PaperlayerView exact revision on arXiv

Source-linked AI summary

StressDream: Steering Video World Models for Robust Policy Evaluation and Improvement

Junwon Seo, Sushant Veer, Ran Tian, Wenhao Ding, Apoorva Sharma, Karen Leung, Edward Schmerling, Marco Pavone, Andrea Bajcsy

arXiv:2606.00267v1cs.CVcs.AIcs.LGcs.RO

TL;DR

Video world models may miss rare, high-impact action outcomes when policy evaluation relies on nominal imaginations. StressDream steers diffusion-based world-model imaginations toward text-specified outcomes while preserving plausibility, improving failure detection recall from 54% to 94% across autonomous-driving and manipulation settings.

  • Problem

    Nominal world-model imaginations can under-explore diverse futures and miss critical, plausible outcomes without prohibitively many samples.

  • Method

    StressDream optimizes diffusion-world-model initial noise using VLM semantic gradients and typical-set plausibility constraints to steer toward text-specified high-impact outcomes.

  • Results

    54% →94%: StressDream substantially increases recall for detecting high-impact outcomes in state-of-the-art autonomous-driving and robotic-manipulation world models.

  • Takeaways & Limitations

    StressDream supports robust policy evaluation and improvement by identifying actions whose plausible imagined futures include potential failures.

  • Takeaways & Limitations

    Its plausibility is limited to outcomes supported by the base world model and does not necessarily imply physical plausibility in the real world.

Abstract

from arXiv · show

Video world models (WMs) have shown promise for policy evaluation and improvement by imagining realistic future observations conditioned on ego-robot actions. While WMs can model distributions over futures, policy evaluation and improvement typically rely on nominal imaginations, which can miss high-impact outcomes of robot actions unless prohibitively many samples are drawn. To enable robust policy evaluation and improvement over WM imaginations, we propose StressDream, which steers imaginations toward high-impact yet plausible outcomes specified at inference time by optimizing the initial noise of diffusion-based WMs. However, optimizing high-dimensional noise is challenging: the optimization must reason about nuanced, scene-dependent target events in generated videos while avoiding out-of-distribution (OOD) noise that yields implausible imaginations. We address this with two complementary objectives: a semantic objective with a Vision-Language Model that provides informative gradients by reasoning about the generated video, and a plausibility objective that prevents the optimized noise from drifting OOD. With state-of-the-art video world models for autonomous driving and robotic manipulation, we show that StressDream effectively steers imaginations toward high-impact yet plausible outcomes specified by text at inference time, such as task failures, enabling robust policy evaluation and improvement by identifying actions whose plausible futures include undesirable outcomes. Video results are available at https://junwon.me/StressDream/.

1 Introduction

StressDream addresses the limitations of nominal and naive-sampling-based policy evaluation by optimizing diffusion world-model initial noise to steer imaginations toward specified, high-impact outcomes while preserving plausibility.

  • Background: Video world models learn distributions over future observations conditioned on ego-robot actions, supporting policy evaluation and improvement with less costly real-world interaction.They model uncertainty in physical interactions and other behaviors.
  • Problem: Nominal imaginations and naive sampling can under-explore diverse outcomes and miss critical events without prohibitively many samples.This limits world models’ use for exploring plausible, high-impact outcomes of robot actions.
  • Approach: StressDream optimizes the initial noise of a diffusion-based world model to steer generation toward inference-time target events, such as spills or collisions, within plausible outcomes.The approach must operate in extremely high-dimensional noise spaces, with examples reaching 1M dims.
  • Contributions: The proposed method steers generations toward high-impact outcomes while preserving plausibility by optimizing high-dimensional Gaussian initial noise.In controlled settings, it steers toward target task-failure events only when those outcomes are plausible.

2 Background: Action-conditioned Video World Models

Action-conditioned video world models model distributions of future observations from observation history and future ego-actions. When instantiated as diffusion models, their generated videos are determined by reverse denoising from an initial noise sample, making that noise a control variable for selecting imagined futures.

  • Video World Models: An action-conditioned video world model models future observations o over horizon H, conditioned on observation history ohist and future ego-action sequence a.Each observation contains n camera images and proprioceptive state q.
  • Diffusion Models as Video World Models: Diffusion-based world models transform observations from the data distribution into standard Gaussian noise across diffusion times τ ∈ {0, 1, . . . , T}.The forward process uses xτ = ατ x + στ ϵ, with decreasing ατ and increasing στ.
  • Diffusion Models as Video World Models: Generation reverses this transformation by repeatedly denoising an initial Gaussian sample xT = ϵ to produce predicted future observations x0 = o.The reverse process is implemented through repeated application of the noise predictor.
  • Initial Noise as a Control Variable: For fixed observation history and actions, the probability-flow formulation makes each imagination a deterministic function o = fθ(ϵ, ohist, a), so initial noise selects the generated video.All stochasticity in generation is governed by the initial noise.

