which does not quantify actual cross-system cost savings. Moreover, relative to G²-Reader, the system changes the embedding model, visual-node representation, and reading prompts, so the overall 2.7-percentage-point improvement cannot be attributed solely to the routing method. Appendix D discloses that some evaluation examples were used for parameter checks and prompt development, and reports limited score changes and unchanged rankings after excluding these 250 examples. This is meaningful mitigating evidence, but complete results on the remaining examples and clarification of whether failure analysis was strictly confined to development examples are still needed. The manuscript also insufficiently specifies residual gain, bridge, ωuv\omega_{uv}, and the selector’s annealing and pruning settings, hindering independent reproduction. Together, these issues limit assessment of the sources of improvement, evaluation independence, and practical efficiency

Questions for authors

1.Could you compare the spectral-aspect and single-sink configurations with a matched termination scale κ\kappa, and report aspect coverage or allocation retention before and after path selection? This would help distinguish the contribution of spectral allocation from that of termination-cost calibration.

2.Could you isolate the effects of content-arc and aspect capacities, and provide a shortest-path comparison with constrained chain length or matched measured token consumption? The current results do not establish which capacities primarily produce the savings, or whether simpler cost controls can achieve a similar accuracy–cost trade-off.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4HH3 (1/3): guarantee after selection, facets, equivalence, and single sink at matched κ (W1, Q1)

We thank the reviewer for carefully assessing the guarantee and the ablations. We ran experiments for both questions and report the results below; all are in the revised paper, marked in blue.

How to read the results below. We use the evaluation sets and the ablation set released by G²-Reader. Accuracy is macro accuracy (the mean of the five subset accuracies), judged by GPT-4o-mini with G²-Reader's prompt. Comparisons are paired on the questions both systems answered, with an exact McNemar test. FlowReader appears in two forms: its reported configuration and a control with κ fixed at 1.3, against which we compare every aspect and capacity ablation.

W1: Does the guarantee survive chain selection, are the question's facets reached, and are spectral aspects equivalent to sub-questions?

After selection. We added a result that extends the guarantee from the routed flow to the chains that are read. Because source and content-to-aspect arcs have unit capacity, the flow decomposes into chains with flows f1≥⋯≥fKf_1\ge \cdots \ge f_K in (0,1], with ∑kfk=F\sum_k f_k=F and the chains ending at aspect a carrying dad_a.

Proposition 2 (retention under selection). Let F=F0F=F_0. For any selected set S of chains, let Ra(S)R_a(S) be the flow of the selected chains ending at aspect a. (i) The unmet allocation equals the flow of the chains not selected, ∑a(da−Ra(S))=∑k∉Sfk\sum_a (d_a-R_a(S))=\sum_{k\notin S} f_k. (ii) If S is the B chains of largest flow, the unmet allocation is at most (1 − B/K)F, and zero when B ≥ K. (iii) Every aspect with da>∑k∉Sfkd_a>\sum_{k\notin S} f_k receives at least one selected chain. With the 1% margin on the routed demand used in the paper, (i)–(iii) hold up to an additive 0.01F00.01F_0.

Facets. The guarantee is stated over spectral directions, so we tested directly whether they are reached as facets of the question. FlowReader reaches 83.3% of the secondary facets against 73.0% for relevance ranking, and reaches more facets than ranking on 140 queries against 7 (p = 3×10⁻³³; Appendix M). We also state the directions: ϕ1\phi_1 is the unit vector maximizing ∑vhv2cos⁡2(ev,ϕ)\sum_v h_v^2\cos^2(e_v,\phi) over the relevant content, da/Fd_a/F is the share of that energy along direction a, and the retained directions are the best rank-A approximation (Eckart–Young).

Equivalence. We agree that p = 0.35 shows only that no difference was detected, so we now test equivalence formally. Two one-sided tests on the same paired data from the evaluation sets give a 90% interval of [−0.47, +1.93] points for sub-questions minus spectral aspects. The two are equivalent within ±2 points (p = 0.04), so spectral aspects match LLM-generated sub-questions without a planning call. Section 4.3 now states the result in these terms.

Q1: Spectral aspects against a single sink at matched κ, and allocation retained after selection

We held κ at 1.3 and collapsed the aspect nodes into a single sink, with everything else unchanged. We ran an even split across the aspects as a second control.

Variant (κ = 1.3)SetSpectral aspectsVariantp
Single sinkablation69.269.61.0
Single sinkevaluation69.368.80.37
Even splitablation69.165.50.11
Even splitevaluation69.369.40.90

This separates the two contributions as the reviewer asked. The 44.8 in Table 3 reflects the effect of the fivefold change in κ, and the revised paper presents it only as a termination-calibration result. At matched κ, the single sink is equivalent in accuracy within ±2 points (paired 90% interval [−1.9, +0.5] on the evaluation sets). The aspects add coverage: with a single sink, the share of questions on which every worker finds nothing answerable rises from 6.7% to 8.3%, and the aspects carry the allocation guarantee of Proposition 2.

The spectral weighting is also the cheaper allocation. An even split reaches the same accuracy on the evaluation sets only with 11% more node visits and 5% more reader tokens.

