acceptodds
Under review as a conference paper at ICLR 2027

World Intelligence Agent: Unifying World Modeling and World Action for Digital Agents

Abstract

Digital agents built on large language models have made rapid progress, with world models emerging as their fundamental predictive engines for understanding complex environments. However, current world models are largely fragmented: they lack a unified data pipeline, are often tied to specific domains, and are treated as passive, isolated simulators. We argue that advancing digital agents requires a true **foundation world model**: a universal cognitive engine that not only masters cross-domain simulation but also yields predictive representations robust enough to extend naturally into downstream applications. To realize this vision of **World Intelligence**, we first construct **WIA-Corpus** (5M continual pre-training examples, 9k SFT trajectories, and 134.6k RL prompts) and **WIA-Bench**, the first paired in-domain (1.8k) and out-of-domain (1.7k) evaluation benchmark spanning seven digital domains. Building on this cohesive data foundation, we develop **WIA-WM**, a foundation world model family (4B / 9B / 27B / 35B-A3B), for digital agents, that masters systematic state-transition modeling. WIA-WM achieves the highest average world-model score among all evaluated models on WIA-Bench, surpassing GPT-5.6-sol and Claude-Opus-4.8 by **** and **** on the seven-domain OOD suite, and empowers downstream agents via simulation SFT, simulation RL, and inference lookahead. Finally, as a natural extension of this predictive engine, we introduce **WIA-WAM**, which couples prediction and execution via Dual-Perspective Inversion (DPI) and Counterfactual Simulation-Guided Policy Optimization (CSPO). This integration allows the agent to implicitly internalize environment dynamics, moving beyond passive simulation to proactive, robust decision-making. Our dataset, code, and models will be released to facilitate future research.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agent, World Model, World Action Model, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TiFVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents WIA, a framework that connects world modeling with action generation for digital agents. It introduces WIA-Corpus and WIA-Bench across seven digital domains, trains WIA-WM at different model scales, and explores its use in simulated training and inference. The authors further propose WIA-WAM, combining DPI and CSPO to jointly learn action generation and environment prediction.

Critical strengths
  1. I find the multi-domain dataset and benchmark to be a valuable contribution. The paper covers seven different environments and provides both in-domain and out-of-domain evaluations. I also appreciate the extensive experiments across different model scales and downstream tasks.

  2. The authors have made a substantial effort to address the concerns raised in the review. In particular, the additional cross-family evaluations, human evaluation, and long-horizon experiments give me more confidence in the world-model results. The new DPI/CSPO ablations are also helpful in understanding the contributions of the proposed components.

Critical weaknesses
  1. I would still be interested in seeing more evidence of long-horizon simulation reliability. The additional experiments are encouraging, although the human evaluation covers a relatively small subset and the multi-step tests are limited to 2–3 transitions. Evaluating longer interactions would further strengthen the paper.

  2. The new ablations make the role of DPI much clearer. The additional improvements from CSPO are relatively modest, however, and I think its contribution could be better understood through further analysis of prediction quality and downstream behavior. It would also be interesting to explore initialization from the fully trained WIA-WM in future work.

Overall, I appreciate the authors' detailed responses and the additional experiments. Most of my original concerns have been addressed, and I am happy to raise my rating from 2 to 3. I believe this is a useful contribution, and I hope the final version can incorporate these additional results and further clarify the remaining limitations.

Questions for authors

None.

10 comments
AuthorsOpenReview ↗

Response 0

Thank you for the careful and constructive review. We will revise the paper to incorporate the additional evaluation, robustness analyses, data-overlap audits, methodological clarifications, and improved presentation of the method figures described below. The revised figures discussed in W2 clarify the training branches, reward flow, and optimization procedure, and we will align the claims and limitations with the resulting evidence. The following responses address W1, W2, Q1, Q2, and Q3 separately.

Completed figure revisions. We have revised the figures to clarify the overall framework and its training and application workflows. The original Figure 2 used three rows: textual modeling examples illustrating the inputs to the three training stages; the CPT–SFT–RL training pipeline; and world-model simulation applications together with the world action model. We have separated this presentation into two figures. The new overview (revised Figure 2) presents the full framework at a higher level: modeling across seven digital domains, world-model training and WIA-Bench, the three world-model simulation applications (simulation SFT, simulation RL, and inference lookahead), and the world action model. The dedicated training figure (revised Figure 3) retains only the input examples and world-model training pipeline from the original Figure 2. Building on the original Figure 3, the revised WIA-WAM figure (now Figure 4) makes the relationships among the CSPO optimization branches, reward propagation, and training objectives more explicit. We have also added an algorithmic workflow diagram for the three world-model simulation applications on the first page of the appendix (Figure 7, Appendix A). Finally, we have improved the appendix visualizations of world-model rollouts across all seven digital domains (Figures 18–24), making the correspondence between textual modeling and visual state transitions clearer in each domain.

AuthorsOpenReview ↗

Response 1 — W1 part1: Cross-family judge robustness

We evaluate the OOD split with two additional judges, Kimi-K3 and Qwen3.8-max.

Kimi-K3 judge.

Table 1. OOD evaluation under the Kimi-K3 judge.