3 Setup: Policy Evaluation and Improvement with Video World Models

The setup evaluates candidate action sequences by imagining video-world-model futures and measuring high-impact outcomes with a test-time criterion. Robust improvement uses inner noise optimization to seek worst-case plausible futures and outer action optimization to remain robust across them, but high-dimensional noise and plausibility make steering difficult.

  • Policy Evaluation and Improvement with Video World Models: Candidate action sequences are evaluated by imagined futures from an action-conditioned video world model using a test-time criterion for high-impact outcomes.The criterion measures whether plausible outcomes relevant to policy evaluation and improvement occur in generated videos.
  • Policy Evaluation and Improvement with Video World Models: Robust policy improvement combines inner noise optimization for worst-case plausible futures with outer action optimization that keeps the criterion low across them.The outer action space may be discrete for sampling-based solvers or continuous for policy optimization.
  • Steering Imagination via Inference-time Noise Optimization: Gradient ascent on initial diffusion noise steers generations toward high-impact outcomes more efficiently than repeated random sampling under limited budgets.Each update uses forward and backward passes through the world model and criterion, while the dense gradient signal improves discovery of plausible critical videos.
  • Challenge: High-Dimensional Noise Space: Video-world-model steering is challenging because noise spaces reach ≈1M dimensions in driving and ≈50K in manipulation while optimization must preserve scene-dependent plausibility.Naive optimization can guide generation toward meaningful events while drifting away from the learned world-model distribution.

4 STRESSDREAM: Steering High-Dimensional Video World Models

STRESSDREAM steers video world-model imaginations by optimizing initial diffusion noise with semantic and plausibility objectives. A VLM scores inference-time target events, while Gaussian-prior regularization keeps optimized noise plausible and score distillation makes optimization efficient.

  • Overall objective: STRESSDREAM combines a VLM-based semantic objective with a plausibility objective to optimize initial noise toward high-impact, inference-time target outcomes without OOD noise.The combined criterion is Ctest = Csem+Cpla.
  • Semantic objective: The semantic objective uses a VLM to differentiably score whether scene-dependent target events described by an inference-time text prompt occur in generated videos.This supports diverse scenes and policy contexts by using the VLM’s general video-understanding capability.
  • Semantic objective: Qwen-VL outputs yes-or-no token probabilities, and Csem is the log-probability difference log pVLM(yes | o, l) − log pVLM(no | o, l).Single-token probabilities provide rich gradients for optimizing high-dimensional noise.
  • Plausibility objective: The plausibility objective keeps optimized noise within the Gaussian prior’s typical set using norm concentration, blockwise isotropy, and spectral whiteness penalties.These terms target global radius deviations, local coordinate correlations, and frequency-domain artifacts, respectively.
  • Efficient optimization: Score distillation approximates gradients with respect to initial noise using gradients at the generated sample, avoiding full backpropagation through iterative denoising.This reduces memory costs and avoids vanishing gradients associated with backpropagating through processes such as 50 denoising steps.

5 Case Study: Steering Dubins Car Video World Model

In a controlled Naughty Dubins Car setting with known stochastic dynamics, StressDream steers one-step world-model imaginations to detect failures that are possible under the system. It achieves high true-positive and true-negative rates, while removing the plausibility objective produces implausible failures and a low true-negative rate.

  • Case Study: Steering Dubins Car Video World Model: The study tests whether world-model steering identifies high-impact failure futures only when those futures are possible under known system dynamics.The controlled setting uses known dynamics and uncertainty to evaluate failure detection.
  • Case Study: Steering Dubins Car Video World Model: The Naughty Dubins Car randomly flips the angular-control sign with probability p = 0.2, creating stochastic dynamics for failure evaluation.The car has state s = [px, py, θ], fixed speed v = 1 m/s, timestep ∆t = 0.05 s, and angular velocity in [−1.25, 1.25] rad/s.
  • Case Study: Steering Dubins Car Video World Model: The evaluation labels 5,000 random initial-state and action-sequence trajectories positive when their ground-truth minimum safety score is below zero.This label indicates that the stochastic dynamics can enter the circular failure set.
  • Case Study: Steering Dubins Car Video World Model: StressDream achieves high true-positive and true-negative rates for detecting plausible failures in world-model imaginations.It correctly identifies trajectories that can fail while avoiding false failure classifications for safe trajectories.

