Interactive World Models: From "Looks Real" to "Useful for Decisions" (Deep Dive)

1. Background: What a World Model Is, and Why "Looking Real" Is Not Enough

An October introductory survey starts from the partially observable Markov decision process (POMDP) and defines a world model as: a model that, given past observations and actions, predicts future observations (or future states) [13]. It can serve three purposes: as a simulator to train agents, as an internal rollout for planners, and as an evaluation environment for policies.

Over the past two years, video world models have mostly developed along the path of generative models, and their evaluation criteria were inherited from video generation: sharpness, perceptual similarity, temporal coherence. Genie 4 represents a new high on this path — realistic human motion, physical phenomena, native audio, and real-time language control [1]. But DeepMind itself writes in its "current limitations" that object permanence drifts over time, and complex physical interactions, real-time audio accuracy, text rendering, spatial accuracy, and long-term consistency remain open problems; the future direction is a "hybrid world model" that combines the structured reasoning of code and language models with pixel-level simulation [1].

The problem is that when used for decisions, the distribution of pixel error does not match the points where decisions are sensitive. A pixel loss treats the background, the robot arm's appearance, and task-irrelevant salient regions all the same, whereas policy learning cares about the states the policy actually visits and actions with large advantage [2]. This is the shared starting point of a batch of October papers.

Technical lineage: how three lines of world-model research are converging

An October introductory survey [13] provides a useful map of the field. Classified by "what is predicted and how it is used," world models have evolved along roughly three lines.

Line 1: latent dynamics for planning and policy learning. "Reconstruction-based" models such as PlaNet and Dreamer learn latent dynamics together with observation reconstruction; JEPA-style methods predict only future latent representations without pixel reconstruction, with V-JEPA learning predictive representations from videos without action annotations, and their key difficulty is "representation collapse," where the encoder maps different observations to nearly identical vectors; and "control-oriented" methods learn dynamics together with decision-relevant quantities such as rewards or values, as when MuZero combines learned latent dynamics with tree search to select actions [13]. Models on this line typically aim not for good-looking frames but for "good enough to decide with."

Line 2: using the learned model as a simulator. The survey notes that instead of planning from scratch at every decision step, an agent can use a learned predictive model to generate synthetic trajectories and then optimize a policy on them with standard RL methods such as policy gradient or actor-critic, reducing the number of interactions with the real environment [13]. This is exactly where DRILL and OpTTA sit [2][3]: the model is generative, but its purpose is to train and evaluate policies.

Line 3: generative interactive environments. The Genie series took a different path: the original Genie infers "latent actions" from unlabeled videos and uses them to condition a generative dynamics model, making video prediction controllable without explicit action labels [13]; DeepMind released Genie 3 in August 2025 and Genie 4 in October 2026, adding native audio and real-time language steering, with Gemini in the loop maintaining coherence [1]. This line aims first for "looks right and can be interacted with."

The convergence point in October 2026. The three lines are now borrowing from each other: generative interactive models are being asked to be "useful for decisions" (DRILL trains a video world model by policy return rather than pixel error [2]); evaluation is testing whether generative models faithfully reproduce physics and failures (WMEval, WorldFailBench [4][5]); and ideas from representation learning are flowing back into video world models, where "making the latent space easier to predict" becomes key to efficiency and long-horizon consistency [7][8]. The survey also flags cross-embodiment sharing as an open problem; Cosmos 3, for example, uses embodiment-specific linear projections to map different robots' action representations into a shared latent space [13]. In one sentence: world models are shifting from "a branch of video generation" to "infrastructure for training and evaluating agents," and the acceptance criterion for infrastructure is decision utility.

