acceptodds
Under review as a conference paper at ICLR 2027

When to Trust Context: A Causal Evidence Gate for Occlusion-Aware Object Detection

Abstract

Transformer-based object detectors aggregate scene context through cross-attention causing query representations to entangle an object's direct visual evidence with contextual co-occurrence priors. Traditionally treated as distinct challenges, we demonstrate that this contextual entanglement underpins two opposing failure modes: (i) co-occurrence bias where models over-rely on spatial surroundings, impairing generalization under distribution shifts and long-tail categories; and (ii) partial observability where context serves as the sole signal for recovering heavily occluded objects. We unify these phenomena through a single causal framework and introduce Evidence-Gated Backdoor Adjustment (EG-BA) for Open-Set Object Detection (OSOD). EG-BA employs a learned per-query evidence score to dynamically interpolate between two regimes: when visual evidence is reliable, it executes a backdoor adjustment to strip confounding context; when evidence is degraded, it leverages context for identity recovery. To ensure the internal mechanism behaves in the exact way, we regularize the latent space via a variational information bottleneck and propose the Gate-Occlusion Calibration Error (GOCE) a diagnostic evaluation tool to understand the gate aligns with true physical occlusion. We investigated that EG-BA improves calibration and unknown-class recall over a vanilla Deformable DETR baseline model on Pascal VOC to MS-COCO dataset and its evidence score tracks synthetic occlusion severity when properly regularized, without this regularization the same information-bottleneck term drives the gate to a collapsed input-independent constant state instead, considered as a failure mode, we report as a central finding of independent interest. We further find that this occlusion-robustness gain does not transfer to real, naturally occurring occlusion and that the mechanism remains steady for co-occurrence-driven performance degradation, the bias it is designed to remove.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
open-set object detection, causal inference, backdoor adjustment, occlusion robustness, co-occurrence bias, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_td5eOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper starts from an interesting tension: context can be harmful when it acts as a shortcut, but useful when direct object evidence is missing. The paper proposes Evidence-Gated Backdoor Adjustment (EG-BA), a module for transformer-based open-set object detection. The motivating hypothesis is that an object query mixes direct visual evidence with scene context. The authors argue that the same context produces two opposing effects: it can induce co-occurrence shortcuts when the object is visible, while helping recognition when the object is occluded.

Critical strengths

The motivating observation about the value and problem of context is worthwhile. Framing context as conditionally useful rather than uniformly harmful is nuanced and makes the paper have an interesting prelude.

The authors report important negative findings, they don't hide them. Some things didn't work and they report it. E.g. gate collapse under the information-bottleneck term, lack of transfer to natural occlusion, and failure on the co-occurrence-shift test. This transparency is valuable.

The ablations include relevant variations on the topic. E.g. evidence supervision, threshold sensitivity, gate distributions, and synthetic occlusion severity. These analyses expose failure modes.

Critical weaknesses

No solid evidence is provided that the learned dictionary identifies confounders rather than generic feature prototypes. That would have been very valuable. But there is no intervention, prototype visualization, or controlled experiment to demonstrate its effects.

The information-bottleneck is presented as a contribution, but it collapses the gate and is set to zero in the final model. The claim of the authors is that an adaptive balance is needed between debiasing and context recovery, but the context-only performs best.

The method fails on co-occurrence bias and natural occlusion, which it target explicitly.

The paper lists Don’t Judge an Object by Its Context, but it does not discuss it seriously, neither do they compare against it. Causal attention and context-debiasing methods are discussed but not taken as baselines.

Occlusion-aware literature seem to be a few and mainly older papers. The paper would be stronger when modern transformer-based approaches are discussed as well.

Questions for authors

What is the relation and difference to "Explicit Modeling of Causal Factors and Confounders for Image Classification", AAAI 2026 by Wei Wu, et al.?

Can you provide a matched non-causal gating baseline with the same architecture and parameter count, so the contribution of the gating can be emphasized?

Can you compare against context-debiasing and occlusion-aware methods under the same Deformable DETR backbone and training protocol?