6 Experiments: Robust Policy Evaluation and Improvement

StressDream is evaluated for robust policy evaluation and improvement with state-of-the-art video world models for autonomous driving and robotic manipulation. It steers imaginations toward specified high-impact outcomes while preserving plausibility, and improves policies by favoring actions whose plausible futures avoid failures.

  • Experimental setup: StressDream is validated on Vista for autonomous driving and Ctrl-World for robotic manipulation.Vista predicts 25 future front-view frames conditioned on waypoints, while Ctrl-World predicts five future frames from three camera views conditioned on joint-position actions.
  • Robust policy evaluation: StressDream steers world-model imaginations toward inference-time safety-critical or failure events that nominal generations miss.Driving evaluation uses target alignment and video quality from WorldModelBench; manipulation generations are labeled success or failure by human judgment.
  • Robust policy evaluation: Removing the typical-set constraint decreases both target-alignment and video-quality scores, showing that the plausibility objective preserves grounded video quality.StressDream is compared with Best-of-N random generation under the same sampling budget.
  • Robust policy evaluation: Steering induces collision outcomes only when they are supported by the world-model distribution.The collision-finetuned Vista can produce collision outcomes, whereas the base Vista cannot imagine them reliably even with steering.
  • Policy improvement: Robust πFT 0.5 favors actions whose plausible outcomes exclude task failures, unlike Nominal πFT 0.5, which can propose risky actions that fail under steering.Robust fine-tuning weights successful steered trajectories at 1.0 and failure trajectories at 0.1; success rates are reported over 20 rollouts per task.
  • Policy improvement: Robust πFT 0.5 achieves higher success by proposing robust actions that avoid potential failures.The policy-improvement experiment fine-tunes π0.5-DROID using 40 successful demonstrations per task and weighted regression.

7 Conclusion … B.4 Score-Distillation for Approximating Initial-Noise Gradients

StressDream steers action-conditioned video world-model imaginations toward inference-time, high-impact outcomes while preserving plausible outcomes through semantic VLM guidance and typical-set noise constraints. The appendices explain the diffusion foundation, typical-set regularizers, domain-specific VLM scoring, and an efficient but approximate initial-noise gradient method, alongside limitations from prompt quality, base-model flaws, and Jacobian assumptions.

  • 7 Conclusion / B Inference-time Steering of Video World Models: StressDream optimizes diffusion world models’ initial noise with VLM semantic objectives and plausibility constraints to target high-impact, inference-time outcomes within plausible futures.The method is intended to support robust policy evaluation and improvement by exposing plausible undesirable outcomes.
  • 7 Conclusion: Prompt quality, reward hacking, limited base-world-model fidelity, and failure to preserve plausibility constrain StressDream’s effectiveness.The conclusion identifies the need for generalizable, robust reward models and stronger base world models.
  • A.2 Diffusion Models: Diffusion world-model sampling deterministically maps initial Gaussian noise ϵ to generated samples through reverse probability-flow ODE updates.The denoiser can use noise-prediction, clean-sample, score, or velocity parameterizations, which are linearly equivalent.
  • A.1 Related Works / A.3 Action-conditioned Video World Models: Action-conditioned video world models generate future observations from histories and robot actions, supporting synthetic data generation, planning, policy evaluation, and policy improvement.Applications include anticipating spills, slips, topples, agent interactions, and safety-critical situations.
  • A.4 Inference-time Steering of Generative Models: Inference-time initial-noise optimization changes only the input noise, allowing criteria to vary without modifying model parameters or repeatedly fine-tuning the generator.Keeping optimized noise plausible under the original Gaussian prior preserves generation under the pretrained model distribution.
  • A.5 Typical Set of High-Dimensional Gaussian Noise / B.1 Regularizing Noise within the Typical Set of High-Dimensional Gaussian Distribution / B.2 Concentration of Gaussian Typical Set Regularizers: Typical-set regularization targets high-probability mass rather than maximum density, constraining Gaussian noise toward a thin shell and discouraging atypical norm, isotropy, and spectral structure.Typical noise is associated with plausible generations, whereas atypical noise can produce artifacts such as blurring or transformations.
  • B.2 Concentration of Gaussian Typical Set Regularizers: Isotropy and spectral-whiteness penalties complement norm concentration by detecting unequal coordinate covariances, correlations, and uneven spatial-frequency energy.Spectral whiteness is small under the Gaussian prior when frequency bins contain sufficiently many coefficients, while large values indicate atypical frequency structure.
  • B.3 Vision-Language Models as General Inference-Time Criterion / B.4 Score-Distillation for Approximating Initial-Noise Gradients: VLMs flexibly score scene- and task-dependent targets, using driving-specialized temporal understanding and multi-view manipulation scoring, while gradient approximation avoids full denoising backpropagation but assumes near-diagonal sampler Jacobians.All three manipulation camera views are essential for reliable failure-event detection, and the approximation avoids storing the full iterative denoising graph.