ModelAndroidMCPOSSearchSWETerminalWebAvgMAvg
GPT-5.6-luna57.3252.4659.9850.1660.2466.8558.2558.2857.90
GPT-5.6-terra57.9851.5961.6450.8761.6467.4858.9759.1058.59
GPT-5.6-sol58.9154.0763.4150.8363.3669.4860.1360.4260.03
Claude-Sonnet-556.8759.2658.7044.0648.0852.9960.1153.4854.30
Claude-Opus-4.857.2463.9259.5150.2260.1264.2660.5958.9459.41
GLM-5.349.0860.6154.1835.4133.0551.9955.5446.9848.55
MiMo-v2.5-pro53.3656.1356.4047.1655.0860.1056.1354.7054.91
Qwen3.8-27B47.1351.1753.6442.1852.5157.6653.1550.9551.06
Qwen3.8-max54.1554.7955.8646.2462.0768.4260.4557.6957.43
DeepSeek-V4-Flash53.1754.7456.4947.7553.3358.9257.2554.3754.52
DeepSeek-V4-Pro54.3154.9357.8749.3654.8262.1157.3755.7955.82
Qwen3.5-4B46.7248.4553.9337.1441.9544.1850.7945.5446.16
Qwen3.5-9B50.6652.5155.8638.6447.5847.1352.7248.6349.30
Qwen3.5-27B54.8555.5258.7140.8253.4659.8457.5653.9954.40
Qwen3.5-35B-A3B54.6353.3755.6538.8850.0554.0254.8651.1551.64
Qwen-AgentWorld52.7656.4660.6254.5961.2764.4058.3258.5258.35
WIA-WM-4B30.7751.8250.2444.1853.6259.4452.0148.8148.87
WIA-WM-9B53.8759.3958.1753.6056.6062.9057.1757.1957.38
WIA-WM-27B56.0360.5858.5951.7262.2965.1258.9058.9159.03
WIA-WM-35B-A3B54.7854.0759.1457.1459.1967.5558.9059.1558.68

Qwen3.8-max judge.

Table 2. OOD evaluation under the Qwen3.8-max judge.

ModelAndroidMCPOSSearchSWETerminalWebAvgMAvg
GPT-5.6-luna46.6745.4949.0436.9650.9854.5951.8348.0547.94
GPT-5.6-terra47.0344.1850.5137.4051.7354.4952.4048.4748.25
GPT-5.6-sol48.6846.4353.3435.5453.3456.5752.7449.5749.52
Claude-Sonnet-545.6652.1555.2533.2040.3240.6651.7144.4445.57
Claude-Opus-4.847.2255.4955.5435.2553.2149.4653.6449.1849.97
GLM-5.342.8750.7648.7227.6230.6944.5548.6140.6841.98
MiMo-v2.5-pro43.8745.9752.9635.2745.9646.6049.1745.3245.68
Qwen3.8-27B38.8341.2747.6131.1844.4746.0144.8741.8742.04
Qwen3.8-max45.0645.7051.5234.2451.3854.1252.9647.8447.85
DeepSeek-V4-Flash44.8245.5251.7835.5644.9547.9448.8045.3245.62
DeepSeek-V4-Pro46.5244.2654.6734.6646.9549.7850.7046.6346.79
Qwen3.5-4B38.1739.0448.8228.4636.8035.6842.2937.9338.47
Qwen3.5-9B41.5441.6750.8828.9141.2837.8044.6340.4140.96
Qwen3.5-27B46.5744.4556.1329.7745.5847.9949.5645.3045.72
Qwen3.5-35B-A3B46.1442.7849.3328.6341.9243.0545.6542.0342.50
Qwen-AgentWorld45.4844.0757.2143.9751.0149.8250.8349.0848.91
WIA-WM-4B25.4543.6946.9037.4048.3447.6045.9442.1342.19
WIA-WM-9B45.2650.5155.0643.3251.1050.9051.3449.3849.64
WIA-WM-27B48.2952.8156.3343.8754.0951.3751.1450.7951.13
WIA-WM-35B-A3B47.4146.1756.7249.0052.1453.6052.4051.3651.06

The absolute scales differ between judges, but WIA-WM remains competitive across scales. WIA-WM-9B records Avg/MAvg of 57.19/57.38 under Kimi-K3 and 49.38/49.64 under Qwen3.8-max; WIA-WM-27B reaches 58.91/59.03 and 50.79/51.13; and WIA-WM-35B-A3B reaches 59.15/58.68 and 51.36/51.06. Among the WIA-WM variants, 35B-A3B has the highest Avg under both judges, while 27B has the highest MAvg. In Avg, 27B is comparable to Claude-Opus-4.8 under Kimi-K3 (58.91 vs. 58.94) and exceeds both GPT-5.6-sol and Claude-Opus-4.8 under Qwen3.8-max (50.79 vs. 49.57 and 49.18). These results indicate that WIA remains highly competitive with frontier models without relying on a single LLM judge for evaluation.

AuthorsOpenReview ↗

Response 2 — W1 part2: Programmatic F1 and human cross-evaluator evidence

As a second part of our W1 response, we report programmatic token F1 and a Codex-assisted human evaluation on WIA-Bench-OOD.

1. Non-LLM F1 evaluation. Token F1 is independent of LLM scoring. The table below reports all 19 evaluated models on the full 1,739 OOD rows; Avg weights examples equally and MAvg weights domains equally.

Table 3. OOD programmatic observation-text token-F1 evaluation.

ModelAndroidMCPOSSearchSWETerminalWebAvgMAvg
GPT-5.6-luna62.6635.6261.5022.9556.7163.4767.0653.6752.85
GPT-5.6-terra64.8233.8463.8724.4759.1963.7067.3254.9153.89
GPT-5.6-sol67.0037.6073.2326.4662.0465.9070.1258.3057.48
Claude-Sonnet-567.6366.6189.0133.6950.9554.6473.3360.4562.27
Claude-Opus-4.866.5468.2986.8832.5562.7263.6575.2863.6965.13
MiMo-v2.5-pro69.5456.3185.5235.2551.8758.2272.9160.5161.37
Qwen3.8-27B65.8844.7177.3133.3750.6356.1263.1955.7455.89
Qwen3.8-max74.5954.3185.3332.6159.6566.6378.8164.2364.56
DeepSeek-V4-Flash70.1758.0685.2037.5048.7059.4172.3160.6461.62
DeepSeek-V4-Pro71.4949.6289.0938.2048.1659.9971.6960.7661.18
Qwen3.5-4B62.3347.1475.6629.7432.2434.4761.5347.6949.02
Qwen3.5-9B65.1349.8682.2029.8537.0237.7364.2550.9052.29
Qwen3.5-27B69.3551.2088.1832.7045.1950.6370.4657.2958.24
Qwen3.5-35B-A3B68.9451.2779.3430.1841.9144.2066.2253.4254.58
Qwen-AgentWorld68.0044.3786.8637.6957.6459.3671.9061.0460.83
WIA-WM-4B4.9451.6780.3838.7537.5658.1065.9247.1948.19
WIA-WM-9B71.6064.3189.3746.4242.0261.6372.3162.5763.95
WIA-WM-27B74.7762.1887.2443.2153.7164.5969.2564.0464.99
WIA-WM-35B-A3B71.6452.5988.5447.0648.6463.6571.5263.0863.38

Among the 19 models, F1 Avg ranges from 47.19 to 64.23. WIA-WM-27B is second in Avg and MAvg, while WIA-WM-9B and 35B-A3B exceed GPT-5.6-sol and Qwen-AgentWorld in both averages. Token overlap is a complementary signal, not a replacement for semantic judging: exact wording can differ despite equivalent state content, and overlap cannot validate fields absent from the reference.

