acceptodds
Under review as a conference paper at ICLR 2027

Can the Evidence Reach the Answer? Diagnostic Reachability over Uncertain Evidence Graphs for Agentic Whole-Slide Reasoning.

Abstract

Whole-slide image visual question answering (WSI-VQA) requires agentic systems to collect diagnostically relevant evidence that is often sparse and spatially dispersed across gigapixel slides. Existing agentic WSI-VQA systems typically decompose reasoning into three stages: evidence acquisition, acquisition sufficiency assessment, and answer generation, yet represent the acquired evidence as unstructured patch level information. Because they do not explicitly preserve the diagnostic evidence pathways that connect morphological observations to an answer through diagnostic concepts, they exhibit two limitations: unreliable pathway agnostic evidence evaluation within stages and diagnostic evidence loss across stages. To address these limitations, we propose ARGUE (Active Reasoning over a Graph of Uncertain Evidence), a training-free agentic WSI-VQA system. Within stages, ARGUE uses diagnostic reachability to enable pathway-aware evaluation by quantifying the probability of realizing evidence pathways under uncertain diagnostic relations. Across stages, ARGUE maintains a shared Uncertain Evidence Graph (UEG) to preserve these pathways, thereby preventing the loss of structured evidence. Experiments on three benchmarks show that ARGUE achieves the best overall performance among all baselines. Ablation studies further validate the effectiveness of ARGUE's core components, while robustness analyses show consistent performance under external noise and different implementation choices.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Agentic systems, WSI-VQA, Whole Slide Images

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7Q91OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ARGUE, a training-free agent for whole-slide image question answering. It represents observed findings, diagnostic concepts, and answer-support/refutation relations in an uncertain evidence graph shared across evidence acquisition, sufficiency assessment, and answer generation. A Monte Carlo diagnostic-reachability score guides patch acquisition and stopping. Experiments on three WSI-VQA benchmarks compare ARGUE with agentic baselines and include graph/component ablations and robustness analyses; the authors report the best aggregate results among the tested systems.

Critical strengths
  1. Preserving explicit links from morphological findings through concepts to candidate answers is a relevant response to the evidence fragmentation of multi-stage WSI agents. The graph is used throughout the pipeline rather than only added as a post-hoc explanation.
  2. The paper tests its central representation through ablations that remove pathway structure and vary relation-probability assignments. This provides more direct evidence for the graph and reachability components than a system-level comparison alone.
Critical weaknesses
  1. The uncertain relations and their validity probabilities are generated by an LLM, yet the paper does not establish that these probabilities are calibrated to diagnostic correctness or that edge dependencies are adequately modeled. Since reachability is the central decision signal for both acquisition and stopping, systematic errors in these estimates can undermine the method's main premise. Ablations against random or text-similarity probabilities do not establish clinical calibration.
  2. Evaluation is based on benchmark answer accuracy and model-driven evidence construction, rather than expert-annotated diagnostic pathways or a prospective clinical decision setting. The results support a promising WSI-VQA agent, but do not yet validate that diagnostic reachability reflects medically sound evidence sufficiency. Existing training-free approaches such as PathNavigate[1] also address evidence acquisition; a stronger comparison on matched budgets and independently assessed evidence quality would clarify ARGUE's incremental contribution.

[1] Yang, Chunze, et al. "PathNavigate: A Training-Free Pathology Agent with Surprise-Guided Scan and Shared Slide Memory for Whole-Slide Image VQA." arXiv preprint arXiv:2605.23559 (2026).

Questions for authors
  1. How are relation-validity probabilities elicited, and have they been calibrated or compared with independent pathologist judgments? Please specify whether the Monte Carlo computation assumes independent edge validity and how sensitive decisions are to violations of that assumption.
  2. Are the main baseline comparisons matched for inspected-patch/token budget and LLM calls? If not, what remains of ARGUE’s advantage under a matched inference budget?
Reviewer_VzzCOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ARGUE, a training-free agent framework for whole-slide image visual question answering (WSI-VQA). The authors argue that existing methods lack modeling of complete diagnostic evidence pathways and can lose evidential relations across different reasoning stages. To this end, this paper introduces the Uncertain Evidence Graph (UEG), which connects morphological observations, diagnostic concepts, and the final answer, and estimates diagnostic reachability via Monte Carlo sampling. This metric is used to guide patch selection, determine whether to terminate the search, and generate the final answer. Experiments demonstrate favorable results across three benchmarks.

Critical strengths

1.The motivation of the paper is clear. Compared to merely storing fragmented patch descriptions, explicitly recording the relationships between observations and diagnostic conclusions is a reasonable direction.

2.The method design is complete. The authors utilize the same evidence graph for evidence acquisition, stopping determination, and answer generation, making the overall pipeline coherent. Guiding the search by seeking evidence that refutes the current working answer is also reasonably well-motivated.

Critical weaknesses

1.The reproduced performance of several baselines is markedly lower than originally reported. For example, PathNavigate achieves 59.42% on SlideBench-BCNB in its original paper, but drops to 37.8% in Table 2; HistoSelect shows a similarly sharp decline on WSI-Bench. Furthermore, Appendix A.2 notes that HistoSelect's LoRA adaptation weights were disabled, which likely degrades its original pipeline.While standardizing the backbone across methods is understandable, the authors should account for these drastic performance drops and report the baselines' performance under their native settings alongside the unified-backbone results. Without this, the current empirical evidence is insufficient to claim superiority over the intact versions of these methods.

