acceptodds
Under review as a conference paper at ICLR 2027

When is seeing enough? Cross-granularity evidence search with Sufficiency verification for High-resolution understanding

Abstract

Although multimodal large language models (MLLMs) have substantially advanced visual understanding, resolving fine-grained details in high-resolution images remains challenging under limited visual input budgets. Existing approaches recover such details through local cropping and iterative region search, but the relevance of individual regions does not establish whether the accumulated evidence is sufficient to support the correct answer. This raises two coupled challenges: acquiring complementary evidence across granularities and determining when the accumulated evidence is sufficient to support an answer. To address these challenges, we propose a novel framework named Cross-Granularity Evidence Search with Sufficiency Verification (CAREER) for high-resolution visual question answering. The core idea is to couple complementary evidence acquisition with explicit sufficiency verification to guide continued observation and answer acceptance. Specifically, we first bootstrap a low-cost global-to-local evidence state and construct a complementary multi-scale candidate atlas that exposes potentially informative observations. Building on this atlas, we adaptively refine or expand views and explore alternative regions according to unresolved evidence requirements, accumulating observations to retain both local details and broader context. To assess evidence sufficiency, we introduce an option-wise verifier that evaluates visual support without access to the generator’s provisional answer, and then incorporate its sufficiency verification to guide further observation and adaptive stopping. Additionally, experiments across three MLLM backbones and three high-resolution QA settings show that CAREER consistently outperforms state-of-the-art visual-search baselines, while also reducing object hallucination on POPE benchmark. The code is available at https://github.com/CAREER-ICLR2027/CAREER.git.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
High-resolution visual question answering, adaptive visual search, multimodal large language models, evidence accumulation, answer verification, object hallucination

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_nr8VOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

CAREER is a training-free framework for high-resolution visual question answering that couples multi-granularity evidence acquisition with sufficiency verification. The key insight is that individually relevant image regions do not guarantee the accumulated evidence is sufficient to distinguish among candidate answers.

The framework proceeds as follows. It first screens the global image and returns immediately if confidence and inter-option margin both exceed thresholds. Otherwise, it constructs a multi-scale candidate atlas combining SAM-based localization with a hierarchical SLIC region tree, ranked by query relevance and visual complexity. An evidence bundle retains all acquired observations for joint reasoning. An independent option-wise verifier then evaluates each answer option against the accumulated bundle, assigning Support, Refute, or Insufficient, without seeing the generator's provisional answer. The controller accepts when absolute support and normalized inter-option gap both pass thresholds, triggers backtracking when progress stalls, and returns the saved global answer as a flagged fallback when the budget is exhausted.

Experiments across three backbones on V*Bench, HR-Bench 4K/8K, and POPE-COCO show consistent accuracy improvements over visual search baselines, with lower latency and memory than CVSearch. Ablations confirm that evidence accumulation, adaptive navigation, and the two-threshold gate each contribute meaningfully to performance.

Critical strengths

S1: Consistent Gains Across Backbones and Settings

CAREER achieves the highest overall accuracy across all nine backbone–benchmark combinations in Table 1. The ablation drops for evidence retention and adaptive navigation are clear and consistent. Results hold across three generators and three high-resolution settings, which provides reasonable confidence that the gains are not cherry-picked.

S2: Accuracy and Computation Are Analyzed Together

Table S6 reports accuracy, mean observations, visual tokens, and latency at each observation budget, comparing CAREER against CVSearch and CVSearch with the same verifier. Across all four nonzero budgets, CAREER achieves higher accuracy and lower latency than both baselines. This joint analysis is more informative than accuracy-only comparisons and makes the efficiency claim concrete.

Critical weaknesses

W1: Limited Novelty Relative to Prior Work

The main contribution of CAREER is the closed-loop coupling of evidence accumulation, option-wise verification, adaptive navigation, and stopping decisions. While this system-level integration is meaningful, the individual components largely have precedents. Local search, hierarchical region partitioning, and visual-complexity-driven candidate ranking are already present in CVSearch, which also employs global screening, expert-assisted localization, semantic region decomposition, and complexity-driven traversal. The genuine incremental value therefore narrows to how the option-wise verifier interacts with the continuously retained evidence bundle — not a fundamentally new visual search paradigm. The paper would be stronger if it more precisely isolated this verifier-evidence coupling as the core novelty, rather than presenting the full pipeline as a novel framework.

W2: Restriction to Closed-Ended QA and Limited Verifier Generalization

The framework requires predefined answer options and calls the verifier once per option at each step. How CAREER generalizes to open-ended or free-form settings remains unclear. The normalized support share ρ t , i ρ t,i ​ and the acceptance gap are both sensitive to the number of options and to how options are worded, yet these sensitivities are not analyzed. In addition, all verifier comparisons are restricted to the Qwen model family. It is therefore unclear whether the reported gains reflect general properties of the verification mechanism or are specific to this family.

W3: Sufficiency Verification Is Not Directly Grounded in Visual Evidence

The acceptance rule is built on the verifier's Support score and the inter-option gap. A verifier can, however, assign high-confidence Support due to language priors, option wording, or its own visual misinterpretation, without the relevant visual evidence being present in the bundle. The reported accuracy and accepted error rates show that the mechanism is useful in aggregate, but they do not directly show that the verifier detects whether critical visual evidence is actually present. The most informative validation would be targeted intervention experiments: occlude the key object, remove necessary context, or substitute the relevant crop, and check whether the verifier transitions from Support to Insufficient and whether search resumes. A comparison against removing irrelevant regions would further isolate whether the verifier responds to genuinely task-critical content.