2. Human review conducted with the assistance of the Codex tool. We invited five human volunteers to conduct a blinded review of 56 cases, eight per domain, with Codex used as a tool to facilitate the review workflow and record the evaluations; the volunteers, rather than Codex, made the judgments. Domain columns and Panel Avg show mean ranks; Pro Avg shows the original DeepSeek-V4-Pro score on the same cases (20–100). Position columns give overall rankings.

Experimental setup. For each case, all 5 volunteers ranked the outputs from the different models, casting ranking votes on the same set of outputs. We first averaged the 5 volunteers' ranks for each model on each case, then averaged these per-case mean ranks across the cases in each domain and across the full 56-case subset.

Table 4. Codex-assisted human-panel rankings and DeepSeek-V4-Pro scores.

ModelAndroidMCPOSSearchSWETerminalWebPanel Avg rankPanel positionPro AvgPro position
GPT-5.6-sol2.7002.9503.2503.2254.1752.5003.4753.182364.293
Qwen3.8-max3.8503.1253.9503.1252.1003.4502.3253.132264.392
DeepSeek-V4-Pro2.8252.9252.9753.0753.3503.6504.0503.264462.824
Qwen3.5-35B-A3B3.1753.3753.6254.5754.3253.8253.9753.839558.435
WIA-WM-35B-A3B2.4502.6251.2001.0001.0501.5751.1751.582189.861

The volunteer panel and Pro scores produce the same overall ordering on these 56 cases: WIA-WM-35B-A3B, Qwen3.8-max, GPT-5.6-sol, DeepSeek-V4-Pro, and Qwen3.5-35B-A3B. Agreement extends across all five positions, including the relative ordering of the four baselines. This agreement supports the reliability of WIA-Bench's model rankings.

Reviewer_PuL5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work introduces the World Intelligence Agent (WIA), a framework designed to integrate world model prediction and actionable decision making for digital agents. The authors first build WIA Corpus, a training corpus consisting of 5M continual pre-training examples that covering seven digital domains, and construct WIA-Bench, an evaluation suite with in-domain (ID) and out-of-domain (OOD) splits. The authors also develop WIA WM, a family of foundation world models trained via three stages, and introduce WIA-WAM, a model that couples action and observation prediction through Dual-Perspective Inversion (DPI) and Counterfactual Simulation-Guided Policy Optimization (CSPO). Extensive experiments show that WIA achieves strong performance across benchmarks.

Critical strengths
  1. Previous works are often domain specific and lack standardized evaluation resources. To tackle this, this work introduces WIA-Corpus and WIA-Bench, which unify seven heterogeneous digital environments under a shared action-observation trajectory format, alongside paired ID and OOD splits. This establishes a standard for evaluating cross-environment generalization and offers a resource for research on general digital world models.
  2. Most existing model based LLM agents keep the world model simulator and action policy decoupled, which introduce inference overhead and fail to propagate task success gradients back into world representations. In this paper, WIA-WAM unifies both paradigms within one parameter space and achieves end-to-end gradient coupling between world model prediction and agent policy via DPI and CSPO.
Critical weaknesses
  1. This work heavily relies on external LLM-as-a-judge models to produce reward signals and almost all quantitative scores for WIA Bench, without objective metrics or human-annotated labels for calibration. This design can introduce validity risks that the RL optimizes against a proxy LLM judge instead of ground-truth environmental feedback, making it vulnerable to reward hacking and unreliable to reflect real-world task performance.

  2. The introduction motivates CSPO by noting that an action with a more accurate predicted future state does not receives stronger reinforcement. However, CSPO rewards world-view samples by the policy’s downstream success, not by fidelity to the true observation. In principle this encourages helpful observations over accurate ones. In Search, for example, the world view could be rewarded for injecting parametric knowledge into a simulated retrieval result. The paper reports neither world-view prediction quality after CSPO nor any fidelity term. Please report the world view’s prediction accuracy before and after CSPO, together with a λ_w = 0 ablation.

Questions for authors
  1. During DPI, given that action generation and world model prediction share the same sequence backbone during joint optimization, how does the method avoid catastrophic forgetting or mutual interference between the two learning targets?

  2. What decontamination techniques are performed between the training sources and the OOD environments, particularly SWE-bench Verified, WebArena, and TerminalBench 2?

  3. How do simulation SFT and simulation RL perform when (a) WIA-WM is replaced by a generic LLM simulator, and (b) the same teacher trajectories are collected in the real AWM environments?

  4. For CSPO: How accurate is the world view after CSPO training, and what happens with λw=0\lambda_w = 0? What values of λp\lambda_p and λw\lambda_w are used? How are prompt groups handled when all policy rollouts receive zero reward?

  5. Could you clearly state your contributions comparing with previous works, such as MuZero, Dreamer, and others?

10 comments
AuthorsOpenReview ↗

Response 0

Thank you for your careful review and constructive suggestions on evaluation reliability, DPI/CSPO, data separation, simulator comparisons, and related work. We will revise the paper to incorporate the additional experiments, overlap audits, and implementation details presented below, and clarify the scope of the conclusions. We will also add comparisons with MuZero and Dreamer to the revised paper. The following responses address your concerns and questions individually.

