acceptodds
Under review as a conference paper at ICLR 2027

Necessity-Bench: A Self-Report-Free Test of Whether Vision-Language Model Answers Are Actually Grounded in Evidence

Abstract

Vision-language models (VLMs) often answer questions about images correctly, but it is unclear whether those answers actually depend on the part of the image that matters. Existing ways of checking this rely on the model's own account of its reasoning: they either reward overlap between a region the model cites and the correct object, or test whether the model's written rationale mentions an object that was removed. Both trust self-report, which can be wrong. We introduce Necessity-Bench, a protocol that uses no self-report at all. For each question, an independent oracle (a scene program or scene graph, never the model) identifies the region the answer depends on and a separate, irrelevant control region. We edit each region in turn and ask the model again, checking three properties of its answer: sufficiency (the relevant region alone reproduces the answer), necessity (removing it changes the answer to the true counterfactual answer, recomputed by re-running the ground-truth program), and specificity (removing the irrelevant region leaves the answer unchanged). The control rules out models that change their answer under any edit, and the recomputed counterfactual is stricter than testing for "any change." The protocol requires no training. On CLEVR (3,000 items; 853 with a well-defined counterfactual), Qwen3-VL-8B scores 0.680 on the strict joint metric versus 0.619 for Qwen3-VL-2B (p=0.0097), a scale effect driven by correctly tracking the counterfactual rather than by answer stability. On real photographs (GQA, 962 items), specificity stays above 0.92 across 2B/8B/32B, but no scale effect appears (p≥0.264). Finally, we ask whether self-report is safe to use as a training signal. Reinforcement learning on a model's own predicted evidence box raises every self-reported metric (2B accuracy 0.617→0.889; 8B box IoU 0.295→0.597). Yet our protocol, run on items excluded from training, shows that the 2B model's true causal grounding falls (necessity 0.656→0.537, p=2.9×10⁻⁶) while the 8B model is unchanged. Self-reported progress and real grounding can move in opposite directions: grounding should be verified by intervention, not by asking the model. Necessity-Bench thus offers a training-free, self-report-free measure of grounding that exposes failures self-report conceals.

Primary area:
datasets and benchmarks
Keywords:
vision-language models, causal intervention, counterfactual evaluation, groundedness, benchmark, model scaling, reinforcement learning, model faithfulness

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_bSLPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes new benchmark called Necessity-Bench to evaluate visual grounding of VLM. Necessity-Bench measures three property including sufficiency, necessity, and specificity. This new benchmark is built upon CLEVR (geometric image) and GQA (natural image). Qwen3-VL evaluations on this benchmark demonstrates that larger models are more visually grounded. The authors further post-train Qwen3-VL on model’s own prediction of target bounding box and shows their metrics are not necessarily increasing despite the increase in the target accuracy.

Critical strengths
  1. The paper proposes simple formulation of three metrics.
  2. Necessity-Bench is evaluated on 3 scales of Qwen3-VL series.
Critical weaknesses
  1. Generally, it reads as a technical report than conference paper. Writing seems like highly AI-generated (enumeration of detailed numbers instead of emphasizing few important numbers with key message), which severely reduces the readability. Lots of unnecessary phrases (e.g., L239-241) and redundancies in writing overall (e.g., L92-107 I don't think this paper has 6 contribution points, Related work can be more compressed).

  2. Although the paper's main contribution is the benchmark. it does not share detailed construction pipeline and only showing single example for CLEVR and GQA. For example, how the target object is extracted from the query? If there are multiple target regions, which one is removed?

  3. The benchmark images in Figure 1 seem like including some artifacts (e.g., rightmost image in GQA sample with blur-like artifacts, simple grey marking in CLEVR sample), which is not a practical settings for VLMs. Is there any chances that VLM misbehaves because of these artifacts?

  4. The paper should include human evaluation and frontier model evaluation on this benchmark to provide comprehensive view. One model series evaluation is not sufficient.

  5. The paper’s main finding in L260-263 is not surprising. Already exists bunch of works [1-3] showing that larger VLM is more visually grounded.

  6. I am bit skeptical about Sec 6 experiment setting and results. It is well known that the quality of data matters for RL post-training as it is trained with few samples (around 10k) at academic level compared to SFT. I am not sure whether training on self-reported region is proper way to apply RL post-training as their labels will be inevitably noisy. Also the coordinate of the target object in image only represents one format of self-reported grounding. VLM can also report the grounding in long caption format (e.g., A ball is behind the cube on the top-left side of the image.), which is not tested.

