PaperlayerView exact revision on arXiv

Source-linked AI summary

Video Generation Meets Agent Harness: Agentic Video Generation, Agentic World Modeling and Recursive Self-Improvement

Liaoyuan Fan, Yuhao Zheng, Dongxing Mao, Hangyu Ran, Siyuan Hu, Weijia Mao, Rui Zhao, Weijia Wu, Guian Fang, Yanzhe Chen, Zeyu Zhu, Yuxin Jiang, Yiren Song, Yuchao Gu, Junchao Yi, Qi Hu, Dan Kong, Alex Jinpeng Wang, Kevin Qinghong Lin, Mike Zheng Shou, Wei Liu

arXiv:2610.40191v1cs.CV

TL;DR

Video tasks require coordinating shots, references, inspection, and edits, while interactive video must support actions grounded in observations. This report examines these settings and recursive model improvement through the agent harness, distinguishing video-output evaluation from evaluation of the harness's role in diagnosing failures and carrying out repairs.

  • Problem

    Video generators alone do not organize the decisions needed to complete video tasks, and interactive video must support continuing actions grounded in observations rather than only finished clips.

    Source: §1.1 · p. 3

  • Method

    The report analyzes production, interactive world modeling, and model development through the agent harness, distinguishing output revisions, updates to harness state, generator parameters and evaluators, and changes to development procedures.

    Source: §1.2 · p. 5

  • Results

    The report presents the harness as linking production, interactive experience, and further model development, and distinguishes video-output evaluation from evaluation of harness diagnosis and corrective interventions.

    Source: §1.2 · p. 5, §5.3 · p. 29

  • Takeaways & Limitations

    Assessing recursive self-improvement requires evaluating video quality and task fulfillment alongside harness diagnosis and repairs, then testing whether revised development procedures yield better updates for a given budget.

    Source: §5.3 · p. 29, §5.3.2 · p. 32

    DirectorBench [255] identifies transitions between units as the lowest-scoring part of current video production.

    Source: §7.1 · p. 39

Abstract

from arXiv · show

Video generation is shifting from end-to-end synthesis toward agentic generation, where planning, tool use, and feedback enable coherent, controllable multi-shot videos. In parallel, video world models provide action-conditioned observations and predicted futures, serving as interactive environments for agents. We study their intersection through the agent harness: an executable runtime that constructs context, exposes tools, maintains state, routes feedback, and manages budgets around video models. This report examines three directions. First, agentic video generation, where harnesses coordinate scripting, shot planning, asset management, production, and editing, alongside the evaluation challenges of these systems. Second, agentic world modeling, where harnesses mediate observation, control, execution, and state. We discuss hybrid designs in which game engines or agent-written programs maintain world state while video models render observations, and how agents and world models can mutually evaluate and improve one another. Third, recursive self-improvement (RSI) in video model development, distinguishing output revisions, updates to harness state, generator parameters, and evaluators, and changes to the improvement procedure itself. We identify how these improvements feed back into subsequent development. Across these directions, we analyze essential agent capabilities, including audiovisual perception, long-horizon planning, memory, code-based interaction, feedback-guided control, skill reuse, and multi-agent coordination. We conclude with challenges in autonomous video production, lifelong memory, programmable hybrid control, autonomous model development, provenance, cost accounting, and evaluation linking video quality and task performance to sustained improvements in models and harnesses.

1 Introduction

The survey examines how agent harnesses organize video production, world-model interaction, and video-model development toward recursive self-improvement. It treats the harness as the shared execution layer connecting these settings through feedback, state, and later improvements.

