acceptodds
Under review as a conference paper at ICLR 2027

EonWeave: Weaving Persistent Evidence into Consistent Long-Horizon World Generation

Abstract

Long-horizon world generation requires a persistent model of the world, not merely short-range temporal continuity: a model must faithfully recover previously observed identities, scenes, and world states after they have disappeared from view for extended periods. We term this capability . Existing autoregressive generators struggle in this regime because short-term conditions progressively accumulate errors, while historical evidence is either underutilized, redundantly stored, or inefficiently retrieved. To address these coupled limitations, we introduce , a unified framework for integrating persistent evidence into long-horizon generation. first calibrates source reliability by selectively re-noising the short-term condition while preserving clean historical evidence, preventing memory bypass and promoting adaptive reliance on persistent context. Building on this reliability asymmetry, curates a fixed-capacity Persistent Scene Atlas through pose–appearance diversity-gain selection and block-adaptive projection, retaining complementary world-state evidence under bounded memory. then converts the curated history into query-aligned references through noise-adaptive appearance–geometry routing and retrieval-to-reference distillation, enabling efficient utilization of heterogeneous historical evidence. Extensive experiments demonstrate state-of-the-art generation quality and revisitation consistency, with 's PSNR advantage widening from approximately dB in single-shot generation to dB under extended autoregressive rollouts. These results establish as a substantive advance toward consistent long-horizon world modeling.

Primary area:
generative models
Keywords:
World Models, Long-Horizon World Generation, Revisitation Consistency, Persistent Memory, Autoregressive Generation, Historical Retrieval

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 4 reviews, above 27% of the others.
Reviewer_bKuMOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper presents EonWeave, a memory-augmented framework for long-horizon video generation comprising three components: ARiSE that encourages reliance on historical evidence through selective condition perturbation, ReCAP that maintains a fixed-capacity scene atlas for preserving diverse observations, and RaDAR that retrieves and aligns historical patches with current queries to recovers previously observed subjects and scenes after prolonged absence.

Critical strengths
  1. The paper studies an important problem: recovering previously observed subjects and scenes after prolonged absence, which goes beyond maintaining short-term temporal continuity;
  2. The proposed framework systematically integrates reliability calibration, bounded memory, and historical retrieval and reports promising reconstruction results on two benchmarks.
Critical weaknesses
  1. The evaluation does not directly establish revisitation consistency. General fidelity and consistency metrics do not show whether content is faithfully recovered after a long absence. A convincing metric would strengthen the claim.
  2. Tables 1 and 2 show that the headline PSNR gains of 3.037 dB (19.426 vs. 22.463) in single-shot generation and 5.773 dB (19.449 vs. 25.222) on Mira-World come from EonWeave-DMD rather than EonWeave Full. Figure 3 labels the results only as “EonWeave,” leaving the variant unclear. It's hard to follow the claims.
  3. In Table 3, w/o ARiSE also improves the drift metric from 0.337 to 0.312, so the claim that it improves rollout stability needs further explanation.
  4. The method relies on clean ground-truth memory during training but stores its own potentially flawed predictions at inference. This leaves a critical training–inference mismatch unaddressed: incorrect identities or scene states may be retained and repeatedly reinforced, yet the paper provides no analysis of this failure mode.
Questions for authors
  1. Why does removing ARiSE improve the reported drift metric?
  2. What are the runtime and memory costs relative to AR-HyDRA at matched resolution, rollout length, and sampling steps?
  3. Is the learned geometry weight monotonic with noise, and does the model actually shift reliance toward atlas evidence?
Reviewer_XezQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses revisitation consistency in autoregressive video generation. The authors aim to keep the scene sctructure and recover the charactors after they leave and re-enter the frame. The authors build their method based on Wan2.1-T2V-1.3B with three components. and with distinct module names... The first one called arise, which corrupts recent conditions while preserving historical evidence during training, and re-noises generated conditions during inference. The second one called recap, which maintains a fixed-capacity atlas using pose–appearance diversity and block-specific projections. The third one called radar and it combines appearance and geometry for historical routing and aggregates retrieved evidence into a fixed-size reference grid. Then follows a four step DMD. However, the evaluation does not directly isolate event-level revisitation accuracy or establish the proposed reliability mechanisms independently.