W4: Fallback Strategy Is Underexplored

When the observation budget is exhausted, CAREER returns the saved global answer as an unverified fallback. The paper does not justify this choice or ablate alternatives. It is not clear why the global answer is preferred over using the accumulated evidence state at termination, which may contain more relevant information than the initial global view. No comparison is reported against, for example, returning the answer with the highest accumulated verifier support, or a weighted combination of global and search-phase predictions. Given that fallback accuracy is substantially lower than accepted accuracy — 62.22% versus 90.91% on HR-Bench 4K — the fallback branch is a meaningful source of error whose design deserves explicit analysis.

Questions for authors

Why does the fallback return the saved global answer rather than the answer with the highest accumulated verifier support at termination? Have alternative fallback strategies been compared?

Have the authors tested the verifier with a model outside the Qwen family? The current comparison covers six Qwen checkpoints but does not establish whether the gains are architecture-specific.

5 comments
AuthorsOpenReview ↗

Response to reviewer nr8V (part1)

Thank you for your detailed review and for noting the value of reporting accuracy alongside inference cost. We appreciate your suggestions on isolating CAREER’s contribution and testing whether verification responds to the presence of key visual evidence.

W1. Novelty relative to CVSearch under matched verification

Thank you for pointing out the similarities between CAREER and CVSearch. CVSearch already uses full-image screening, expert-assisted localization, hierarchical region partitioning, and complexity-guided search. CAREER extends this approach by verifying each answer option against accumulated evidence and using the verification results to control the search.

  1. The methods assess different evidence. CVSearch assesses whether the current view is sufficient to answer the question. CAREER retains the full image and valid local observations from previous rounds. The verifier uses the entire evidence bundle to determine whether it supports each option, refutes it, or provides insufficient evidence. Table 3 of the submitted manuscript shows that removing prior local observations lowers HR-8K accuracy from 79.5% to 74.0%.
  2. Option-wise verification and evidence accumulation form a search loop. The generator proposes the next observation based on the current evidence bundle. Before taking that action, the verifier checks the leading option’s absolute Support and its margin over the runner-up. When both thresholds are met, CAREER stops and returns the answer. Otherwise, it continues searching and backtracks when needed. Each new observation goes into the same evidence bundle for both navigation and verification in the next round.
  3. The comparison with a shared verifier supports this design. In Table S4 of the submitted manuscript, CVSearch retains its original candidate construction and traversal procedure and uses the same stepwise verifier, acceptance thresholds, and fallback rule as CAREER. Full CAREER achieves gains of 3.1, 3.5, and 4.5 percentage points on V*Bench, HR-4K, and HR-8K, respectively.

W2. Extending CAREER to open-ended question answering on FineRS-4K

To assess whether CAREER extends to open-ended question answering, we compared methods using the same backbone on the 744 open-ended test questions in FineRS-4K. In this setting, CAREER generates candidate short answers from the accumulated visual evidence. The verifier assesses each candidate, and the controller uses the results to decide whether to continue observing. CAREER generates all candidates during inference.

Under the same answer scoring protocol, full-image direct answering reaches 59.14%, and full-image generation followed by verification reaches 60.48%. CVSearch reaches 58.33%, increasing to 61.83% with verification after search. With the same maximum local observation budget, CAREER reaches 65.05%, outperforming the baseline that uses the same verifier only after search. This comparison supports the benefit of integrating answer verification into search for high-resolution, open-ended short-answer questions.

AuthorsOpenReview ↗

Response to reviewer nr8V (part2)

W3. Testing sensitivity to adding, removing, and rephrasing answer options

Thank you for noting that normalized Support depends on the set of answer options. We tested how changes to the option set affect final answer accuracy on V*Bench, HR-4K, and HR-8K.

  • Changing the number of options. For questions with four original options, we either remove one distractor using a fixed random seed (41) or add a distractor of the same type as the existing options and manually verify that it is incorrect. For questions with two original options, we test distractor addition only. For example, in a color question, we add a color absent from the target object. We preserve the wording of the correct answer and all retained options.
  • Paraphrasing the options. We paraphrase the correct option and the distractors while preserving the number and order of the options and the identity of the correct answer. For example, we change left to on the left side and dark blue to a dark shade of blue. We manually check that each paraphrase preserves the original attribute, direction, or numerical value and that the correct answer remains unique.

We rerun the full procedure for each setting and report final accuracy. For each question, we keep the image, question text, models, search budget, and acceptance thresholds fixed across the applicable settings. On V*Bench, we evaluate distractor removal on the 107 questions with four original answer options.

Table 1. Answer accuracy under changes to the answer options

Option settingV*BenchHR-4KHR-8K
Original options91.62%78.00%79.50%
Remove one distractor91.59% (98/107)78.50%80.00%
Add one distractor90.58%77.25%78.75%
Paraphrase the options91.10%77.75%79.25%

With the acceptance thresholds held fixed, adding a distractor reduces final accuracy by 0.75 to 1.04 percentage points across the three evaluations. Paraphrasing changes accuracy by at most 0.52 percentage points.

W4. Comparing Qwen, Gemma, and Phi verifiers under matched settings

To test whether option-wise verification transfers across model families, we evaluated Gemma 3 4B IT and Phi-4-multimodal-instruct as verifiers. We vary the verifier while holding the generator, verification prompt, observation budget, and acceptance thresholds fixed.

Table 2. Answer accuracy across verifiers

VerifierV*BenchHR-4K subsetHR-8K subset
Qwen2.5-VL-3B89.5%76.5%77.5%
Qwen3-VL-2B91.1%77.0%78.0%
Qwen2.5-VL-7B91.1%77.5%78.5%
Qwen3-VL-4B (default)91.6%78.5%79.5%
Qwen3-VL-8B92.1%79.0%80.0%
Qwen3-VL-32B92.7%79.5%79.0%
Gemma 3 4B IT91.1%78.0%79.0%
Phi-4-multimodal-instruct90.6%77.5%78.5%

With the acceptance thresholds from the submitted configuration, Gemma and Phi remain within 1.0 percentage point of Qwen3-VL-4B (default) on all three evaluations. Both verifiers accept a slightly smaller share of answers and have a slightly higher error rate among accepted answers. These results show that CAREER's option-wise verification and search control can operate with the two non-Qwen families tested while maintaining comparable answer accuracy.