Reward-hacking mitigation. We designed the reward and evaluation protocol following the separation principle described in Qwen-AgentWorld: Language World Models for General Agents [1] (§§3.4.1–3.4.2 and 4.1–4.2). That work combines open-ended multi-dimensional rubric judging with executable rule-based verifiers, using the verifier as an objective correctness anchor; it also strictly extracts the predicted observation before judging and distinguishes deterministic content through domain-specific criteria. We borrow this design principle in our judge prompts and evaluation protocol: the judge receives the isolated predicted observation, applies domain-specific criteria, and scores against the corresponding ground-truth OOD observation. These constraints reduce opportunities for self-praise, prompt-keyword matching, and other proxy exploits, thereby mitigating reward hacking, although they do not make open-ended judging completely immune to it.

Additional cross-judge robustness. We have further incorporated an evaluation of OOD splits, employing Kimi-K3 and Qwen3.8-max as independent judge models; the corresponding results are presented in Tables 7 and 8 of the manuscript's appendix. Under Kimi-K3, GPT-5.6-sol obtains 60.42 Avg / 60.03 MAvg and Claude-Opus-4.8 obtains 58.94 / 59.41, while WIA-WM-9B obtains 57.19 / 57.38, WIA-WM-27B obtains 58.91 / 59.03, and WIA-WM-35B-A3B obtains 59.15 / 58.68. Under Qwen3.8-max, GPT-5.6-sol obtains 49.57 / 49.52 and Claude-Opus-4.8 obtains 49.18 / 49.97, while the corresponding WIA-WM results are 49.38 / 49.64, 50.79 / 51.13, and 51.36 / 51.06. The larger WIA-WM variants are comparable to the frontier baselines under Kimi-K3, and WIA-WM-27B and WIA-WM-35B-A3B exceed both baselines in Avg and MAvg under Qwen3.8-max. This cross-judge consistency supports the reliability of the reported OOD conclusions beyond any single judge model.

Evaluation-reliability evidence. Responses 0 and 1 jointly address W1's concern. Response 0 establishes reward-hacking mitigations and cross-judge robustness; Response 1 complements rubric scores with full-OOD programmatic token-F1 and a blinded human cross-evaluator audit, and compares the resulting model rankings with the original DeepSeek-V4-Pro scores.

[1] Qwen Team. Qwen-AgentWorld: Language World Models for General Agents. arXiv preprint arXiv:2606.24597v1, 2026.

AuthorsOpenReview ↗

Response 1 — W1: Objective F1 and cross-evaluator checks

We address judge bias with full-OOD token F1 and a cross-evaluator audit, separating evaluation evidence from training-reward validity.

1. Non-LLM F1 evaluation. We additionally report F1 on the full WIA-Bench-OOD set, providing a complementary metric independent of LLM scoring. Scores are on a 0–100 scale; Avg averages examples and MAvg averages domains.

Table 1. Full WIA-Bench-OOD programmatic token-F1 results.

ModelAndroidMCPOSSearchSWETerminalWebAvgMAvg
GPT-5.6-luna62.6635.6261.5022.9556.7163.4767.0653.6752.85
GPT-5.6-terra64.8233.8463.8724.4759.1963.7067.3254.9153.89
GPT-5.6-sol67.0037.6073.2326.4662.0465.9070.1258.3057.48
Claude-Sonnet-567.6366.6189.0133.6950.9554.6473.3360.4562.27
Claude-Opus-4.866.5468.2986.8832.5562.7263.6575.2863.6965.13
MiMo-v2.5-pro69.5456.3185.5235.2551.8758.2272.9160.5161.37
Qwen3.8-27B65.8844.7177.3133.3750.6356.1263.1955.7455.89
Qwen3.8-max74.5954.3185.3332.6159.6566.6378.8164.2364.56
DeepSeek-V4-Flash70.1758.0685.2037.5048.7059.4172.3160.6461.62
DeepSeek-V4-Pro71.4949.6289.0938.2048.1659.9971.6960.7661.18
Qwen3.5-4B62.3347.1475.6629.7432.2434.4761.5347.6949.02
Qwen3.5-9B65.1349.8682.2029.8537.0237.7364.2550.9052.29
Qwen3.5-27B69.3551.2088.1832.7045.1950.6370.4657.2958.24
Qwen3.5-35B-A3B68.9451.2779.3430.1841.9144.2066.2253.4254.58
Qwen-AgentWorld68.0044.3786.8637.6957.6459.3671.9061.0460.83
WIA-WM-4B4.9451.6780.3838.7537.5658.1065.9247.1948.19
WIA-WM-9B71.6064.3189.3746.4242.0261.6372.3162.5763.95
WIA-WM-27B74.7762.1887.2443.2153.7164.5969.2564.0464.99
WIA-WM-35B-A3B71.6452.5988.5447.0648.6463.6571.5263.0863.38

Among these 19 models, F1 Avg ranges from 47.19–64.23. Qwen3.5-4B scores 47.69 and GPT-5.6-sol 58.30, showing that a limited aggregate spread is not unique to rubric judging. WIA-WM-27B ranks second in Avg at 64.04, only 0.19 below Qwen3.8-max, and second in MAvg at 64.99, versus Claude-Opus-4.8's 65.13. WIA-WM-9B and 35B-A3B also exceed GPT-5.6-sol and Qwen-AgentWorld in both averages. These results support WIA's competitiveness under a non-LLM scorer; token overlap is complementary to, not a validation or replacement of, semantic judging.

2. Human review conducted with the assistance of the Codex tool. We invited five human volunteers to conduct a blinded review of 56 cases, eight per domain, with Codex used as a tool to facilitate the review workflow and record the evaluations; the volunteers, rather than Codex, made the judgments. Domain columns and Panel Avg show mean ranks; Pro Avg shows the original DeepSeek-V4-Pro score on the same cases (20–100). Position columns give overall rankings.

Experimental setup. For each case, all 5 volunteers ranked the outputs from the different models, casting ranking votes on the same set of outputs. We first averaged the 5 volunteers' ranks for each model on each case, then averaged these per-case mean ranks across the cases in each domain and across the full 56-case subset.

