acceptodds
Under review as a conference paper at ICLR 2027

Ego-Forge: Text and Geometric-Attention Free Exo-to-Egocentric Video Generation

Abstract

Exo-to-egocentric video generation aims to synthesize what a person sees from their own viewpoint given third-person footage and a target head trajectory. The task requires transferring appearance and semantics across large viewpoint changes while hallucinating content never observed by the exocentric camera. Existing approaches either impose additional input requirements, such as a ground-truth initial egocentric frame or multiple synchronized exocentric views, or remain limited to category-specific settings. EgoX is the first to address cross-activity and in-the-wild generalization, but requires a human-provided caption of the non-existent egocentric view at inference and introduces a computationally expensive geometry-guided attention bias that can propagate reconstruction errors and suppress textual and visual context. We therefore propose Ego-Forge, a caption-free and bias-free framework for exo-to-egocentric generation. It introduces Dynamic Captioning, which derives conditioning tokens directly from the model's hidden states and adapts them to the diffusion timestep and network depth, replacing external text conditioning. By scaling training by an order of magnitude and using all available exocentric viewpoints, Ego-Forge learns cross-view correspondence implicitly and eliminates the need for geometry-guided attention, requiring only a lightweight depth prior. Ego-Forge achieves state-of-the-art performance on Ego-Exo4D, runs faster end-to-end, requires no external annotation at inference, and generalizes to in-the-wild scenes, including cases where over-reliance on geometry blocks appearance inference. Our model and source code will be made publicly available.

Primary area:
generative models
Keywords:
Exo-to-egocentric video generation, Cross-view video synthesis, In-the-wild generalization, Diffusion models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_sBEQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Building on EgoX, Ego-Forge generates a first-person video from a third-person video + head trajectory, using a pre-trained video model. Instead of relying on manual captioning Ego-Forge introduces a learned captioner “R”. It also claims that EgoXs geometry guided attention can be dropped when enough training data is available.

Critical strengths

The problem is interesting and relevant. Being caption-free at inference is a huge advantage.

Critical weaknesses

My main concern is that the paper leaves out a lot of crucial information of the architecture. The authors should add an architecture description of “R” - in its current state it is not possible to re-implement R as stated. For example, in Eq. 3 Q^{exo} and Q^{ego} have their separate attention calls while in Figure 3 they have the same (?) projections. Which is it? Where is the “value” part, it is missing in both Eq3 and Fig3. How is hih_i fed into R?

Figure 2 and Figure 3 are way too small. In general, Figure 3 is difficult to read, the two colors for K^{exo} and K^{ego} are almost the same. The toKto_K and toQto_Q should also get different colors to not make them look similar to the mask. What positional embedding was used?

Some claims need to be softened, i.e.

  • The paper states that LL053-055 “[EgoX] … can propagate reconstruction error ...“ but this is also true for Ego-Forge - it uses a depth-reprojected prior which comes with its own bias. Also crucially, this is not ablated in any of the experiments.
  • The paper states (in its title) that GGA can be dropped (when training is scaled) - but they do not show any evidence for that! No ablation varies data size to measure this.
Questions for authors

Is the room-camera part of the canvas noised? In Figure 1 it looks like only the egocentric part is noised and the exo-part is kept clean, but in 3.1 it seems that both VexoV^{exo} and VegoV^{ego} are noised.

Figure 7: what are the differences in “Rendering” between EgoX and Ego-Forge? Should that not be the same operation in both pipelines? Why is it so much slower in EgoX? Is this run on 25 frames or 49 frames on EgoX? How many denoising steps where used for EgoX vs Ego-Forge?

1 comment

Q1. Architecture description of R, Figures 2-3, and the positional embedding.

We thank the reviewer for the careful reading and for pointing out where our description of R was unclear.

  • We have redesigned Figure 3 for clarity, enlarged Figures 2–3, expanded the architecture description in the text, and added the full PyTorch implementation of R (including how h_i is fed to R) to the Appendix G.
  • No positional embedding is used: the queries are learned, and h_i already carries positional information from the backbone's rotary self-attention.

We appreciate the reviewer's constructive feedback and the opportunity to improve this part of paper. The corresponding changes are highlighted in red in the revised PDF for ease of reference.

Q2. Error propagation from the depth-reprojected prior (LL053–055).