1.1 Motivation

  • Video tasks require agents to turn goals into decisions, use observed outcomes to guide revisions, and maintain state across temporally coupled actions and shots.A harness makes this control operational by constructing context, exposing tools, maintaining state, routing feedback, and managing permissions and budgets.Source: §1.1 · p. 3

1.2 Contributions and Positioning

  • For agentic video generation, the survey examines how harnesses coordinate planning, execution, tools, memory, editing, and watermarking around shared project state.It considers how feedback changes generation and editing decisions and affects video quality, consistency, and efficiency.Source: §1.2 · p. 4
  • For agentic world modeling, the survey examines harness interventions in observation, control, execution, event organization, and state management, including engine-held world state.It considers sustained interaction, control fidelity, state continuity, and agent-driven evaluation.Source: §1.2 · p. 4
  • The survey’s harness-centered framework connects video models as production synthesizers, interactive environments, and objects of development.It applies a shared execution loop to production and world-model interaction, including designs where an engine or program holds world state.Source: §1.2 · p. 5
  • The survey distinguishes current-output repair and updates to harnesses, generators, or evaluators from changes to the development procedure; no surveyed system yet revises that procedure.It identifies channels through which component gains can improve later failure diagnosis, data curation, training, and validation.Source: §1.2 · p. 5
  • The survey specifies evaluation needs for production and reciprocal agent–world-model testing, alongside video-specific agent capabilities and remaining challenges.It links task execution and world interaction, which expose failures, to development that carries selected evidence into later tasks.Source: §1.2 · p. 5

2 Foundations, Preliminaries, and Framework

The framework separates video-generation methods, agent decisions, and harness execution, then uses a shared intervention view to connect production, world interaction, and model development. It distinguishes immediate feedback-driven revisions from persistent changes to development procedures that must be independently validated.

2.1 Video Generation Foundations

  • Video generation can condition on diverse inputs that controllers construct or revise, while streaming is an execution mode that exposes chunks without itself guaranteeing interactive control.Streaming makes latency, temporal-state retention, and control-update intervals important interface properties.Source: §2.1 · p. 7

2.2 Agents and Harness Engineering

  • An agent selects consequential actions from observations and outcomes, while its harness executes and constrains operations by constructing context, managing tools and state, routing feedback, and enforcing budgets and permissions.The controller proposes actions; the harness determines what runs and persists, while tools or the environment produce consequences.Source: §2.2 · pp. 8–9
  • Feedback roles differ by routing: internal critics inform the next action, external evaluators report evidence without controlling a run, reward models provide signals, and editors change artifacts.Improvements may arise from changes to the generator, controller, harness, or budget; runtime state updates alone do not establish persistent cross-task improvement.Source: §2.2 · p. 10

2.3 A Unified View of Harness Intervention

  • The unified framework treats agentic video production, interactive world modeling, and video-model development as distinct harness-intervention settings, with connections requiring identifiable evidence transfer or accepted updates.Production actions alter project artifacts, world-interaction actions alter continuing conditions, and development actions select data, training, evaluation, or validated changes.Source: §2.3 · pp. 10–11
  • Recursive self-improvement requires validated changes to procedures governing later development rounds, with independent assessment showing benefit beyond the feedback used to select changes.Generation and world-interaction traces can inform development, while accepted model and harness changes can improve later tasks; useful transfers require selecting traces, grounding corrections, and validating updates.Source: §2.3 · p. 11

3 Agentic Video Generation

Agentic video generation turns a creative brief into an edited audiovisual work through planning, shot production, and editing. An agent harness connects these stages and maintains project state so decisions and revisions can be coordinated across artifacts.

3.1 Script and Shot Planning

  • Planning spans script development, shot and task planning, and camera staging; higher-level changes can invalidate more downstream work, while executed geometric camera controls are more exact than loose text instructions.Script changes can affect multiple scenes, whereas camera changes preserve the story and assets. Camera quality should be assessed in the form actually executed.Source: §3.1 · pp. 12–13

3.2 Asset Management

  • Asset management creates references, preserves identity, and tracks cross-shot state; explicit records are inspectable and reusable but can propagate stale references, unlike consistency held within a generator’s bounded context.Records should distinguish stable identity from changing pose, location, or clothing, and separate appearance references from recent state and narrative facts.Source: §3.2 · pp. 13–14