Table 2. Codex-assisted human-panel rankings and DeepSeek-V4-Pro scores.

ModelAndroidMCPOSSearchSWETerminalWebPanel Avg rank ↓Panel position ↓Pro Avg score ↑Pro position ↓
GPT-5.6-sol2.7002.9503.2503.2254.1752.5003.4753.182364.293
Qwen3.8-max3.8503.1253.9503.1252.1003.4502.3253.132264.392
DeepSeek-V4-Pro2.8252.9252.9753.0753.3503.6504.0503.264462.824
Qwen3.5-35B-A3B3.1753.3753.6254.5754.3253.8253.9753.839558.435
WIA-WM-35B-A3B2.4502.6251.2001.0001.0501.5751.1751.582189.861

The volunteer panel and Pro scores produce the same overall ordering on these 56 cases: WIA-WM-35B-A3B, Qwen3.8-max, GPT-5.6-sol, DeepSeek-V4-Pro, and Qwen3.5-35B-A3B. Agreement extends across all five positions, including the relative ordering of the four baselines. This agreement supports the reliability of WIA-Bench's model rankings.

AuthorsOpenReview ↗

Response 2 — W2: CSPO fidelity and implementation

Controlled comparison. We ran a matched four-way comparison from the same WIA-CPT-4B initialization on WebShop and Search-R1. When λw=0\lambda_w = 0, only the policy-view objective remains, and CSPO reduces to standard GRPO with rollout_n=8. DPI + GRPO is the closest matched control to disabling the CSPO/world branch during RL, while BC + CSPO tests whether CSPO can compensate for the absence of DPI. The table reports the final task metrics and a programmatic world-view token-F1; the latter is a complementary lexical measure, not a replacement for semantic evaluation.

Table 3. Matched WIA-CPT-4B task-performance comparison of BC/DPI and GRPO/CSPO on WebShop and Search-R1.

MethodWebShop Score ↑WebShop success ↑Search Accuracy ↑Search Avg Turns ↓
BC + GRPO91.1275.0050.395.14
DPI + GRPO92.3377.3451.565.11
BC + CSPO91.2574.2050.395.07
DPI + CSPO (WIA-WAM)93.2178.1351.954.98

Table 4. Matched Obs-text F1 evaluation for the four WAM training variants.

MethodWebShop Obs-text F1 ↑Search Obs-text F1 ↑
BC + GRPO0.00480.0556
DPI + GRPO0.47150.3949
BC + CSPO0.02670.0559
DPI + CSPO (WIA-WAM)0.47290.3992

Obs-text F1 is token-level F1 between generated observation text and reference observation text; it measures lexical overlap/fidelity and is not the semantic task score.

The main pattern is consistent across the two domains. Holding GRPO fixed, adding DPI raises WebShop score/success by +1.21/+2.34 points and Search accuracy by +1.17 points, while increasing Obs-text F1 from 0.0048 to 0.4715 on WebShop and from 0.0556 to 0.3949 on Search. Holding DPI fixed, adding CSPO further improves WebShop score/success by +0.88/+0.79 points and Search accuracy by +0.39 points; it also reduces Search turns from 5.11 to 4.98. In contrast, BC + CSPO changes WebShop success from 75.00 to 74.20 and leaves Search accuracy unchanged, with only negligible changes in Obs-text F1. This interaction is why we view DPI as the mechanism that establishes a usable shared world-view representation, and CSPO as the mechanism that turns that representation into better decisions.

Fidelity versus usefulness. The controlled results do not support the stronger claim that continuation reward alone guarantees higher surface-level observation fidelity. In fact, the CSPO increment in Obs-text F1 is modest (+0.0014 on WebShop and +0.0043 on Search), whereas the task gains are clear. We therefore interpret CSPO as improving decision-useful world predictions: a sampled observation is reinforced when the policy can continue from it and obtain a better task outcome, rather than merely when it has higher lexical overlap. The large BC-to-DPI F1 jump, followed by the DPI-to-CSPO task gain, is consistent with this division of labor.

Search-specific evidence. Search is a useful stress test because correctness depends on multi-step retrieval and temporal/source grounding. WIA-WAM reaches 51.95 accuracy and 4.98 average turns, compared with 50.39/5.14 for BC + GRPO; the corresponding Obs-text F1 is 0.3992, compared with 0.0556 for the policy-only baseline. Thus, the gain is not obtained by simply taking longer trajectories: accuracy increases while the average number of turns decreases.

Implementation and reward handling. DPI renders each expert trajectory once in the agent view and once in the world view, with a 1:1 mixture; action tokens are optimized in the agent branch and observation tokens in the world branch. In the reported CSPO implementation, each prompt uses 8 policy rollouts, 4 world rollouts, and 4 predicted-observation continuations for 200 update steps. We set λp=λw=1.0\lambda_p = \lambda_w = 1.0, assigning equal weight to the two distinct optimization objectives in this experiment. The policy coefficient λp\lambda_p weights the policy-view objective, which updates the reasoning, action, and answer tokens in policy rollouts. The world coefficient λw\lambda_w weights the world-view objective, which optimizes the observation tokens in world rollouts using rewards obtained from predicted-observation continuations. These continuation rewards are treated as fixed signals: gradients are not backpropagated through the continuation generation process into the world predictions. The two branch losses are computed separately and jointly update the shared model parameters.

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

The paper presents unified foundation world models for digital agents. The authors collect a corpus WIA-Corpus covering continual pre-training (CPT), SFT, and RL data in 7 domains, fine-tune various Qwen3.5 models from 4B to 35B-A3B on WIA-Corpus, and present a series of WIA-WM models. It also introduces a world-bench evaluation setting including in-domain and out-of-domain test data. WIA-WM models outperform their backbone models and both closed-source and open-source LLMs. The improvement on in-domain test data is more significant. The improvement in world models is also reflected in improved agent task performance.

Critical strengths
  1. The paper collected large-scale training data for the CPT, SFT, and RL stages in 7 digital domains. The data can benefit future research on digital agents and world models.