B.5 Implementation Details · B.6 Time Complexity · C Naughty Dubins Car

STRESSDREAM stabilizes noise optimization with gradient clipping, typical-shell projection, and verifier-guided updates that retain the best generation. Its runtime is dominated by video generation, taking about 30 minutes for 20 steering iterations in the reported Vista setup.

  • B.5 Implementation Details: The method projects noise back to the typical Gaussian shell when its norm deviates by more than 3.0 and omits momentum to limit out-of-distribution drift.The ±3.0 interval contains more than 99.99% of typical Gaussian samples, given an approximately 0.707 standard deviation of Gaussian norm.
  • B.5 Implementation Details: The procedure initializes Gaussian noise and iteratively generates future observations conditioned on history and actions.Each iteration evaluates the generated video with a VLM criterion.
  • B.5 Implementation Details: The algorithm returns the worst-case generation associated with the highest recorded verifier score.The best score is tracked during optimization and the corresponding generation is returned.
  • B.5 Implementation Details: The semantic score is the VLM log-probability difference between “yes” and “no” for the prompted generated video.The score is written as log pVLM(yes | oi, l) −log pVLM(no | oi, l).
  • B.5 Implementation Details: Optimization combines VLM and regularization gradients, clips each coordinate to [−0.3, 0.3], and updates the initial noise by gradient ascent.The algorithm specifies gtotal = gvlm + greg and ϵi+1 = ϵi + η · gtotal.
  • B.6 Time Complexity: A deterministic video generation costs Tgen = K × Tdenoise, while STRESSDREAM adds verifier-gradient computation at each of N steering iterations.The gradient approximation avoids backpropagating through the full denoising trajectory, so the added cost is mainly the verifier backward pass.
  • B.6 Time Complexity: 30 minutes is the reported runtime for noise optimization with N = 20 iterations using Vista with K = 50 on a single H100 GPU.A single generation takes approximately 1–2 minutes, and video generation is the dominant cost.

C.1 Implementation Details

The Dubins Car experiments use a compact latent-diffusion video world model, autoregressive evaluation with 10,000 true-dynamics rollouts per condition, and five-step inference. They compare StressDream with an approximate classifier-guidance baseline and show that plausibility-aware steering matches ground-truth minimum safety outcomes.

  • Video World Model: The Dubins Car world model uses a smaller-scale SVD-style latent diffusion architecture trained with EDM on 128 × 128 × 3 RGB observations.Each observation renders the vehicle and circular failure set at the environment center.
  • Video World Model: Inference uses one-step latent prediction with H = 1, latent and noise dimension 1024, and an Euler sampler with 5 denoising steps.The world-model hyperparameters are summarized in Table 2.
  • Evaluation Setup: Evaluation rolls out the world model autoregressively from each trajectory’s initial image under its action sequence and labels outcomes using 10,000 Monte Carlo trajectories of the true uncertain dynamics.A trajectory is labeled positive if at least one sampled trajectory meets the stated positive criterion.
  • Baseline: Classifier Guidance: Classifier guidance uses a one-step clean-latent approximation because the failure predictor is trained on clean latents, while exact guidance would backpropagate through every denoising step.The approximation avoids denoiser backpropagation but can be biased and produce high-failure latents inconsistent with action conditioning or learned dynamics.
  • Results: Pessimistic Steering: StressDream steers imagined trajectories toward lower safety scores that closely match the ground-truth system’s minimum achievable scores, unlike CG and StressDream without plausibility.The latter two predict implausibly low safety scores, whereas StressDream provides meaningful policy-evaluation signals.

C.2 Ablation: Optimistic Steering

The paper also evaluates optimistic steering by reversing the steering objective to seek imagined trajectories with higher safety scores, comparing them against the maximum achievable under ground-truth stochastic dynamics.