3.3 Shot Production

  • Shot production uses learned video generation, code-based animation, or 3D scenes and simulation; each route needs feedback suited to its failure points, from inspecting takes to checking rendered execution or scene operations.Prompt checks cannot reveal synthesis failures; running code can still produce incorrect explanations or mistimed narration, while 3D systems validate at different stages of scene execution.Source: §3.3 · p. 14

3.4 Editing and Post-production

  • Editing includes story-driven cutting, timeline operations, and frame-content changes; reliable revisions track source timecodes and operations so only intended intervals or content change.Timeline inspection, rendered-output review, and persistent records support diagnosis and resumption, while semantic edits and restoration must preserve unaffected content and update every timeline use of a shot.Source: §3.4 · pp. 14–15

3.5 Watermarking and Provenance

  • Watermarks can mark generated takes, edited clips, or final exports, while provenance records trace asset origins and edits; locally regenerated shots need their own identity and marking record.Detecting an alteration and locating it are distinct capabilities, and robustness to compression does not establish detection of unauthorized edits.Source: §3.5 · p. 15

4 Agentic World Modeling

This section examines agentic world modeling, where video generation supplies action-conditioned observations for sustained interaction and an agent harness organizes control, context, and feedback. It also considers explicit program- or engine-held world state and evaluation that separates model reliability from agent capability, including checks in independent environments.

4.1 From Video Generation Models to Interactive World Models

  • Interactive world models turn action-conditioned predictions into observations for subsequent decisions, while the harness organizes action selection, context, and feedback around task goals.Unlike one-shot video generation, interaction requires consequences to persist across steps; memory can extend context beyond the current view. Physical world models predict environmental evolution, while world action models also generate actions.Source: §4.1 · pp. 16–17

4.2 Harness Interventions

  • Semantic interventions translate intent into explicit event conditions, as LingBot-World 2.0 [4] uses a VLM to propose events and a video generator to realize changes.Its Director–Pilot design distinguishes character interaction, SAM-assisted object tracking across chunks, and textual interventions; demonstrations support the design, but report no isolated quantitative harness gain.Source: §4.2 · p. 18
  • State and feedback interventions retrieve relevant history and respond to deviations, but stored information only supports state when the generator can consume and use it.WorldMem and WorldPlay [144] integrate memory with generation; harness-level tracking, logs, retrieval, and corrective controls are possible designs, not uniform features. The evidence for harness-only gains remains thin.Source: §4.2 · p. 18

4.3 Explicit World State and Generative Rendering

  • Engine- or program-held state separates simulation from rendering: rules update the world, and video models render the resulting structure, as in Matrix-Game 4.0 [156] and EdenLive.Agents may edit rules or state upstream of rendering; the resulting program state is directly readable, and off-screen objects persist independently of generator context.Source: §4.3 · pp. 19–20
  • Explicit-state systems combine a world program, an appearance anchor, and timed event text, requiring the renderer to reconcile structure, appearance, and event timing.Program checks can catch errors before synthesis, but cannot verify the renderer’s final pixels. Explicit state also leaves appearance consistency and adherence unresolved, and worlds remain limited by their programmed rules.Source: §4.3 · p. 21