  2. The paper fine-tuned LLMs on both action prediction and observation prediction via Dual-Perspective Inversion. The ablation shows that training the same model for both action and observation prediction outperforms training world models alone.

Critical weaknesses
  1. The author claims that the main motivation for the proposed foundation world model is that current world models are fragmented (tied to specific domains). But the experiments do not compare the trained foundation world models with other world models trained on a specific domain, such as a sub-domain in the WIA-Corpus. In order to justify the benefits of having a foundation world model, it should outperform a world model trained on a specific domain. It does not answer the question of why web data improves world models in terminal use.

  2. The metric in WIA-bench may not be discriminative enough. From Table 1, the weakest model Qwen 3.5-4B is able to achieve a score of 40.86 on average, while much more powerful models such as GPT-5.6-sol and GLM-5.3, as well as the state-of-the-art fine-tuned models, only score below 55. What does the breakdown metric reveal when comparing these models as world models for digital environments?

Questions for authors

See the weakness section.

6 comments
AuthorsOpenReview ↗

Response 0

Thank you for your careful review and constructive suggestions on domain specialization, cross-domain transfer, and the discriminative power of WIA-Bench. We will revise the paper to incorporate the specialist comparisons, complementary token-F1 evaluation, and six-dimensional analysis presented below. We will include the full specialist setup and cross-domain difference matrix in the revised appendix, and clarify the interpretation and limitations of these results. The following responses address your concerns and questions individually.

AuthorsOpenReview ↗

Response 1 — W1&Q1: Benefits over domain specialists and cross-domain transfer

Setting. We compare jointly trained WIA-WM-9B, trained with RL across all seven domains, with five domain-only RL specialists: Android-only, Web-only, Search-only, SWE-only, and MCP-only. Each specialist starts from the shared WIA-SFT-9B initialization and performs one epoch over its own domain-specific RL training set.

All rows use the same 1,739-case WIA-Bench-OOD evaluation subset scored by DeepSeek-V4-Flash. Avg weights examples equally; MAvg weights the seven domains equally. The MCP subset differs from the original paper's Table 1 subset, although both contain 121 TAU-2 examples.

The Oracle-routed MoE is an assembled reference that selects the highest-scoring specialist separately for each evaluation domain.

Table 1. Cross-domain comparison between single-domain RL specialists and all-domain WIA-WM-9B.

Model / RL training domainAndroidMCPOSSearchSWETerminalWebAvgMAvg
WIA-SFT-9B47.1134.2564.0840.5852.1155.6454.0250.5649.68
Android-only RL50.9031.3666.0134.1043.8851.1454.8948.0747.47
Web-only RL51.0434.7165.4935.6845.9056.1857.2550.0649.46
Search-only RL47.3135.3755.2951.3147.1855.4755.0050.6849.56
SWE-only RL49.1934.9958.0741.1461.0462.0955.2352.9951.68
MCP-only RL51.2536.8363.8740.1148.8059.7455.6451.6050.89
Oracle-routed MoE (best specialist per domain)51.2536.8366.0151.3161.0462.0957.2556.4855.11
WIA-WM-9B (all-domain RL)53.5037.0269.4149.3761.2263.0758.4657.3056.01

Benefit over specialists. Jointly trained WIA-WM-9B achieves the highest Avg/MAvg (57.30/56.01), exceeding the strongest specialist, SWE-only (52.99/51.68), by 4.31/4.33 points. Its lead under both sample-weighted and domain-balanced averaging shows that the aggregate advantage persists across the two weighting schemes.

Advantage over oracle-routed MoE. WIA-WM-9B exceeds the favorable oracle-routed MoE by 0.82 Avg and 0.90 MAvg. This is a strong result: even when each evaluation domain is assigned its best specialist checkpoint, a single jointly trained WIA-WM-9B achieves higher sample-weighted and domain-balanced performance without domain-dependent model switching.

Strong performance on specialist domains. On Android, Web, SWE, and MCP, the joint model exceeds the corresponding domain specialist by 2.60, 1.21, 0.18, and 0.19 points, respectively. Joint training therefore combines broad coverage with competitive in-domain performance. Search-only reaches 51.31, while the joint model scores 49.37 and leads across the other six domains. On OS and Terminal, the joint model reaches 69.41 and 63.07, exceeding the best single-domain variant in each column by 3.40 and 0.98 points.

Cross-domain gains. Relative to the shared WIA-SFT-9B reference, SWE-only improves Terminal by 6.45 points, MCP-only improves Android and Terminal by 4.14 and 4.10, and Web-only improves Android by 3.93. These observed gains extend beyond each variant's RL training domain. The joint model also scores above the SFT reference in all seven domains, with differences ranging from 2.77 to 9.11 points and aggregate gains of 6.74 Avg and 6.33 MAvg. The largest differences are in SWE (+9.11), Search (+8.79), and Terminal (+7.43), spanning code execution, retrieval, and command-line interaction.

Conclusion. The results support the practical benefit of a reusable cross-domain world model: WIA-WM-9B combines the strongest aggregate scores with leading performance in six of seven domains and exceeds the corresponding specialist in four of the five specialist training domains. Together with the observed cross-domain gains, this provides empirical support for joint multi-domain training as an effective way to consolidate broad observation-prediction capability in one checkpoint.

AuthorsOpenReview ↗

Response 2 — W2&Q2 part1: Score range and complementary F1 evaluation

We address the aggregate range here and the capability breakdown in the next response.

Motivation. This experiment directly addresses the concern that the WIA-Bench metric may not be sufficiently discriminative. In the paper's Table 1, the weakest model, Qwen3.5-4B, reaches 40.86 Avg, while substantially more capable models such as GPT-5.6-sol and GLM-5.3, together with state-of-the-art fine-tuned models, remain below 55. We therefore add a complementary, programmatic observation-text token-F1 evaluation to test whether the apparent score compression is specific to rubric judging or reflects a broader limitation of the benchmark. F1 measures directly verifiable observation-text agreement, providing an independent check on model separation and on whether WIA-WM's advantage extends to explicit state-prediction fidelity. It is intended to complement the semantic WIA-Bench score rather than replace it.