[1] Lee, K. I., Kim, M., Yoon, S., Kim, M., Lee, D., Koh, H., & Jung, K. (2025, April). Vlind-bench: Measuring language priors in large vision-language models. In Findings of the Association for Computational Linguistics: NAACL 2025 (pp. 4129-4144).

[2] Jing, L., & Du, X. (2024). Fgaif: Aligning large vision-language models with fine-grained ai feedback. arXiv preprint arXiv:2404.05046.

[3] Wang, X., Cui, Y., Yao, X., Wang, S., Hu, G., & Qin, X. (2025). CHARTHAL: A Fine-grained Framework Evaluating Hallucination of Large Vision Language Models in Chart Understanding. arXiv preprint arXiv:2509.17481.

Questions for authors
  1. How well the proposed benchmark scores aligned with previous hallucination or visual grounding benchmarks?
  2. Does Necessity-Bench can be extended to long text generation (hallucination) benchmarks in addition to CLEVR and GQA?
2 comments
AuthorsOpenReview ↗

Response Part 1

We thank the reviewer for a detailed and constructive review. Every point led to a concrete change, most to a new experiment, including a human annotation study. (Part 1 of 2: W1–W4)

W1. Writing, redundancy, six contributions, long related work. We agree and rewrote the paper.

  • Contributions: six became three (end of §1, p. 2).
  • Related work (§2, pp. 2–3): shortened from 828 to 648 words while adding the FPVG/VisFIS/SwapMix comparison requested by Reviewer AsTm.
  • Results (§4, pp. 5–6): each paragraph now leads with its message, and the detail is in tables.
  • L239–241: the InternVL compatibility remark and similar sentences were removed.

W2. Construction pipeline; target extraction; multiple targets. The full pipeline is in App. A (p. 13), with pseudo-code and per-family statistics (Table 11). §3 (pp. 3–4) summarises it, and Figs. 2–4 (pp. 15–16) show further examples.

  • CLEVR. Each question's functional program is executed on the ground-truth scene graph. The evidence set is every object the final node depends on: the counted set (count), the tested set (exist), the referent (query_*), and both operands for comparisons.
  • Multiple targets. With two evidence objects (902 of 3,000 items), both are removed together, and the counterfactual is recomputed with both excluded.
  • Control. A non-evidence object more than 90 px from all evidence.
  • GQA. The semantic program names the scene-graph object at each step, and their union (≤2) is removed together.

W3. Could artifacts make the VLM misbehave? We answer with human annotation and three controls. Figure 1 (p. 8) now shows each example under both edit mechanisms.

  1. Human annotation (new App. D, p. 15; Table 14, p. 17). An annotator judged 100 random items per tier.
    • GQA: the evidence was removed in 82% of items, and in 40 of the 41 that our detector check marks as removed. Where removal succeeded, the original answer was no longer supported in 85%.
    • CLEVR: where the object was gone, the oracle answer was correct in 95%. Humans also found that the learned inpainter often leaves a coloured remnant, which our residue score detects (AUC 0.91).
    • Screened subset. We fixed a screening threshold from the human labels alone, before computing any model result. It keeps 349 strict items, about 91% removal-correct (95% CI 76–98%), and the main CLEVR numbers are reported on them (Table 5, p. 7).
    • Result. On screened items every model is more grounded than on unscreened ones (necessity +0.12 to +0.26): remnants were hiding grounding, not inflating it.
  2. Paired control (§4, p. 5). The identical edit on an irrelevant object leaves 85–95% of answers unchanged, for every model, tier and edit.
  3. Artifact-only ring control (§4, p. 5; Table 12). Inpainting a ring around the evidence without removing it changes 6–7% of GQA answers and 0.4–1.9% of CLEVR answers, against 40–45% when the evidence is removed.
  4. A second, learned inpainter (LaMa; §3, p. 4). GQA results are unchanged (joint 0.389 vs 0.372, p = 0.48). The submitted grey CLEVR patch overstated necessity, and we report that correction (Table 6, p. 8).

GQA edits remain visible (5% judged natural). Together, the controls reduce concern about editing artifacts: visible editing alone changes only a small fraction of answers.

W4. Human evaluation; frontier models; one model series.