2 comments
AuthorsOpenReview ↗

Response to Reviewer td5e: dictionary analysis, baselines, limitations

Thank you for the detailed review, and for recognising the negative findings. All new numbers come from one training run per configuration (single seed); we flag where this limits the conclusion.

W1 (no evidence that the dictionary identifies confounders). Fair. We ran a prototype analysis on 3,894 confident predictions (K=20): attention is diffuse (entropy 2.75 vs max 3.00; mean max weight 0.15 vs 0.05 uniform), all 20 prototypes are used, and NMI between the top prototype and predicted class is 0.16. We will not claim the prototypes are semantic confounders and will describe the dictionary as a learned context basis. We have no intervention experiment.

W2 (IB as a contribution; context-only performs best). Agreed on both. At lambda_IB=0.1 the gate collapses to a constant (mean 0.0179, std 0.0000); the headline uses lambda_IB=0 plus a small anti-collapse term. We will reframe IB as a documented negative finding. Fixed e=0 has higher mAP (52.82 vs 51.20) and U-Recall (0.0689 vs 0.0573); the adaptive gate is better on WI and A-OSE. We claim only this trade-off.

W3 (fails on co-occurrence and natural occlusion). Agreed: Table 7 shows parity and Table 8 shows no meaningful gap reduction. We will say in the abstract and conclusion that we do not demonstrate bias mitigation or natural-occlusion gains. 8 of 10 co-occurrence pairs involve person. Re-analysis. Table 8's EG-BA entry (-0.1344) came from the collapsed lambda_IB=0.1 checkpoint, not the lambda_IB=0 headline model; this was our error. For the headline model the mean gap is -0.1121 vs -0.1186 for vanilla (matched MLP variant: -0.1185). Excluding person, pairs re-mined from the remaining classes (10 pairs) give -0.0736 (EG-BA), -0.0867 (vanilla), -0.0703 (MLP variant). So excluding person shrinks the gap by 27-41% depending on the model, consistent with an occlusion/crowding effect, but a negative gap remains in all models, EG-BA is only 0.013 narrower than vanilla, and the MLP variant is equally narrow. Negative pairs: 8/10 in all three models (the submission's 9/10 was a miscount) and 7/10, 6/10, 6/10 without person. We therefore do not claim bias mitigation.

W4 (Don't Judge an Object by Its Context; causal baselines). We have not compared against CONTA/CATT-style methods and could not reproduce them under our protocol in the rebuttal period. Singh et al. (2020) decorrelates a category's features from its co-occurring context via a feature split for multi-label image classification; it works at the image level with no per-query evidence signal and no occlusion recovery. We will discuss it properly in Related Work; applying a feature split to detection queries would be a natural baseline that we have not run.

W5 (occlusion literature is old). We will add recent transformer-based occlusion work.

Q1 (relation to Wu et al., AAAI 2026). We will read and discuss it in the revision; we do not want to claim a difference we have not verified.

Q2 (matched non-causal gating baseline). Done: same gate, loss, recipe and data; the dictionary attention is replaced by an MLP of the query with matched context parameters (268,042 vs 268,288). Result (single seed): the MLP variant is worse on known mAP (46.90 vs 51.20), WI (0.0193 vs 0.0097) and A-OSE (1417 vs 1224), and better on U-Recall (0.0767 vs 0.0573). Because the other re-run configurations (K=10, 50) also fall 5-7 mAP points below our K=20 run, seed variance may be as large as these gaps, so we do not claim the dictionary beats a generic MLP; we will report this as a mixed result. The full module adds 0.565M parameters (28.70M vs 28.13M), which we had wrongly reported as zero.

Q3 (comparison under the same backbone). Not available within the rebuttal period, as above. Table 9 compares against published numbers from other protocols; we will state this and not claim superiority.

Also found: Table 2 (K ablation) used lambda_IB=0.1; we re-ran K=10/50 at lambda_IB=0. K ablation re-run at lambda_IB=0 (mAP / WI / A-OSE / U-Recall avg): K=10: 45.94 / 0.0174 / 1349 / 0.0865; K=20: 51.20 / 0.0097 / 1224 / 0.0573; K=50: 44.10 / 0.0075 / 1319 / 0.0887. WI falls as K grows, U-Recall rises with K, and K=20 is best on mAP and A-OSE (single seed). We will replace the Table 2 discussion with these runs, keeping Table 2 as an explicitly labelled lambda_IB=0.1 result.