AuthorsOpenReview ↗

Response to reviewer nr8V (part3)

W5. Testing how blurring key visual evidence affects verification and search

Thank you for suggesting a direct intervention on the visual evidence. We reuse the 200 multiple-choice questions from the FineRS-4K region perturbation experiment and their annotations for key evidence and irrelevant regions. For each eligible question, we select the first decision step in the original trajectory at which local verification correctly accepted the answer, with observation budget remaining and a feasible next action available.

For each selected instance, we hold the question, options, and search state before the intervention fixed. We compare the original evidence bundle with two versions in which we blur either the key evidence region or the matched irrelevant region. Both interventions use local Gaussian blur with the same area and strength. We apply the blur to the full image and every local view containing the affected region. The verifier rescores all options, and the controller uses the original rule to decide whether to continue observing. We record whether the previously accepted correct option is reclassified as Insufficient and whether the system obtains at least one new valid observation.

Table 3. Paired evidence occlusion experiment

Evidence conditionMean Support for the originally accepted correct optionReclassified as Insufficient after occlusionSearch resumed and obtained a new observation
Original evidence bundle0.86——
Blur the key evidence region0.4758%72%
Blur the matched irrelevant region0.836%9%

Blurring key evidence lowers mean Support for the originally accepted correct option from 0.86 to 0.47, and 58% of those options are reclassified as Insufficient. When the matched irrelevant region is blurred, mean Support stays at 0.83, and 6% are reclassified as Insufficient. The controller resumes search and obtains a new observation in 72% of instances after blurring key evidence, compared with 9% after blurring the irrelevant region. At decision steps where the system originally stopped with the correct answer, blurring question-relevant evidence more often leads the verifier to judge the evidence insufficient and the controller to continue searching.

W6. Testing whether Support tracks answer correctness and responds to missing evidence

To assess whether Support is suitable as an answer acceptance signal, we compare the verifier's Support probability with the generator's confidence in the same candidate on V*Bench. The candidate answer and visual evidence for each question are held fixed. Higher Support bins have higher observed accuracy. Support distinguishes correct from incorrect candidates with an AUROC of 0.85, compared with 0.75 for generator confidence.

The paired experiment in W5 (Table 3) also shows that blurring key evidence more often changes the verifier's judgment from Support to Insufficient than blurring a matched irrelevant region. The system also resumes search and obtains a new observation more often. The first experiment tests whether Support distinguishes correct from incorrect answers, and the second tests whether it responds to the loss of key visual evidence.

We will expand the appendix section on acceptance, error sources, and verifier choice to report the paired AUROC comparison and the evidence occlusion experiment described in W5.

Reviewer_p8bDOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CAREER, a training-free framework for high-resolution visual question answering. The method uses existing tools, including SAM 3, SLIC, and CLIP, to construct and rank candidate image regions. It then adaptively selects regions for zooming, expansion, and evidence accumulation. Finally, the authors employ a frozen Qwen3-VL-4B-Instruct model as a verifier to guide multi-stage observation and answer acceptance. The authors evaluate their method on V*Bench, HR-Bench, and POPE-COCO using three different MLLM backbones, reporting improvements over existing visual-search methods in both accuracy and efficiency.

Critical strengths

The paper is well written. The proposed framework demonstrates improvements over existing visual-search methods across different MLLM backbones and benchmarks. The authors also provide several ablation studies to analyze the effects of individual components. The idea of retaining previously observed image regions and combining local details with broader context is reasonable for high-resolution visual understanding, especially for questions involving spatial relationships between objects.

Critical weaknesses

My major concern is that the method is a little limited in novelty. The method mainly combine extern tools and heuristics region-ranking methods. Such methods is also quite used in agent/harness related frameworks. In particular, the candidate-ranking formula uses Sobel edge density as one of its measures of visual information. However, regions with stronger gradients or more complex textures are not necessarily more informative for answering a specific question. This heuristic may also be sensitive to image noise or irrelevant textures. Also this paper involves serval hyperparameters, although there is some sensitivity analyses, the generalizability of these design choices is not demonstrated. Another concerns raise from the use of verifier. It is unclear that the probability will reflect the visual evidence or only confidence about the model's answer. A deeper study about such relationship is preferrable. Lastly, there is limited evalluation about more recent MLLM models, suck as qwen3.5vl,etc.

Questions for authors
  1. How can such method general beyonds selection problem. Current methods is limited to Closed-ended QA.
  2. How will such methods generalize to recent mllms?
  3. How do the authors justify using the SUPPORT token probability as a measure of visual evidence sufficiency?
  4. Why does the method return the initial global prediction when verification fails and the observation budget is exhausted? Have the authors compared this fallback strategy with returning the latest prediction obtained from the accumulated evidence?
5 comments
AuthorsOpenReview ↗

Response to reviewer p8bD (part1)

Thank you for your constructive review and for noting the value of retaining local detail and broader context. Your questions helped us clarify CAREER’s contribution and test its ranking and verification decisions more directly.

W1. Novelty of search guided by accumulated evidence and verification

Thank you for raising this point. CAREER uses existing tools to construct candidate regions and ranks them with heuristics. Our contribution is a search loop that links candidate acquisition, evidence accumulation across rounds, and option-wise verification.