In EgoX, depth errors affect generation through two pathways: they corrupt the ego prior and introduce errors into attention biases via GGA, compounding their effects. Ego-Forge avoids this second pathway by removing GGA, while its backbone can recover the correct viewpoint even from poor priors, as shown in our supplementary videos. We revised the abstract, highlighted in red, to clarify this distinction.

Q3. Evidence that GGA can be dropped at scale.

Please refer to our common reply, where we address this question with the requested data-scaling ablation and provide the corresponding analysis, supporting the argument that GGA can be removed under training at scale.

Q4. Is the exocentric part of the canvas noised?

We agree that this was an inconsistency in the formulation. Only the egocentric part is noised, while the exocentric part remains clean as conditioning, as shown in Figure 2. We have corrected Sec. 3.1 accordingly (highlighted in red).

Q5. Inference time in Figure 7: rendering, frame count, and denoising steps.

  • Frames: As already mentioned in Line 377, all timings in Figure 7 are for 25 frames, for both methods.

  • Denoising steps: EgoX uses 50 steps, while Ego-Forge uses 40 (as already mentioned in Line 363). We also tested 50 steps for Ego-Forge and observed no further improvement, which contributes to its speed-up.

  • Why rendering differs: The two pipelines use fundamentally different rendering procedures. EgoX (official repository, EgoX-EgoPriorRenderer; scripts/render_vipe.sh for rendering) first fuses a static background point cloud across frames, then adds per-frame dynamic objects and rasterizes the merged cloud using PyTorch3D. This multi-frame fusion is the main computational bottleneck. In contrast, Ego-Forge directly renders each frame from a single MoGe-V2 point cloud, without cross-frame fusion, completing the entire prior pipeline in under 15 seconds.

Please note that Figure 7 excludes EgoX's captioning time, since its exo- and ego-view captions require human or VLM input at inference, with varying associated costs. We exclude this overhead; accounting for it would further increase Ego-Forge's speed-up.

We again thank the reviewer for the thoughtful feedback, particularly the insightful questions in Q1 and Q4, which helped us further clarify and improve the paper.

Reviewer_Qhk4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Ego-Forge is based on EgoX, which is a exo-to-ego video generation model. It drops two parts of the model and simplifies it.

  1. It drops the ego view human-written caption and replaces with a dynamic captioner which is conditioned with the model hidden states.
  2. It drops the GGA bias, where the authors found isn't needed if you train the model on all four exo views. The model shows small gains over EgoX and also a 4.3x inference speedup.
Critical strengths

The work improves upon the original work - small performance gain and speed up. Also the work improves the practical aspects of the model, where caption free pipeline is cheaper and useful for the user. Also the ablation show that Dynamic Captioner is effective, showing improvements in performance -PSNR 13.93 to 16.37).

Critical weaknesses

The main claims and the results are mixed. The authors claim that GGA can be dropped when training is scaled, but there's no exact ablation for this. It would be EgoX on the same 100k clips, only without the GGA module. Also it needs Ego-Forge with GGA module Also the depth model is different. EgoForge uses MoGe-v2, and this will be part of the 4.3x speedup. Captioner, data, extra finetuning is also different, so it is hard to know where the gain comes from.

Also the gains over EgoX are very marginal, only FVD clearly improves.

Questions for authors
  1. Authors will need more refined ablation study. For example, can authors train EgoX w/o GGA on the same data, depth model, and Ego-Forge with GGA added back, plus a data-scale curve?
  2. Where did the numbers in Table 1 come from? Are the EgoX numbers copied, and which captions were used?
  3. Could you add confidence intervals and evaluations with multiple seeds.
  4. Could you specify where the 4.3x speedup came from. Break it down to different sources - depth model, GGA removal and captioner overhead.
3 comments

A) "The authors claim that GGA can be dropped when training is scaled, but there's no exact ablation for this."

We thank the reviewer for this observation. Please refer to our common reply, where we address this question with the requested data-scaling ablation and provide the corresponding analysis, supporting the argument that GGA can be removed under training at scale.

B) "It would be EgoX on the same 100k clips, only without the GGA module."

Ego-Forge vs. EgoX without GGA at 100K would also differ in captioning, so it would not isolate GGA. A captioned model without GGA on 100K clips is already in our Table 2, row 2: "Text caption".

C) "Also it needs Ego-Forge with GGA module."