2. Release Update: Genie 4's Four Capabilities and "Gemini in the Loop"

According to DeepMind's official blog [1], Genie 4's four core capabilities are:

1. Shaping worlds in real time: change the weather, move the camera, or add people or objects in natural language; Gemini rewrites prompts according to scene context, and between inputs Gemini also monitors and steers the scene so the world keeps progressing coherently.
2. Realism of physics and human motion: it can simulate physical scenarios such as aerodynamics and collisions, and people move naturally, react to their environment, and can be controlled in real time.
3. Native audio: the first real-time world model in the Genie series to generate audio natively, including ambient sound and character dialogue that can be interacted with in real time.
4. Simulating complex events with Gemini: Gemini 3.8 Flash, as the guiding agent in the loop, continuously observes the scene, breaks goals into steps, checks new frames to confirm whether each step succeeded, and adjusts the plan immediately when the world changes or the user intervenes. DeepMind also showed Gemini 4 Argon working with Genie 4 to simulate Apptronik's Apollo humanoid robot playing checkers.

As for availability, Genie 4 is first offered as a research preview to trusted testers, then to Google AI Ultra subscribers (Project Genie) and early partners on the Gemini API and Google AI Studio; all generated video and audio carry an invisible SynthID watermark [1].

Our reading: What is most noteworthy about Genie 4 is not image quality but the division of labor between "a generative model + a reasoning model" — Gemini handles structure and planning, Genie handles immediate physical and sensory rendering. This echoes the research community's push toward "decision usefulness": a pure pixel model can hardly guarantee causality and long-term consistency on its own; it needs explicit structure and goals as constraints.

3. Algorithm ①: DRILL — Training World Models for "Usefulness to the Policy"

One-line intuition: When training a robot policy in imagination, the world model's objective should not be "does the next frame look right" but "after learning inside it, does the policy perform well in the real environment?" DRILL writes this objective as a bilevel optimization and offers a cheap approximation [2].

Analogy: A flight simulator's value lies not in how realistic the scenery outside the window is, but in whether the maneuvers pilots practice in it work on a real plane. A simulator that paints beautiful clouds but gets the effect of crosswinds wrong is "pretty but harmful."

Step by step (2026-10-22, a NeurIPS 2026 paper):

1. Inner level: on imagined trajectories generated by the world model, perform one GRPO update of the vision-language-action (VLA) policy.
2. Outer level: update the world model so that "the policy after the inner update" achieves higher return in the target simulator.
3. DRILL-IMG: differentiate through the inner update directly (a meta-gradient) via truncated Hessian-vector products, requiring a small number of simulator samples.
4. DRILL-VWI: a cheap closed-form substitute — instead of penalizing pixel error uniformly, weight errors by "policy visitation frequency × advantage × model confidence," with no extra simulator samples. The paper proves that, under certain local assumptions, VWI is a first-order approximation of IMG's meta-gradient.

Key formula (DRILL-VWI weights):

w(o_t, a_t) = d̃(o_t) · sgn(Â(o_t, a_t)) · |Â(o_t, a_t)|^α · s(o_t)
w̄ = stopgrad( normalize( clip(w, −w_max, w_max) ) )
L_VWI(ϕ) = Σ_t  w̄(o_t, a_t) · ‖ Ŵ_ϕ(o_≤t, a_t) − o_{t+1} ‖²

• o_t / a_t: observation and action at step t; Ŵ_ϕ: the world model with parameters ϕ, predicting the next frame.
• d̃(o_t): "policy visitation density" estimated with k-nearest neighbors in the world model's latent space — places the policy often visits get larger weights.
• Â: group-relative advantage; sgn takes its sign; α controls how much the advantage magnitude matters.
• s(o_t): a model-confidence weight (greater than 0).
• clip / normalize / stopgrad: clip to ±w_max, normalize within the batch, and treat weights as constants when updating the world model.
• The squared-error term in the formula is illustrative; see the original for the exact form of the prediction loss.

