acceptodds
Under review as a conference paper at ICLR 2027

When Seeing Becomes Saying: Nonmonotonic Answer Dependence in Visual Chain-of-Thought

Abstract

A numerical value can continue to influence a vision-language model after its written positions have been blocked. In visual chain-of-thought, separate and joint masks test answer dependence, while numerical substitutions and state transfer trace the value that remains influential. Individual masks miss joint dependencies on 51 of 119 Qwen and 73 of 119 InternVL TabMWP traces under blank-image continuation. Enlarging a mask can also restore an answer, with complete-answer controls showing how recovery changes across intervention operators. Replacing the source value while blocking all digit positions tests which number still reaches the answer. Across 800 controlled scenes, nonnumeric-state transfer favors the used replacement over its unused alternative by 3.56 points for Qwen and 0.94 for InternVL. Access to a written number and the influence of its value therefore require different tests. These findings motivate Gate, which chooses between retaining CoT computation and answering afresh. Gate improves over voting by 1.57 points across development benchmarks. On 600 new ChartQA images, frozen Gate reaches 85.00%, matching candidate scoring and improving over voting by 1.67 points.

Primary area:
foundation or frontier models, including LLMs
Keywords:
vision-language models, chain-of-thought, visual reasoning, diagnostic evaluation, answer reuse, self-consistency

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wFHSOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the distinction between "accessing the position of numbers" and "being influenced by those numerical values" in visual Chain-of-Thought reasoning. Keeping the existing reasoning fixed, the authors show through individual and joint masking of number positions that independent masking can miss joint dependencies, and expanding the masking region can even paradoxically restore the original answer. Further numerical replacement and non-numerical state-transfer experiments reveal that even when number positions are masked, numerical influence can still propagate via internal states at other locations. Based on these findings, the authors propose Gate, which selects between CoT voting answers and independent direct answers based on response features. The approach is validated through internal interventions on two primary models, 800 synthetic scenarios, 400 natural error trajectories, and answer-selection experiments across multiple models and tasks. Gate outperforms voting on new ChartQA images and matches an answer scorer using matched features.

Critical strengths

Uncovers interesting and diagnostically valuable phenomena. The paper demonstrates that a visual CoT's reliance on numerical information cannot always be determined by independently masking corresponding positions: masking two sets of numbers separately may not affect the answer, whereas joint masking changes it; expanding the masking region can even restore the original answer. These phenomena serve as a timely reminder to interpret models' information usage with caution when relying solely on local ablation results.
Provides systematic analysis and experimental validation. The authors combine various masking operations, numerical replacements, non-numerical state transfers, and paired replacement-value controls, further testing them across independent synthetic scenarios and natural error trajectories. In particular, the state-transfer experiments provide more direct evidence of numerical propagation than simply observing whether the answer is preserved, elevating the findings beyond isolated case studies.

Critical weaknesses

Insufficient independent advantage, cross-scenario stability, and cost-benefit justification for Gate. Compared to voting, Gate incurs the extra cost of generating direct answers, yet its advantages over more competitive selection strategies are limited. On new ChartQA images, both Gate and the answer scorer using matched features achieve 85.00%, with identical sample-by-sample correctness. Its advantage over a fixed fallback is also modest, and it fails to consistently outperform the fixed fallback in classification and localization extension tasks. Thus, while the results support the value of learned answer selection, they fall short of demonstrating clear, stable additional benefits specific to Gate.
Limited mechanistic explanation and guidance value for real-world errors. Although the paper performs attention position masking, value zeroing, and state transfer, the analysis primarily characterizes output changes under intervention. It does not clearly explain the specific formation mechanisms underlying joint dependencies and non-monotonic recovery, nor does it establish the necessity of the identified pathways in natural responses. Across 400 natural error trajectories, non-numerical state transfer detected a positive numerical-following effect in only 3 images, with the overall confidence interval spanning zero. Consequently, while the paper convincingly proves that these phenomena can occur, there remains insufficient evidence regarding their importance for real-world visual reasoning failures or whether they can guide effective error diagnosis and correction.