  • C.2 Ablation: Optimistic Steering: Optimistic steering changes the objective sign to optimize imagined trajectories toward higher safety scores.Figure 16 reports the resulting maximum safety score for each imagined trajectory and shows the ground-truth stochastic-dynamics maximum as a horizontal reference.

C.3 Ablation: Impact of plausibility objective … D.3 Autonomous Driving: Evaluation Data Curation

StressDream’s plausibility objective keeps steered generations near the training distribution, while fine-tuned world models and curated evaluation data support testing rare, interaction-dependent outcomes. The experiments specify driving and manipulation setups, task criteria, steering prompts, and retrieval-based clip curation for 2.5-second futures.

  • C.3 Ablation: Impact of plausibility objective: The ablation evaluates noise typicality using norm concentration, isotropy, and spectral whiteness, plus latent OOD scores for generated samples.Higher norm-concentration values indicate more typical Gaussian noise, while lower isotropy and spectral-whiteness regularizer values are better; higher latent OOD scores indicate greater distributional distance.
  • C.3 Ablation: Impact of plausibility objective: StressDream with all regularizers keeps optimized noise near nominal statistics and produces low-OOD generations, unlike unregularized steering and classifier guidance.Unregularized steering yields atypical noise and larger OOD scores; classifier guidance preserves initial-noise statistics but still increases the OOD score.
  • D.1 Video World Models: Vista generates 25 future frames at 10Hz from an 8-dimensional action representing four future ego-trajectory waypoints, while Ctrl-World uses 5Hz videos and joint-position and gripper actions.Tables 5 and 6 summarize the corresponding world-model and noise-optimization hyperparameters.
  • D. Experiment Details: The experiment details fine-tune video world models because base checkpoints can have degraded dynamics and visual quality on specific scenarios, especially for rare events such as collisions or spills.Driving fine-tuning uses PAI-AV, nuScenes, and Nexar data; manipulation fine-tuning uses successful and failed teleoperation trajectories.
  • D.2 Robotic Manipulation: Task Details: The manipulation experiments define success and failure using task-specific object outcomes and binary VLM prompts across block stacking, knife placement, utensil picking, and pouring or bag-placement tasks.Tasks test events including blocks remaining stacked, knives staying in bowls, spoons remaining on plates, and coffee beans or candy not spilling.
  • D.2 Robotic Manipulation: Task Details: Open candy-bag placement is lower-risk than open coffee-bag placement because gummy bears are stickier and heavier, reducing the risk of spilling.The candy task succeeds when gummy bears remain inside the bag and fails if any spill out; the coffee-bag task treats any bean spill as failure.
  • D.3 Autonomous Driving: Evaluation Data Curation: The autonomous-driving evaluation set is curated from PAI-AV scenarios with uncertain or interaction-dependent futures, including pedestrian crossings, vehicle merging, traffic-light changes, and stop-sign behavior.Table 8 provides event categories and verifier prompts for Qwen-VL and X-CLIP, with X-CLIP’s first description serving as the positive event description.
  • D.3 Autonomous Driving: Evaluation Data Curation: Candidate driving clips are retrieved by embedding Alpamayo-R1 reasoning traces with Sentence-BERT, then human-verified for event occurrence and timing relative to the model’s 2.5-second future.The curation selects frames based on the annotated event time so clips align with the world model’s generated horizon.

D.4 Evaluation Metrics · D.5 Fine-tuning π0.5-DROID Policy · E Additional Results

The paper evaluates steered generations using held-out measures of target alignment and plausibility, with distinct protocols for autonomous driving and robotic manipulation. It also fine-tunes π0.5-DROID on success- and failure-weighted imagined trajectories to improve policy performance.

  • D.4 Evaluation Metrics: Autonomous-driving evaluation uses held-out metrics to separately assess whether generated videos depict the target event and remain physically realistic.These metrics are not used during steering, preventing optimization from merely increasing the steering score without meaningful or plausible video changes.
  • D.4 Evaluation Metrics: WorldModelBench reports autonomous-driving scores for instruction following (0–3), physics adherence (0–5), and commonsense consistency (0–2).WMB is selected because its evaluation distribution includes driving data and evaluates both prompt matching and visual plausibility.
  • D.4 Evaluation Metrics: An external Gemini VLM judge provides a 0–10 target-alignment score during evaluation but is not used for steering.The judge scores adherence to the text instruction while ignoring expected world-model artifacts such as blur, flickering, warping, and low resolution.
  • D.4 Evaluation Metrics: For robotic manipulation, human evaluation is the primary metric for task success or failure, supplemented by Robometer reward-model scores.Human judgments are preferred because reward models provide continuous scores and require prompt engineering to infer task success.
  • D.5 Fine-tuning π0.5-DROID Policy: The policy is fine-tuned for 10k steps on a single H100 GPU with joint-position actions, and inference uses an open-loop action horizon of 16.Joint-position actions are used because the controller does not support the base policy’s original joint-velocity actions.