  1. Region relevance and evidence sufficiency guide different decisions. The ranking determines which regions to inspect first. Even when two regions are relevant to the question, the current view can lack the context needed to determine their relative positions. CAREER retains the full image and all valid observations from previous rounds and uses the accumulated evidence to assess whether it can distinguish the answer options.
  2. Verification directly controls the search. The generator proposes the next observation based on the accumulated evidence and earlier verification feedback. The verifier scores each option against the same evidence bundle. CAREER stops when the leading option meets both the absolute Support and relative margin thresholds. Otherwise, it continues the search and adds the new observation to the evidence bundle for the next round.
  3. The search loop improves accuracy with the same external tools and ranking. In Table 3 of the submitted manuscript, visiting candidates in rank order with the same candidate regions, ranking, and verification rule yields 73.0% accuracy on HR-8K. A separate ablation that removes evidence accumulation across rounds yields 74.0%, compared with 79.5% for full CAREER. These two comparisons isolate the contributions of evidence-guided navigation and evidence accumulation across rounds, respectively.

W2. Testing Sobel ranking and the impact of irrelevant texture on accuracy

1. What does Sobel add to the ranking? We report the ablation from Table S5 of the submitted manuscript, which varies only the ranking cues. Table 1 shows that visual cues alone perform worse than query relevance alone on all four evaluations. Adding Sobel edge density to query relevance and SAM feature dispersion improves accuracy on all four, including a gain from 76.5% to 79.5% on HR-8K.

Table 1. Ablation of candidate-region ranking cues

Ranking cuesV*BenchHR-4KHR-8KPOPE-Adv
Query relevance only89.0%75.5%74.5%86.50%
Visual cues only (feature dispersion + Sobel)87.4%73.5%72.5%85.70%
Query relevance + feature dispersion (no Sobel)91.1%77.0%76.5%87.10%
Full CAREER91.6%78.0%79.5%88.33%

2. Does irrelevant texture drive the gains? We use 200 FineRS-4K questions answerable from target attributes. For each image, we locate the key evidence with the official mask and select a similarly sized, highly textured irrelevant region. We blur either region with matched area and strength or enhance texture in the irrelevant region. Both variants use the same questions, option order, and budget of eight local observations.

Table 2. Region perturbation experiment on FineRS-4K

Input conditionFull CAREERNo Sobel
Original image74.5%71.0%
Blur the key evidence region48.5%47.5%
Blur the matched, highly textured irrelevant region74.0%71.0%
Enhance texture in the irrelevant region73.0%70.5%

In Table 2, CAREER with Sobel retains 73.0–74.0% accuracy under the two irrelevant-texture perturbations, close to 74.5% on the original images and above the variant without Sobel in both cases. Blurring the key evidence reduces accuracy to 48.5%. In this test, CAREER is more sensitive to changes in evidence needed to answer the question than to changes in irrelevant texture.

AuthorsOpenReview ↗

Response to reviewer p8bD (part2)

W3. Measuring sensitivity to query relevance and visual cue weights

Thank you for raising this point. The submitted manuscript analyzes the acceptance thresholds and local observation budget. We additionally evaluated two weights used to rank candidate regions. We use the same five values for each parameter across V*Bench, HR-8K, and POPE-Adv. We vary one parameter at a time and hold the generator, verifier, and local observation budget fixed at K=8K=8.

First, we fix λ=0.50\lambda=0.50 (default) and vary the query relevance weight ww to assess sensitivity to the balance between semantic relevance and visual cues. w=0.70w=0.70 (default) gives the highest mean accuracy across the five tested values. At w=0.60w=0.60 and w=0.80w=0.80, each accuracy differs from the corresponding result at w=0.70w=0.70 by at most 0.52 percentage points.

Table 3. Sensitivity to the query relevance weight ww

wwV*BenchHR-8KPOPE-AdvMean accuracy
0.5090.05%77.50%87.57%85.04%
0.6091.10%79.00%88.13%86.08%
0.70 (default)91.62%79.50%88.33%86.48%
0.8091.10%79.25%88.20%86.18%
0.9090.58%78.00%87.77%85.45%

We then fix w=0.70w=0.70 (default) and vary the visual cue weight λ\lambda to assess the balance between feature dispersion and Sobel edge density. Each endpoint retains only one of the two visual cues. Query relevance is included in every setting.

Table 4. Sensitivity to the visual cue weight λ\lambda

λ\lambdaV*BenchHR-8KPOPE-AdvMean accuracy
0.00 (Sobel as the only visual cue)90.6%77.5%87.40%85.17%
0.2591.10%78.88%88.07%86.02%
0.50 (default)91.62%79.50%88.33%86.48%
0.7591.62%79.88%88.43%86.64%
1.00 (feature dispersion as the only visual cue)91.1%76.5%87.10%84.90%

λ=0.50\lambda=0.50 (default) yields the second-highest mean accuracy, 0.16 percentage points below the best value. On each of the three evaluations, accuracy changes by less than 0.7 percentage points when either weight is moved to the nearest tested value on either side. This supports using the same pair of weights across all three evaluations.

W4. Testing transfer to stronger generators with a fixed verifier

We evaluated Qwen3.5-9B and Gemma 4-31B-IT. Qwen3.5-9B extends the evaluation to the Qwen3.5 series, while Gemma 4-31B-IT tests transfer to a larger model from a different family. For each backbone, we compare direct answering from the full image, CVSearch, and CAREER. The two search methods share a maximum budget of K=8K=8 local observations. CAREER uses the controller and Qwen3-VL-4B verifier from the submitted configuration.