Questions for authors

Cost-benefit and stability of Gate. On new ChartQA images, Gate performs identically to the matched-feature answer scorer, and its advantage over a fixed fallback is minor. What results support the independent methodological value of Gate? When accounting for the extra direct generation calls, is there a comparison under the same inference budget demonstrating that it outperforms simpler strategies like increased sampling?
Internal mechanisms behind the phenomena. To what extent can the current attention intervention and state-transfer experiments explain joint dependencies and non-monotonic recovery? Specifically, how can one disentangle the respective roles of numerical information propagation, competing numerical influences, and attention normalization changes caused by masking?
Diagnostic value for real-world errors. In the 400 natural error trajectories, non-numerical state transfer detected a positive numerical-following effect in only 3 images. How can one distinguish whether this stems from limited phenomenon coverage or from precise numerical matching missing computationally transformed effects? What practical error identification or correction use cases are supported by the current findings?

1 comment

Thank you for recognizing the diagnostic value and raising questions about Gate, mechanism and practical use.

1. Gate: selection value and cost.

Gate learns when to retain CoT-supported Vote or answer afresh with Direct. Matched shares the candidates, 68 features and data splits, but predicts candidate correctness rather than Direct-minus-Vote correctness gain.

EvaluationGateMatchedFixed fallback
Grouped development, 19-cell macro82.18%82.08%80.64%
Frozen ChartQA, 600 images, 2 models85.00%85.00%84.42%
Benchmark holdout, 19-cell macro81.92%81.89%80.64%

Gate exceeds fallback by 1.54 pp [1.19,1.89] in development and 1.27 [0.92,1.62] under benchmark holdout; it exceeds a training-fold-tuned agreement threshold by 0.47 [0.20,0.76] in development. Gate-minus-Matched is +0.10 [−0.01,0.21] and +0.02 [−0.14,0.18], respectively (all 95% CIs). Frozen correctness is identical for Gate and Matched; the prespecified Gate-minus-fallback difference is +0.58 [−0.67,1.83] (97.5%). Grouped and benchmark-holdout results support learned selective reuse over fallback; the gain objective's extra accuracy over Matched remains unestablished.

We measured six strategies separately on the same first 100 ChartQA images for both 8B models, using one RTX A6000, batch size 1. Time includes image loading, visual encoding, prefill, generation and selection, excluding model loading and warmup. Values are accuracy (%) / mean seconds per image; IF denotes all-invalid Direct fallback.

StrategyQwenInternVL
Gate83 / 15.6885 / 13.76
Matched83 / 15.6685 / 13.74
Fixed fallback82 / 15.7388 / 13.78
Vote-3+IF81 / 15.7285 / 13.62
Vote-4+IF82 / 20.9685 / 18.06
Adaptive fourth CoT82 / 16.9785 / 14.38

Adaptive sampling adds a fourth CoT under Gate's non-unanimity condition and votes over four, using Direct if all are invalid. Gate and adaptive sampling average 3.16/3.14 calls (Qwen/InternVL); token counts differ, so this matches calls, not exact compute. Gate's mean time is 25.2%/23.9% lower than Vote-4+IF and 1.30/0.62 seconds lower than adaptive sampling. Timing is newly measured; accuracy replays the existing fixed seed.

Across five seed/temperature conditions, Gate exceeds Vote-4+IF by 0.5–1.0 pp; all exploratory 95% paired intervals reach zero (Response Evidence Table E5).

2. What the intervention controls establish.

Earlier prefill blocking preserves TabMWP joint dependence, including with original images. Value zeroing retains keys and softmax normalization at the intervened call; later states can evolve. Paired value transfer moves nonnumeric states from an edited run into the recipient while preserving its text, testing continued influence of the replacement value.

The submitted numerical-competition control compares a competing-value mask with an equal-size control on selected blank-image cases:

ModelCasesRecovery difference (pp)Holm-adjusted exact p
Qwen30+50.000.0001221
InternVL25+28.000.06543

Only Qwen passes the adjusted exact test; the joint two-model criterion is not met. Mask locations and content also differ. This constrains the competition explanation in selected cases without fully decomposing propagation, competition and normalization. Recovery is intervention-sensitive and cannot alone establish numerical irrelevance.

