- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents EAGLE-360, a framework for target localization in 360-degree panoramas. The model uses the full panorama to understand the scene and can request additional perspective views before predicting the target direction.
The authors introduce three components: AtlasRoPE for spherical positional encoding, Spherically Stratified Replay for improving directional coverage in training data, and Spherical Direction Regularization for continuous direction supervision. They also construct a dataset of 21,884 reference trajectories and train the model using progressive SFT followed by GRPO-based reinforcement learning.
Experiments are conducted on a direction-balanced test set of 713 questions and H* Bench. The final model achieves 84.01% localization accuracy on EAGLE-360 and 75.00% object-search success on H* Bench. The paper also provides component ablations and experiments on the effect of visual feedback.
Critical strengths
-
The paper addresses a meaningful problem in panoramic visual understanding. The authors consider several difficulties specific to 360-degree images, including horizontal seam continuity, polar distortion, and uneven directional coverage. The direction-balanced evaluation is useful because it shows that models performing well on horizontal directions can still struggle with polar targets. EAGLE-360 achieves a clear improvement over its SFT baseline, particularly for upward and downward localization.
-
The experimental work is fairly extensive. The authors construct a multi-turn training dataset, evaluate several open-source and proprietary models, and include transfer experiments on H* Bench. I also appreciate the matched component ablations and the visual feedback controls. In Table 4, real tool observations outperform mismatched observations under the same initial panorama conditions. This provides evidence that the model benefits from relevant visual feedback rather than only from generating additional reasoning steps.
Critical weaknesses
- I am not fully convinced by the novelty and independent contribution of the proposed components. AtlasRoPE builds on existing spherical positional encoding approaches, while SSR is mainly a direction-aware data replay strategy and SDR adds an auxiliary direction prediction loss. These components are reasonable, but the overall method seems to combine several existing ideas rather than introduce a clearly new learning framework.
More importantly, the ablations in Table 3 show that SSR contributes most of the improvement. Plain SFT achieves 76.30% accuracy, and adding SSR alone increases it to 83.03%. The full AtlasRoPE+SSR+SDR configuration reaches 84.43%, while AtlasRoPE alone slightly decreases overall accuracy, although it improves angular error. The additional contributions of AtlasRoPE and SDR are relatively small. The paper would benefit from stronger evidence showing why these components are necessary beyond SSR.
- The contribution of the final training pipeline is not sufficiently isolated. The main EAGLE-360 model includes additional interaction-focused SFT and GRPO, but Table 3 evaluates the geometry-aware components before these stages. As a result, it is difficult to determine how much GRPO contributes to the final localization accuracy and whether reinforcement learning improves tool selection or coordinate refinement. The full SFT configuration in Table 3 reaches 84.43%, while the final system reports 84.01%, although these are not directly matched training runs.
I also have some concerns about the evaluation setting. The direction-balanced test is constructed partly through image reorientation, which changes the relationship between visual content and gravity and does not introduce independent scenes. The paper acknowledges this limitation, as well as test-set reuse during development and single-run experiments. H* Bench provides useful transfer evidence, but its task and evaluation protocols are not fully matched across methods. These factors make it harder to judge how well the reported improvements would generalize to new panoramic environments.
Questions for authors
-
Can the authors provide a direct comparison between the interaction-focused SFT checkpoint and the final GRPO checkpoint under the same evaluation setting? It would be useful to report localization accuracy, tool-call frequency, and performance on questions requiring multiple observations. This would help clarify the contribution of reinforcement learning.
-
SSR provides most of the improvement in Table 3. Can the authors further explain the benefit of AtlasRoPE and SDR relative to SSR alone? In particular, do these components provide consistent improvements on seam-adjacent and polar examples across different training seeds?
-
Can the authors provide more evidence that the method generalizes to independent panoramic scenes without evaluation-time reorientation? It would also help to separate the benefit of visual feedback for improving the estimated target direction from its benefit for producing valid final answers, since the current feedback ablation shows improvements in both accuracy and answer validity.