On retention, as reported under W1: after selection, 90.0% of the allocation is met unit for unit, and aspects allotted at least 1 unit are read in 99.8% of cases and receive 98.5% of their allocation (per-aspect table in Appendix K). These results are in Section 4.3(d) and Appendix J of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 4HH3 (2/3): capacity isolation, comparison with G²-Reader, and evaluation independence (W2, Q2)

Q2: Which capacities produce the saving, and can a simpler cost control match it?

We lifted each family of capacities separately on the ablation set, with κ, the costs, the total flow, the number of chains, and the selector unchanged. We also ran the simpler control the reviewer suggests: shortest path over the same network with every chain capped at 2 nodes.

Variant (κ = 1.3)FlowReaderVariantChain lengthReader tokensEvery worker refusesp
FlowReader, all capacities69.2—1.3945.7k7.6%—
Aspect quotas lifted69.269.61.2943.9k7.6%1.0
Content capacities lifted69.069.42.1054.0k12.9%1.0
Shortest path, chains ≤ 2 nodes69.163.01.2243.6k—0.014

The saving comes from the unit content capacities. Lifting them keeps accuracy but lengthens chains from 1.39 to 2.10 nodes, raises reader tokens by 18%, and narrows the evidence from 17.3 to 13.6 distinct nodes. Lifting the aspect quotas changes neither accuracy nor cost; they guarantee allocation and facet coverage.

Capping chain length does not reproduce the trade-off. The capped shortest path matches FlowReader's token cost (43.6k against 45.7k) but loses 6.1 points (63.0 against 69.1, p = 0.014). The capacitated flow achieves the lower cost while maintaining accuracy. These results are in Section 4.3(b) and Appendix J of the revised paper.

W2: Figure 2 and the comparison with G²-Reader

We now measure both systems under the same conditions. We ran G²-Reader's released code, unmodified, on our content graphs, using the same embedder, visual-node representation, backbone, and judge; the only difference is the reasoning workflow.

SystemAccuracyDistinct nodesModel calls per questionReader tokensLatency (mean, median)
FlowReader71.017.118 (16 in parallel)45.9k20 s, 15 s
G²-Reader, planning graph62.58.48.5 (sequential)41.6k118 s, 70 s

FlowReader is 8.5 points ahead on the ablation set (38 vs 17 discordant, p = 0.006), reads twice as many distinct nodes at a similar token cost, and answers about 6 times faster. Figure 2 now shows G²-Reader's measured node count alongside the configured maximum.

W2: Evaluation independence

We follow G²-Reader's protocol: its released ablation set, data, and parses; the same backbone, judge, and decoding; and its evaluation prompt, byte for byte. We ran all our ablations, the constants sweep, and the prompt-wording sweep on the ablation set only.

Three earlier choices were made by comparing two or three alternatives on the full sets: the worker prompt family (+2.5 points on two subsets), the reasoner prompt (+0.6 macro points), and a refusal fallback added after an audit of one full-set run. The FetaTab example we disclosed came from that audit, and its effect is bounded by 0.2 macro points. Appendix D now lists these with their effect sizes.

Results on the questions outside the ablation set, as requested:

ConfigurationAll questionsOutside the ablation set
FlowReader (reported)68.668.3
FlowReader, κ = 1.369.369.7
Single sink, κ = 1.368.668.5
Even split, κ = 1.369.269.4
Centered SVD, κ = 1.368.368.5
MMR-1666.766.7

Each configuration moves by at most 0.35 points, and each comparison made in the paper keeps its direction or moves toward FlowReader: the single sink, the even split, and the centered SVD are each compared with the κ = 1.3 control, and MMR with the reported configuration. Per-subset results are in Appendix N.

AuthorsOpenReview ↗

Response to Reviewer 4HH3 (3/3): reproducibility (W2)

W2: Reproducibility

The revised paper now defines each of these, as implemented in the released code (Appendix A.2 for the costs, Appendix C for the selector). Here u, v are unit node embeddings, q is the query, and N is the number of nodes.