We appreciate this thoughtful suggestion. In fact, we already tested reintroducing GGA and found that the weaker depth estimates lead to inaccurate geometric guidance, causing GGA to fail. This has been clarified in Section 4.2, in blue.

D) "Captioner, data, extra finetuning is also different, so it is hard to know where the gain comes from."

We isolate the contributions through dedicated ablations: data scaling in the common reply's table and Dynamic Captioning in Sec. 4.2. The remaining differences are deliberate components of our design:

  • Captioner. Stage 1 uses open-source Qwen captions, which may be even less detailed than EgoX's paid API captions.
  • Data. We use more training data because removing GGA and captions requires the backbone to learn geometry and semantics implicitly.
  • Extra finetuning. Stages 2–3 progressively remove the need for captions.

E) "Also the gains over EgoX are very marginal, only FVD clearly improves."

PSNR, SSIM, and LPIPS capture only certain aspects of video quality and do not fully reflect perceptual quality and temporal consistency, which is also consistent to the observations reported in [1-4]. For example, we observed that using fewer than our 40 denoising steps produces blurrier videos despite yielding higher PSNR, illustrating the limitations of these metrics. FVD, which better captures video-level quality, shows a clear improvement over EgoX.

[1] High-Resolution Image Synthesis with Latent Diffusion Models - Rombach et al., CVPR 2022

[2] Image Super-Resolution via Iterative Refinement - Saharia et al., listed as TPAMI 2022

[3] Emu Video: Factorizing Text-to-Video Generation by Explicit Image Conditioning - Girdhar et al., 2023

[4] StateSpaceDiffuser: Bringing Long Context to Diffusion World Models - Savov et al., 2025

Q1.

  • EgoX without GGA. It is already reported in our paper. Please refer to Table 2, row 2.
  • Ego-Forge with GGA. Please see C).
  • Data-scale curve. Please see A).

Q2.

We thank the reviewer for their careful attention to this matter. The EgoX numbers are not copied; we independently re-evaluate EgoX following its exact protocol. Specifically, we evaluate on 25 frames, as now explicitly clarified in the revised PDF (Line 364 in blue). We use the released seen/unseen sample lists and captions, which we kindly obtained from the authors upon request, and run the official ViPE preprocessing and evaluation code. We sample every second frame from the original 49-frame clips, retaining both the first and last frames so that each caption continues to describe the entire clip. We appreciate the reviewer’s feedback, which prompted us to clarify these important evaluation details in the experimental section of the revised paper (in blue).

Q3.

We evaluate Ego-Forge on the unseen split with 3 sampling seeds (4, 42 (main paper), 422). We report the mean ± standard deviation across seeds and, in brackets, the 95% confidence interval of the mean (t-distribution).

  • PSNR: 14.58 ± 0.06 [14.42, 14.74]
  • SSIM: 0.482 ± 0.003 [0.475, 0.489]
  • LPIPS: 0.537 ± 0.004 [0.528, 0.547]
  • FVD: 318.6 ± 5.4 [305.3, 332.0]

On every metric, EgoX lies outside Ego-Forge's 95% confidence interval, so the gains are beyond seed variation.

Per seed (PSNR / SSIM / LPIPS / FVD), for reference:

  • seed 4: 14.63 / 0.479 / 0.540 / 317.46
  • seed 42: 14.60 / 0.484 / 0.533 / 313.95
  • seed 422: 14.51 / 0.483 / 0.539 / 324.51

Q4.

This breakdown is already reported in Figure 7 of the paper; for clarity, we also give it as a table below.

All timings are per 25-frame sample.

StageEgoXEgo-Forge
Depth (reconstruction + rendering)435 s (260 s + 175 s)15 s
Denoising (*): GGA80 s0
Denoising (*): video backbone180 s145 s
Total695 s160 s

We have also added these statistics to the Appendix H of the revised paper (in blue). Please note that Figure 7 excludes EgoX's captioning time; accounting for it would further increase Ego-Forge's speed-up.

We sincerely thank the reviewer for all the helpful feedback, which motivated us to add Table 6 and further clarify the experimental details and speed-up analysis in the revised paper.

Reviewer_Qhk4OpenReview ↗

Response to rebuttal

Thank you for the rebuttal. Some are answered, but my main concern is only partially addressed.