Worked example (illustrative; numbers hypothetical): A batch contains two kinds of frames. Type A frames have large areas of flickering background — large pixel error, but the policy makes almost no key decisions there (low visitation density, advantage near 0), so their weight is near 0. Type B frames show the gripper about to touch an object — small pixel error, but large advantage and frequent visits, so their weight is amplified. After training, the model is "blurrier" on type A frames and overall pixel error rises; but it predicts type B frames more accurately, and the grasping behavior the policy learns is more reliable in the real simulator.

Results (per the paper): Across 5 public robot-arm manipulation simulators and 2 VLA base models, DRILL-IMG's average policy success rate is 13.1 percentage points higher than WMPO and 7.9 points higher than RLVR-World, while its world model's one-step pixel error on the same training data is 17% higher; the gains are consistent across random seeds and grow with the length of imagined trajectories [2].

Why it works, and when it fails: It moves the world model's "attention" from pixels to where decisions matter — a direct implementation of "useful for decisions." Failure risks: the weights depend on the current policy, so if early exploration is insufficient, the world model may predict poorly in regions the policy "has not visited yet but will need later"; the IMG version needs the target simulator and significant backpropagation memory.

4. Algorithm ②: OpTTA — "Optimistic" World-Model Adaptation at Test Time

One-line intuition: When the deployment environment differs from training, the world model must adapt online. Adapting with "prediction error" alone is not enough; OpTTA adds a "can the planner reach the goal" cost, nudging the model toward explanations that are "optimistic about reaching the goal" [3].

Analogy: A new driver in an unfamiliar city not only corrects their understanding of road conditions (prediction) but also leans toward the understanding under which "this road gets me to my destination," planning accordingly — and if that optimism is wrong, new observations soon correct it.

Step by step (2026-10-26):

1. Collect a small number of real transitions (observation, action, next observation) in the test environment.
2. A planner (such as CEM or gradient-based planning) generates candidate trajectories in the current world model and computes the cost of reaching the goal.
3. Minimize two terms jointly over the world model's parameters: the prediction loss on real transitions and the goal cost of the planner's candidate trajectories.
4. Keep planning and acting with the adapted model, with no extra environment interaction dedicated to "exploration."

Key formula:

L(ϕ) = L_pred(ϕ; real transitions) + λ · min_{τ ∈ candidates} J_goal(τ; ϕ)

• ϕ: world model parameters; L_pred: the model's prediction error on actually observed transitions.
• τ: a candidate trajectory the planner rolls out in the model; J_goal: distance from the trajectory's endpoint to the goal (cost of reaching it).
• λ: weight of the optimism term. The idea comes from "reward-biased maximum likelihood" in adaptive control.

Worked example (illustrative): A robot arm moves to a more slippery tabletop. Adapting only on prediction error, the model learns "objects slide," but may remain vague about "how to push to the goal position"; with the optimism term, among several parameter settings that explain the observations equally well, the model favors the one where "a feasible path to push to the goal exists," so the planner finds usable actions faster. If that optimistic judgment is wrong, the next real observations pull it back through L_pred.

Results (per the paper): Under visual and environmental shifts in manipulation, reaching, and navigation tasks, when adapting only the predictor, OpTTA raises the final planning success rate by 13.5 percentage points on average over the frozen model, versus 6.9 points for AdaJEPA, which uses prediction loss only; when the encoder is adapted as well, the figures are 13.5 and 7.5 points; the experiments cover 3 JEPA world models [3].

Why it works, and when it fails: It brings "task" information into adaptation, reducing adjustment wasted on goal-irrelevant details. The risk is that optimism bias can lead to overconfidence in high-stakes settings and must be paired with safety constraints.

5. Algorithm ③: WorldPager — Long-Term Memory on a Single Consumer GPU

One-line intuition: An interactive world model must remember "what a place should look like when we return to it," often by using past segments' KV caches as retrieval memory — but this memory quickly exhausts RAM. WorldPager finds that the bottleneck of such memory is capacity, not bandwidth, and that because the camera pose is an input, which memory will be read is known before computation — so it can be tiered and prefetched [9].