  • Residual gain. gain(u,v)=clip(cos⁡(v,q−(qTu)u),0,1)gain(u,v)=clip(\cos(v, q-(q^Tu)u), 0, 1): how much of the query not covered by u is explained by v. It is conditioned only on the immediate predecessor, which keeps arc costs additive.
  • Bridge. bridge(u,v)=∑t∈Tu∩Tvidf(t)bridge(u,v)=\sum_{t\in T_u\cap T_v} idf(t), normalized by its maximum over arcs. TvT_v are the keyword tokens of v (content tokens if it has none), idf(t)=log⁡((N+1)/(dft+0.5))idf(t)=\log((N+1)/(df_t+0.5)), and only tokens with dft≤⌈N⌉df_t \le \lceil\sqrt{N}\rceil count.
  • Edge score. puv=min⁡(1,0.7⋅gain+0.3⋅bridge+0.2⋅muv)p_{uv}=\min(1, 0.7\cdot gain+0.3\cdot bridge+0.2\cdot m_{uv}), where muv=1m_{uv}=1 if the two nodes differ in modality and 0 otherwise. This is Eq. 9 with β = 0.30 and γ = 0.20.
  • Agreement. ωuv=clip(cos⁡(u,v),0,1)\omega_{uv}=clip(\cos(u,v), 0, 1) divided by its relevance-weighted mean over arcs. The traversal cost is then wuv=(1−puv)−Δ−(ωuvhv−hˉ)w_{uv}=(1-p_{uv})-\Delta-(\omega_{uv}h_v-\bar{h}), as in Eq. 2.
  • Replicator selector. Chain quality is Qi=1∣πi∣∑v∈πihv2Q_i=\sqrt{\frac{1}{|\pi_i|}\sum_{v\in \pi_i} h_v^2}, with payoff Mii=Qi2M_{ii}=Q_i^2 and Mij=QiQj(1−Rij)M_{ij}=\sqrt{Q_iQ_j}(1-R_{ij}), where RijR_{ij} is the Jaccard overlap of the node sets. Shares start at xi∝Qix_i\propto Q_i and are updated by x←x⊙(Mx)/(xTMx)x\leftarrow x\odot (Mx)/(x^TMx).
  • Annealing and pruning. After each update, the shares are blended towards the uniform distribution over the n surviving chains with weight ϵt=0.2⋅0.9t\epsilon_t=0.2\cdot 0.9^t. The lift of a chain is its share above the blend's uniform level, ℓi=max⁡(0,xi−ϵt/n)\ell_i = \max(0, x_i - \epsilon_t/n). A chain is removed when ℓi<max⁡(5×10−4,0.05max⁡jℓj)\ell_i < \max(5 \times 10^{-4}, 0.05 \max_j \ell_j). Iteration stops at convergence (190–305 iterations in practice) or after 1,000.
Reviewer_DosqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an exact minimum-cost flow formulation for selecting query-relevant evidence from multimodal knowledge bases in retrieval-augmented generation (RAG). By casting evidence selection as a resource-allocation problem, the method achieves accuracy comparable to approaches that use an LLM to decompose a query into sub-tasks, while allocating the retrieval budget more efficiently and reducing retrieval costs. The authors evaluate FlowReader on all five subsets of VisDoMBench. It outperforms the strongest prior system, G²-Reader, in overall accuracy and maintains strong performance under a reduced retrieval budget, with only a modest decline relative to its full-budget setting. These results demonstrate the effectiveness and efficiency of FlowReader for multimodal evidence selection.

Critical strengths
  1. Novelty of the framework. The paper proposes a capacitated flow network algorithm that automatically allocates the retrieval budget under budget constraints, and then selects and filters evidence chains via minimum-cost flow. This effectively conserves retrieval resources and ensures that, when retrieval is effective, the useful information retrieved is maximized to support downstream tasks. This directly addresses a key weakness of prior RAG systems: expensive and unstable LLM-based decomposition and iterative retrieval. The formulation is theoretically clean, practically efficient, and removes a major source of latency and cost.

  2. An efficient search strategy and a concise, well-structured methodology framework. In FlowReader, routing and path selection are completed before any evidence is read. Consequently, all selected evidence chains can be dispatched to the vision-language model (VLM) for reading in a single parallel pass, entirely eliminating the iterative dependency of conventional approaches, in which the system reads, judges sufficiency, re-plans, and reads again.

Critical weaknesses

1. SVD-Allocated Search Budget: More Spectral Coverage, Less Semantic Coverage

In the method proposed in the paper, the singular value decomposition (SVD) step extracts the directions of variation within the relevance-weighted embeddings. However, the information encoded in the feature vectors captured along these directions may not be precisely and exclusively aligned with the semantic scope covered by the user's query; that is, it may fail to fully correspond to the semantics the user actually needs to obtain.

This limitation becomes particularly pronounced in complex tasks, or in cases where the required evidence is sparse and can only be surfaced through precise retrieval: there is no guarantee that the ultimately retrieved information stream can completely cover the specific information demanded by the user.

From this perspective, whether the proposed method remains effective in real-world scenarios that demand precise retrieval warrants further verification and clarification.

2. Insufficient diversity in the types of tested models and validation datasets

The paper's evaluation is restricted to VisDoMBench, which comprises only academic documents (tables, charts, and slides), and is therefore limited in both scale and diversity. In real-world deployment, documents frequently involve noisy web pages, multilingual content, complex commercial contracts, and a wide range of domain-specific data formats.

Moreover, the evaluation pipeline relies on a single VLM backbone (Qwen3-VL-32B), a single embedding model, and a single LLM-based judge. The set of models validated experimentally is thus rather narrow, and the absence of human-calibrated evaluation leaves the reliability of the results without stronger empirical validation and support.

Questions for authors
  1. Could the authors provide quantitative or qualitative evidence demonstrating that the spectral directions extracted by SVD indeed correspond to human-interpretable query intents? If such spectral directions are decoupled from the underlying semantic facets, how can the quality of answers to complex questions be assured?

  2. The paper acknowledges that the finally selected subset is not guaranteed to preserve the theoretically allocated budget. In the actual evaluation, how large is the deviation between the facet distribution of the 16 chains ultimately selected and the theoretical capacity allocation dad_a? Could the authors provide quantitative data on this?

  3. The paper employs GPT-4o-mini as the final judge and reports that "judge agreement across repeated scoring is 96–99%." However, GPT-4o itself may exhibit systematic preferences for particular layouts, linguistic styles, or response lengths, and this form of LLM bias is not discussed in the paper.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Dosq (1/2): spectral directions, allocation, and evaluation breadth (W1, W2, Q1, Q2)

We thank the reviewer for the positive assessment. We answer the three questions and the two weaknesses below; the new results are in the revised paper, marked in blue.