Table 6 lacks a +GGA arm. compared with Table 1 GGA-free Stage 1 needs ~25x more data to match EgoX. Most of the speedup comes from the depth pipeline and reduced (50 40) steps. GGA removal gives ~1.13x. metric omits CLIP-I, where EgoX is better on both splits

AuthorsOpenReview ↗

Response to Reviewer Qhk4

Thank you for the response. We address each point individually below.

Table 6 lacks a +GGA arm.

We implicitly explained this in our previous message part C. Ego-Forge removes GGA as the main bottleneck to its training performance and scale-up. Adding it back would reintroduce that bottleneck, contradicting the method’s design.

Compared with Table 1, GGA-free Stage 1 needs ~25× more data to match EgoX.

The practicality of scaling is our fundamental thesis (see paragraph 3 of introduction). The scalability is not merely about the efficiency, but also the ability to handle imperfect data.

The raised concern therefore does not hold as is, within the scope of our paper. That being said, we further clarify why over-commitment to EgoX (as suggested) isn't the way forward.

  1. Scaling EgoX is infeasible in practice. EgoX’s preprocessing alone takes 515 s per clip (see table below): ViPE reconstruction, rendering and GGA. For 100K clips, that is over 1.6 GPU-years before any training starts, plus the cost of captioning with a paid API. Our full pipeline needs about 21 GPU-days (Our Table A, Appendix).

Cost of EgoX's preprocessing (per 25-frame clip, one H200 GPU):

StepTime
3D reconstruction260 s
Ego-prior rendering175 s
GGA computation80 s
Total515 s
  1. Even with 100K clips, we argue EgoX still faces fundamental limitations.
  • Trained on its own data design (well-aligned views only), it would still never see poorly aligned cameras, so more data would not help it handle diverse camera configurations at inference.

  • Trained on our data design (diverse views, including poor exo-ego overlap), GGA itself becomes the obstacle. It is reliable only when geometric cues are present and accurate. Where they are missing, the cosine similarities in EgoX’s bias approach −1, and the logarithm drives the bias toward −∞: (EgoX, Eqs. 4,5)

sm,n′=sm,n+log⁡(g(q^m,k^n)⋅λg),g(a^,b^)=cos⁡sim⁡(a^,b^)+1s'_{m,n} = s_{m,n} + \log (g(\hat{q}_m, \hat{k}_n) \cdot \lambda_g), \qquad g(\hat{a}, \hat{b}) = \cos\operatorname{sim}(\hat{a}, \hat{b}) + 1

As cos⁡sim⁡→−1\cos\operatorname{sim} \rightarrow -1, g→0g \rightarrow 0 and log⁡g→−∞\log g \rightarrow -\infty, so the bias suppresses the corresponding attention entries. The bias then does not merely fail to help; it suppresses those attention entries outright, so the model cannot learn to recover the missing content from the visual stream itself.

Our design therefore provides something that EgoX cannot obtain from training at scale alone.

Most of the speed-up comes from the depth pipeline and fewer steps (50 vs 40); removing GGA alone gives ~1.13×.

The depth pipeline and GGA cannot be separated. GGA needs accurate geometry, which is why EgoX runs heavy multi-frame reconstruction (ViPE and background fusion). Removing GGA is what allows Ego-Forge to replace that pipeline with a single, lightweight depth pass that is tolerant to noise; so most of the speed-up follows from removing GGA. This is clearly discussed in the third paragraph of our Introduction, and Figure 7 breaks the times down. Geometry can be cached for benchmarking, but in real applications it must be recomputed for every new video, so it belongs to the pipeline time.

On the inference steps:

  • Running Ego-Forge with 50 steps brings no further improvement.
  • Running EgoX with 40 steps notably degrades its metrics (table below), so EgoX cannot match our speed by simply using fewer steps.

Unseen scenes, varying inference steps:

MethodStepsPSNR ↑SSIM ↑LPIPS ↓CLIP-I ↑FVD ↓
EgoX4014.110.4330.5790.843479.65
Ego-Forge (Ours)4014.600.4840.5330.864313.95

As stated in our first reply, EgoX’s captioning time was excluded from the comparison. Including it, the speed-up ranges:

  • Worst case for us: Ego-Forge at 50 steps, EgoX without captioning: 695 / 195 = 3.6×.
  • Actual setting: Ego-Forge at 40 steps, EgoX with GPT-4o captioning (~35 s per clip): 730 / 160 = 4.6×.

The metrics omit CLIP-I, where EgoX is better on both splits.