Reviewer_td5eOpenReview ↗

Response to Rebuttal

Dear authors, I appreciate the honesty and openness in your responses. It helps to better understand this paper's choices, limitations, and implications. It does not cause me to change my score. My reasoning is as follows: most of the problems were acknowledged by the authors (thank you), some of them were addressed in textual changes (e.g. scope and precision of claims, related works, more precisely defined contributions), but my main concerns were not resolved during the rebuttal. It's fair of the authors to say that additional experiments regarding very related works as baselines, were not feasible in the rebuttal period. The authors' confirmation about limitations of the method, and that baseline experiments would have been helpful, confirms my earlier assessment that the paper does not reach ICLR quality at this stage.

Reviewer_Pb7mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Evidence-Gated Backdoor Adjustment (EG-BA), a framework for open-set object detection that aims to reconcile two competing roles of contextual information: contextual bias under distribution shifts and contextual compensation when visual evidence is degraded by occlusion.

Experiments on the VOC-to-COCO benchmark show improvements in Wilderness Impact and unknown-class recall compared with vanilla Deformable DETR. The paper also investigates gate collapse, synthetic occlusion, natural occlusion, and co-occurrence shifts.

However, the experimental findings reveal substantial limitations: the information-bottleneck regularizer can cause gate collapse, improvements under synthetic occlusion do not consistently transfer to natural occlusion, and the proposed mechanism does not reduce the measured co-occurrence performance gap.

Critical strengths
  • An interesting and practically relevant formulation of contextual information usage.

This paper identifies an important tension in object detection, as contextual information can introduce spurious correlations when direct visual evidence is sufficient yet may become useful when objects are partially occluded. The proposed per-query adaptive mechanism provides a coherent and potentially useful way to investigate this tension.

Critical weaknesses
  • The experiments do not substantiate the two central objectives of the proposed framework.

The paper's primary motivation is to simultaneously mitigate contextual co-occurrence bias and improve detection when direct visual evidence is degraded. However, the experiments directly evaluating these objectives provide limited or contradictory evidence.

Table 7 shows that improvements on synthetic occlusion do not translate into consistent gains on naturally occluded objects. Performance on Occluded-COCO and Separated-COCO remains close to the vanilla Deformable DETR baseline, with some metrics showing slight degradation.

Table 8 reports a mean co-occurrence AP gap of -0.1344 for EG-BA compared with -0.1186 for the baseline. Rather than reducing this gap, the proposed method exhibits a slightly larger performance disparity.

  • The causal interpretation and necessity of the adaptive evidence gate remain insufficiently validated.

Table 3 shows that the proposed information-bottleneck objective causes the evidence gate to collapse, requiring the headline configuration to disable this regularizer by setting its weight to zero.

Table 4 raises an additional concern regarding the necessity of adaptive gating. The fixed context configuration achieves higher known-class mAP (52.82 vs. 51.20) and higher unknown-class recall (0.0689 vs. 0.0573) than the adaptive gate, although the latter achieves better WI and A-OSE.

Therefore, while adaptive gating may provide a different calibration trade-off, the current experiments do not clearly demonstrate that evidence-conditioned routing is necessary for the claimed improvements, nor do they isolate the contribution of the proposed causal interpretation from that of a more general feature-fusion mechanism.

Questions for authors

Q1. Can the authors provide more direct evidence that EG-BA actually mitigates contextual bias?

Q2. What is the precise relationship between evidence supervision, information-bottleneck regularization, and the reported headline results?

Q3. Can the authors better isolate the benefits of adaptive gating from simpler context-fusion alternatives?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Pb7m: gate necessity, matched baseline, co-occurrence re-analysis

Thank you for the precise review. All new numbers come from one training run per configuration (single seed); we flag where this limits the conclusion.