How to read the results below. We use the evaluation sets and the ablation set released by G²-Reader. Accuracy is macro accuracy (the mean of the five subset accuracies), judged by GPT-4o-mini with G²-Reader's prompt. We compare paired questions answered by both systems using an exact McNemar test.

W1 and Q1: Do the spectral directions correspond to what the question asks for?

We checked this in three ways.

  • Examples. We inspected the evidence each direction gathers for 8 queries, selected automatically as the questions on which the most directions received a chain (4 are shown in the new Appendix L).
  • Coverage. FlowReader reads at least one aligned node for 83.3% of the secondary facets, against 73.0% for relevance ranking and 80.9% for MMR. On the evaluation sets, it reaches more facets than ranking on 140 questions against 7 (p = 3×10⁻³³). It does so with almost no loss in relevance (0.683 against 0.697 for top-16).
  • Against LLM-written sub-questions. Replacing the spectral aspects with sub-questions generated by the backbone gives equivalent accuracy within ±2 points (p = 0.04), so the directions do the job of explicit, human-readable intents without a planning call.

On precise retrieval: on the evaluation sets, FlowReader's final answer is a refusal on 1.3% of questions against 12.6% for top-16 selection, and every worker comes back empty on 7.0% of questions against 7.8–11.4% for ranking and MMR.

On how answer quality is assured if a direction is off: accuracy does not rest on the directions alone. The query-conditioned costs select the evidence, and with the whole budget on the leading direction accuracy is unchanged (69.6 against 69.2, p = 1.0). The aspects add coverage and a guaranteed allocation.

Q2: How far do the 16 chains read deviate from the allocation dₐ?

Very little. After selection, 90.0% of the allocation is met unit for unit. Aspects allotted at least 1 unit are read in 99.8% of cases and receive 98.5% of their allocation. Almost all aspects left without a chain (99.9%) are directions allotted less than 1 unit, which is too little to fill a chain. We also added Proposition 2 (Section 3.4), which bounds this loss; the per-aspect table is in Appendix K.

W2: Diversity of benchmarks and models

We added a second benchmark, MMLongBench-Doc. Its documents go well beyond academic material: research and financial reports, brochures, guidebooks, administrative and industry files, and tutorials, each tens of pages long. With the same configuration and no retuning, FlowReader reaches 47.2 against 45.4 for the 16 most relevant nodes (130 vs 111 discordant; Appendix N).

On models: we added a second judge (see Q3), and our embedder is an open 768-dimensional model where the strongest baseline uses a proprietary 1,536-dimensional one. We keep the backbone of the reported baselines, Qwen3-VL-32B-Instruct, so that every system in Table 1 is compared on the same reader.

AuthorsOpenReview ↗

Response to Reviewer Dosq (2/2): judge bias (Q3)

Q3: Could the judge be biased towards certain styles or answer lengths?

Every number in the paper, ours and the baselines', is scored by the same judge with the same prompt, and the judge sees only the question, the gold answer, and the prediction. We ran two checks.

A second judge. We re-scored five systems on the ablation set with GPT-4o and the same prompt. Each cell is FlowReader / the other system:

FlowReader againstGPT-4o-miniGPT-4o
G²-Reader, planning graph71.0 / 62.563.3 / 58.5
G²-Reader, no planning70.6 / 65.363.3 / 60.0
Top-k71.0 / 66.563.3 / 58.9
MMR71.0 / 66.963.3 / 61.7

The two judges agree on 85–92% of answers (Cohen's κ of 0.68–0.83). GPT-4o is stricter for every system, and FlowReader stays ahead of every system under both judges.

Answer length. FlowReader's answers are longer on average (46 words against 22–42), so we checked whether the judge rewards length. It does not: within FlowReader, length is not associated with being judged correct (r = −0.07), and its longest answers score lower than its mid-length ones (67.9 against 78.3).

Both checks are in Appendix D and Table 13.

Dear Reviewer,

As the discussion period draws to a close, we would appreciate your feedback on whether our responses and additional results have addressed your concerns and whether any points still need clarification.

Best Regards Authors

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

The paper proposes FlowReader, an evidence selection method for question answering over long multimodal documents that formulates the selection as a single minimum-cost flow problem. The authors argue that ranking retrieval units independently under a fixed budget tends to retrieve near-duplicates of the facet most similar to the query, while planner-based systems such as G2G^2-Reader do not guarantee an optimal allocation of the budget across sub-questions.

FlowReader represents the documents as a graph whose nodes are document elements (passages, tables, figures, and slides) and whose edges connect related elements. Given a query, it applies SVD to the embeddings of the elements relevant to the query and treats the main directions as latent query aspects. It divides the evidence budget among the query aspects in proportion to how much of the relevant content each direction explains, as measured by its squared singular value. It then solves a minimum-cost flow problem on this graph that selects short paths of relevant and mutually consistent elements, called chains, as evidence, under the constraint that each query aspect receives its allotted share. A subset of these chains is read in parallel by a vision-language model without a planning step, and a reasoner combines the findings into an answer.