3. Natural errors and diagnostic utility.

On 400 original-image Qwen ChartQA error traces, changing the replacement value changes numerical outputs on 86 images under source editing and 54 under nonnumeric-state transfer; 73/51 respectively have no exact replacement match. Exact transfer yields three positive images and a net used-minus-unused effect of 0.38 pp (97.5% CI [0.00,0.88]). This distinguishes output sensitivity from literal numerical following.

An exploratory AI-assisted audit of edited prefixes found 33 agreed direct mappings, 184 provisionally ineligible cases and 183 unresolved cases. No non-identity transformation met confirmation criteria. Many edits replace a completed result rather than an upstream operand, preventing a unique transformed target without additional assumptions. The transformed-following effect therefore remains unestimable, rather than showing that transformations do not occur (Response Evidence, Section 6).

The supported use is controlled diagnosis of answer dependence: identifying joint reliance missed by local ablations and testing continued influence of substituted values. Effective localization or correction of natural visual errors remains a future application.

Our diagnostic experiments establish how local interventions can mischaracterize numerical dependence; Gate evaluates the practical choice between reusing CoT computation and answering afresh.

Reviewer_R3KyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at how numbers written in a visual CoT affect the final answer of VLMs. The first part uses interventions on a fixed CoT trace. It masks the digit positions at answer time, and then replaces the source value and transfers hidden states at nonnumeric positions. The main findings are that masking two groups of digits together can change the answer even when each mask alone does not, that enlarging a mask can sometimes restore the original answer, and that a substituted value can still influence the answer when all digit positions are blocked. The second part proposes Gate, a ridge regressor on agreement and response features that decides between keeping a majority-voted CoT answer and answering afresh with Direct prompting.

Overall, the paper is careful in places, but I don’t think the main claims are well supported, and the two parts don’t connect.

Critical strengths
  • The text-only replay result is useful on its own. When the CoT is given back as text without the image, the model reproduces wrong answers 63–95% of the time (Table 11b), so a reading error written in the trace tends to get locked in. Together with the finding that CoT is still needed with exact transcriptions (100% vs 1.7–6.7% for Direct), this gives a clear picture: CoT helps computation but carries reading This finding is a useful, practical takeaway.

  • The experiments and evaluation are comprehensive. Especially, it tests 11 models and many tasks. Some results like CoT vs Direct grid alone is informative.

Critical weaknesses
  • The main finding is largely expected from the intervention design. The masks only block digit keys at answer time, while earlier states are left unchanged, so the value is already in the other positions’ hidden states. Transferring those states then carries the same information. This shows that position access and value in states differ, but that is how a transformer works.

  • The non-monotone recovery also changes a lot with the operator, which might suggest a mask artifact.

  • The effects are small. In the 800-scene study, the used value is followed 6.4% vs 1.1% for Qwen and 1.7% vs 0.4% for InternVL. With the real image on ChartQA, transfer gives 0.38 points. The larger numbers come from DVQA cases selected for recovery (n=23, 14), and most results use blank images.

  • Gate is weakly connected to the analysis and only adds a little.

Questions for authors
  • What happens if you intervene on the digits earlier, during the prefill of the CoT, instead of only at answer time? This would separate position access from value in states more cleanly.
  • Do any Gate features come from the interventions? If not, what is the actual link between the two parts?
1 comment

Thank you for suggesting earlier intervention. We ask when local numerical ablations provide reliable evidence about answer dependence; we do not claim that information propagation beyond a token's position is itself novel.

Matched timing and operator controls. We use the same recorded CoT tokens and 32 images per task/model as the submitted complete-answer controls. Answer-stage hard masking starts at the answer tag; prefill masking starts at the first CoT query while re-encoding the same fixed text. Both continue through answer generation. We also compare answer-stage value zeroing, which retains keys and softmax normalization at the intervened call. Other settings remain fixed.

Event / imageModelAnswer hard maskPrefill hard maskAnswer value zero
TabMWP joint / originalQwen598
TabMWP joint / blankQwen162715
TabMWP joint / originalInternVL263
TabMWP joint / blankInternVL192821
DVQA recovery / blankQwen1213
DVQA recovery / blankInternVL502

Counts are out of 32 images per row; one invalid unmasked InternVL/DVQA blank continuation cannot define an event. Pooling models within image, prefill versus answer-stage hard masking increases joint dependence by 12.50 pp [3.12, 23.44] on original images and 31.25 pp [17.19, 46.88] on blank images; blank-DVQA recovery decreases by 25.00 pp [12.50, 37.50] (exploratory 95% paired-image CIs).