4.4 Evaluation in Agentic World Modeling

  • Evaluation must test interaction, not just video quality, across long-horizon coherence, intervention sensitivity, and constraint consistency [30].Revisits and controlled action changes can expose forgotten state or incorrect consequences that plausible short clips may conceal.Source: §4.4 · p. 22
  • Harness-guided active tests adapt actions to observations, while reliable comparisons require documented controls, matched budgets, verification records, and held-out tasks.PlayWorld [8] uses Agent Players for long-horizon objectives; SCOPE [154] separates development feedback from held-out assessment and freezes the selected policy.Source: §4.4 · p. 22
  • World models can also serve as environments for agent evaluation, as SIMA 2 [175] tests generalization in held-out scenes generated by Genie 3.Evaluation is reciprocal: agents probe generated worlds, and generated worlds test agents’ perception, planning, and action across scenes and objectives.Source: §4.4 · p. 22
  • Evaluation must separate world-model reliability from agent capability; WorldArena [176] compares policy success in action-controlled world models with RoboTwin outcomes.Weak agents may fail to reach states that reveal defects, while incorrect transitions can make ineffective actions appear successful.Source: §4.4 · pp. 22–23
  • Use paired comparisons that hold the harness fixed across world models and match world models across agents, reporting transition fidelity, task completion, and policy-ranking agreement.Independent simulators or real environments can verify results; evaluation asks whether generated worlds withstand interactive testing and whether measured performance reflects actual capability.Source: §4.4 · p. 23

5 Toward Recursive Self-Improvement in Video Generation

The section distinguishes improvements to video generators and their harnesses from recursive self-improvement: RSI requires evidence that the development procedure itself has improved subsequent rounds. It organizes this question around foundation models, agent harnesses, and evaluation.

Toward Recursive Self-Improvement in Video Generation

  • RSI requires a validated change to the development procedure that makes later generator-development rounds more effective, and surveyed systems have not yet demonstrated this.The literature describes mechanisms for improving generators and agent systems, alongside emerging persistent changes to parts of development.Source: §5 · p. 23
  • The discussion examines RSI through foundation-model improvement, agent-harness improvement, and evaluation of feedback and evidence for validating updates.These perspectives cover data curation and training; planning, tools, memory, and reusable procedures; and feedback used to propose updates.Source: §5 · p. 23
  • Most existing methods update development state, including model parameters and retained harness components, under a fixed procedure; reusing improved components alone does not establish RSI.Evidence must show that the procedure producing updates has itself improved.Source: §5 · p. 23

5.1 Foundation Model Improvement

Foundation-model improvement spans data curation and model training, with development decisions shaping the data, supervision, and optimization used to improve video generators. These feedback loops can connect model use to later development, but may still follow fixed procedures for data collection, optimization, and candidate acceptance.

5.1.1 Data Curation

  • Data curation organizes improvement into acquisition and retrieval, annotation and synthesis, and refinement of video data and supervision.These interventions select training samples, construct supervision, and filter or correct data.Source: §5.1.1 · p. 23
  • Repeatedly reusing generated training samples can accumulate errors and omissions, while recursive training studies report losses in coverage, quality, or diversity.The provenance and mixture of synthetic and real data matter when assessing later rounds; EdenLive synthesizes training pairs from agent-built proxy videos.Source: §5.1.1 · p. 24

5.1.2 Model Training

  • Model training comprises pre-training, supervised fine-tuning, reinforcement learning and reward modeling, and distillation to improve generation efficiency.Distillation transfers teacher knowledge to students; approaches also adapt pretrained models with task-specific supervision and optimize them using evaluative feedback.Source: §5.1.2 · pp. 25–26
  • Improved generators, evaluators, and production feedback can guide further development, but these mechanisms may operate under fixed data-collection, optimization, and candidate-acceptance rules.Reward reliability matters because evaluators can shape the samples and objectives used to optimize generators.Source: §5.1.2 · pp. 25–26
  • JAVEdit retains validated pipeline fixes, but whether such changes improve later development requires measuring subsequent update quality and cost.Evaluation must go beyond the data or model produced in the current round to assess effects on training and candidate validation.Source: §5.1.2 · p. 26

5.2 Agent Harness Improvement

Agent harnesses support both production and development: they organize generation and editing, while development harnesses diagnose failures, curate data, train, and validate candidates. Production mechanisms can improve later work when retained changes strengthen the procedures for diagnosing failures, constructing updates, and validating candidates.

