acceptodds
Under review as a conference paper at ICLR 2027

WORLDTOKEN: TIME-FIRST SEQUENCE MODELING FOR ROBOTIC IMITATION LEARNING

Abstract

Vision-language-action policies inherit both capabilities and input representations from pretrained vision-language models, showing great potential for robotic manipulation across diverse industrial settings. As these policies increasingly use interaction history, organizing the representation of historical observations determines how experiences enter temporal context and how relationships across time are modeled, which is a generally ignored challenge in previous works. In this work, we believe solving this challenge requires an architectural reconstruction and propose WorldToken. WorldToken encodes each timestep's observations into one world token, processes the resulting history with a causal Transformer, and generates action chunks with a diffusion action head. Unified token enables long horizon tasks while relieving the memory requirement of the temporal backbone, hence improving performance. This design also offers high interpretability and allows advances in language modeling, such as pre-training and scaling, to be transferred to robot interaction policies. In RoboCasa experiments, WorldToken successfully handles most tasks with 85M parameters and achieve 59.4% mean closed-loop success close to with 3.35B parameters. On the memory benchmark RMBench Blocks Ranking, WorldToken can reach the context of two minutes and achieve a success rate of 95%. In addition, WorldToken’s holdout action RMSE is well described by power-law fits, and its closed-loop success rate improves consistently with increasing training data size in a study involving approximately 350,000 closed-loop evaluation episodes across 50 trained policies on RoboCasa, showing its scaling potential. We also conduct experiments to analyze information preservation and history use in WorldToken, providing empirical grounding for future work.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning, Imitation Learning, Sequence Modeling, Multimodal Representation Learning, History-Conditioned Policies

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_vhPnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents WorldToken, a new robot learning policy architecture emphasizing on history information processing where a causal transformer encodes history information into a few latent tokens (dubbed world token in the paper) and DiT decodes the latent tokens into continuous action through iterative denoising. Simulation experiments ablated various model hyperparameter choices.

Critical strengths
  1. Timely topic. Memory for robot learning is becoming more and more important as a way of improving upon the mainstream current-state-only policies.
Critical weaknesses
  1. Certain architecture design choices are ill-motivated and not ablated. It is fairly standard to use an encoder-decoder architecture for encoding history information, however conditioning the DiT action decoder solely on one world token and expect them to encode the full current and history visual, proprioception, and task command information is ill-motivated and unabated (e.g. against a design where the action decoder also conditions on full current state information). Further, since the training supervision completely comes from the action diffusion objective, nothing in the architecture design explicitly prevents the causal transformer encoder from learning a degenerated world tokens (say that simply copies or averages past history). This concern is also not addressed in ablation studies. Finally, with the default 10-step context length and the ~0.2s difference between each context step used in most of the experiments from the paper, the model can encode ~2s of history, and whether such a short history length provides enough benefits is unclear. Although the paper shows one long context task (RMBench Blocks Ranking), the authors fail to establish consistent benefits from their method through evaluation on multiple tasks, and the RMBench Blocks Ranking training loss is manually engineered as discussed in Appendix F.3.

  2. Missing baselines and incomplete comparison. Given that there is plenty of memory for robot learning works (and policies without explicit memory design), it is unusual that the paper’s experiments compare the authors method against a very limited set of external baselines, making it hard to gauge the contribution of the paper’s architecture design. In addition, the paper does not provide real world experiments to validate the results in simulation. Finally, some experiment design is questionable in rigor. For example, in the RMBench Blocks Ranking experiment at Table 5 the authors train the model with a long context length but evaluate the model with different context length, whereas the previous experiments in Table 3 show that mismatch between training and inference time context history length can drastically hurt performance.

Questions for authors
  1. Can the authors provide additional justification or ablations on various architectural designs including design to condition the action decoder on only the world tokens?

  2. Can the authors explain whether or why the causal transformer encoder does or does not learn a degenerated mapping between history information to the world tokens through additional experiments?

  3. How does the WorldToken policies compare to other robot learning baselines with or without explicit memory design?

3 comments
AuthorsOpenReview ↗

Response to Reviewer vhPn

Thank you for your comments. WorldToken is a compact, time-first VLA architecture rather than a dedicated memory module. We clarify its information flow, design choices, and evaluation scope below.