Analogy: A librarian who knows which shelf a reader will visit next carries those books from the stacks to the front desk in advance, so the reader never has to wait in line.

Step by step (2026-10-26):

1. Diagnosis: on a machine with one RTX 4090, 32 GiB of host memory, and one NVMe SSD, Matrix-Game 2.0's KV memory exhausts host memory after 45 seconds of video at bf16 precision and 2.8 minutes at int4; the model's positional-encoding horizon is 4.3 minutes.
2. Modeling: build an analytical model that predicts when memory "overflows" and whether stalls will occur; on held-out configurations, the overflow point is predicted to within one segment.
3. Tiered storage: store int4 KV across three tiers — GPU memory, host memory, and NVMe — instead of recomputing it.
4. Pose-based prefetching: using extrapolated camera poses, move the memory the next segment needs to the GPU in advance; on a miss, do not wait, and write back in the background.
5. Position handling: a training-free position handling scheme that breaks through the 4.3-minute positional horizon.

Key formulas:

usable duration ≈ available memory / (KV bytes produced per second of video)
no-stall condition: prefetch time ≤ compute time to generate one segment

• KV bytes per second: determined by frame rate, tokens per frame, number of layers, head dimension, and precision (bf16 or int4).
• Prefetch time: time to read the next segment's memory from NVMe or host memory.

Worked example: Per the paper, host memory is exhausted after 45 seconds at bf16, extended to 2.8 minutes at int4 (about 3.7×, close to the 4× compression from 16-bit to 4-bit). Adding NVMe to the storage hierarchy removes the host-memory capacity limit; as long as "the time to read the next segment's memory" is less than "the time to compute one segment," there are no stalls.

Results (per the paper): WorldPager generates 37.5 minutes of continuous video with zero stalls and constant host-memory usage, at about 9.5 frames per second; on a closed test set, int4 storage shows no measurable quality loss, and on long-interval revisits, image quality drops by about 1 MUSIQ point [9].

Why it works, and when it fails: It exploits a structural feature of interactive world models — the access pattern is determined by the input pose and can be known in advance. It fails where poses are hard to extrapolate (e.g., sudden teleportation or sharp turns), when prefetching misses and the system must fall back to on-demand loading.

6. Evaluation: Can a World Model Serve as the "Examiner"?

WMEval (2026-10-24) [5]: uses Unity physics simulation as the "answer key," executes the same action sequences in 9 scenes, and compares the world model's states with the simulator's, evaluating action controllability, scene and character consistency, prompt following, kinematics, fluid dynamics, physical causality, and lighting. The authors point out that many key dimensions are "invisible" to conventional perceptual metrics. According to the paper, in the collision-dependent billiards scene the two evaluated models track the ball closely, with final displacement errors of only 0.83 m and 0.45 m, yet their robust mechanics errors are 26.8% and 23.3%, about five times Unity's own 4.9%; in the racing scene, Project Genie executes the commanded turn but overshoots the endpoint by 9.39 m. Large viewpoint changes expose problems too: Project Genie's character consistency in the billiards scene drops from 76.2% under small perturbations to 29.8% over a full 360° orbit. The authors conclude that generating action-responsive motion is far easier than reproducing its momentum and energy.

WorldFailBench (2026-10-24) [4]: asks a sharper question — if a world model is used to evaluate robot policies, can it faithfully reproduce how the policy fails, at which stage it fails, and under what conditions it turns from success to failure? The benchmark covers LIBERO, RoboTwin simulation, and real G1 humanoid-robot tasks, and tests 7 world models. Per the paper, in LIBERO's failure-boundary evaluation, the world models' order accuracy (OA, defined below) in judging "which direction of change makes the policy less reliable" is only 0 to 0.175, versus 0.85 to 0.95 for the simulator's own repeated controls; world models also cannot reliably quantify the effect of scene changes on success rates.