Critical strengths
  1. The method provides an framework for managing persistent historical context.
Critical weaknesses
  1. The evaluation does not directly measure the revisitation consistency. The principal consistency metrics are sequence level DINO and CLIP similarities. They do not directly quantify whether a particular returning instance recovers its established identity, location, or state.
  2. Training and inference mismatch: Training uses clean ground-truth atlases, whereas deployment progressively inserts generated observations that may contain structured errors.
  3. All the experiments are done in a low quality synthetic scenario with synthetic charactors. Remains a question if the method works on real world scenarios. Experiments are too toy.
  4. Abaltions are not well controlled. Arise, recap and radar are all removed as a whole during ablations, not well evaluated within each module.
Questions for authors
  1. Would the revisitation score still hold if a real human charator is being evaluated with face ID matching? It is quite basic for a work in late 2026.
Reviewer_S1DgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on long-horizon and revisit consistency (when subjects leave and re-enter the camera view) in autoregressive camera-conditioned video generation. The authors propose a framework with three main parts: 1. ARiSE, which perturbs short-term conditioning latents during training and re-injects noise during inference to reduce exposure bias and over-reliance on recent frames. 2. ReCAP, a fixed-capacity persistent latent memory (scene atlas) updated via greedy diversity selection over camera poses and appearance features. And 3. RaDAR, a patch-level retrieval and soft-routing module that aggregates historical memory into reference tokens using geometry and appearance matching. A few-step distilled variant (DMD) is also trained for faster inference. The method is tested on HM-World (single-shot and 3-step AR) and a newly curated Mira-World dataset (9-step AR), showing gains in global PSNR, SSIM, LPIPS, and VBench metrics.

Critical strengths

The paper focuses on a practical bottleneck in autoregressive video and world modeling, namely balancing short-term continuity with long-term memory under a fixed memory budget. Across both HM-World and Mira-World, the full system achieves numerical improvements over the compared camera-controlled video generation baselines, and the DMD variant shows good potential for efficient few-step generation.

Critical weaknesses

1.Weak validation of the core revisit consistency claim, insufficient supplementary videos, and missing world model baselines. (1) The main claim of this paper is preserving subject and scene identity across long-term exit and re-entry events, yet neither the quantitative metrics nor the qualitative videos properly validate this. Quantitatively, the paper relies on whole-sequence averaged PSNR/SSIM/LPIPS and global VBench Subject Consistency (which only compares global DINO features across frames without tracking the exiting and re-entering objects). Even though HM-World provides ground-truth subject bounding boxes and exit/re-entry timestamps, no targeted event-level metrics after re-entry are reported. (2) Qualitatively, in terms of video evidence, the supplementary material provides only three very short video clips. In two of them, the subject disappears for only about 1 second before re-entering, and the longest disappearance is just around 3 seconds, which barely requires a long-term persistent memory. Furthermore, the supplementary material only shows results from the authors' own method without any side-by-side video comparisons against baseline methods. (3) Finally, while framing the method around world modeling and long-horizon memory, the paper omits comparisons with several important recent world model baselines, such as HY-World, the Yume series, LingBot-World, and Infinite-World.