  1. Architecture design and action conditioning (Weakness 1 and Question 1)

Our contribution is the time-first VLA architecture rather than the encoder–decoder structure itself. The encoder fuses the current images, proprioception, and task condition into one world token. A causal Transformer contextualizes the token sequence, and the DiT uses the resulting history representation. Because the encoder receives no past observations, past timesteps remain individually represented rather than compressed into one token.

About the architectural alternatives, our paper reports three diagnostics. A current-observation bypass lets every DiT block access all 50 fused current-observation tokens, increasing RMSE and reducing success from 48.12% to 46.20% and from 48.38% to 42.00%. Passing 50 rather than one token per timestep improves success by only 2.09 points and RMSE by 3.3% at D300, while requiring 55.31 times as many temporal-backbone FLOPs. Moving history processing to denoiser cross-attention yields mixed results. WorldToken has 6.4%–8.7% lower RMSE, while the DP-like variants have 1.4–3.0 points higher success.

Regarding the scope of our long-history evidence. A practical advantage of WorldToken is that each timestep adds only one temporal token, keeping long histories compact. This allows the policy to use a 608-step, approximately 146-second context on Blocks Ranking without a separate memory module. The interventions detailed in Question 2 show that this history is behaviorally useful. We present Blocks Ranking as a single-task long-history case study.

The Blocks Ranking objective affects only downward button execution and uses no history supervision. Under the standard objective, the policy reaches the correct arrangement in 99/100 episodes but has only 13 official successes because most executions miss the button threshold, obscuring long-context evaluation. Although different code revisions prevent a clean loss-only interpretation, all history interventions fix the modified checkpoint and vary only visible history. Reducing history from approximately 146 to 8 seconds lowers official/strict success from 95%/94% to 28%/24%, demonstrating long-context utilization.

  1. Comparisons and evaluation scope (Weakness 2)

Our comparisons address two questions: whether WorldToken provides effective control relative to a same size RoboCasa baseline, and whether WorldToken can reach success rates comparable to substantially larger pretrained VLAs. We therefore select BC-Transformer as the official, reproducible, from-scratch baseline and use pi0.5 as a large pretrained reference. On the same 23 tasks, D300 demonstrations, and 1,150 episodes, WorldToken achieves 46.83% mean success versus 31.28% for BC-Transformer. At D2900, the 85.3M-parameter WorldToken reaches 59.45%, close to 62.1% for the approximately 3.35B-parameter pretrained pi0.5. Published 23-task means are 51.11% for GR00T N1 and 64.96% for GR00T N1.5. Also, The K-token diagnostics described above provide controlled internal comparisons. Dedicated-memory results use different backbones, data, and protocols, so we do not treat them as same-protocol controls or claim superiority.

Our simulation experiments directly establish WorldToken’s effectiveness and its data-scaling, token-interface, and history-utilization trends under controlled conditions. These are the architectural questions studied in this paper. Sim-to-real robustness is a separate deployment question. We view real-robot experiments as an important direction for future work.

Table 5 tests whether a fixed policy uses information beyond shorter history windows, so the checkpoint must remain fixed. Training separate policies with matched context lengths would change weights and training dynamics, preventing attribution to visible history alone. That adaptation question is addressed in Table 3. Thus, the train–test context mismatch is an intended intervention rather than an uncontrolled variable. We will clarify that Table 5 measures history utilization, not optimal short-context performance.

  1. Degenerate history mapping (Question 2)

As explained above, each world token is independently encoded from the current observation, so the causal Transformer does not map history into world tokens. It only contextualizes the token sequence for action prediction, and the fixed-checkpoint history interventions reported above show that this contextualization does not ignore history.

  1. Baselines with and without memory (Question 3)

As described above, WorldToken is not intended to be a dedicated memory mechanism. Its one-token-per-timestep interface supports long explicit contexts and provides memory-like functionality, as validated by the fixed-checkpoint history interventions.

Reviewer_vhPnOpenReview ↗

Acknowledgement of Rebuttal

Thank you for providing a rebuttal that clarifies the methodology and provides useful context. However, since my main concerns about limited comparison with existing temporal model architectures and lack of compelling evidence for the methodology's advantage remain, I would like to maintain my original rating.