5.2 Agent Harness Improvement

  • Production harnesses organize planning, execution, and recovery, while development harnesses diagnose failures, curate data, train, and validate candidates.Many production mechanisms revise only the current output; retained memory, rules, skills, or controller weights update harness state ξk. Their relevance to recursive self-improvement depends on whether they improve later development procedures.Source: §5.2 · p. 26

5.2.1 Planning & Orchestration

  • Planning and orchestration translate creative intent into production steps, coordinate tool and agent dependencies, and route feedback into workflow revision.Examples include shot planning in MovieAgent [85], multi-turn production trajectories in VideoGen-Agent [5], and editing-DAG restructuring in VideoAgent.Source: §5.2.1 · pp. 26–27

5.2.2 Tool Use & Control

  • Tool-use methods select and condition generators, then edit or repair results while balancing correction units, preservation constraints, and costs.Approaches range from prompt refinement and requirement checks in VISTA [3] and SCMAPR [110] to region-preserving repair in VideoRepair [232] and cut, audio, and transition edits in CutCraft.Source: §5.2.2 · pp. 27–28

5.2.3 Memory & Experience

  • Memory supports consistency within a project and reuses evidence from past actions, though cross-collection accumulation remains untested for EditDuet [128].DreamFactory [236] carries character descriptions and visual references across scenes; AesopAgent [237] and A2RD [238] retain feedback or scored prompt rewrites for later decisions.Source: §5.2.3 · p. 28

5.2.4 Skills & Reusable Procedures

  • Reusable prompts, rules, and skills encode production guidance and routines, with VideoWeaver [13] extending reuse to procedures that build workflows for unseen categories.SkillPE [242] keeps its selected library fixed during deployment, while VideoWeaver evaluates evolved composition skills and skill creators on unseen categories.Source: §5.2.4 · p. 28
  • Adaptation ranges from revising the current attempt to retaining cross-task changes in rules, prompts, skills, or trained controller policies.Parametric methods include supervised trajectories with GRPO in VideoGen-Agent [5] and trajectory-level Flow-GRPO in NEWTON [222]; retained experience still requires independent validation, and AESR uses test-set feedback for its playbook.Source: §5.2.4 · pp. 28–29
  • Production feedback can inform model or harness development only when it identifies the failed capability, conditions, and a suitable intervention.Retaining traces or preferences alone does not establish that development-harness revisions improve successive generator-development rounds; this remains an open evaluation need.Source: §5.2.4 · p. 29

5.3 Evaluation in Self-Improvement

Evaluation for video self-improvement must assess both generated outputs and the harness that produces them, while separating feedback for development from independent evidence of improvement. It should track output revisions, retained capabilities, and changes to the update procedure across subsequent development.

5.3 Evaluation in Self-Improvement

  • Evaluation pairs video-output assessment with harness assessment, separating development feedback from independent validation and distinguishing current-output gains, retained capabilities, and changes to the update procedure.For agentic generation, detecting a defect is insufficient: feedback should connect visual evidence to decisions or representations that can be revised.Source: §5.3 · p. 29

5.3.1 Video Output Evaluation

  • Video-output evaluation combines metrics and benchmarks, learned judges, human reliability checks, and agentic pipelines, but observable defects alone do not establish production causes or corrections.When judges also supply learning feedback, independent evaluation helps detect reward overoptimization; held-out requests help check whether gains persist beyond optimized tasks or judges.Source: §5.3.1 · pp. 29–30