On VisDoMBench, FlowReader achieves a macro accuracy of 68.9, which is 2.7 points higher than that of G2G^2-Reader, while reading 17.5 document elements per query on average. Based on ablations, the authors conclude that the cost design drives the accuracy gains, while the capacity constraints preserve the accuracy with 73% of the reader tokens required by shortest-path routing without them.

Critical strengths

S1. The paper tackles the practical problem of evidence selection under a limited reading budget with a single minimum-cost flow formulation that requires no LLM planning calls.

S2. In the controlled ablations, the flow-based selection achieves higher accuracy than top-16 and MMR selection within the same pipeline.

Critical weaknesses

W1. The connection between the proposed mechanism and the motivating failures is not sufficiently validated.

W1.1 The motivating failures of redundant retrieval and missing necessary evidence (Section 1) are not directly evaluated. The analyses of repeated node reads, refusal rates, and "Reached" (Section 4.3 and Appendices G and H) provide useful but indirect evidence. Fewer repeated reads do not imply lower semantic redundancy among distinct nodes relative to top-kk, MMR, or G2G^2-Reader, and "Reached" measures coverage of the system's own top-16 nodes rather than recall of supporting evidence. The authors should compare the semantic redundancy of the evidence selected by FlowReader, top-kk, MMR, and G2G^2-Reader at similar budgets. Where the benchmark or its source datasets annotate supporting evidence, the authors should also report how much of it each system retrieves.

W1.2 As the authors note in Sections 3.3 and 3.4, the coverage guarantee concerns spectral directions rather than semantic facets, and the selected chains are not guaranteed to preserve the allocation. The paper does not report how much of the allocation remains in the 16 chains that are actually read. The authors should report the number of aspects represented and the flow retained for each aspect before and after chain selection.

W1.3 The spectral directions are not shown to correspond to distinct query aspects. The leading direction receives a median of 73.7% of the budget, while the remaining budget is distributed across a median of 13 directions (Appendix A.4). Since the matrix of relevance-weighted embeddings appears to be built without centering (Section 3.2), the leading direction may correspond to a component shared by all relevant content rather than to a distinct facet. Section 4.3 compares the proportional split of the budget with an equal split and replaces the aspects with LLM-generated sub-questions, but neither experiment shows that the directions correspond to distinct aspects. The authors should provide examples of the evidence assigned to each aspect and compare FlowReader with a variant that gives the whole budget to the leading direction while keeping κ\kappa and the rest of the configuration unchanged.

W2. The added value of the proposed mechanism over budget-matched baselines and G2G^2-Reader is not fully isolated.

W2.1 Top-kk and MMR are evaluated only at 16 nodes (Table 3), whereas FlowReader is evaluated at three average budgets (Table 2). FlowReader at an average of 12.9 nodes (Table 2) still exceeds top-16 (Table 3) on all five subsets, but the reader-token cost of top-kk and MMR is not reported (Table 5). Without accuracy--budget curves for these baselines, the efficiency advantage of FlowReader over simple alternatives cannot be fully assessed. The authors should report the accuracy of top-kk and MMR at average node or reader-token budgets comparable to those in Table 2.

W2.2 The shortest-path comparison removes all capacity constraints at once, so it does not isolate the effect of the aspect capacities (Section 4.3(b)). The paper attributes the reduction in reader tokens to the capacity constraints as a whole, so it remains unclear how much of this reduction comes from the aspect capacities rather than from the other capacities. The single-sink ablation does not isolate the aspect capacities either, since it also merges the aspect terminals and, as the authors note, changes κ\kappa approximately fivefold. The authors should report a variant that relaxes only the aspect-to-sink capacities while keeping the other capacities, the costs, κ\kappa, the total flow, the number of chains read, and the selection procedure unchanged.

W2.3 The comparison with G2G^2-Reader does not isolate routing. FlowReader differs from G2G^2-Reader in the embedder and the visual-node prompt as well as in its workflow (Sections 3.1, 3.4, and 4.1), and the baseline results are transcribed from prior work (Table 1). Within the FlowReader pipeline, routing improves over top-16 selection (68.6 versus 65.7, Table 3), but the 2.7-point gap compares complete systems and does not show that routing is more effective than the planning of G2G^2-Reader. In addition, Figure 2 compares the measured node count of FlowReader with the configured maximum of G2G^2-Reader, although the latency of sequential planning is part of the motivation. The authors should compare the two systems with the same embedder and visual-node prompt, report the measured calls, tokens, and latency of both, and state the differences that remain.

Questions for authors

Q1. At similar budgets, how do FlowReader, top-kk, MMR, and G2G^2-Reader compare in the semantic redundancy of the selected evidence and in how much of the annotated supporting evidence they retrieve? How do these differences relate to accuracy? (Ref W1.1)

Q2. After chain selection, how many of the retained aspects are represented, by how many chains, and what fraction of each aspect's flow remains? What kind of evidence does the leading direction typically select? (Ref W1.2 and W1.3)

Q3. How does the accuracy of top-kk and MMR compare with that of FlowReader at comparable average node or reader-token budgets, such as those in Table 2? (Ref W2.1)

Q4. When only the aspect-to-sink capacities are relaxed and the rest of the configuration is kept unchanged, how do the accuracy and the reader tokens change? (Ref W2.2)

Q5. What are the measured node count, calls, tokens, and latency of G2G^2-Reader per query, and how does the accuracy gap change when both systems use the same embedder and visual-node prompt? (Ref W2.3)

Q6. Appendix D reports that excluding the 250 ablation examples changes each score by at most 0.44 points and preserves the rankings. Were the remaining examples also excluded from every development decision, including the failure analysis used for prompt refinement?