Complementary F1 results. We report F1 results on the same OOD evaluation set used in the paper. Scores are reported on a 0–100 scale. Avg averages all examples; MAvg equally weights the seven domain scores.

Table 2. OOD programmatic observation-text token-F1 results.

ModelAndroidMCPOSSearchSWETerminalWebAvgMAvg
GPT-5.6-luna62.6635.6261.5022.9556.7163.4767.0653.6752.85
GPT-5.6-terra64.8233.8463.8724.4759.1963.7067.3254.9153.89
GPT-5.6-sol67.0037.6073.2326.4662.0465.9070.1258.3057.48
Claude-Sonnet-567.6366.6189.0133.6950.9554.6473.3360.4562.27
Claude-Opus-4.866.5468.2986.8832.5562.7263.6575.2863.6965.13
MiMo-v2.5-pro69.5456.3185.5235.2551.8758.2272.9160.5161.37
Qwen3.8-27B65.8844.7177.3133.3750.6356.1263.1955.7455.89
Qwen3.8-max74.5954.3185.3332.6159.6566.6378.8164.2364.56
DeepSeek-V4-Flash70.1758.0685.2037.5048.7059.4172.3160.6461.62
DeepSeek-V4-Pro71.4949.6289.0938.2048.1659.9971.6960.7661.18
Qwen3.5-4B62.3347.1475.6629.7432.2434.4761.5347.6949.02
Qwen3.5-9B65.1349.8682.2029.8537.0237.7364.2550.9052.29
Qwen3.5-27B69.3551.2088.1832.7045.1950.6370.4657.2958.24
Qwen3.5-35B-A3B68.9451.2779.3430.1841.9144.2066.2253.4254.58
Qwen-AgentWorld68.0044.3786.8637.6957.6459.3671.9061.0460.83
WIA-WM-4B4.9451.6780.3838.7537.5658.1065.9247.1948.19
WIA-WM-9B71.6064.3189.3746.4242.0261.6372.3162.5763.95
WIA-WM-27B74.7762.1887.2443.2153.7164.5969.2564.0464.99
WIA-WM-35B-A3B71.6452.5988.5447.0648.6463.6571.5263.0863.38

Among the 19 models reported here, F1 Avg ranges from 47.19–64.23. Qwen3.5-4B scores 47.69 and GPT-5.6-sol 58.30, showing that a limited aggregate spread is not unique to rubric judging. WIA-WM-27B ranks second in Avg at 64.04, only 0.19 below Qwen3.8-max, and second in MAvg at 64.99, versus Claude-Opus-4.8's 65.13. WIA-WM-9B and 35B-A3B also exceed GPT-5.6-sol and Qwen-AgentWorld in both averages. These results support WIA's competitiveness under a non-LLM scorer; token overlap is complementary to, not a validation or replacement of, semantic judging.

Related benchmark context. Qwen-AgentWorld [1] evaluates AgentWorldBench across seven domains and 2,170 cases using five rubric dimensions. Its Table 5 reports overall scores of 42.88–58.71: GPT-5.4 reaches 58.25, and the 35B-A3B/397B-A17B world models reach 56.39/58.71.

Baseline selection matters. The comparison models in Table 1 are post-trained models released in 2026, including frontier proprietary and open-weight models. Importantly, our ablation evaluates Qwen3.5-4B-Base at 35.74 OOD Avg, below the post-trained Qwen3.5-4B's 40.46. WIA-Bench therefore assigns a lower score to this weaker pretrained baseline; the approximately 40–55 interval is not a fixed scoring floor or ceiling. The observed range should be interpreted in light of this recent, post-trained comparison set.

[1] Qwen Team. Qwen-AgentWorld: Language World Models for General Agents. arXiv:2606.24597, 2026, Section 4 and Table 5.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Manuscript Revisions

We thank the reviewers for their constructive feedback. We have revised the manuscript to clarify the method presentation and add evaluation, robustness analyses, simulator-transfer experiments, and training-data overlap audits. The revised text and captions are marked in blue. The main changes are summarized below; all figure and table numbers refer to the revised manuscript.

  1. Clearer framework and training figures (Figures 2–4). We reorganized the original three-row method figure into a framework overview and a dedicated world-model training figure. Figure 2 now presents the seven digital domains, WIA-WM training and WIA-Bench evaluation, the three simulator applications, and WIA-WAM at a higher level. Figure 3 retains the textual input examples and CPT–SFT–RL training pipeline, making the supervision at each stage easier to follow. The revised WIA-WAM illustration in Figure 4 clarifies the DPI views, CSPO optimization branches, reward propagation, and training objectives. We also updated the corresponding captions and main-text references.

  2. Explicit simulator workflows (Figure 7). We added a workflow diagram on the first page of the appendix covering simulation SFT, simulation RL, and inference lookahead. It distinguishes frozen and trainable components, shows the supervision and reward flow, and explains how predicted observations support training or action selection while real-environment feedback updates the history during deployment.

  3. Broader evaluation of world-model predictions (Tables 7, 8, 13, 14, and 18). We added OOD evaluations using Kimi-K3 and Qwen3.8-max to examine sensitivity to the judge family. We also report reference-based token-F1 independently of LLM judging and a Codex-assisted human review of 56 selected OOD cases, with eight cases per domain. In addition, we disaggregate the original evaluation into Format, Factuality, Consistency, Realism, Quality, and Efficiency, making the sources of the aggregate score differences visible.

  4. Joint versus domain-specialized training (Table 12). We added a comparison between joint WIA-WM-9B training and five domain-specialized RL variants sharing the same WIA-SFT-9B initialization. The results characterize cross-domain transfer and specialization: joint training achieves the highest aggregate scores and leads on six of seven domains, while the Search specialist performs best on Search.

  5. Multi-step prediction evaluation (Table 15). We extended the evaluation beyond individual next-observation predictions by reporting successive prediction scores over two- or three-step rollouts in Android, Search, SWE, Terminal, and Web. The experiment includes eight models and 50 trajectories per domain, allowing readers to inspect how prediction quality changes over successive simulated transitions.