5.3.2 Agent Harness Evaluation

  • Harness evaluation covers effectiveness and efficiency, diagnosability and repairability, generalization, and continual adaptation, examining how prompts, memory, skills, tools, and control logic support production and improvement.The evaluation setting should specify what changes, what persists, and what feedback is available, since fixed-harness sampling, sequential refinement, shared harness evolution, and task-specific harness scaling differ in update scope.Source: §5.3.2 · pp. 30–31
  • Effectiveness and efficiency comparisons should control model and tool versions, feedback access, and resource budgets, while reporting development and production costs separately.Assess rewritten scripts against the original brief, use matched sampling budgets for ablations, and report generator, critic, and controller calls, time, and computational or monetary cost.Source: §5.3.2 · p. 31
  • Diagnosability and repairability tests should measure actionable diagnosis, repair success and locality, preservation of correct content, and recovery cost against controlled faults and fixed-budget retry or regeneration baselines.Targeted interventions should verify proposed causes, because visual localization alone does not establish a causal diagnosis.Source: §5.3.2 · p. 31
  • Generalization tests keep the selected harness and shared state fixed, using new examples or task categories without letting test feedback drive candidate selection or persistent updates.Validation can guide whether a skill is broadly retained or category-specific, but final testing should remain independent of those decisions.Source: §5.3.2 · p. 31
  • Continual-adaptation evaluation tracks quality, completion, cumulative cost, gains, and regressions across updates and earlier task families, separating adaptation data from evaluation feedback.Additional tools, skills, or agents can impair earlier-task performance even with fixed model weights; online tasks should be scored before their feedback informs later updates.Source: §5.3.2 · pp. 31–32
  • To establish recursive self-improvement, compare fixed and revisable development procedures under matched components, feedback access, and budgets, then test subsequent updates independently for quality, cost, and retained capabilities.The revised procedure must be used in later rounds; evaluation asks whether it produces better updates for a given budget or comparable quality at lower cost.Source: §5.3.2 · p. 32

6 Agent Design for Video Generation

This section examines video-agent design, connecting perception and reasoning to generation interfaces, feedback, reusable experience, and collaboration across tasks ranging from short clips to long narratives and continuing interactions. It considers how agents balance control precision, computation, observations, skills, and collaboration, while using traceable state, failure evidence, resource use, and retained improvements to assess system-level benefits.

6.1 Foundational Agent Capabilities

  • Long-horizon agents must combine task-relevant perception, selective memory, and dependency-aware planning while balancing evidence quality and planning detail against acquisition cost and downstream repair.Sparse observations can support semantic checks, but transient motion or sound-timing defects may need denser inspection; global plans suit known dependencies, while local planning adapts to stochastic outputs.Source: §6.1 · pp. 32, 33, 35

6.2 Agent–Environment Interaction

  • Agents can invoke models and tools, write executable controls, or operate GUIs; hybrid workflows preserve explicit structure while delegating appearance to neural synthesis, then verify the rendered result.Programmatic controls support precise timing, geometry, and recoverable execution; GUI interaction broadens application coverage but is sensitive to layout and hidden state, so effects must be checked.Source: §6.2 · pp. 35–36

6.3 Feedback-Guided Agent Control

  • Feedback-guided control separates evaluation, diagnosis, recovery, and acceptance, using the smallest effective intervention while accounting for collateral changes, unresolved constraints, and remaining budget.Stopping requires recording whether the artifact was accepted, remains unresolved, or needs user judgment; comparisons should include rejected candidates and generation and evaluation costs.Source: §6.3 · pp. 36–37

6.4 Agent Experience and Skill Reuse

  • Agents reuse experience through contextual examples, refined skills, and deployable packages, but must validate transfer on distinct tasks and retain earlier capabilities after updates.Skills should specify assumptions, resources, operations, constraints, and failure cases; tests after tool or generator changes should check portability and whether outputs still meet acceptance criteria.Source: §6.4 · pp. 37–38