AuthorsOpenReview ↗

Response to Reviewer vhPn

To address the remaining concern, we would like to clarify two points: a comparison with externally developed temporal systems and a controlled evaluation of our proposed temporal interface.

Regarding why we did not compare with other temporal systems, we do not claim exhaustive superiority over the rapidly emerging family of dedicated memory systems, since our study does not focus on that area. Rather, long-context scalability is an architectural capability enabled by WorldToken’s one-token-per-timestep interface. Therefore, our experiments are designed to demonstrate that the proposed method has great potential in long-context settings. Meanwhile, we agree that adding comparison experiments with recently developed temporal systems would better demonstrate WorldToken’s potential. We therefore plan to conduct experiments using Mamba for full-history recurrent compression and add the results to the appendix.

The claimed architectural advantage of WorldToken is specifically its compact one-token-per-timestep interface. We support this claim in three ways. First, WorldToken processes a 608-step, approximately 145-second context without a separate memory module. Second, fixed-checkpoint history interventions reduce success from 95% to 28% when the visible history is shortened from 146 to 8 seconds, confirming that the policy actually uses distant information. Third, retaining 50 tokens per timestep increases temporal-backbone FLOPs by a factor of 55.31 while providing only a 2.09-point improvement in success at D300. At D2900, the single-token model matches or exceeds the 50-token variant in terms of success while achieving lower RMSE. These results demonstrate a favorable trade-off among per-timestep spatial detail, temporal coverage, and computation under the evaluated settings.

Beyond these empirical findings, we further clarify the broader architectural motivation behind WorldToken’s one-token-per-timestep design.

We rethink history modeling at the architectural level because we believe memory is even more fundamental to robotics than commonly assumed—much as in daily life, a blind person can still live independently and accomplish many complex tasks, whereas someone with severe memory impairment struggles to complete even basic multi-step actions. Looking back at the history of NLP, mainstream recurrent memory mechanisms must progressively compress history into a fixed state, which creates inherent bottlenecks for long-horizon modeling and scaling. As a result, it was ultimately superseded by Transformers. We therefore adopt a structural paradigm similar to LLMs, aiming to replicate the methodological evolution and scaling trajectory of LLMs in robotics. Along this direction, WorldToken makes a deliberate trade-off between temporal and spatial information: allocating more representational and computational budget to extending the temporal sequence rather than preserving fine spatial granularity within each timestep. Our experiments demonstrate the feasibility of this trade-off: comparing N3 (K=1) in Table 1 with N2 (K=50) in Table2 shows that at D2900, the single-token N3 matches or surpasses the N2 (K=50) in success rate while achieving lower RMSE. This indicates that the key to performance is not retaining more raw information per timestep, but using a stronger encoder (and model capacity) to compress each timestep effectively. We conjecture that this also holds for more complex observations. On one hand, images possess low-dimensional structure. On the other, a closed-loop policy, much like the human eye, can retain only focal details at the current moment and adjust its gaze on demand in the next timestep to inspect other locations.

Going one step further, this one-token-per-timestep structure aligns robot sequence modeling directly with language. Just as humans can speak with their eyes closed, a robot model should have the same capability. Due to current training data scale limitations, our paper passes a CLIP text vector at each timestep, but our underlying design intent is to pass one text token per timestep. Under that formulation, masking out the other sensory inputs at a given timestep naturally reduces the model to a language model, opening up possibilities for multi-source data training and multimodal capability fusion in the future.

Reviewer_9Np9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the WorldToken architecture, which aims to address the question of whether the representation granularity inherited from vision-language models should be retained in history-conditioned robot policies. The WorldToken architecture encodes observations at each time step, adds 4 readout tokens, encodes them into a single WorldToken via a network, processes the resulting tokens into a historical sequence through a causal transformer, and finally generates action chunks via a diffusion action head. This model successfully accomplishes most tasks with only 85 million parameters, achieving an average closed-loop success rate of 59.4%, with performance approaching that of π0.5 which uses 3.35 billion parameters; on Blocks Ranking, a sub-task of the memory capability benchmark RMBench, WorldToken supports a context length of up to two minutes and achieves a success rate of 95%. Additionally, the paper puts forward 4 questions to evaluate the WorldToken architecture, examining its performance and scalability.

Critical strengths

The motivation is sound, and the proposed paradigm appears promising. The simulation experiments are extensive and reasonably thorough.