  6. Downstream simulator-transfer comparison (Table 16). We added a controlled policy-training comparison using a common Qwen3.5-4B initialization and the same 2,310-task AWM pool, with evaluation on TAU-2. The table compares WIA-WM with alternative simulators, real-AWM training, and a no-additional-RL baseline. It reports Airline, Retail, and Telecom pass@1 scores separately, together with their macro average, to expose both overall transfer utility and domain-specific differences.

  7. Decoding and seed robustness (Table 17). We added evaluations of all four WIA-WM scales under additional decoding settings and random seeds. The table reports the original results alongside seed2 and seed42 runs, including per-domain scores and both aggregate measures. The accompanying text specifies the temperature and top-p settings and reports the observed score variation.

  8. Training-data overlap audits (Table 19 and accompanying text). We added a comparison of all 3,547 WIA-Bench examples against the 5,032,034-record CPT corpus, reporting the maximum token-F1 overlap distribution separately for ID and OOD. The revised text also reports exact-match checks against the released SFT and RL data using the audited last-observation and ground-truth fields. These additions make the audit scope and matching criteria explicit.

  9. Improved seven-domain modeling visualizations and methodological context (Figures 18–24). We replaced the appendix world-model rollout visualizations across Android, MCP, OS, Search, SWE, Terminal, and Web. The revised panels and captions make the correspondence between textual observations, actions, predicted observations, and visual state transitions clearer.

AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair,

We thank the reviewers for recognizing WIA's seven-domain data and benchmark, extensive experiments, and integration of world modeling with agent learning. Through complementary evaluations, controlled ablations, simulator comparisons, and methodological clarifications, we addressed their central questions and strengthened the paper. All three reviewers now give a positive rating of 3: rELH and PuL5 explicitly maintained their positive ratings, while TiFV raised their rating from 2 to 3, stating that "Most of my original concerns have been addressed."

  1. Cross-family LLM judges (TiFV; PuL5: judge reliability). To establish cross-family evaluation reliability, we added Kimi-K3 and Qwen3.8-max evaluations. The larger WIA models remain competitive under Kimi-K3, while WIA-WM-27B and 35B-A3B exceed GPT-5.6-sol and Claude-Opus-4.8 in both averages under Qwen3.8-max. These results support WIA's strong performance across distinct judge families.

  2. Programmatic token-F1 and human evaluation (PuL5, TiFV, rELH). To complement LLM judging with independent evidence, we added programmatic token-F1 on all 1,739 OOD examples; WIA-WM-27B ranks second in both averages. The rebuttal also reports a blinded evaluation by five human volunteers on 56 cases, with Codex assisting the workflow. Human judgments and DeepSeek-V4-Pro produce the same overall ordering of all five models. These checks support evaluation reliability.

  3. Six-dimensional score breakdown (rELH). Gains span Format, Factuality, Consistency, Realism, Quality, and Efficiency. WIA-WM-27B and 35B-A3B rank in the top three on every dimension, supporting broad improvements rather than reward hacking through a single scoring dimension.

  4. Decoding and seed robustness (TiFV). We evaluated all four fixed WIA-WM checkpoints under additional decoding settings and inference seeds. Avg spreads are only 0.07, 0.37, 0.25, and 0.10 points for 4B, 9B, 27B, and 35B-A3B, respectively, and the Avg ordering remains unchanged. This supports stable aggregate performance.

  5. Benefits of unified cross-domain modeling (rELH). We compared joint WIA-WM-9B with five domain-specific RL specialists sharing the same SFT initialization. The joint model achieves 57.30 Avg (sample average) and 56.01 MAvg (equal-domain macro average), exceeding the best individual specialist by 4.31 and 4.33 points, respectively, and an oracle reference selecting the best specialist for each evaluation domain by 0.82 and 0.90 points. The joint model leads on six of seven domains, while specialist-to-other-domain gains provide concrete evidence of transfer.

  6. Distinct contributions of DPI and CSPO (PuL5, TiFV). We added TiFV's requested BC+GRPO, DPI+GRPO, BC+CSPO, and DPI+CSPO comparisons on WebShop and Search-R1, starting from the same WIA-CPT-4B initialization. Under GRPO, DPI increases observation F1 from 0.0048 to 0.4715 on WebShop and from 0.0556 to 0.3949 on Search while improving policy performance, supporting compatible action/world learning. Adding CSPO after DPI further improves WebShop success by 0.79 points and Search accuracy by 0.39 points, with improved observation F1 and fewer Search turns. These comparisons clarify the complementary roles: DPI establishes observation-prediction capability, and CSPO translates it into decision utility.

  7. Reward flow and implementation (PuL5, TiFV). We clarified continuation-derived world rewards, shared-parameter updates, loss weights, rollout counts, zero-reward handling, and WIA-CPT initialization. These details explain how DPI establishes joint action/world learning and CSPO uses continuation outcomes to optimize policy utility.

  8. Multi-step prediction (TiFV) and simulator transfer (PuL5). We added 250 trajectories with 650 prediction targets across five domains, evaluating successive predictions over 2-3 transitions. WIA outperforms its same-size backbone at every tested step in SWE, Terminal, and Web across all four scales. A common-initialization RL comparison on the same 2,310-task AWM pool evaluates generic simulators, Qwen-AgentWorld, real AWM, and WIA-WM. WIA-WM achieves the highest TAU-2 macro pass@1 of 50.12%, exceeding Qwen-AgentWorld by 10.65 points and the no-additional-RL reference by 3.43 points. This connects stronger predictive modeling to tangible downstream agent gains.

  9. Data audits (PuL5, TiFV) and presentation (TiFV). We reported source separation, SFT/RL exact-match checks, and a lexical audit of 5,032,034 CPT records against 3,547 benchmark examples. Revised figures address TiFV's presentation request; MuZero/Dreamer comparisons address PuL5's positioning question.

Overall, the rebuttal reinforces WIA as a cohesive contribution spanning resources, predictive models, and actionable learning. The expanded evidence and positive feedback from all three reviewers support its value as a foundation for further research on general digital agents.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.