Table 5. Answer accuracy across generator backbones

GeneratorMethodV*BenchHR-4KHR-8K
InternVL2.5-8BFull-image direct answering69.1%66.0%57.5%
CVSearch88.5%77.5%76.5%
CAREER91.1%79.0%78.0%
LLaVA-OV-7BFull-image direct answering75.4%63.0%60.0%
CVSearch89.5%76.0%73.0%
CAREER91.1%80.0%77.0%
Qwen2.5-VL-7BFull-image direct answering71.2%68.5%65.0%
CVSearch89.0%76.5%76.0%
CAREER91.6%78.0%79.5%
Qwen3.5-9BFull-image direct answering87.96%75.75%73.75%
CVSearch90.58%78.00%77.75%
CAREER93.72%80.00%81.25%
Gemma 4-31B-ITFull-image direct answering89.53%77.25%76.00%
CVSearch92.67%79.75%79.25%
CAREER94.76%81.25%81.00%

Both newer generators improve the accuracy of full-image direct answering and CVSearch. CAREER still outperforms CVSearch on all three evaluations with each of these backbones. On HR-8K, CAREER improves accuracy from 77.75% to 81.25% with Qwen3.5-9B and from 79.25% to 81.00% with Gemma 4-31B-IT. Among the two CAREER configurations, Gemma achieves higher accuracy on V*Bench and HR-4K, and Qwen3.5 achieves slightly higher accuracy on HR-8K.

AuthorsOpenReview ↗

Response to reviewer p8bD (part3)

W5. Comparing verification during and after search on open-ended questions

We evaluate 744 open-ended questions from FineRS-4K using the same backbone and answer scoring protocol. The search methods share a maximum local observation budget. CAREER generates candidate short answers from accumulated visual evidence and verifies them during search.

Table 6. Open-ended answer accuracy on FineRS-4K

MethodAccuracy
Full-image direct answering59.14%
Full-image generation + verification60.48%
CVSearch58.33%
CVSearch + final verification61.83%
CAREER65.05%

CAREER outperforms CVSearch with the same verifier applied after search by 3.22 percentage points, supporting the use of verification during search for open-ended questions.

W6. Comparing verifier Support and generator confidence for answer acceptance

To assess whether Support is suitable as an answer acceptance signal, we compare the verifier's Support probability with the generator's confidence in the same candidate on V*Bench. The candidate answer and visual evidence for each question are held fixed. The verifier independently assesses whether the evidence supports the candidate, refutes it, or is insufficient. Higher Support bins have higher observed accuracy. Support distinguishes correct from incorrect candidates with an AUROC of 0.85, compared with 0.75 for generator confidence. This paired comparison supports using Support as an acceptance signal. We will expand the appendix section on acceptance, error sources, and verifier choice to describe the paired evaluation and report the accuracy by Support score bin and the AUROC comparison.

The specific images and analysis can be found in the revised version.

W7. Using accumulated search evidence to improve fallback answers

Thank you for suggesting this comparison. The submitted manuscript uses the initial full-image answer as a conservative fallback. We compare this rule with CAREER-Plus (CAREER+ in the revised manuscript) on the three backbones profiled for efficiency. We hold the search trajectories, observation budget, and verification thresholds fixed. When fallback is triggered, CAREER-Plus uses the same generator to answer from the evidence accumulated by the end of search.

As shown in Table 7, CAREER-Plus improves accuracy in seven of the nine backbone–benchmark combinations and ties in the other two. It improves HR-8K accuracy for all three backbones, with the largest gain for Qwen2.5-VL-7B, from 79.5% to 80.5%. This comparison shows that accumulated local evidence can help the generator correct some initial answers even when search ends before the verification thresholds are met.

Table 7. Overall accuracy under different fallback strategies

BackboneMethodV*BenchHR-4KHR-8K
InternVL2.5-8BCVSearch88.5%77.5%76.5%
CAREER91.1%79.0%78.0%
CAREER-Plus91.6%79.0%78.5%
LLaVA-OV-7BCVSearch89.5%76.0%73.0%
CAREER91.1%80.0%77.0%
CAREER-Plus91.1%80.5%77.5%
Qwen2.5-VL-7BCVSearch89.0%76.5%76.0%
CAREER91.6%78.0%79.5%
CAREER-Plus92.1%78.5%80.5%

Table 8. End-to-end efficiency on one H200 (CAREER → CAREER-Plus)

BackboneHR-8K gain (pp)Latency (s/question) ↓Throughput (questions/min) ↑Peak GPU memory (GiB) ↓
InternVL2.5-8B+0.515.36 → 16.753.91 → 3.5830.4 → 30.9
LLaVA-OV-7B+0.513.97 → 15.344.29 → 3.9028.5 → 29.1
Qwen2.5-VL-7B+1.014.81 → 16.194.05 → 3.7231.7 → 32.2

CAREER-Plus makes one additional generator call on the accumulated evidence only when fallback is triggered. The HR-8K gains come with higher latency and peak GPU memory and lower throughput. See Appendix E.10 (Table S20) of the revised manuscript for the paired measurements.

Thank you again for encouraging us to evaluate CAREER with newer generators and on open-ended questions.

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

This paper studies question answering on high-resolution images and proposes CAREER (Cross-granularity Evidence Search with Sufficiency Verification), a training-free framework. The authors argue that a local region being relevant to the question does not mean the evidence collected so far is sufficient to answer it.

The pipeline has four stages:Global screening.

  1. The backbone first answers using only the global image. If both the top-1 confidence and the top-1/top-2 gap exceed their thresholds, the answer is returned directly.
  2. Candidate atlas construction. Otherwise, the method builds a multi-scale candidate atlas from query-directed SAM 3 proposals and a recursive SLIC region tree. Candidates are ranked by a weighted combination of CLIP relevance, SAM feature dispersion, and Sobel edge density.
  3. Evidence accumulation and search. An evidence bundle keeps the global image together with all previously observed crops. The generator chooses the next action from ZOOM / EXPAND / SPLIT / NEXT.
  4. Independent verification. A separate frozen Qwen3-VL-4B verifier judges each option as Support / Refute / Insufficient. An answer is accepted only when both the absolute support and the normalized margin exceed their thresholds. When the search stalls, the method backtracks (RECOVER). When the budget is exhausted, it returns the global answer and flags it as "unverified."