  • Human evaluation: see W3.1.
  • Models: we went from one family to nine models: Qwen3-VL 2B/4B/8B (+32B on GQA), InternVL3.5 2B/8B, Qwen2.5-VL-7B, LLaVA-OneVision-7B and Pixtral-12B (Table 2, p. 5; Table 5, p. 7). On CLEVR, accuracy does not order grounding: Qwen2.5-VL-7B is the most accurate (0.980) but not the most grounded; Qwen3-VL-4B is (0.774).
  • Frontier models: we evaluated the latest generation of leading open-weight VLM families (Qwen3-VL, InternVL3.5, Qwen2.5-VL, LLaVA-OneVision, Pixtral) at 2B–32B, four families and nine models, chosen so that every number is reproducible from public weights. Across this range the key findings do not depend on scale or family. Grounding on real photographs is flat from 2B to 32B (GQA joint 0.31–0.39). In every model, 20–31% of correct answers are not grounded. Because these patterns hold across a 16× parameter range and four independent families, we expect them to persist at larger scale.
AuthorsOpenReview ↗

Response Part 2: W5, W6, Q1, Q2, and a new real-photo result

A strict counterfactual on real photographs (new; §4, pp. 5–6; Table 3, p. 6). For GQA "Are there any X?" questions with exactly one X, the answer after removal must change to no. On 425 detector-checked items:

  • Models change to "no" only 53–62% of the time.
  • The two inpainters agree.
  • 8B is no more grounded than 2B (joint 0.553 vs 0.567).

W5. "Larger VLMs are more grounded" is not surprising. We accept the point. The submission did present the CLEVR scaling effect as its central positive finding, and the revision shifts the emphasis to where scale and grounding diverge. [1] measures language priors, [2] aligns models with fine-grained AI feedback, and [3] evaluates chart hallucination. None tests whether an answer causally depends on the oracle evidence; we now cite and discuss all three (§4, p. 5). On screened interventions:

  1. Domain-dependent. Scale helps on CLEVR in both families tested (Qwen3-VL 2B vs 8B, p = 7e-8; InternVL3.5, p = 0.003) but not on real photographs: no joint difference from 2B to 32B on GQA (paired McNemar, all p ≥ 0.12), and none on the strict real-photo tier.
  2. Not monotonic. Qwen3-VL-4B is at least as grounded as 8B (0.774 vs 0.745, p = 0.25).
  3. Decoupled from accuracy item by item. In every model, 20–31% of correct answers are not causally grounded (§4, p. 5).
  4. Separable by training. For the 2B model, self-report RL improved accuracy and every self-reported metric without a statistically detectable improvement in grounding (§6.1, pp. 7–8; Tables 7–8).

W6. RL setting: data, noisy labels, one self-report format.

  • The labels are exact. The reward compares the model's box with the oracle evidence box from the CLEVR program. "Self-reported" refers to what is optimised, the model's own claim of where it looked (§6.1, p. 7).

  • Data budget. We agree the two recipes are not a controlled ablation: they differ in training data, output format and steps, and training and evaluation share scenes (App. F, p. 18). What we can show:

    • With similarly small budgets, the box reward improved 2B's accuracy and self-reported metrics without a statistically detectable improvement in grounding. On held-out screened removals, necessity went 0.565 → 0.529 (p = 0.18). Under the submitted edit it significantly lowers it (0.538 → 0.475, p = 7e-5; Table 8).
    • The intervention reward raises it (screened removals: 0.679 → 0.751, p = 1e-4; §6.2, p. 9).

    So the two recipes produce different outcomes on grounding. We do not attribute this to the reward signal alone.

  • Other self-report formats (new; §6.3, p. 10; Table 10, p. 10). The model states in one sentence which object it used:

    • It names the oracle object in 75% (2B) / 87% (8B) of items.
    • Naming predicts grounding (p ≤ 0.002), but 28–31% of items are named but not grounded.
    • The irrelevant control object is mentioned in 27–31% of evidence sentences.

    This is one additional format, not a long-form caption evaluation.

Q1. Alignment with hallucination and grounding benchmarks.

  • POPE (App. E, p. 17; Table 15). The eight models fall in a narrow POPE band (F1 0.84–0.90). We found no significant association between POPE and grounding across the eight tested models (Spearman ρ = −0.19, p = 0.65). Qwen3-VL-2B and 4B tie on POPE (0.890 vs 0.896) yet differ by 0.19 in grounding (p = 6e-11).
  • FPVG (App. B, p. 14). Adding our checks to FPVG's conditions rejects 14–22% of its positives.

Q2. Extension to long-text generation. Yes, in principle: the three conditions apply to a claim. The claim should be made from the evidence-only image, retracted or corrected when the evidence is removed, and kept when an irrelevant object is removed, with entailment replacing string equality. Our one-sentence self-report test is a first step, not a long-form hallucination benchmark. The Conclusion (§7, p. 10) states this as future work.

We hope the human annotation, the stricter metric and the new controls address the soundness concerns.