Q7. Appendix I states that alternative values of the termination balance bb reduce accuracy by 2.3–3.1 points, whereas Table 8 reports a reduction of 0.4 points for b=0.55b=0.55. How was 0.634 initially chosen as the reference value of this local sweep, and which statement is correct?

Q8. At the reported operating point, how are the total flow, the number of candidate paths after flow decomposition, the number of chains read, the number of distinct nodes, and the number of reader calls related, for example for a single query?

10 comments
AuthorsOpenReview ↗

Response to Reviewer etEa (1/5): redundancy and facet coverage (W1.1, Q1)

We thank the reviewer for the detailed and constructive review. We ran every requested experiment and report the results below; all are included in the revised paper and marked in blue.

How to read the results below. We use the evaluation sets and the ablation set released by G²-Reader. Accuracy is macro accuracy (the mean of the five subset accuracies), judged by GPT-4o-mini with G²-Reader's prompt. Comparisons are paired on the questions both systems answered, with an exact McNemar test. FlowReader appears in two forms: its reported configuration and a control with κ fixed at 1.3, against which every aspect and capacity ablation is compared.

W1.1 and Q1: Is the selected evidence less redundant, and does it cover more of what the question needs?

We measured both directly, as suggested. All selectors use the same evidence budget on the full evaluation sets. G²-Reader is measured on the ablation set, as in the matched comparison (see W2.3), with the nodes it read recovered from its retrieval logs; all 7,116 retrieved units were matched to nodes in our graphs.

Redundancy. We measure it between the evidence units the reader receives, i.e., between chains (for ranking and MMR, each unit is a single node). Nodes within a chain are meant to agree, so we report that separately.

SelectorNodesNear-duplicate pairs between chains (cos ≥ 0.9)Cosine between chainsCosine within chains
FlowReader17.515.3%0.7700.871
Top-16 by relevance16.016.5%0.786—
Greedy chains17.715.2%0.7690.872
MMR-1616.04.5%0.684—
G²-Reader, planning graph (ablation set)8.424.8%0.811—

FlowReader's evidence is less redundant than relevance ranking's and much less redundant than G²-Reader's: on the same ablation-set questions, G²-Reader has 7.0 points more near-duplicate pairs and a higher overall cosine (+0.032), even though it reads half as many nodes. Within a chain, FlowReader's nodes are closely related (cosine 0.871, against 0.770 between chains), which is what the agreement term is designed to produce.

On how this relates to accuracy: MMR removes the most duplicates but is less accurate than FlowReader (66.7 against 68.6 on the evaluation sets; 131 vs 167 discordant, p = 0.04), because it gives up relevance to do so. G²-Reader has the most redundant evidence and the lowest accuracy in the matched comparison (62.5 against 71.0). Removing duplicates is therefore not enough; FlowReader lowers redundancy while keeping the evidence relevant.

Coverage. To test whether ranking concentrates on the leading facet and misses the others, we reconstructed each query's spectral aspects on the ablation set and checked what every selector actually read against them. A secondary facet is any direction after the first that is allotted at least 1 unit of flow.

SelectorSecondary facets reached (best node aligned ≥ 0.3)Mean best alignment per secondary facetNodes read from outside the top-16Mean relevance of nodes read
FlowReader83.3%0.39922.7%0.683
Top-16 by relevance73.0%0.3670%0.697
MMR-1680.9%0.38642.8%0.657
Greedy chains83.3%0.39923.4%0.681

Relevance ranking leaves 27.0% of secondary facets without any aligned node, against 16.7% for FlowReader. This is the coverage failure described in Section 1. FlowReader reaches these facets by reading 22.7% of its nodes from beyond the top 16, with almost no loss in relevance (0.683 against 0.697); MMR goes much further outside (42.8%) and loses more relevance for less coverage.

The same holds on the full evaluation sets, compared query by query with a sign test:

FlowReader againstQueries where FlowReader reaches more facets / fewerpQueries where FlowReader aligns better / worsep
Top-16 by relevance140 / 73×10⁻³³359 / 1212×10⁻²⁸
MMR-1663 / 308×10⁻⁴714 / 1322.5×10⁻⁹⁷
AuthorsOpenReview ↗

Response to Reviewer etEa (2/5): allocation after selection and spectral directions (W1.2, W1.3, Q2)

All selectors put a similar share of their alignment on the leading direction (79.5–80.8%). The difference between them is therefore whether they reach the secondary facets at all, not how the evidence is spread overall.

The released benchmark files mark only the gold document for each question, not the supporting evidence inside it, so recall of annotated evidence cannot be computed. As the closest available measure, we report how often every worker finds nothing answerable in what it was given: 7.0% of questions for FlowReader, against 7.8% for top-16, 8.3% for top-13 and 11.4% for MMR. The redundancy and facet tables are in Appendix M of the revised paper, and the refusal rates are in Appendix G.

W1.2 and Q2: How much of the allocation survives chain selection?

Most of it: 90.0% of the allocated flow remains in the chains that are read, and almost every aspect allotted a full unit is read.