Experiments cover three 7–8B backbones on V*Bench, HR-Bench 4K/8K, and POPE. CAREER outperforms ZoomEye, RAP, and CVSearch in every setting, with about 25% lower latency than CVSearch.

Critical strengths
  1. Clear problem formulation. The distinction between relevance and sufficiency is valuable. Fig. 1 and the cases in Appendix F.1/F.2 clearly show that relational questions need fine-grained detail and global context at the same time.
  2. Table 4 includes controls such as verifier-only, generator + verifier, and CVSearch with the same verifier. Table 6 provides a budget sweep, and Table 7 reports the accepted and fallback branches separately.
  3. Honest about limitations. The authors explicitly acknowledge that, under the default thresholds, the theoretical bound does not guarantee a low error rate (βv=0.3625). They also admit that quantities such as ε、δ、κ are not estimated in the experiments. Failure cases are included as well (F.5).
Critical weaknesses
  1. The verifier produces the final answer, which seriously confounds the claim of cross-backbone gains. According to line 9 of Algorithm 1, on acceptance the method returns the option with the highest verifier share, not the generator's answer. On fallback, it returns the global answer. The final answer therefore has only two possible sources: Qwen3-VL-4B when verification passes, and the backbone's global answer when it does not. In practice, the backbone only navigates during the intermediate steps. Qwen3-VL-4B is a generation newer than all three backbones and has stronger high-resolution capability on its own. Table 4 supports this concern. "Generator + verifier" with no cropping at all already reaches 76.0/76.5 on HR-4K/8K, on par with the original CVSearch (76.5/76.0). A substantial share of the gains in Table 1 therefore comes from adding a stronger model, not from the search strategy. For the same reason, the results on three backbones do not show that the method generalizes across backbones, since all three share the same verifier.

  2. The comparison with baselines is unfair. None of ZoomEye, RAP, or CVSearch uses an external verifier. The POPE baselines (VCD, MFCD, PAI, AllPath) are single-forward-pass decoding or intervention methods. CAREER, by contrast, additionally uses SAM 3, CLIP, and a 4B verifier, and runs M option-wise scoring passes per round. Its latency is roughly 10× that of the backbone (14.8 s vs. 1.56 s). Given this compute gap, claiming SOTA on POPE means little.

  3. The theoretical part is largely decorative. The core assumptions of Theorem 3.1 are that the verifier shares agree with the history-conditioned posterior in sup-norm (ε_v, δ_v), and that reachability κ_t holds. The authors themselves acknowledge that these quantities are neither estimable nor certified. Under such assumptions the risk bound is close to: if the scores are aligned with the posterior, the error rate upon acceptance is naturally low. This section contributes little to understanding or improving the method, and the space would be better used for additional experiments.

Questions for authors
  1. What is the accuracy of Qwen3-VL-4B used alone as the backbone, both with the global image only and with CVSearch, on V*Bench, HR-4K, and HR-8K? If it alone comes close to CAREER, the paper's core contribution does not hold.
  2. Please add Qwen3-VL-4B as a second opinion (e.g., through ensembling or verifier re-ranking) on top of the final answers of ZoomEye, RAP, or CVSearch. Does CAREER still lead under this setting?
4 comments
AuthorsOpenReview ↗

Response to reviewer Mzph (part1)

Thank you for your careful review and for highlighting the distinction between finding relevant regions and gathering enough evidence to answer. We appreciate your questions about the verifier’s contribution, the fairness of the comparisons, and the role of the theory.

W1. Isolating the benefit of search with a shared vision-language model

We fix Qwen3-VL-4B-Instruct across the first four methods in Table 1 and evaluate all 191 V*Bench questions and 800 instances each from HR-4K and HR-8K. For full-image verification, the verifier scores each option separately and selects the option with the highest Support, following Table S4 of the submitted manuscript. CAREER uses the same model for generation and verification and follows the submitted search, acceptance, and fallback procedure. We also report the submitted CAREER configuration from Table S7, which uses Qwen2.5-VL-7B as the generator and Qwen3-VL-4B as the verifier.

Table 1. Answer accuracy with a shared model

MethodV*Bench (%)HR-4K (%)HR-8K (%)
Full-image option-wise verification76.9670.0067.50
Full-image direct answering76.9670.2567.50
CVSearch86.9174.5071.75
CAREER (shared model)89.5378.0078.50
CAREER91.6278.0079.50

With Qwen3-VL-4B held fixed, CAREER improves over CVSearch by 2.62, 3.50, and 6.75 percentage points on V*Bench, HR-4K, and HR-8K, respectively. The largest gain is on HR-8K, where accuracy increases from 71.75% to 78.50%. This comparison shows that CAREER's combination of local evidence search and stepwise verification improves on CVSearch when both methods use the same model.

W2. Testing CAREER with stronger generators and the same verifier

We evaluated Qwen3.5-9B and Gemma 4-31B-IT. Qwen3.5-9B extends the evaluation to the Qwen3.5 series, while Gemma 4-31B-IT tests transfer to a larger model from a different family. For each backbone, we compare direct answering from the full image, CVSearch, and CAREER. The two search methods share a maximum budget of K=8K=8 local observations. CAREER uses the controller and Qwen3-VL-4B verifier from the submitted configuration.