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

The paper presents Necessity-Bench to study a VLM causality branch. It uses image interventions to measure sufficiency, necessity, and specificity. Experiments are conducted on CLEVR and GQA. RL is also explored, and the authors claim improvements in conventional grounding metrics while reducing evidence dependence.

Critical strengths
  1. A good combination of independently derived object-removal counterfactuals, correct-answer verification, and paired irrelevant-removal controls.
  2. The RL results showing that better accuracy and localisation do not necessarily mean better grounding are interesting.
Critical weaknesses

Similar self-report-free, evidence-intervention methods already exist. I feel the paper does not sufficiently establish its added value over these approaches. Some relevant papers are:

  • Reich et al., Measuring Faithful and Plausible Visual Grounding in VQA, Findings of EMNLP 2023. - closest reference. Relavent but not as close as the FPVG above:
  • Reich & Schultz, Uncovering the Full Potential of Visual Grounding Methods in VQA, ACL 2024.
  • Ying et al., VisFIS, NeurIPS 2022.
  • Gupta et al., SwapMix, CVPR 2022.
Questions for authors
  1. What does Necessity-Bench offer beyond FPVG and similar counterfactual evaluation methods? Could the authors provide a direct comparison?
  2. How can the benchmark be used to improve VLMs?
1 comment

We thank the reviewer for recognizing our combination of recomputed counterfactuals, answer verification and paired controls, and for the prior-work pointers. We now discuss and directly compare against all four: §2 (p. 2), Table 1 (p. 3), App. B (p. 14) and Table 13 (p. 17).

W1/Q1. Added value over FPVG, Reich & Schultz, VisFIS and SwapMix

All four intervene on the detector object features (Faster R-CNN) consumed by classical region-based VQA models. Generative VLMs read pixels through a ViT and have no object slots, so these interventions do not apply to them as defined.

  • FPVG (Reich et al., 2023), the closest. It counts a question as grounded if the answer is kept with relevant-only features and changes in any way with irrelevant-only features. FPVG already shows across several VQA architectures that accuracy and grounding can diverge, as we do. Our additions:
    • (i) Pixel-level intervention on generative VLMs, with removal checks calibrated against human judgement (App. D).
    • (ii) The new answer must change to the recomputed counterfactual, obtained by re-executing the question program on the edited scene (§3, p. 3). Human annotation finds this oracle correct in 95% of CLEVR items where the evidence was removed.
    • (iii) A paired control: the identical edit on an irrelevant object must not change the answer.
  • Reich & Schultz (2024) show that grounding methods are underestimated when relevant content is missing from imperfect detector features. Pixel-level removal avoids that feature-coverage issue by construction.
  • VisFIS (Ying et al., 2022) combines feature-importance supervision with right-for-the-right-reason objectives and also defines evaluation metrics. Its interventions act on object features, and it does not score the answer against a recomputed counterfactual or pair it with a same-edit control.
  • SwapMix (Gupta et al., 2022) measures reliance on context by swapping irrelevant objects' features. That is the counterpart of our specificity test. It does not test whether the relevant evidence is necessary.

Direct, matched comparison on identical model answers (App. B; Table 13). The criteria are not nested, so we compare both directions on the 349 screened CLEVR removals, all eight models:

  • Forward. Adding counterfactual correctness and the paired control to FPVG's own conditions rejects 14–22% of the items FPVG accepts. Of these, 3–12% changed to the wrong answer and 8–13% also flip under the irrelevant control.
Of FPVG+ items:changed to a wrong answerflip under the controlrejected
Qwen3-VL-2B10%13%22%
Qwen3-VL-8B4%10%14%
LLaVA-OV-7B12%8%20%
all 8 models3–12%8–13%14–22%
  • Reverse. 2–10% of the items our joint metric accepts fail FPVG. All of them fail FPVG's relevant-only condition: the evidence-only answer differs from the model's own original answer. Our sufficiency metric is scored against the ground truth and reported separately (§3, p. 3).

So the added checks remove a consistent 14–22% of FPVG's positives: models that react to the edit without tracking the evidence, or flip under any edit. That is our measurable added value: verified pixel interventions, correct counterfactual answers, and a paired control.