We also made the guarantee formal. The new Proposition 2 (Section 3.4) extends it from the routed flow to the selected chains: the allocation lost in selection equals the flow of the chains left out; it is at most (1 − B/K)F when the B chains of largest flow are kept, and an aspect can go unread only if its allocation is no larger than the flow left out.

Per aspect, on the ablation set (κ = 1.3), using the aspect at which each selected chain ends:

Aspects, by rankAspects per queryShare of the allocationRead by ≥ 1 chainChains read per aspectAllotment met by the chains read
Leading direction1.0074.1%100.0%12.1299.3%
2nd–4th directions2.9915.9%94.4%1.0492.2%
5th and later (all < 1 unit)7.9810.0%9.3%0.0917.1%
All aspects allotted ≥ 1 unit1.8781.1%99.8%7.1098.5%
All aspects allotted < 1 unit10.1118.9%26.8%0.2753.4%

Aspects allotted at least 1 unit are read in 99.8% of cases and keep 98.5% of their allocation. What is lost falls almost entirely on directions allotted less than 1 unit, which is too little to fill a whole chain: these make up 99.9% of the aspects left without a chain. The result does not depend on the spectral weighting, since an even split retains 91.3%. The full table is in Appendix K of the revised paper.

W1.3: Do the directions correspond to distinct aspects, and does the uncentered SVD matter?

We ran the variant requested: the whole budget on the leading direction, with κ held at 1.3 and everything else unchanged. We also ran three related controls.

Variant (κ = 1.3)SetSpectral aspectsVariantChain lengthReader tokensp
Leading direction only (single sink)ablation69.269.61.2843.9k1.0
Even split across aspectsablation69.165.51.5247.5k0.11
Directions < 1 unit droppedablation69.169.91.3444.8k0.84
Leading direction only (single sink)evaluation69.368.81.3544.0k0.37
Even split across aspectsevaluation69.369.41.6248.4k0.90
Centered SVDevaluation69.468.51.5347.2k0.15

For reference, the spectral aspects have chain length 1.46 and 46.1k reader tokens on the evaluation sets.

What the directions hold. We inspected the evidence assigned to each direction for 8 queries, selected automatically as the questions on which the most directions received a chain; 4 are shown in the new Appendix L. The leading direction gathers the content on the question's main theme. In 7 of the 8 queries, one of the 2nd–4th directions, with 5–9% of the demand, isolates a distinct part of the question's own document, for example the numerical study behind the figure a question asks about, or the sensitivity analysis a timing question depends on. A later direction can also land on one of the distractor documents, but the spectral weighting gives it only 3–9% of the demand, against 69–76% for the leading direction. The coverage results in W1.1 give the quantitative side: FlowReader reaches more secondary facets than ranking on 140 queries against 7.

AuthorsOpenReview ↗

Response to Reviewer etEa (3/5): aspect controls and matched budgets (W1.3, W2.1, Q3)

Leading direction only. Accuracy is equivalent within ±2 points (68.8 against 69.3 on the evaluation sets; paired 90% interval [−1.9, +0.5]). The aspect allocation therefore costs no accuracy, and what it adds is coverage: with a single sink, the share of questions on which every worker finds nothing answerable rises from 6.7% to 8.3%.

Spectral against even weighting. An even split reaches the same accuracy on the evaluation sets only by reading more: longer chains (1.62 against 1.46 nodes), 11% more node visits and 5% more reader tokens. Weighting by σ² is the cheaper allocation.

Centering. As the reviewer notes, the uncentered leading direction is the component shared by the relevant content. This is intended, since that component is the question's main theme. Removing it with a centered SVD does not help: 68.5 against 69.4.

The tail. Dropping the directions allotted less than 1 unit and redistributing their share leaves accuracy unchanged (69.9 against 69.1) at 2% fewer reader tokens, so the method does not rely on the small directions. The revised Section 4.3 and Appendix J state each component's role.

W2.1 and Q3: How do top-k and MMR compare with FlowReader at matched budgets?

We ran top-k, MMR, and FlowReader at the three budgets of Table 2, on the ablation set:

Nodes readFlowReaderTop-kMMRFlowReader vs top-k (discordant, p)Reader tokens (FlowReader / top-k / MMR)
963.762.958.926 vs 24, p = 0.8922.6k / 7.3k / 22.3k
1365.764.966.126 vs 24, p = 0.8934.3k / 9.8k / 31.0k
16–17 (reported)71.066.566.929 vs 18, p = 0.1445.9k / 11.6k / —

FlowReader is ahead of top-k at every budget, and the margin grows with the budget: +0.8 at 9 and 13 nodes, +4.5 at the reported setting. This is what the method predicts. At small budgets there is little to divide, and the leading direction takes a median 95% and 88% of the demand at 9 and 13 nodes; at the reported budget it takes 74%, leaving room for the secondary facets. On the full evaluation sets at the reported budget, FlowReader reaches 68.6 against 65.7 for top-16 (Table 3).

Against MMR, FlowReader leads at 9 nodes and at the reported budget, and the two are level at 13.

On tokens: FlowReader uses more than top-k because each chain is read in a separate call. Compared with other systems that read in several calls, it is the cheaper one: 20 s per question against 118 s for G²-Reader (W2.3), and 27% fewer reader tokens than shortest-path routing without capacities (Section 4.3(b)).