6.5 Multi-Agent Organization

  • Multi-agent teams help when specialization, parallelism, or independent review outweigh routing and synchronization costs; delegation, shared state, ownership, and update propagation must be explicit.Team size should follow task demands, and project examples of agent swarms do not establish improved final-video quality; comparisons should use a single controller with comparable tools and budget.Source: §6.5 · p. 38
  • Collaborative review can expose overlooked errors through independent perspectives, but correlated critics and repeated debate may add cost without improving judgment.Reviewer roles should specify the criterion, affected content, required evidence, and authority; decision rules are needed for disagreements over subjective creative goals.Source: §6.5 · p. 39
  • Review should yield a verifiable revision or explicit acceptance, update shared state after constraint checks, and use escalation or termination when debate cannot resolve the task.Evaluation should compare review and decision procedures at matched budgets, measuring latency, unresolved disagreement, and collateral changes.Source: §6.5 · p. 39
  • Video-agent architectures should make accepted state, failure evidence, resource use, and retained improvements traceable while selectively using explicit representations, richer observations, reusable skills, and collaboration.Explicit representations suit inspectable geometry, timing, or content; richer observations help when uncertainty can change a decision; skills and collaboration warrant their overhead when procedures recur or responsibilities are separable.Source: §6.5 · p. 39

7 Future Work and Challenges

Future progress depends on harnesses that can complete and improve video workflows autonomously while maintaining coherent state, controlling costs, and making outputs and development processes verifiable. Evaluation must measure sustained task performance and improvement, not video quality alone.

7.1 Fully End-to-end Video Production

  • End-to-end production requires a harness to coordinate the full pipeline, repair specific failures while preserving accepted content, and decide when work is ready for delivery.DirectorBench [255] finds transitions between units to be the lowest-scoring part of current long-form workflows; acceptance and stopping criteria tied to creative intent and budget remain undeveloped.Source: §7.1 · p. 39

7.2 Lifelong Memory

  • Lifelong memory must preserve reliable project constraints and transferable experience while retrieving relevant evidence and discarding obsolete or conflicting information.Records should retain provenance, temporal scope, and artifact versions; evaluation should test whether memory improves new projects and revisited environments without simply adding context or transferring outdated assumptions. StoryMem [302] uses dynamically updated visual memory, and WorldMem retrieves memory frames with associated states.Source: §7.2 · pp. 39–40

7.3 Hybrid Code-Pixel Generation

  • Hybrid code-pixel systems need callable controls and verification to keep program state, agent memory, and generated observations aligned over extended generation and interaction.Editable physical representations and executable programs can make generation conditions inspectable, but current systems often use one structural representation, only some run in real time, and paired data come from few engines. The Genie 4 announcement identifies hybrid world models as a future direction; agents should verify that state updates are realized in video and correct deviations.Source: §7.3 · p. 40

7.4 Autonomous Human-free Development

  • Autonomous model development must diagnose weaknesses, create targeted data and interventions, and improve the procedures used to discover failures and validate updates.Failures in video production and world interaction could guide supervision, but gains must be attributed, generated-supervision errors contained, and generator and critic capabilities preserved. Independent validation and recoverable versions are needed to test for durable progress rather than adaptation to learned evaluation signals.Source: §7.4 · p. 40

7.5 Efficiency

  • Efficiency requires common accounting of quality, time, and resources across all generation, inspection, repair, rejected candidates, verification, and model-development rounds.Harnesses should allocate resources, reuse accepted assets and intermediate results, and stop refinement when another call is not worth its cost.Source: §7.5 · p. 41

7.6 Provenance and Safety

  • Auditable video production requires signed provenance for each accepted version and safety constraints that follow identity assets and derived outputs across tools and projects.Harnesses should bind reuse permissions to identity assets, check generated and edited outputs against restrictions, and limit access to production tools and generated code; marks attached earlier may not describe later edits. C2PA Content Credentials provide an example for edit histories.Source: §7.6 · p. 41

7.7 Emerging Tasks

  • Benchmarks should test complete video systems across production, persistent interaction, failure recovery, and successive development rounds—not only video quality or isolated models.They should measure task completion, controllability, consistency, horizon, complexity, and resource constraints; RSI evaluation should track gains, regressions, and cumulative costs across updates and test transfer on independent tasks.Source: §7.7 · pp. 41–42

Related papers

Loading 2610.40191v1…