Algorithm ④: Two Failure-Boundary Metrics — OA and GE

One-line intuition: To test whether a world model can serve as the "examiner," it is not enough to compare overall success rates; we must see whether it reproduces "when a condition changes slightly, which way the success rate moves, and by how much" [4].

Analogy: To evaluate a driving simulator, it is not enough to check whether novices' "overall pass rate" matches the real road test; we must also check whether directions such as "pass rates drop on wet roads and rise with good visibility" match reality.

Step by step:

1. Fix the task, the policy, and the base scene, and choose several pairs of adjacent conditions (xᵢ, xⱼ) that differ only slightly in object position, orientation, or size.
2. Run the policy many times in the reference environment (simulator or real robot) and in the world model, obtaining the success rate S(x) under each condition.
3. Compute the change in success rate Δ(i, j) = S(xᵢ) − S(xⱼ).
4. Use OA to measure whether the directions of change agree, and GE to measure the error in their magnitude.

Key formulas:

Δ_r(i, j) = S_r(x_i) − S_r(x_j),   r ∈ {ref, wm}
OA = (1/|B|) · Σ_{(i,j)∈B} 1[ Δ_ref(i, j) · Δ_wm(i, j) > 0 ]
GE = (1/|B|) · Σ_{(i,j)∈B} | Δ_ref(i, j) − Δ_wm(i, j) |

• S_r(x): the policy's empirical success rate in environment r under condition x; ref is the reference environment, wm the world model.
• B: the set of selected condition pairs; |B|: the number of pairs.
• 1[·]: 1 if the condition holds, otherwise 0; the two changes must have the same sign to count as "directionally consistent," and if either is zero it counts as 0.
• OA (order accuracy): higher is better; GE (gap error): lower is better.

Worked example (illustrative): Three pairs of conditions. In the reference environment, Δ is +0.3, −0.2, and +0.1; in the world model, +0.1, +0.05, and 0. Only the first pair has the same sign, so OA = 1/3; GE = (0.2 + 0.25 + 0.1)/3 ≈ 0.18. Even if the world model's "average success rate" is close to the reference environment's, it has learned almost nothing about the failure boundary.

Why it works, and when it fails: OA makes the question "can it be used for decisions" concrete as "can it give the correct direction of sensitivity" — exactly the information most needed to screen risks before deployment. The limitation is that stable success-rate estimates require many repeated runs, which is costly on real robots.

AnyGoal (2026-10-19) [6]: examines compositional generalization — every state factor and action lies within the training support, but the combinations are rare in expert trajectories. Per the paper, several current world-model families fail consistently under such recombination.

Dual Exam (2026-10-22) [12]: examines agents (with fixed environment responses) and world models (with fixed action interventions) separately. Per the paper, in the agent track GPT 6 Astra has the highest weighted index at 31.3, and in the world-model track GPT-Image-2 has the highest total score at 21.7 (the tasks and metrics differ, so the two cannot be compared directly); in a 150-case pilot of a coupled "agent + world model" pair, the number of successful tasks was 0.

Together, these evaluations show that there is a huge gap between "looks right" and "fails right" — and the latter is precisely what matters most when world models are used for policy evaluation and risk scenario analysis.