Critical weaknesses
  1. The paper contains no real-robot experiments, so it remains unclear whether the method is feasible in real-world settings.
  2. The paper provides no direct, same-protocol comparison against representative existing approaches: BC-Transformer is the only reproduced baseline, and the π0.5 comparison is quoted from prior work on a different task set.
  3. Increasing model capacity beyond 218.8M parameters yields no consistent improvements, leaving the scalability of WorldToken to larger models insufficiently established.
  4. Several appendix tables present extensive numerical results across experimental configurations without consistently highlighting the most informative comparisons or key findings. Improving the visual emphasis on important results would make the experimental evidence easier to interpret.
Questions for authors
  1. The 23 evaluation tasks in this article do not include OpenDoubleDoor in the 24-task benchmark used for π 0.5, but the higher success rate of WorldToken on the other three switch gate tasks is still not enough to judge the impact of the lack of this task on the overall average success rate. Therefore, whether to provide the average success rate of the two methods on the same task for consistent comparison?
  2. The paper studies the number of tokens for each time step and the utilization of historical information by the strategy. Could the author further explain: how many tokens should be retained for each time step when the strategy needs to utilize a longer history?
1 comment
AuthorsOpenReview ↗

Response to Reviewer 9Np9

Thank you for recognizing the potential of our method and the breadth of our experiments. We clarify our contributions and address your concerns and questions below.

1、Real-robot validation (Weakness 1)

We agree that real-robot experiments are needed to establish deployment feasibility. Our primary question, however, is architectural: how time-first organization affects multitask control, scaling, token granularity, and history use. Controlled simulation allows these factors to be varied systematically, so our conclusions are scoped to simulation-based imitation learning rather than sim-to-real robustness. We will view real-robot validation as important future work.

2、Baseline comparisons and shared-task evaluation (Weakness 2 and Question 1)

Our comparisons address two questions: whether WorldToken provides effective target-domain control relative to an official from-scratch RoboCasa baseline, and whether a compact WorldToken policy can reach success rates comparable to substantially larger pretrained VLAs. We therefore select BC-Transformer as the official, reproducible, from-scratch baseline and use pi0.5 as a large pretrained reference.

Using BC-Transformer's official implementation and native recipe, we evaluate both methods on the same 23 tasks, D300 demonstration identities, and 1,150 evaluation episodes. WorldToken achieves 46.83% mean success versus 31.28%.

We use 23 tasks because OpenDoubleDoor contains only 1,500 official demonstrations, insufficient for the common D2900 endpoint of our scaling study. This data-driven exclusion is applied consistently, and BC-Transformer is evaluated on exactly the same tasks.

For the cited pi0.5 run, task-level results are unavailable. Its rounded 24-task mean of 62.1% bounds its mean over our 23 shared tasks to 60.40%–64.86%. WorldToken reaches 59.45% with 85.3M parameters, within 0.95–5.41 percentage points of the approximately 3.35B-parameter pretrained model. Other published task-level results give exact 23-task means of 51.11% for GR00T N1, and 64.96% for GR00T N1.5. Although their protocols differ, these comparisons demonstrate the competitiveness of WorldToken as a substantially smaller policy.

3、Gains from increasing model capacity (Weakness 3)

Our experiments establish systematic scaling within the tested regime. Across all five model sizes, every increase in training data reduces RMSE and improves success rate. Moreover, increasing capacity from 44.3M to 218.8M parameters reduces RMSE by 5.17% at D50 and by 22.73% at D2900, showing that the benefit of additional capacity grows with data. This data–capacity interaction is consistent with language-model scaling studies (Kaplan et al., 2020; Hoffmann et al., 2022). The absence of consistent gains beyond 218.8M parameters with at most 2,900 demonstrations per task suggests that performance may be data-limited, rather than that WorldToken has reached an inherent architectural limit. These results demonstrate WorldToken's scaling potential within the available data regime and motivate further evaluation as larger and more diverse datasets become available.

4、Presentation of appendix results (Weakness 4)

We apologize that the breadth of our appendix obscures some findings. Its extensive ablations and diagnostics characterize WorldToken's properties and trade-offs and provide complete evidence. We will reorganize them around the main experimental questions. Each subsection and caption will state the purpose, variables changed and held fixed, and principal finding. Matched configurations and key differences will be visually emphasized.