Over our three seeds (4, 42, 422), Ego-Forge’s unseen CLIP-I is 0.868 ± 0.016 [0.852, 0.883] . We attribute EgoX’s CLIP-I advantage to its use of the "caption of the non-existent egocentric view", which we argue is impracticable to assume to have (during inference) in line 089 of the main paper. Since this caption describes the target ego video, strong guidance pulls generation toward its content, which CLIP-I rewards. Please note that, Ego-Forge clearly outperforms EgoX in PSNR, SSIM, LPIPS and FVD, and our previous reply showed that these gains exceed seed variation.

Reviewer_zy7bOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to address the reliance of existing exo-to-ego video generation methods on manually provided textual descriptions and computationally expensive geometry-guided attention. Motivated by the need for a more practical and efficient generation framework, the paper argues that cross-view correspondence can be learned implicitly from large-scale training data without explicitly imposing geometric attention constraints.

The author proposes Ego-Forge, a text- and geometry-guided-attention-free framework built upon Wan2.1-I2V-14B. Its key component is Dynamic Captioning, which replaces external text conditioning with visual representations derived from the model's internal hidden states. The framework adopts a three-stage training strategy: text-conditioned adaptation, cross-attention output distillation, and caption-free joint fine-tuning. Additionally, by leveraging all available exocentric viewpoints during training, the model learns cross-view correspondence without geometry-guided attention.

Evaluations are conducted on the Ego-Exo4D dataset. The proposed method achieves improvements over existing approaches on several image and video quality metrics while providing a 4.3× end-to-end inference speedup compared to baseline.

Critical strengths
  • Overall, I appreciate the motivation of this paper and its understanding of the practical limitations of existing exo-to-egocentric video generation methods, particularly their dependence on target-view captions and computationally expensive geometric attention.
  • The proposed Dynamic Captioning mechanism is an interesting idea. Conditioning the generated representations on both diffusion timesteps and transformer block indices provides a reasonable way to adapt the conditioning information.
  • The proposed framework achieves promising quantitative results while substantially reducing inference time, demonstrating its potential for more efficient exo-to-ego video gen.
Critical weaknesses

My main concerns lie in the training data. The improvements may be partially attributed to the substantially larger training data rather than the proposed architectural changes. Ego-Forge is trained on all available exo views, approximately 100K clips, while EgoX uses only around 4K clips. This difference makes it difficult to isolate the contributions of the proposed method from those of increased training data. It would be beneficial to provide a controlled comparison under the same training data budget, particularly an ablation of geometry-guided attention with the same amount of training data.

Moreover, it remains unclear how robust the proposed framework is to inaccurate depth estimation or noisy camera trajectories. The experiments seem to use GT head trajectories. Additional experiments would help support the claimed advantages of removing explicit geometric attention.

Questions for authors

Please see the weaknesses.

1 comment

Q1. Are the gains due to more training data rather than the method?

We thank the reviewer for raising this interesting question, which we also find intriguing. We believe the gains are due to the scale enabled by our design. Unfortunately, we are unable to have a conclusive answer because this would require running EgoX on 100k training clips which is estimated to require over 1.6 GPU-years of computation, compared to just 21 GPU-days for our method (please refer to Appendix, Table A).

  • GGA is fragile. It fails (1) when the 3D estimate is poor, due to limited exo-ego coverage or inaccurate due to depth errors, and (2) when over-reliance on geometry blocks appearance inference, even with correct geometry (Figs. 1 and A: the mirror reflection cannot be inferred from geometry). Removing GGA requires the backbone to infer geometry itself, which is only possible with training at scale.
  • Ego captions do not exist at inference. It is a real limitation rather than a convenience issue: describing the ego view requires the ego video, which is what we generate. Any such caption is a hallucination by an off-the-shelf VLM or a human, not the backbone’s own capability. Dynamic Captioning instead infers these semantics from the model’s hidden states, which likewise requires training at scale.

On the requested ablation:

  1. EgoX at our scale is infeasible: its preprocessing takes 515 s per clip (table below), i.e. 1.6 years for 100K clips on one GPU, before any training, plus the cost of captioning them with a paid API.