Weakness 1 (the two objectives are not substantiated). We agree. Table 7 (Occluded/Separated-COCO) shows parity with the baseline, and Table 8 shows no meaningful reduction in the co-occurrence gap. We will state in the abstract and conclusion that occlusion-evidence tracking is demonstrated on synthetic occlusion only and that we do not demonstrate bias mitigation. 8 of 10 mined pairs involve person, which confounds co-occurrence with occlusion. Re-analysis. Table 8's EG-BA entry (-0.1344) came from the collapsed lambda_IB=0.1 checkpoint, not the lambda_IB=0 headline model; this was our error. For the headline model the mean gap is -0.1121 vs -0.1186 for vanilla (matched MLP variant: -0.1185). Excluding person, pairs re-mined from the remaining classes (10 pairs) give -0.0736 (EG-BA), -0.0867 (vanilla), -0.0703 (MLP variant). So excluding person shrinks the gap by 27-41% depending on the model, consistent with an occlusion/crowding effect, but a negative gap remains in all models, EG-BA is only 0.013 narrower than vanilla, and the MLP variant is equally narrow. Negative pairs: 8/10 in all three models (the submission's 9/10 was a miscount) and 7/10, 6/10, 6/10 without person. We therefore do not claim bias mitigation.

Weakness 2 (necessity of the adaptive gate; causal interpretation vs generic fusion). Agreed. Table 4 shows fixed e=0 has higher mAP (52.82 vs 51.20) and U-Recall (0.0689 vs 0.0573), while the adaptive gate has better WI and A-OSE. We therefore claim only a calibration trade-off.

Q1 (direct evidence of mitigating contextual bias). We do not have it: Table 8 is negative, and a prototype analysis (3,894 confident predictions, K=20) shows diffuse attention (entropy 2.75 vs max 3.00; mean max weight 0.15 vs 0.05 uniform) with NMI 0.16 between top prototype and predicted class. We will not claim the prototypes are semantic confounders.

Q2 (evidence supervision vs IB vs headline results). Precisely: the VOC headline uses lambda_IB=0, a small anti-collapse term 0.05*(mean e - 0.5)^2, and no occlusion supervision. Evidence supervision (lambda_ev=1) is used only in the COCO diagnostic run, where it reduces GOCE from 0.5632 to 0.0991 with ECE unchanged (0.0072 vs 0.0099). At lambda_IB=0.1 the gate collapses to a constant (mean 0.0179, std 0.0000); at 0 it stays distributed (0.4890, std 0.1804). IB is therefore a documented negative finding, not a contribution.

Q3 (isolating adaptive gating from simpler context fusion). We trained a matched non-causal baseline: same gate, loss, recipe and data, with the dictionary attention replaced by an MLP of the query (268,042 vs 268,288 context parameters). Result (single seed): the MLP variant is worse on known mAP (46.90 vs 51.20), WI (0.0193 vs 0.0097) and A-OSE (1417 vs 1224), and better on U-Recall (0.0767 vs 0.0573). Because the other re-run configurations (K=10, 50) also fall 5-7 mAP points below our K=20 run, seed variance may be as large as these gaps, so we do not claim the dictionary beats a generic MLP; we will report this as a mixed result. A fixed-gate comparison is already in Table 4 (e=0, e=1).

Corrections found while preparing this reply: Table 2 (K ablation) was run at lambda_IB=0.1 (collapsed); we re-ran K=10/50 at lambda_IB=0. K ablation re-run at lambda_IB=0 (mAP / WI / A-OSE / U-Recall avg): K=10: 45.94 / 0.0174 / 1349 / 0.0865; K=20: 51.20 / 0.0097 / 1224 / 0.0573; K=50: 44.10 / 0.0075 / 1319 / 0.0887. WI falls as K grows, U-Recall rises with K, and K=20 is best on mAP and A-OSE (single seed). We will replace the Table 2 discussion with these runs, keeping Table 2 as an explicitly labelled lambda_IB=0.1 result. Appendix A.1 claimed zero parameter overhead; measured +0.565M (+2.0%), 86.5 vs 93.3 FPS, ECE 0.0255 vs 0.0187. We will fix this and limit "improves calibration" to WI and U-Recall.