Joint dependence thus persists under earlier blocking, including with original images. Recovery changes substantially with timing and operator, so recovery alone is not a reliable indicator of numerical irrelevance. Prefill re-encodes fixed text; it neither regenerates reasoning nor erases information already expressed in remaining tokens. Value zeroing retains keys and normalization at the intervened call, while later states can evolve. These controls narrow alternatives but do not fully decompose numerical competition from subsequent attention changes.

Effect size and cohort selection. The synthetic study includes all source-eligible scenes, not only recovery cases. Under blank images and blocked digits, source-edit following of used versus unused replacements is 6.44% versus 1.12% for Qwen (800 images) and 1.70% versus 0.44% for InternVL (795 images). For nonnumeric-state transfer, rates are 4.81% versus 1.25% and 1.38% versus 0.44%, a net used-value preference of 3.56 and 0.94 pp. All four prespecified synthetic contrasts have positive lower bounds at the Bonferroni-adjusted 98.75% level. In the separate 400-image Qwen natural-error cohort with original images, transfer has three positive images and a net effect of 0.38 pp (97.5% CI [0.00, 0.88]); broad prevalence is not established.

What transfer establishes. Separate masks can miss joint dependence, and unchanged or recovered answers do not establish numerical irrelevance. Paired value transfer changes nonnumeric states from an edited run while preserving recipient text, testing whether replacement identity still influences the answer when digit positions are blocked. Together, joint restrictions, timing/operator controls and transfer distinguish written-position access from continued numerical influence, without claiming complete mechanistic decomposition or natural-pathway necessity. We cite attention knockout (Geva et al., 2023), activation patching (Zhang & Nanda, 2024), and CoT value editing (Zhu et al., 2025).

Connection to Gate. The link is selective computation reuse, not a direct mapping from diagnostic labels to routing. Text-only replay and exact-transcription results show that CoT can preserve useful computation but also carry forward an erroneous visual observation. This motivates choosing between retaining the Vote supported by existing CoTs and answering afresh with Direct. Gate predicts the correctness gain of Direct over Vote from output-derived agreement, dispersion, length, lexical and task/model features; intervention outcomes are not features or training labels.

On 600 frozen ChartQA images (1,200 model–image records), Gate replaces 28 valid votes: 10 errors corrected and 2 correct answers spoiled, net +8 (+0.67 pp) beyond all-invalid fallback. In the fixed first-100-image same-trace analysis, eligible subsets contain 46 Qwen and 43 InternVL records across 49 images. Gate selects Direct once per model and leaves Vote's total correct count unchanged; numerical-group subsets contain one joint-dependence event and one recovery. These sparse observations do not establish a stable diagnostic-label/routing association. Gate's selection benefit is evaluated separately.

We hope these controls clarify why joint dependence persists under earlier intervention, why recovery is intervention-sensitive, and how the diagnostic findings motivate selective reuse without serving as its causal validation.

Reviewer_s4CuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks how the numbers a vision-language model writes in its chain-of-thought affect its final answer. The authors take a CoT the model has already written, keep it fixed, and let the model produce only the final answer again while blocking it from looking at some of the numbers in the CoT. They report three findings. Blocking two sets of numbers one at a time can leave the answer unchanged while blocking both changes it, and blocking more numbers can sometimes bring the original answer back. When they edit a number in the CoT and block every digit, the answer can still follow the edited value, which they take to mean the value is carried by the tokens around it. Also they propose a method to improve task accuracies of vision language models. The proposed Gate method will sample three CoTs, keep the answer if all three agree, and otherwise let a small regression model choose between the majority answer and a direct answer given without reasoning.