2.Confounded mechanism attribution, coarse ablations, and unclear test-time conditioning protocols. The experiments show that the full pipeline gets higher numbers, but fail to prove that the proposed components work for the claimed reasons. (1) In Table 8, the model already outperforms the strongest baseline by +4.475 dB at the very first segment, where the persistent memory is empty and no autoregressive rollout has happened yet. Also, in Appendix G.2, adding ARiSE improves the first segment by +3.998 dB, but its advantage shrinks to +1.889 dB by the final segment (and shows higher drift in Table 7: 0.337 vs. 0.312). This contradicts the claim that ARiSE mainly suppresses long-horizon error accumulation, and the noise sweep in Table 9 omits the 0.0 noise baseline needed to separate training perturbation from inference re-noising. (2) Removing ReCAP or RaDAR entirely only shows that having a memory path helps, not that these specific designs are better than simpler alternatives. There is no fixed-capacity comparison of ReCAP against simple strategies like FIFO, uniform sampling, or pose-only selection, nor are RaDAR's internal components ablated individually alongside actual runtime latency and VRAM overhead. (3) The inference algorithm requires per-segment text prompts and future camera poses at each rollout step, but it is unclear where these future prompts come from during testing on Mira-World and whether all baselines received the exact same per-segment conditioning.

Questions for authors

1.Can you provide side-by-side qualitative video comparisons against baselines on longer sequences where subjects leave the view for multiple segments? Also, how does your method compare against recent world model baselines such as HY-World, the Yume series, LingBot-World, and Infinite-World?

2.Using HM-World's exit/re-entry timestamps and bounding boxes, can you report targeted event-level metrics specifically within the post-re-entry window (such as masked PSNR/LPIPS on the re-entering subject and DINO identity similarity before exit vs. after re-entry)? What is the actual distribution of subject disappearance durations in Mira-World?

3.During multi-step evaluation on Mira-World and HM-World, where do the segment-level text prompts and camera poses come from at each rollout step? Are the texts extracted from ground-truth future video clips, and did all compared baselines receive the exact same per-segment text and camera inputs?

4.Can you provide a 2x2 ablation for ARiSE separating training-time perturbation (on/off) and inference-time re-noising (on/off, including 0 noise)? For ReCAP and RaDAR, how does ReCAP compare against FIFO, uniform sampling, and pose-only selection under the same capacity (18 slots), what is the individual contribution of RaDAR's sub-modules (learned schedule vs. fixed weight, dustbin, soft aggregation), and what is the actual wall-clock latency and peak VRAM per segment compared to AR-HyDRA?

Reviewer_DeKrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies revisitation consistency in long-horizon video generation: preserving previously established identities, appearances, and spatial states when subjects or scenes reappear after prolonged absence. The proposed framework, EonWeave, combines short-term video conditioning with a fixed-capacity historical memory during autoregressive generation. It comprises three components: ARiSE encourages historical evidence utilization through asymmetric condition perturbation; ReCAP selects and retains historical records based on appearance and camera-pose diversity; and RaDAR retrieves evidence using appearance and geometry, aggregating it into query-aligned references. Experiments on HM-World and Mira-World evaluate single-shot and autoregressive generation, reporting improvements in PSNR, subject and background consistency, and resistance to long-horizon degradation.

Critical strengths

The paper addresses a meaningful challenge in long-horizon video generation: maintaining consistency when subjects or scenes reappear after prolonged absence. The framework integrates historical evidence retention, utilization, and retrieval alignment, aiming to preserve both short-term continuity and long-term world states under a fixed memory budget.

Critical weaknesses

The evaluation relies mainly on GT-based reconstruction metrics, which do not directly establish revisitation consistency with the model’s own history; the qualitative examples also need clearer historical correspondence. Training and distillation chunk settings, the five-frame local window, and practical runtime costs require clarification. Baseline adaptations, particularly ReCamMaster and AR-HyDRA, need more detail to assess fairness. Finally, module-level ablations do not isolate the contributions of individual design choices.

Questions for authors
  1. Could you explain how the current evaluation establishes revisitation consistency and the model’s ability to retain and use historical information? Please also clarify how each baseline is configured and adapted to the evaluation protocols.
  2. Please clarify the number of chunks used during training and evaluation, the number of video frames and latent temporal positions per chunk, and how the chunk size relates to the five-frame local window.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.