2.During candidate reranking, the model does not visually inspect candidate patches; rather, it attaches the best-matching textual evidence gap description as a hypothetical finding to compute reachability gain. Consequently, if multiple candidate patches match the same evidence gap, they yield the exact same temporary graph and identical reachability gains (assuming deterministic graph construction), leaving their relative ranking entirely dependent on CONCH cosine similarity.It is therefore difficult to disentangle the benefit of LLM-based query expansion (generating evidence gaps) from that of reachability-based reranking. A crucial ablation is missing: evaluating Stage I with gap relevance alone (λ=1\lambda = 1, without reachability reranking).

Questions for authors

On Baseline Reproduction: What accounts for the substantial performance gap between your reproduced results for PathNavigate, PathAgent, and HistoSelect and their originally published numbers? Could you report results under their native, intact configurations—particularly HistoSelect with its original LoRA adaptation enabled?
On Stage I Reranking: Could you provide an ablation with λ=1\lambda = 1 to isolate the performance of using evidence-gap retrieval alone? Furthermore, when multiple candidate patches match the exact same evidence gap description, how does the predicted reachability gain differentiate among them?
On Probabilistic Graph Modeling: What happens to overall performance if all edge probabilities are fixed to πt(e)≡1\pi_t(e) \equiv 1 (reducing UEG to deterministic graph connectivity)? Have you compared Monte Carlo reachability against simpler heuristics, such as path counting or maximum-probability path scoring?
On the Stopping Criterion and Variance: In practice, what percentage of queries terminate early via Equation 5 versus hitting the hard cap of maximum rounds (T=5T = 5)? Have you quantified how often premature stopping leads to incorrect final answers? Finally, could you report multi-run means and standard deviations on a representative data subset to confirm that the smaller ablation gains (e.g., 1–2%) hold above run-to-run variance?

Reviewer_RW2NOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ARGUE is a training-free system for whole-slide image question answering. It maintains an uncertain evidence graph that connects observed morphology to concepts and then to support or refutation of a working answer. Monte Carlo estimates of graph reachability guide patch acquisition, answer updates, stopping, and final generation. Experiments on WSI-Bench, WSI-VQA, and SlideBench-BCNB show substantial gains over the evaluated baselines. The paper also studies graph structure, stage-specific strategies, alternative backbones, and noisy relation probabilities.

Critical strengths
  1. The performance gains are substantial across benchmarks. WSI-VQA accuracy increases from the strongest baseline's 0.459 to 0.549, and SlideBench-BCNB overall accuracy increases from 0.467 to 0.563 (Table 2). ARGUE also leads on WSI-Bench (Table 1). The gains remain with shared Qwen3 backbones and a Patho-R1 observer (Table 4 and Figure 5), supporting the value of the overall framework.
  2. The efficiency analysis shows modest additional cost. ARGUE uses slightly more tokens than PathAgent and fewer than BEACON (Figure 3). Its inference time is 115.51 seconds per WSI, compared with 106.22 for PathAgent and 115.94 for BEACON (Figure 9). The improvement in answer quality therefore comes at a comparable time and token cost.
  3. The ablations support the graph relations and stopping criterion. Removing diagnostic relations reduces WSI-P/WSI-R from 0.423/0.461 to 0.354/0.349 (Figure 1(c), Appendix A.3). Replacing the proposed stopping rule with an LLM judgment reduces these scores to 0.338/0.408 (Table 3). These results provide clear evidence that both components contribute to the final performance.
Critical weaknesses
  1. Node merging does not address semantic duplication. Appendix A.9 merges findings and concepts only when their normalized strings match exactly. Equivalent descriptions with different wording can remain separate nodes and create extra paths for the same evidence. Since distinct edges are treated as independent, this can inflate support or refutation reachability and affect answer updates and stopping without adding new diagnostic information. The paper does not evaluate the effect of such repeated concepts or findings.
  2. The reliability of generated concepts and relations remains unclear. The default system relies on general-purpose Qwen models for visual findings, concepts, relations, and edge probabilities. The expert study on 10 questions (Figure 7, Appendix A.4) provides useful checks of graph coherence and probability plausibility, but does not separately measure concept correctness or establish relation quality across the benchmarks. Figure 5 tests Patho-R1 as the visual observer; the effect of a medical language model on concept generation, relation construction, and edge scoring remains unexplored.
  3. Visual errors can propagate through the text-based reasoning process. Graph construction and edge scoring operate on the visual observer's descriptions (Sections 3.1 and 3.3). An incorrect description can therefore produce an internally consistent evidence chain for the wrong conclusion. Reachability does not itself check whether the finding is present in the image. The current analyses do not show how often this occurs or whether later rounds detect and correct the underlying observation.
Questions for authors
  1. Initialization and open-ended answers: How is the initial working answer obtained before the first reachability reranking step, and what visual evidence is available at that point? Appendix A.9 allows the answer to change, but how does support/refutation of a single working answer cover other plausible answers, especially those that are not directly incompatible with it?
  2. Graph size and stopping behavior: What are the typical and largest numbers of nodes and edges in the final graphs on each benchmark, and how do these counts grow across rounds? On average, how many rounds are needed to meet the stopping criterion, and what fraction of questions instead reach the five-round limit?
1 comment

I appreciate the detailed reply from the authors and the effort made during the rebuttal phase. My concerns are now addressed and I would like to maintain my recommendation of acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.