Critical strengths

None

Critical weaknesses
  1. The paper is very hard to follow, and its main ideas cannot be understood from the main text. The abstract never says what problem the paper solves. It opens with a claim, jumps to counts such as "51 of 119 Qwen ... traces," and uses phrases like "nonnumeric-state transfer favors the used replacement over its unused alternative" that a reader cannot decode before Section 3. The introduction does the same with "mask," which appears many times before Section 3.1 explains what is masked and how. Later, blocking all digits is called All, union, numerical masking, Masked and U in different places. Section 3.2 even defines one result relative to the "source answer" and the next relative to the "native answer," which are not always equal. Section 3 lists steps but rarely says why they are taken. The most important design choice is that blocking applies only while the final answer is generated, so the words written after a number have already seen it. Section 5.3 depends on this, yet it gets one sentence. A figure showing where blocking applies, and how the two blocking methods differ, would help a lot. The group names Answer, Other, All and None are ordinary words, which makes sentences like "All combines these groups, and None leaves the trace unmasked" hard to parse. The related work section lists papers in one clause each without saying how this paper differs from them. Some links between sections also do not hold. Section 5.4 shows that a larger accuracy gap between real and blank images does not mean better answers, then says this "motivates" Gate, which does not follow.

  2. The findings are not clearly new, the method does not build on them, and the evidence for its gains is weak. Blocking attention to certain positions and copying internal activations from one run into another are standard tools in interpretability, usually called attention knockout and activation patching. The paper does not cite this work. It does cite Zhu et al. (2025), who already showed that changing numbers in a CoT changes the later computation, which is close to the paper's own editing experiment, but it never compares against them. I also do not find the main result surprising. In a decoder, every token after a number can read that number, so blocking the number's own position at answer time will not remove it from the words that follow. Outside hand-picked cases the effect is small: about 3.6 and 0.9 percentage points on 800 synthetic charts, and 0.38 points on real ChartQA errors. Gate uses none of these findings. In effect it is majority voting with a learned fallback, and it barely beats simpler versions: by 0.47 points over a tuned agreement rule and by 0.10 over a scorer with the same inputs, with identical results on the new ChartQA images. The statistics cannot support differences this small. Tables 1 to 3 report no variance or confidence intervals, and nothing is repeated across random seeds, although Vote and Gate depend on sampling and Gate's margins range from 0.1 to 1.6 points. Temperature and the number of samples are never varied. The comparisons in Table 4 change several things at once, so it is unclear where the gain comes from, and Vote does not get the fallback Gate uses when all samples are invalid. Finally, every model has at most 9B parameters and the main experiments use only two 8B models. Gate needs only model outputs, so it could easily have been tested on stronger models.

Questions for authors
  1. How often does Gate choose the direct answer instead of the vote, and how accurate is each choice when it is made?
  2. How much do the results vary across random seeds, and how do they change with sampling temperature and the number of samples?
  3. If Vote also falls back to the direct answer when all samples are invalid, how much of Gate's gain remains?
  4. Did you compare editing the source value without blocking the digits (as in Zhu et al., 2025) against editing with all digits blocked? The difference would show how much of the effect passes through the indirect route. Relatedly, how does your contribution go beyond the known information-leakage limitation of attention knockout discussed by Geva et al. (2023)?
  5. Have you tried replacing every digit with a neutral placeholder, recomputing the CoT and then blocking? Comparing this with blocking alone would show how much of the effect passes through the words after each number.
1 comment

Thank you. We test what local digit ablations establish about numerical dependence.

Definitions. Recorded CoT tokens are fixed and re-encoded, not regenerated. Masks start at the answer tag or first CoT query (prefill) and continue through answering. Answer/Other partition answer-value/remaining digit-bearing positions; All is their union. Retention refers to the original CoT answer (source); joint dependence/recovery to the valid unmasked answer (native); accuracy to ground truth. See Sections 3.1–3.2 and Fig. 2(d). We cite Geva et al. (2023), Zhang & Nanda (2024) and Zhu et al. (2025).

Uncertainty. Paired-image CIs: Table 4/App. E.1 (Gate), Table 14/B.3 (grouping), Table 27/C.2 (synthetic). Value zeroing minus hard masking changes blank-DVQA recovery by −28.13 [−43.75,−12.50] pp (Qwen) and −9.38 [−21.88,3.13] (InternVL), descriptive 95% CIs (B.5).