Table 2. Answer accuracy across generator backbones

GeneratorMethodV*BenchHR-4KHR-8K
InternVL2.5-8BFull-image direct answering69.1%66.0%57.5%
CVSearch88.5%77.5%76.5%
CAREER91.1%79.0%78.0%
LLaVA-OV-7BFull-image direct answering75.4%63.0%60.0%
CVSearch89.5%76.0%73.0%
CAREER91.1%80.0%77.0%
Qwen2.5-VL-7BFull-image direct answering71.2%68.5%65.0%
CVSearch89.0%76.5%76.0%
CAREER91.6%78.0%79.5%
Qwen3.5-9BFull-image direct answering87.96%75.75%73.75%
CVSearch90.58%78.00%77.75%
CAREER93.72%80.00%81.25%
Gemma 4-31B-ITFull-image direct answering89.53%77.25%76.00%
CVSearch92.67%79.75%79.25%
CAREER94.76%81.25%81.00%

Both newer generators improve the accuracy of full-image direct answering and CVSearch. CAREER still outperforms CVSearch on all three evaluations with each of these backbones. On HR-8K, CAREER improves accuracy from 77.75% to 81.25% with Qwen3.5-9B and from 79.25% to 81.00% with Gemma 4-31B-IT. Among the two CAREER configurations, Gemma achieves higher accuracy on V*Bench and HR-4K, and Qwen3.5 achieves slightly higher accuracy on HR-8K.

W3. Comparing verifier families with the generator and thresholds held fixed

To test whether option-wise verification transfers across model families, we evaluated Gemma 3 4B IT and Phi-4-multimodal-instruct as verifiers. We vary the verifier while holding the generator, verification prompt, observation budget, and acceptance thresholds fixed.

Table 3. Answer accuracy across verifiers

VerifierV*BenchHR-4K subsetHR-8K subset
Qwen2.5-VL-3B89.5%76.5%77.5%
Qwen3-VL-2B91.1%77.0%78.0%
Qwen2.5-VL-7B91.1%77.5%78.5%
Qwen3-VL-4B (default)91.6%78.5%79.5%
Qwen3-VL-8B92.1%79.0%80.0%
Qwen3-VL-32B92.7%79.5%79.0%
Gemma 3 4B IT91.1%78.0%79.0%
Phi-4-multimodal-instruct90.6%77.5%78.5%

With the acceptance thresholds from the submitted configuration, Gemma and Phi remain within 1.0 percentage point of Qwen3-VL-4B (default) on all three evaluations. Both verifiers accept a slightly smaller share of answers and have a slightly higher error rate among accepted answers. These results show that CAREER's option-wise verification and search control can operate with the two non-Qwen families tested while maintaining comparable answer accuracy.

AuthorsOpenReview ↗

Response to reviewer Mzph (part2)

W4. Comparing verification during search with reranking after search

All methods use Qwen2.5-VL-7B as the backbone. For final reranking, CVSearch, RAP, and ZoomEye complete their original searches, then Qwen3-VL-4B scores the options using the collected observations and selects the one with the highest Support. These three settings share the verification prompt and input budget. CAREER follows the submitted procedure and verifies accumulated evidence during search. Stepwise CVSearch follows Table S4 of the submitted manuscript and checks the acceptance thresholds after each valid local observation.

Table 4. Answer accuracy when the same verifier is added to the search baselines

MethodV*Bench (%)HR-4K (%)HR-8K (%)
CVSearch89.0176.5076.00
CVSearch + stepwise verification88.4874.5075.00
CVSearch + final reranking89.5376.2576.50
RAP84.8274.5076.00
RAP + final reranking86.9175.2577.00
ZoomEye85.3472.5070.00
ZoomEye + final reranking87.4372.2572.00
CAREER91.6278.0079.50

Final reranking improves several baseline results. For example, RAP improves from 76.00% to 77.00% on HR-8K. CAREER leads the strongest reranking baseline, CVSearch with final reranking, by 2.09 percentage points on V*Bench and 1.75 on HR-4K. On HR-8K, CAREER leads RAP with final reranking, the strongest reranking baseline there, by 2.50 percentage points.

W5. Comparing POPE accuracy against baselines with matched inference time

Thank you for pointing out the difference in inference cost on POPE. We agree that the accuracy of the single-pass baselines and CAREER should be interpreted alongside their runtimes. We therefore added multiple-pass variants of VCD, MFCD, PAI, and AllPath on the Random, Popular, and Adversarial subsets of POPE.

Each baseline retains its original decoding or intervention mechanism. It answers using a fixed set of input combinations drawn from the original image, its horizontally flipped version, and semantically equivalent question paraphrases. The final Yes/No answer is selected by majority vote. On a separate calibration set, we chose the smallest number of calls for each baseline that matched or slightly exceeded CAREER's mean runtime per question, then fixed that configuration before evaluation. All methods use Qwen2.5-VL-7B and run on the same H200 GPU with a batch size of 1. End-to-end timing includes input preprocessing, any search and verification performed by the method, and final answer generation.

Table 5. POPE accuracy at similar inference times

MethodAnswering callsRandom acc.Popular acc.Adversarial acc.Time (s/question) ↓Throughput (questions/min) ↑
Full-image direct answering184.40%83.47%81.83%1.5638.46
Original VCD188.67%87.63%86.47%2.9820.13
VCD + voting590.10%89.27%87.17%14.954.01
Original MFCD185.33%84.47%84.13%4.9812.05
MFCD + voting487.27%86.30%85.50%14.974.01
Original PAI186.43%85.60%84.17%2.1627.78
PAI + voting788.20%87.07%85.80%15.163.96
Original AllPath190.60%88.60%86.60%1.6636.14
AllPath + voting991.67%90.13%87.40%15.004.00
CAREER (K=8K=8)—92.77%91.23%88.33%14.814.05

All four baselines improve with the larger inference budget. AllPath is the strongest baseline and reaches 87.40% on the Adversarial subset at 15.00 seconds per question. CAREER reaches 88.33% at 14.81 seconds per question, a gain of 0.93 percentage points. CAREER also improves over AllPath by 1.10 percentage points on each of the Random and Popular subsets. CAREER thus retains its accuracy advantage on all three subsets when the baselines receive similar or slightly longer inference times.