5、Tokens per timestep for longer histories (Question 2)

Our results favor retaining one token per timestep for longer histories and allocating additional capacity to representation dimensionality and model size. K=4 provides no consistent benefit over K=1. K=50 improves RMSE by 1.7%–4.2% and success rate by 1.6–6.4 percentage points across the five data scales, but these gains do not increase systematically with data. In contrast, increasing the single-token model from N2 to N3 provides larger RMSE gains at higher data scales. At D2900, N3 reduces RMSE by 7.93%, compared with 2.55% for K=50, both relative to N2 with training seed 0.

The computational difference grows with history length. With KV caching, increasing the visible history from 10 to 500 timesteps raises the K=50 model’s cost per decision from 24.72 to 39.77 GFLOPs, while the single-token N3 model increases only from 28.29 to 28.30 GFLOPs. Thus, one token per timestep provides the better tradeoff between performance and computation for long histories under our tested settings.

Reviewer_GY4VOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

WorldToken compresses each policy timestep’s multiview images, proprioception, and task condition into a single learned token. A causal Transformer processes the token history, and a diffusion head generates action chunks. The paper evaluates this architecture across five dataset sizes and five model capacities on RoboCasa, compares different token counts, and studies history dependence through truncation and matched-context training. A separate Blocks Ranking experiment evaluates longer histories. The central contribution is an empirical study of the trade-off between compact observation representations, temporal context, and policy performance.

Critical strengths
  1. The architecture is simple and modular. Separating within-timestep perception, temporal processing, and action generation provides a clear interface for studying history-conditioned policies. Reusing encoded observations also avoids repeatedly processing historical images.

  2. The RoboCasa evaluation is substantial. The five-by-five data-and-capacity sweep includes two training seeds per configuration and evaluates final checkpoints without selecting them by rollout performance. This provides useful evidence about data scaling within the proposed architecture.

  3. The history analysis includes an important control. The authors distinguish truncating a trained policy’s history from training a policy with shorter context. Short-context training recovers much of the performance lost under inference-time truncation, demonstrating that reliance on history does not establish that a task requires it.

  4. The token-count experiments quantify a useful trade-off. At larger data scales, passing one token per timestep gives performance close to the 50-token variant while substantially reducing temporal-backbone computation.

Critical weaknesses
  1. My main concern is the limited conceptual contribution. WorldToken combines established components—per-timestep observation compression, causal temporal modeling, and diffusion-based action generation—into a coherent policy architecture. However, the paper does not sufficiently establish a novel mechanism or a generalizable insight beyond the effectiveness of this particular combination. The extensive experiments characterize the proposed system, but the absence of closely matched architectural baselines makes it difficult to determine which findings are specific to the proposed “time-first” organization. I therefore view the work as a useful engineering study whose scientific contribution is currently insufficiently established for acceptance.

  2. The comparisons do not isolate the proposed contribution. The BC-Transformer comparison changes the action head, observation and execution cadence, optimization, and training budget. The comparison with π₀.₅ uses different demonstration counts, task sets, and pretraining. Furthermore, all token-count variants preserve timestep grouping and causal temporal order. These experiments therefore do not establish whether the proposed organization offers an advantage over closely related policies under matched conditions.

  3. The strongest long-history result requires substantial qualification. The reported 95% Blocks Ranking success uses a task-specific loss modification encouraging additional motion during downward actions. Appendix F reports 13% success with the standard-loss procedure, although that policy reaches the target block geometry in 99% of trials. Different code revisions further confound the comparison. This does not invalidate the history interventions on the fixed modified-loss checkpoint, but it makes the headline result difficult to attribute to the general architecture.

  4. History utilization is not the same as a distinctive memory capability. Truncating a long-context policy demonstrates dependence on its available history, but does not establish superiority over appropriately trained short-context or recurrent policies. The paper itself illustrates this distinction on RoboCasa. Blocks Ranking lacks corresponding matched-context training and memory-policy baselines, and its shorter-context failures primarily involve execution and placement rather than incorrect swap order. Consequently, the evidence for a specific long-term reasoning advantage remains limited.

Questions for authors
  1. Conceptual contribution: What new modeling principle distinguishes WorldToken from prior policies that compress each timestep’s observations before temporal processing? Which experimental finding specifically supports this distinction?