Reviewer_Q1BwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates two failure modes with detecting unknown objects. It traces the failure to a backdoor path: scene context influences both the query representation and the co-occurrence context. The authors introduce a novel framework EG-BA to estimate context and remove it. Integrating EG-BA into Deformable-DETR improves calibration and unknow-class recall, demonstrating the effectiveness of the proposed method.

Critical strengths
  1. The motivation is reasonable. The hypothesis that query is not purely affected by its own appearance but also scene or co-occurrence context is reasonable.
  2. The calibration metric improvement is convincing. Table 1 illustrates that EG-BA improves U-Recall and WI with clear margins.
Critical weaknesses
  1. Concerns about its effectiveness. Mitigating co-occurrence bias is an important motivation of this paper, while the results in Table 8 show that EG-BA does not close the performance gap between co-occurring and isolated object instances.
  2. In Table 4, forcing e_i = 0 achieves higher U-Recall (0.0689) and known mAP (52.82) than the proposed adaptive gate. In this case, using pure context leads to even better results. This raises concerns about the necessity of the adaptive gate. It seems that the adaptive gate merely trades known mAP for calibration, without a clear advantage on both metrics. In addition, Table 4 is missing the down-arrow for WI.
  3. In Section 3.1, co-occurrence context is viewed as a confounder that the backdoor adjustment removes, while in Section 3.2 it is retained for replacement when evidence is missing. It seems that the formula does not align well with the mechanism, and the paper does not describe when to use context and when to remove it.
  4. The presentation and the notation need thorough revision. For example, "under perform" should be "underperform". The lowercase c denotes the value of the confounder in Eq. 2, but c_i in Eq. 3 is a d-dimensional attention estimate. d is used both for the feature dimension (d = 256) and for the debiased vector d_i in Eq. 5. The symbol σ serves as the sigmoid function in Eq. 4 and as the Gaussian standard deviation σ_i in Eq. 8. Z_i denotes the evidence variable in the causal graph, while z_i denotes the sampled latent in Eq. 9. These issues make the manuscript hard to follow.
Questions for authors

See the weaknesses above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Q1Bw: corrections, matched baseline, narrowed claims

Thank you for the careful review. We agree with the main concerns and respond to each below. All new numbers come from one training run per configuration (single seed); we flag where this limits the conclusion.

W1 (Table 8: co-occurrence gap not closed). Correct. We do not claim bias mitigation and will say so in the abstract and conclusion. 8 of the 10 mined pairs involve person, which mixes co-occurrence with physical occlusion. Re-analysis. Table 8's EG-BA entry (-0.1344) came from the collapsed lambda_IB=0.1 checkpoint, not the lambda_IB=0 headline model; this was our error. For the headline model the mean gap is -0.1121 vs -0.1186 for vanilla (matched MLP variant: -0.1185). Excluding person, pairs re-mined from the remaining classes (10 pairs) give -0.0736 (EG-BA), -0.0867 (vanilla), -0.0703 (MLP variant). So excluding person shrinks the gap by 27-41% depending on the model, consistent with an occlusion/crowding effect, but a negative gap remains in all models, EG-BA is only 0.013 narrower than vanilla, and the MLP variant is equally narrow. Negative pairs: 8/10 in all three models (the submission's 9/10 was a miscount) and 7/10, 6/10, 6/10 without person. We therefore do not claim bias mitigation.

W2 (Table 4: fixed e=0 beats the adaptive gate on mAP and U-Recall). Correct: e=0 gives mAP 52.82 vs 51.20 and U-Recall 0.0689 vs 0.0573; the adaptive gate is better on WI (0.0097 vs 0.0141) and A-OSE (1224 vs 1366). It is a trade-off, not dominance. We will remove "sole architecture capable" and add the missing down-arrow on WI. To test whether the dictionary and gate matter, we also trained a matched non-causal baseline (same gate, loss and recipe; dictionary attention replaced by an MLP of the query, 268,042 vs 268,288 context parameters). Result (single seed): the MLP variant is worse on known mAP (46.90 vs 51.20), WI (0.0193 vs 0.0097) and A-OSE (1417 vs 1224), and better on U-Recall (0.0767 vs 0.0573). Because the other re-run configurations (K=10, 50) also fall 5-7 mAP points below our K=20 run, seed variance may be as large as these gaps, so we do not claim the dictionary beats a generic MLP; we will report this as a mixed result.