AuthorsOpenReview ↗

Response to reviewer Mzph (part3)

W6. Theoretical analysis of acceptance and fallback under a finite observation budget

Thank you for raising these concerns. We agree that a high posterior probability for an accepted answer implies low conditional error. The proof and the experiments address two complementary questions: what the acceptance rule guarantees under posterior alignment, and where prediction errors arise when search has a finite budget.

1. What do the two thresholds guarantee under posterior alignment?

The alignment assumption relates normalized verifier Support to the answer posterior along the search trajectory. With bounded total Support, the proof derives a lower bound from each threshold at the round where CAREER stops. It takes the larger of the two bounds and subtracts the alignment error to lower bound the posterior of the accepted answer. This makes the contribution of each threshold to the accepted-answer risk bound explicit.

Table 3 of the submitted manuscript tests this design in the implemented controller. On HR-8K, using both thresholds yields 79.5% accuracy, compared with 77.0% for absolute Support alone and 75.5% for the relative margin alone, gains of 2.5 and 4.0 percentage points, respectively. In a separate paired V*Bench test, verifier Support also discriminates correct from incorrect candidates better than generator confidence (AUROC 0.85 vs. 0.75). These results show the benefit of the joint rule and quantify how well Support distinguishes correct answers in practice.

2. How do acceptance coverage and fallback affect overall error?

Under the reachability and threshold-compatibility assumptions, Theorem 3.1 bounds the probability of reaching acceptance within the observation budget. The appendix then decomposes overall error into the conditional errors of accepted and fallback answers, each weighted by how often that branch is used. The alignment error and reachability rates remain unestimated in our experiments, and the risk bound remains loose at the absolute Support threshold of 0.65 (default) and the relative margin threshold of 0.15 (default).

Table S7 of the submitted manuscript measures the two contributions directly. On HR-4K, 360 of 800 instances fall back (45%). Fallback accounts for 136 of the 176 errors (about 77%), while acceptance accounts for 40. These branches contribute 17.0 and 5.0 percentage points to overall error, respectively. The measured breakdown identifies fallback as the larger source of error and motivates improving how CAREER uses accumulated evidence when the budget runs out.

Thank you again for prompting us to clarify the contribution of search under matched settings and the scope of the theoretical guarantees.

AuthorsOpenReview ↗

Gentle Reminder: Please Review and Respond to Our Rebuttal

Dear Reviewer Mzph,

Thank you for your time and thoughtful review. As the author–reviewer discussion period is about to close, we would appreciate it if you could take a final look at our rebuttal and let us know whether it addresses your concerns. If the clarifications resolve your concerns, we hope you might reconsider your overall assessment in light of them. If you have any other questions, feel free to ask.

Thank you again for helping us improve the paper.

Comments on the paper

1 comment

Dear Area Chairs and Reviewers,

Thank you to the Area Chairs and reviewers for your careful evaluation and constructive feedback. We are encouraged by the reviewers’ comments on CAREER’s motivation, design, and experimental evaluation. In particular, the reviewers highlighted the following strengths.

  • “Clear problem formulation. The distinction between relevance and sufficiency is valuable.” (Reviewer Mzph)
  • “The idea of retaining previously observed image regions and combining local details with broader context is reasonable for high-resolution visual understanding, especially for questions involving spatial relationships between objects.” (Reviewer p8bD)
  • “CAREER achieves the highest overall accuracy across all nine backbone–benchmark combinations in Table 1.” (Reviewer nr8V)
  • “This joint analysis is more informative than accuracy-only comparisons and makes the efficiency claim concrete.” (Reviewer nr8V)

During the rebuttal, we responded to each reviewer and conducted additional experiments. The main changes are summarized below.

  • Controlled baseline comparisons (Reviewers Mzph and nr8V). With Qwen3-VL-4B held fixed, CAREER outperformed CVSearch on all three high-resolution benchmarks, with the largest gain on HR-8K (6.75 percentage points). CAREER also retained its advantage on all three POPE subsets when baselines received comparable inference time.
  • Generalization across models and tasks (Reviewers Mzph, p8bD, and nr8V). We tested stronger generators, evaluated Gemma and Phi as alternative verifiers, and extended CAREER to open-ended QA. On 744 open-ended FineRS-4K questions, CAREER achieved 65.05% accuracy, compared with 61.83% for CVSearch using the same verifier after search.
  • Visual grounding of the stopping decision (Reviewers p8bD and nr8V). At steps where CAREER had correctly accepted an answer, blurring key visual evidence caused search to resume in 72% of cases, compared with 9% after blurring a matched irrelevant region.
  • Fallback using accumulated evidence (Reviewers p8bD and nr8V). Following the reviewers’ suggestion, we introduced CAREER-Plus, which uses the evidence gathered during search when the observation budget runs out. It improved accuracy in seven of nine backbone–benchmark combinations and matched CAREER in the other two.

We especially appreciate Reviewer p8bD’s follow-up recognizing the gains from CAREER-Plus and the results on open-ended questions and stronger backbones, and their decision to raise their rating from 2 to 3 (Weak Acceptance). We thank the Area Chairs and reviewers for helping us clarify CAREER’s contribution and improve the method itself.

Best regards,
The authors of the CAREER submission

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.