7. Representation and Architecture: Making World Models Easier to "Predict"

• Predictable latents (2026-10-26) [7]: video world models usually predict in the latent space of an autoencoder "trained for reconstruction," which is not optimized for dynamics modeling. The authors find that adding a temporal representation-learning objective, combined with a "past-conditioned" decoder, yields — at the same training budget — up to 5× cheaper inference, more than 7× faster training, and 15%–24% lower pixel and perceptual error out of distribution and over long horizons; generated results also agree better with real-world policy evaluation.
• CADRE (2026-10-21) [8]: a family of causal diffusion representation autoencoders released by Kyutai and others to answer "where, and how much, should a world model compress?" Per the paper, for manipulation tasks, temporal compression is more effective in the autoencoder than in the world model; manipulation and navigation need opposite compression strategies — temporal compression is cheap when the camera is static, and spatial compression is cheap when the camera moves.
• Mix Forcing (2026-10-24) [10]: when converting a bidirectional video diffusion model into an autoregressive interactive model, plain teacher forcing destroys the pretrained visual prior. It trains jointly with "bidirectional + autoregressive mixed attention" (MoBA) and then distills into a 4-step autoregressive generator, supporting per-segment text interaction and camera control; the paper reports no quality degradation over an hour of generation.
• World-in-Context (2026-10-23) [11]: puts all kinds of control conditions into the context and uses positional encodings that change with accumulated control; per the paper, a 5B model matches the 22B EchoWM on WBench navigation (81.5 vs. 81.6), with a viewpoint-switching success rate of 93.8% versus 18.8% for a same-size baseline.

8. Side-by-Side Comparison

9. Our Views

First, the evaluation standard for world models is shifting from "generation quality" to "decision utility." DRILL's "worse pixels, better policy" [2] is a landmark result: it shows the two objectives are not only different but sometimes in conflict. Over the next year, we expect more world models to report both "visual metrics" and "downstream task metrics."

Second, "are failures faithfully simulated" is the key threshold for using world models in evaluation. WorldFailBench's near-zero order accuracy [4] shows that replacing simulators with world models for policy evaluation still calls for great caution. A world model that can only "simulate success" will make immature policies look ready to ship.

Third, hybrid architectures are the pragmatic direction. Genie 4's "Gemini plans + Genie renders" [1] and AW-0's joint generation of code, video, and actions [14] both use structured reasoning to make up for pixel models' weaknesses in causality and long-term consistency.

Fourth, systems engineering determines usability. WorldPager's 37.5 minutes with zero stalls on a single consumer GPU [9] and the several-fold efficiency gains from predictable latents [7] show that "can it run for a long time at reasonable cost" matters as much as model capability.

10. Outlook

6–12 months:

• World models for robotics and agent training will widely adopt "decision-weighted" training objectives or fine-tuning stages.
• World-model evaluations will add dimensions such as "failure faithfulness" and "physical-state consistency"; perceptual metrics alone will no longer suffice to demonstrate usability.
• Long-horizon interactive world models on consumer hardware will appear more often in game prototyping, film previsualization, and education.

1–3 years:

• "Hybrid world models" may become mainstream: language or code models maintain explicit state and rules, pixel models handle sensory rendering, and the two interact through structured interfaces.
• World models may become agents' "training ground" and "exam hall," provided calibration procedures against real environments are established.
• The risk is "realistic errors": the more realistic a world model, the harder its errors are for humans to notice; before using world models for high-stakes decisions, independent validation against real data is needed.

11. Implications for Financial Risk Management and Enterprise Deployment

Financial institutions also use "world models" — stress-test scenarios, macroeconomic projections, and customer-behavior simulations are all, in essence, rolling out "what if…" inside a model. This research offers several transferable lessons:

1. Evaluate simulation models by decision utility. A scenario-generation model should be judged not only by historical fit but by how the strategies built on it (such as credit limits, pricing, and collection strategies) perform in real backtests.
2. Focus on "failure boundaries." Whether a simulation model can reproduce the conditions under which a strategy fails matters more than reproducing average performance. Borrowing WorldFailBench's approach, systematically vary key factors and compare whether the direction of "turning from success to failure" agrees between model projections and real data.
3. Watch compositional generalization. Scenarios in which each factor has been seen but the combination is rare (e.g., a particular combination of interest-rate and industry shocks) are often where models are least reliable, and should be the focus of stress tests.
4. Separate rules from generation. Following the hybrid world-model idea, keep regulatory rules and business constraints in an explicit rule layer and let the generative model only project within those constraints, making it easier to audit and explain.