W3 (Sec. 3.1 vs 3.2: context removed vs retained). Fair point. Backdoor adjustment is the intended behaviour when visual evidence is reliable (e_i near 1); context compensation when it is degraded (e_i near 0); e_i interpolates. This is a design heuristic motivated by the causal graph, not derived from it, and the original text did not say when each applies. We will state this explicitly in Sec. 3.2 and in the Discussion.

W4 (notation and presentation). Agreed. We will rename the clashing symbols (confounder value c vs estimate c_i, feature dimension d vs debiased vector d_i, sigmoid sigma vs std sigma_i, causal variable Z_i vs latent z_i), fix "underperform", and add the WI arrow. While preparing this reply we also found that Eq. 4 is described as a clamp, but the implementation uses a smooth tanh bound; we will fix the wording.

Further corrections we found: Table 2 (K ablation) was run at lambda_IB=0.1 (collapsed gates); we re-ran K=10/50 at lambda_IB=0. K ablation re-run at lambda_IB=0 (mAP / WI / A-OSE / U-Recall avg): K=10: 45.94 / 0.0174 / 1349 / 0.0865; K=20: 51.20 / 0.0097 / 1224 / 0.0573; K=50: 44.10 / 0.0075 / 1319 / 0.0887. WI falls as K grows, U-Recall rises with K, and K=20 is best on mAP and A-OSE (single seed). We will replace the Table 2 discussion with these runs, keeping Table 2 as an explicitly labelled lambda_IB=0.1 result. Appendix A.1 claimed zero parameter overhead; measured 28.70M vs 28.13M (+2.0%), 86.5 vs 93.3 FPS, and ECE 0.0255 vs 0.0187. We will correct this and restrict our calibration claim to WI and U-Recall (A-OSE and ECE are worse than the baseline).

We hope the narrowed claims address the concern about effectiveness.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of changes: corrections, new experiments, and claims

We thank the reviewers for their invaluable comments. All three reviews found real problems, and in preparing our replies we also found errors of our own. This comment summarizes them; individual replies are posted under each review.

Corrections to the submission

  1. Table 8: the EG-BA entry (-0.1344) came from the collapsed lambda_IB=0.1 checkpoint, not the lambda_IB=0 headline model. The headline model gives -0.1121 (vanilla: -0.1186). The "9/10 negative pairs" count was also wrong; it is 8/10 for all models.
  2. Appendix A.1 claimed zero parameter overhead. Measured: 28.70M vs 28.13M parameters (+2.0%), 86.5 vs 93.3 FPS, ECE 0.0255 vs 0.0187.
  3. Table 2 (K ablation) was run at lambda_IB=0.1, where the gate collapses to a constant. We re-ran K=10 and K=50 at lambda_IB=0.
  4. Eq. 4 is a smooth tanh bound, not a clamp. Several symbols clashed and are renamed.

New experiments (single seed per configuration)

  • Matched non-causal baseline (dictionary attention replaced by an MLP of the query, 268,042 vs 268,288 context parameters): worse than EG-BA on known mAP (46.90 vs 51.20), WI (0.0193 vs 0.0097) and A-OSE (1417 vs 1224), better on U-Recall (0.0767 vs 0.0573). Because the K=10 and K=50 runs also fall 5-7 mAP below our K=20 run, seed variance may be as large as these gaps, so we do not claim the dictionary beats a generic MLP.
  • Co-occurrence analysis excluding person: gaps shrink by 27-41% in all models but stay negative; EG-BA is only 0.013 narrower than vanilla, and the MLP baseline is equally narrow.
  • Prototype analysis: attention over the K=20 prototypes is diffuse (NMI with predicted class 0.16), so we no longer describe them as semantic confounders.

Claims we have narrowed

  • The information bottleneck is reported as a negative finding, not a contribution.
  • We claim an improvement in WI and U-Recall only, at a cost in known mAP, A-OSE and ECE.
  • We do not claim bias mitigation or natural-occlusion gains.

The revised PDF and code are uploaded. Details and remaining limitations are in the individual replies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.