The reported demand was fixed before any sweep and is the same on all five subsets and on MMLongBench-Doc. The table is in Appendix N of the revised paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the three reviewers for their careful and constructive reviews. We ran every experiment requested, and the revised paper includes all of them, with changes marked in blue. The reviews share four themes; below we give the main result for each and where it is in the revision. Details are in the individual responses.

1. The guarantee after chain selection, and whether the question's facets are reached (etEa, 4HH3, Dosq)

  • A new Proposition 2 extends the allocation guarantee from the routed flow to the chains that are read, with an explicit bound (Section 3.4; proof in Appendix B).
  • Measured after selection, 90.0% of the allocation is met unit for unit, and aspects allotted at least 1 unit are read in 99.8% of cases (Appendix K).
  • FlowReader reaches 83.3% of the secondary facets, against 73.0% for relevance ranking and 80.9% for MMR, with less redundant evidence than ranking and G²-Reader (Appendix M). Examples of what each direction gathers are in Appendix L.
  • Spectral aspects are equivalent to LLM-generated sub-questions within ±2 points, now tested formally (p = 0.04; Section 4.3).

2. What the aspects and the capacities each contribute (etEa, 4HH3)

  • With κ held fixed, a single sink is equivalent in accuracy to the spectral aspects. The 44.8 row of Table 3 is therefore presented as a termination-calibration result (Section 4.3(d), Appendix J).
  • Lifting each family of capacities separately shows that the reader-token saving comes from the unit content capacities with the calibrated termination cost, while the aspect quotas provide the guarantee and the facet coverage (Section 4.3(b), Appendix J).
  • Shortest path with chains capped at 2 nodes matches FlowReader's cost but not its accuracy (63.0 against 69.1, p = 0.014).
  • Contributions (1)–(2) and the conclusion are restated in these terms.

3. Matched comparisons with G²-Reader and with simple baselines (etEa, 4HH3)

  • G²-Reader's released code, run unmodified on our graphs with the same embedder and backbone, reaches 62.5 against 71.0 for FlowReader (p = 0.006), and FlowReader answers about 6 times faster. Figure 2 now shows G²-Reader's measured node count (Section 4.2).
  • Top-k and MMR are evaluated at 9, 13 and 16 nodes; FlowReader is ahead of top-k at every budget (Appendix N).

4. Breadth and independence of the evaluation (etEa, 4HH3, Dosq)

  • A second benchmark, MMLongBench-Doc, with long non-academic documents: 47.2 against 45.4 for the 16 most relevant nodes, with no retuning (Section 4.2, Appendix N).
  • A second judge, GPT-4o: FlowReader stays ahead of every system under both judges, and answer length is not associated with being judged correct (Appendix D).
  • The choices made on the full sets are listed with their effect sizes (Appendix D), and results outside the ablation set are reported per subset; every configuration moves by at most 0.35 points (Appendix N).

Clarifications and corrections

  • Appendix I: the sentence on the termination balance b is corrected, and the origin of 0.634 is stated.
  • Section 4.1: per-query accounting of flow, candidate chains, chains read, nodes, visits, and calls.
  • Appendices A.2 and C: definitions of the residual gain, the bridge term, ωuv\omega_{uv}, and the selector's annealing and pruning settings.
AuthorsOpenReview ↗

Summary of the discussion phase

We appreciate the detailed and helpful discussion from the reviewers. Below is a brief summary:

We added new experiments based on the reviewers' feedback. These updates are included in the revised paper and highlighted in blue.

  1. We directly compared FlowReader to G²-Reader using its original code, running both on the same graphs, embedder, backbone, and judge. FlowReader scored 71.0 versus 62.5 (p = 0.006), with similar token cost and about six times faster speed.
  2. We included controls with matched κ (single sink, even split) and ran separate ablation studies for content capacities and aspect quotas.
  3. Proposition 2 extends the allocation guarantee to the chains selected for reading, and we measured retention in these cases.
  4. We measured facet coverage directly, ran a formal equivalence test with LLM-generated sub-questions, included results beyond the ablation sample, added a second judge, used another benchmark (MMLongBench-Doc), and made budget-matched comparisons with reader-call counts.

Where the reviewers stand

  • Reviewer 4HH3 stated that the new controls "substantially address" the main concerns about attribution and cost, and raised their rating to weak accept. We made the two requested wording changes.
  • Reviewer Dosq also rated the paper as weak accept. The questions raised were addressed with the results listed above.
  • Reviewer etEa found the extra experiments helpful and noted that several concerns were addressed, but kept the original rating. Two points remain: aspect quotas improve facet coverage without affecting accuracy or cost, and FlowReader uses more reader tokens and calls than top-k.

Our position on the two remaining points

We agree with both points, and the revised paper acknowledges them (see Section 4.3(b) and Appendix J for quotas; Section 4.2 and Appendix N for cost). The quotas provide coverage (84.1% versus 70.8%, p = 8×10⁻⁹) and an allocation guarantee at no extra cost, without a planning call. For cost, FlowReader's advantage over top-k is accuracy. Compared to the planner-based state of the art, it is more accurate at a similar token cost and much faster.

The discussion helped us define the paper's claims more precisely, and we appreciate it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.