Hongxi AI has long worked on scorecards and rule engines, and we believe that whether for robots or for finance, the principle is the same: "simulation serves decisions, and decisions must be testable."

12. Conclusion

Genie 4 shows a new high for interactive world models in realism and interactivity, while concurrent research reminds us that the ultimate standard for a world model is whether it helps agents make better decisions and whether it faithfully simulates failure. Moving from "looks real" to "useful for decisions" is the path world models must take from demo to production.

Visit Hongxi AI at https://www.hxallai.com.

References

1. [1] Raia Hadsell, Diego Rivas, Hang Qi. Genie 4: A new generation of world simulation. Google DeepMind, 2026-10-27. https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/
2. [2] Zhaolu Kang, Tailong Luo, Lei Wei, Guangyuan Dong, et al. DRILL: Training World Models to Improve Policies, Not Predict Pixels. NeurIPS 2026 / arXiv, 2026-10-22. https://arxiv.org/abs/2610.30814
3. [3] Akshay Mete, Dileep Kalathil, P. R. Kumar. Adapt for Planning, Not just Prediction: Optimistic Test-Time World Model Adaptation. Texas A&M University, 2026-10-26. https://arxiv.org/abs/2610.35636
4. [4] Zheng Zhang, Xiaoyang Tang, Shuaiqi Cheng, Jiakai Huang, et al. WorldFailBench: Do World Models Understand Policy Failures? arXiv, 2026-10-24. https://arxiv.org/abs/2610.33565
5. [5] Lanxiang Hu, Yixin Huang, Abhilash Shankarampeta, Zilin Dai, et al. WMEval: Evaluating Interactive World Models Grounded on Physical Simulations. arXiv, 2026-10-24. https://arxiv.org/abs/2610.33894
6. [6] Heejeong Nam, Divyat Mahajan, Quentin Le Lidec, Randall Balestriero, et al. AnyGoal: On World Model Generalization Beyond Expert Rollouts. arXiv, 2026-10-19. https://arxiv.org/abs/2610.26844
7. [7] Mattia Scardecchia, Sherry Yang. Video World Models Need Predictable Latents. arXiv, 2026-10-26. https://arxiv.org/abs/2610.34663
8. [8] Adrien Ramanana Rahary, Yutong Bai, Nicolas Dufour, Patrick Pérez, et al. Rethinking Video Autoencoders for Interactive World Models. Kyutai, et al., 2026-10-21. https://arxiv.org/abs/2610.30129
9. [9] Bingnan Liu, Ruini Xue. WorldPager: Zero-Stall Tiered Memory for Long-Horizon World Models on a Single Consumer GPU. arXiv, 2026-10-26. https://arxiv.org/abs/2610.35529
10. [10] Zelin Gao, Hao Ouyang, Jiapeng Zhu, Qiuyu Wang, et al. Mix Forcing: Causal Pre-Training via Mixture Attention for Interactive World Models. arXiv, 2026-10-24. https://arxiv.org/abs/2610.33589
11. [11] Harold Haodong Chen, Rui Peng, Rongjin Guo, Lingteng Qiu, et al. World-in-Context: Interactive Video World Modeling with In-Context Conditioning. arXiv, 2026-10-23. https://arxiv.org/abs/2610.33169
12. [12] Zhefan Rao, Meng Chu, Xuan Zhang, Jingfan Chen, et al. Dual Exam: Controlled Evaluation of Multimodal Agents and World Models. arXiv, 2026-10-22. https://arxiv.org/abs/2610.31066
13. [13] Zhihui Zhu. An Introduction to World Models for Generalist Agents. arXiv, 2026-10-22. (Background reference.) https://arxiv.org/abs/2610.30789
14. [14] Zhuoyang Liu, Dantong Niu, Botong Zhang, Boning Shao, et al. Agentic World Action Models. arXiv, 2026-10-21. https://arxiv.org/abs/2610.30136

Research · Frontier AI Deep Dive · Great Unicorn Intelligence