E.1 Robust Policy Evaluation via Video World Model Steering … E.4 Robust Policy Improvement

StressDream steers video world-model imaginations toward plausible, text-defined failures and rare events, enabling pessimistic policy evaluation and improvement. Across driving and manipulation experiments, steering improves failure discovery while plausibility constraints preserve realistic outcomes.

  • E.1 Robust Policy Evaluation via Video World Model Steering: STRESSDREAM generates plausible task-failure futures from the same observation history and action sequence when nominal imaginations miss failures.This supports robust policy evaluation through pessimistic world-model imaginations.
  • E.1 Robust Policy Evaluation via Video World Model Steering: Across candidate action sequences, pessimistic steering reveals potential failures or safety-critical events while robust sequences remain safe under pessimistic imagination.This enables robust action selection for planning and control.
  • E.1 Robust Policy Evaluation via Video World Model Steering: STRESSDREAM with noise optimization increases target-event alignment more effectively than baselines while also reporting WorldModelBench video-quality metrics.Driving evaluation uses WorldModelBench and Gemini-3.0; the supplied passage does not provide the individual metric values.
  • E.1 Robust Policy Evaluation via Video World Model Steering: STRESSDREAM detects substantially more failure-prone robotic-manipulation trajectories than nominal generation or Best-of-N sampling.Failure detection uses Qwen3-VL scores from task-failure prompts against trajectory-level success and failure labels.
  • E.2 STRESSDREAM Can Effectively Discover Rare, Long-Tailed Events: STRESSDREAM discovers rare, long-tailed driving events more effectively than random sampling, despite random sampling using 40 samples versus 20 optimization steps.Evaluation covers pedestrian crossing, traffic-light change, and collision using Gemini-3.0 target-alignment scores.
  • E.3 STRESSDREAM Steers into Plausible Outcomes Only: A collision-fine-tuned world model correctly imagines collisions, whereas the base model hallucinates the other vehicle disappearing from the scene.The result supports steering toward outcomes represented in the model’s predictive distribution when optimized noise remains near the Gaussian prior.
  • E.4 Robust Policy Improvement: Robust fine-tuning with pessimistic video world-model imaginations improves π0.5-DROID policy success rates across tasks compared to nominal fine-tuning.The Robust πFT 0.5 policy achieves higher success rates than the Nominal πFT 0.5 policy, suggesting pessimistic imaginations can downweight failure-prone demonstrations.

F Limitations & Discussions

StressDream’s steering is a local refinement rather than a global search, and its plausibility is limited by the base world model. Despite sample efficiency, inference remains computationally expensive and can exhibit reward hacking.

  • Suboptimality of Noise Optimization: StressDream’s optimized noise may be suboptimal because gradient-based steering locally refines initial noise rather than globally searching all possible futures.Exhaustive search in the high-dimensional noise space is infeasible, while typical-set constraints make updates reliable mainly within a moderate local region.
  • Suboptimality of Noise Optimization: Large optimization updates can push noise out of distribution and produce implausible generations; combining Best-of-N search with gradient refinement is a proposed extension.The average initial-to-optimized noise distance is far smaller than the distance expected from independent Gaussian samples, indicating local refinement.
  • Preserving Plausibility: Regularization keeps steering within outcomes supported by the world model but cannot correct physical implausibility already present in the base model.StressDream defines plausibility relative to the learned world-model distribution and assumes Gaussian-noise generations are plausible under that model.
  • Computational Inefficiency: Several-minute video world-model generation remains the main runtime bottleneck despite StressDream’s sample efficiency for finding plausible target outcomes.Step-distilled or shortcut models are identified as a potential way to reduce inference cost.
  • Reward Hacking: Vision-Language-Model gradients can cause reward hacking, increasing the criterion without meaningful video changes; multiple rewards or quality metrics can mitigate this.StressDream can incorporate these prior mitigation strategies.
Loading 2606.00267v1…