Cost of EgoX's preprocessing (per 25-frame clip, one H200 GPU):

StepTime
3D reconstruction260 s
Ego-prior rendering175 s
GGA computation80 s
Total515 s
  1. Ego-Forge + GGA is incompatible with our design. Ego-Forge drops GGA to enable large-scale training with a lightweight 3D estimator. We already tested reintroducing GGA and found that the weaker depth estimates lead to inaccurate geometric guidance, causing GGA to fail.

  2. To isolate GGA's contribution, captions must be retained, as in EgoX. Ego-Forge Stage 1 provides exactly this setting. We train it on 4K clips and on 10K clips (11 h and 1 day on 1 GPU). Its 100K counterpart already appears in the paper (Table 2, row 2), completing the data-scale curve. Please refer to our common reply, where we show data scaling is what enables GGA removal.

Q2. Robustness to inaccurate depth and noisy camera trajectories.

Inaccurate depth. We acknowledge that we did not make this sufficiently visible in the main paper. Ego-Forge is designed to tolerate noisy 3D estimates, as already demonstrated by extreme cases of poor or misaligned ego priors in our supplementary material, e.g. 4.gif, 10.gif, where the correct ego view is still generated. We will make these visualizations more prominent in the camera-ready version.

Camera trajectories. Estimating camera trajectories from exocentric videos is an interesting direction for future work. However, following prior work, we take ground-truth head trajectories as input to the generation model to keep the evaluation protocol consistent.

We appreciate the reviewer's valuable feedback, which motivated the additional ablation (Table 6) and helped strengthen the paper.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Common Reply

General Response: GGA can be dropped with scaled training.

We thank all reviewers for raising this important point, which was missing from the original paper. To investigate it directly, we train Ego-Forge without GGA at increasing data scales.

To isolate the contribution of GGA, captions must be retained, as in EgoX. Since Ego-Forge removes both GGA and captions, ablating the full model would conflate two effects: learning geometry implicitly and learning to replace caption semantics. Stage 1 provides exactly the controlled setting we need: it retains captions and the 3D prior but removes GGA. We train it on 4K and 10K clips (11 hours and 1 day, respectively, on a single GPU). Its 100K counterpart is already reported in the paper (Table 2, row 2).

ModelTraining clipsPSNR ↑SSIM ↑LPIPS ↓FVD ↓
Ego-Forge (Stage 1)4K14.710.5200.547273.24
Ego-Forge (Stage 1)10K15.210.5270.531249.30
Ego-Forge (Stage 1), Tab. 2 (row 2)100K16.240.5420.498187.73

Performance improves consistently across all metrics as data scales, certifying that given sufficient training data, the backbone can learn geometry implicitly without GGA. We have added this ablation to the revised paper (Table 6, highlighted in green).

Note on the revised manuscript: Throughout our responses, all references to line numbers refer to the newly revised version of the paper. For clarity, we have highlighted all modifications in color: red, blue, or green, with different colors used to distinguish the changes. Any text that remains in black is unchanged from the original submission.

Area_Chair_Jb9COpenReview ↗

A Gentle Nudge: Let's Keep the Discussion Going!

Dear Reviewers,

The authors have now posted their responses, addressing your concerns and incorporating your suggestions. We are already halfway through the discussion period!

One of the distinguishing features of ICLR is its interactive review process, where constructive discussions not only help strengthen individual papers but also contribute valuable insights to the broader scientific community.

So, let's not let the discussion be a one-way street! I encourage you to revisit the authors' responses, acknowledge their efforts, and engage in further discussion where needed. After all, a good scientific discussion is often as valuable as the final verdict!

-AC

AuthorsOpenReview ↗

Request for Reviewer Follow-up

Dear AC,

Thank you very much for your time and efforts throughout the ICLR 2027 review process.

As we approach the final day of the rebuttal period, we would greatly appreciate your help in encouraging the reviewers to follow up on our rebuttal and share any remaining concerns or feedback.

Thank you again for your time and consideration.

Sincerely,
The Authors

Area_Chair_Jb9COpenReview ↗

ICLR rebuttal follow-up before the discussion period closes

Dear reviewers,

With only a few hours left in the ICLR discussion period, this paper is unfortunately one of the few in my batch that has yet to receive a response to the rebuttal. We would appreciate it if you could join the discussion or at least acknowledge that you have read the author's response. The authors have worked hard to address your concerns, and it would be unfair for the discussion period to close without any acknowledgment of that effort.

Thank you for your service to the community. We appreciate the time and effort you devote to making ICLR a successful conference.

Yours anxious AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.