Q4. Edits and diagnostic value. Prefixes, blank images, checkpoints, seeds and budgets are matched; only answer-stage digit access changes. Readable/blocked used-value following is 99.81/6.44% (Qwen, n=800) and 99.31/1.70% (InternVL, n=795). Differences are 93.38 [92.06,94.62] and 97.61 [96.73,98.43] pp (95% paired CIs). All 3,190 new readable continuations and their archived blocked counterparts are valid. This is a behavioral contrast, not an additive mediation estimate.

Under blocked digits, source-edit outputs follow unused replacements at 1.12%/0.44% (Qwen/InternVL). Nonnumeric-state transfer tests value identity:

ModelUsed (%)Unused (%)Net (pp)
Qwen4.811.253.56
InternVL1.380.440.94

We average two replacement runs per image. All source-eligible scenes are included, without recovery selection. All four prespecified contrasts have positive Bonferroni-adjusted 98.75% lower bounds. On 400 Qwen natural errors with original images, transfer gives 3 positive images; net 0.38 pp (97.5% CI [0.00,0.88]). Broad prevalence is unproven.

Separate masks can miss joint dependence; recovery does not imply numerical irrelevance. Geva's leakage caveat is known; unlike continued-CoT value editing, our transfer preserves recipient text. On matched 32-image cohorts, answer-stage→prefill joint-dependence counts are 5→9/2→6 (Qwen/InternVL) with original TabMWP images and 16→27/19→28 with blank images. Blank-DVQA recovery falls 12→1/5→0. Joint dependence persists under earlier blocking, while recovery is timing-sensitive; fixed text can still carry information.

Q5. Neutralization. Each digit-bearing token becomes one # token; other text and prefix length stay fixed. After re-encoding, answer-stage masking uses the original positions. Blank-image source-answer retention (%):

Model/tasknOriginal blockedNeutral unmaskedNeutral blocked
Qwen/DVQA9870.4113.277.14
InternVL/DVQA10047.0011.006.00
Qwen/TabMWP9952.539.098.08
InternVL/TabMWP9735.0511.347.22

Original unmasked retention is 100%. Much of the drop precedes masking; this control does not isolate all information in later words.

Gate/Q1. Gate retains valid unanimous CoT answers; otherwise it predicts the correctness gain of fresh Direct over Vote from output features, without intervention-derived features or labels. On 600 shared ChartQA images/two models (1,200 records), Direct is selected 44 times: 29/44 correct, versus 991/1,156 for retained Vote (different subsets). Switching corrects 22 and spoils 2, net +20; staged execution requests Direct 226 times.

Q3. Vote+IF uses Direct only when all CoTs are invalid and Direct is valid. Gate gains 1.04 pp [0.66,1.42] on development (95%) and 0.67 [0.17,1.25] on frozen images (exploratory 95%). Of 22 corrections, 12 are invalid fallback; replacing 28 valid votes corrects 10 and spoils 2 (net +8). Gain over tuned agreement is 0.47 [0.20,0.76] pp (95%). Frozen Gate and Matched both score 85% with identical correctness; the prespecified Gate–fixed-fallback interval includes zero (common response).

Q2. Sampling. Mean accuracy (%) over Qwen3-VL-8B and InternVL3.5-8B on the same first 100 frozen images; A/B/C are independent pools; Direct is greedy.

Seed/TGateVote-4+IFGate–fallback CI (pp)
A/0.784.083.5[−3.5,1.5]
B/0.784.584.0[−4,1.5]
C/0.786.085.0[−2.5,0]
A/0.584.584.0[−2.5,2]
A/1.083.583.0[−5.5,−0.5]

All CIs here are exploratory 95% image-paired. Gate–Vote-4+IF intervals are [0,1.5], except C/0.7 [0,2.5]. Gate–Matched is 0 [0,0] in four rows and −0.5 [−1.5,0] at A/1.0; fixed fallback ties or exceeds Gate by 0–2.5. Details: Response Evidence Table E5. This is not a full factorial.

Frozen InternVL3.5-14B (100 images, InternVL family alias): Gate/Matched/Vote+IF 84%, Direct/fallback 86%; Gate selects Direct 0/100. Costs: wFHS reply. Gate applies selective reuse; diagnostics rest on intervention evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.