  2. Matched comparisons: Can you compare against a closely related policy while holding perception, the diffusion action head, observation cadence, and training budget fixed, varying only the temporal representation or organization?

  3. Long-history capability: How do shorter-context policies trained specifically for Blocks Ranking perform? Can a recurrent or existing memory-based baseline achieve comparable results under the same training conditions?

  4. Task-specific training: Can you isolate the effect of the modified Blocks Ranking loss using identical code and training settings, reporting block arrangement, button pressing, and strict sequence-following outcomes separately?

1 comment
AuthorsOpenReview ↗

Response to Reviewer GY4V

We thank the reviewer for the constructive assessment and for recognizing our modular design and extensive evaluation. We clarify below our contribution, comparisons, and Blocks Ranking results.

1、Conceptual contribution

The distinguishing principle is not per-timestep compression itself, but aligning the top-level sequence with physical policy time: each position represents one observation–decision–replanning event, independently of the number of perceptual tokens used within that timestep. WorldToken therefore separates within-timestep multimodal fusion, cross-timestep causal modeling, and diffusion action generation. Unlike approaches that compress timesteps and then use a separate memory or recurrent module, WorldToken keeps past timestep representations directly accessible through causal attention without an additional memory mechanism.

The token-interface experiment supports the practical value of this organization. At D300, passing all 50 fused observation tokens to the temporal backbone improves mean success only from 47.48% to 49.57% and reduces RMSE by 3.3%, while requiring 55.31 times the analytic temporal-backbone FLOPs. On Blocks Ranking, truncating the same policy's context from 608 to 32 timesteps reduces strict success from 94% to 24%, showing that the compact temporal sequence is used in control. These results support its performance–computation trade-off rather than universal superiority over alternative sequence organizations.

2、Matched comparisons To evaluate WorldToken as a complete policy, we compare it with BC-Transformer under each method’s native training recipe. We match the raw observations, D300 demonstrations, 23 tasks, and 1,150 evaluation episodes while retaining BC-Transformer's official action head, cadence, optimization, and training schedule. These differences are intentional: replacing the baseline's native components with ours would no longer reproduce BC-Transformer, whereas forcing WorldToken into the baseline interface would remove part of the proposed system. We therefore interpret this as an end-to-end comparison rather than a causal ablation of temporal organization. WorldToken achieves 46.83% mean success versus 31.28% for our BC-Transformer reproduction. It also achieves a numerically similar reported success rate to pi0.5 with substantially fewer parameters (59.45% at 85.3M versus 62.1% at approximately 3.35B). Although the experimental protocols differ, this demonstrates the competitiveness of WorldToken as a substantially smaller policy.

Our paper also reports controlled diagnostics. The K-token variants share the visual stems, fusion encoder, temporal backbone, diffusion head, cadence, data, and optimizer-step budget while varying the temporal token interface. At D2900, K=1 and K=4 achieve 59.1% and 59.9% success and RMSE values of 0.091 and 0.090, while K=4 requires 4.03 times the temporal-backbone FLOPs.

3、Task-specific training

The task-specific objective addresses an issue with button execution that is specific to the benchmark. With the standard objective, the policy reaches the correct block arrangement in 99 of 100 episodes but succeeds in only 13 because most button presses do not cross the required threshold. The modified objective adjusts only the loss for downward button-press motion and uses no history labels, yielding 95% official success and 94% strict sequence-and-placement success. Although different code revisions prevent a clean loss-only attribution, all history experiments use the same modified-loss checkpoint. Reducing visible history from approximately 146 seconds to 8 seconds lowers official success from 95% to 28% and strict success from 94% to 24%. Therefore, this decrease cannot be explained by the modified objective.

4、Long-history capability

We apologize that our use of “memory” was ambiguous. We use this term in a functional sense: WorldToken retains past timestep representations in its context and directly accesses them through attention when predicting the current action. It therefore provides context-based memory without a separate memory bank or recurrent hidden state. In Blocks Ranking, reducing the same policy's visible history from approximately 146 seconds to 8 seconds lowers official success from 95% to 28% and strict success from 94% to 24%, demonstrating that this context provides a memory-like function. We will revise the paper to avoid implying superiority over recurrent or dedicated memory mechanisms.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.