Q2. How can the benchmark improve VLMs? (§6.2, p. 9; Tables 7–9, pp. 9–10). The oracle supplies the correct answer for every edited image, so the interventions form a verifiable RL reward with no self-report. With GRPO on CLEVR questions disjoint from evaluation:

  • Qwen3-VL-2B, screened removals: necessity 0.679 → 0.751 (p = 1e-4) and joint 0.587 → 0.650 (p = 7e-4). The necessity gain holds under an edit unseen in training and in three seeds (Table 16, p. 18). GQA is unchanged (joint 0.391 vs 0.389).
  • A self-reported box reward (§6.1, pp. 7–8; Table 8) improved accuracy and every self-reported metric without a statistically detectable improvement in grounding. On held-out screened removals, accuracy rose 0.748 → 0.827 (p = 4e-4); necessity went 0.565 → 0.529 (p = 0.18), which establishes neither degradation nor equivalence. Under the submitted edit it even lowers necessity (0.538 → 0.475, p = 7e-5).
  • Qwen3-VL-8B improves under the training edit (600 steps: 0.765 → 0.824, p = 4e-9), but not significantly on screened removals.
  • Limits. Training and evaluation share scenes, and the two recipes differ in more than the reward (App. F). We therefore report them as two recipes with different outcomes on grounding, not as an isolated reward effect.

Beyond the metric, none of the four calibrates its interventions against human judgement, studies how grounding changes with model scale, or contrasts self-report with intervention rewards. We hope this clarifies the added value.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To the reviewers, Area Chairs, Senior Area Chairs and Program Chairs

We sincerely thank the reviewers for careful and constructive reviews. We have uploaded a revised PDF with every change in blue.

What is new. Necessity-Bench is a self-report-free grounding test for generative VLMs that intervenes on pixels, not detector features. It differs from prior intervention tests (FPVG, VisFIS, SwapMix) in three ways. (1) The answer must change to a recomputed counterfactual, not just change. (2) Every item is paired with the identical edit on an irrelevant object. (3) Removals are checked against human judgement. On identical model answers, these checks reject 14–22% of the items FPVG accepts as grounded (App. B), for each of eight models.

Human validation. Both reviews asked whether the edited images and their oracle answers can be trusted. We answered with human-annotated data rather than argument. An annotator inspected 200 randomly sampled items image by image: the original, the evidence-removed edit, the control-removed edit and the oracle answer. We then calibrated an automatic removal screen on those labels, re-scored every model, and recomputed every statistical test. This took most of the discussion period. We chose to finish it rather than respond earlier with weaker evidence, and we apologise for the late response.

What the human annotation shows (App. D, p. 15; Table 14, p. 17)

  • The oracle is right. Where the evidence was removed, the oracle answer was judged correct in 95% of CLEVR and 85% of GQA items.
  • Our automatic checks agree with humans. The GQA detector check is 98% precise (40/41). The CLEVR residue score predicts human judgement with AUC 0.91.
  • Headline results use screened removals (Table 5, p. 7). The threshold was fixed from human labels before any model result was computed; screened items are about 91% removal-correct (95% CI 76–98%).

Why this matters for validity. The human labels confirm the two assumptions the benchmark rests on: the oracle answers are correct, and our automatic checks identify real removals. The bias we found runs against us: unverified remnants made models look less grounded. Our grounding estimates are therefore conservative, not inflated.

A stricter metric. While re-auditing, we tightened strict necessity to require an actual answer change to the correct counterfactual (§3, p. 3). Every number was recomputed, and every main conclusion holds.

How we addressed the reviews

  • Prior work (AsTm): paper-by-paper differentiation and a direct, two-way FPVG comparison (§2; App. B).
  • Improving VLMs (AsTm): the interventions are an RL reward with no self-report; on screened items it raises a 2B model's necessity (0.679 → 0.751, p = 1e-4).
  • Soundness (bSLP): human annotation, a paired control (85–95% of answers unchanged), an artifact-only control (6–7% of GQA answers change vs 40–45% for real removal), a second inpainter, and a strict real-photo tier.
  • Breadth (bSLP): nine current-generation open-weight models from four families (2B–32B), POPE, and a free-text self-report test.
  • Presentation (bSLP): three contributions instead of six, a shorter related work, the full pipeline with pseudo-code.

What the field learns

  • Accuracy ≠ grounding. On screened removals, 20–31% of correct answers are not causally grounded, in every model.
  • Scale depends on domain. It improves grounding on CLEVR in both families tested (Qwen3-VL p = 7e-8; InternVL3.5 p = 0.003), but not on real photographs (2B–32B).
  • POPE. We found no significant association between POPE and grounding across the eight tested models.
  • Self-report training. For a 2B model, optimising a self-reported box improved accuracy and every self-reported metric without a statistically detectable improvement in grounding.

Necessity-Bench measures whether an answer depends on its evidence, for any pixel-input VLM, with four forward passes and no training. We release the benchmark, the annotation sheets and the removal checks so that every number can be verified.

Thank you again for your time and guidance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.