acceptodds
Under review as a conference paper at ICLR 2027

State Forcing: Telic Instruction Following for Streaming Video Generation

Abstract

Recent advances in autoregressive video generation have enabled efficient streaming beyond the native clip lengths of pretrained models. Yet visually coherent rollouts can still exhibit progress regression and re-execution of completed actions. We study telic instruction following, which requires completing a requested state transition without regression and preserving its consequence. Our empirical study shows that verified terminal states can guide progress toward completion, and support consequence retention in a frozen generator. Building on this finding, we introduce State Forcing, a state-privileged on-policy self-distillation (OPSD) framework. Through complementary Advance and Hold conditions, a frozen teacher uses offline terminal states to supervise progress and consequence retention along student-generated trajectories. The student retains the original instruction and streaming pipeline, without additional inputs or computation at inference. We also introduce TelicBench to jointly evaluate completion and temporal validity over full rollouts. State Forcing outperforms the evaluated baselines in completion and strict success, achieving 37.33% strict success compared with 24.93% for the strongest baseline. These gains are achieved with overall video quality comparable to the distilled streaming baselines.

Primary area:
generative models
Keywords:
Video generation, Instruction following, Streaming video generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yp4iOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies telic instruction following in streaming video generation, where a model should complete a requested state transition, avoid progress regression, and preserve the achieved consequence. The paper introduces State Forcing, a state-privileged on-policy self-distillation framework built on a pretrained streaming video generator. During training, a frozen teacher receives a verified terminal video chunk through two complementary conditions: ADVANCE uses the terminal chunk to guide an unfinished event toward completion, while HOLD uses it to supervise consequence retention after completion. The student follows its standard streaming pipeline and matches the teacher’s denoising predictions along student-generated trajectories, requiring no additional inputs or computation at inference. The paper also introduces TelicBench, containing 150 instructions evaluated using Completion Rate, Telic Violation Rate, and Telic Success Rate. Component ablations and a limited human study provide additional validation.

Critical strengths
  • Important and underexplored problem formulation. The paper identifies a meaningful limitation of current streaming video generators: long-term visual coherence does not guarantee successful task execution. Explicitly separating completion, non-regressing progress, and consequence retention provides a useful perspective that is not adequately captured by standard long-video quality metrics.

  • Practical terminal-state supervision with no inference-time overhead. Using verified terminal chunks as privileged teacher context is intuitive and practically attractive. The ADVANCE and HOLD conditions target two distinct failure modes, and the ablations provide initial evidence that they play complementary roles. The trained student retains the original architecture, sampler, prompt, and cache policy at inference.

Critical weaknesses
  • The methodological novelty is limited relative to prior privileged on-policy distillation. The core training objective, student-generated rollout, frozen teacher, stop-gradient cache update, and denoising prediction matching largely follow existing OPSD-style methods. The primary technical modification is how the teacher context is initialized using terminal chunks. Consequently, the method may be viewed as a task-specific application of OPSD-V rather than a fundamentally new distillation framework. The paper lacks a matched comparison against OPSD-V or standard on-policy distillation using the same telic data, initialization, and training budget.

  • The experiments do not conclusively establish the claimed mechanism. The central claim that terminal-state privilege teaches progress and consequence retention is not supported by sufficient counterfactual ablations. In particular, the paper does not compare correct terminal chunks against random future chunks, non-terminal chunks, mismatched terminal chunks, or terminal text alone. Moreover, TVR does not penalize pausing or inactivity, so reduced motion or post-completion freezing may improve the reported metrics without demonstrating genuine consequence retention. Generic VBench scores do not directly rule out this shortcut on telic prompts. This concern is amplified because the same VLM family is used for prompt generation, training-data verification, terminal-state localization, and automatic evaluation.

  • The visual results and method presentation need improvement. In Figures 1 and 2, the first and last frames exhibit noticeable inconsistencies in visual details, such as changes in the color of the girl’s skirt. Moreover, the LongLive results appear reasonably strong, making the qualitative advantage of State Forcing less convincing. In Figure 3, the overall framework appears highly similar to OPSD-V, and the paper should more clearly highlight its methodological differences.

These limitations prevent the current evidence from clearly separating the proposed mechanism from simpler explanations such as telic-specific fine-tuning, stronger state anchoring, or motion suppression.

Questions for authors
  • How many source rollouts were generated to obtain the 3,000 accepted training records? What proportion was rejected because of incomplete execution, incorrect initial states, or subject/scene drift? This information is important for determining whether the method mainly learns from easy cases that the base model can already complete.

  • How are ADVANCE and HOLD records constructed and balanced? Can the same source rollout provide both branches, and how are the unfinished prefix, terminal boundary, and terminal chunk (G) selected? The current description does not establish how much variation exists in the training states.

  • In ADVANCE, the future terminal chunk replaces the initial sink while being assigned the sink’s temporal positions. How are positional encodings and cache timestamps handled in this operation? Why does placing future content at past temporal positions not introduce temporal aliasing or encourage direct appearance copying?

  • How does State Forcing compare with standard OPSD-V, ordinary on-policy distillation, and supervised fine-tuning when all methods use the same warm-up checkpoint, telic training data, and optimization budget? Without these matched baselines, it is difficult to attribute the gains specifically to terminal-state forcing.

7 comments
AuthorsOpenReview ↗

Author Response to Reviewer yp4i (Part 1/5)

Response to Reviewer yp4i

Thank you for recognizing the importance of telic instruction following and the practical value of State Forcing. We address your concerns and questions in detail below.

Weakness 1 / Question 4 — Methodological novelty and matched comparisons

Positioning and distinction from existing OPSD methods.

OPSD was originally formulated for LLM post-training (Zhao et al., 2026a), where a single model acts as teacher and student under different contexts: the teacher receives privileged information, such as reference reasoning traces, while the student retains its standard question context. Subsequent OPSD methods adapt this principle to different tasks and modalities through task-specific teacher-context construction. For example, Learning from the Self-future (Luo et al., 2026) uses self-generated answers as privileged suffix context for diffusion LLMs, while D-OPSD (Jiang et al., 2026) conditions the teacher on multimodal features incorporating the target image for diffusion-model adaptation. In these methods, the privileged context shapes the teacher’s predictions and therefore the corrective supervision provided along student-generated trajectories. Within this paradigm, State Forcing introduces verified terminal states as teacher privilege for telic instruction following. ADVANCE uses terminal-state evidence to guide unfinished execution toward completion, while HOLD constructs supervision for preserving the achieved consequence. Our contribution lies in identifying terminal-state evidence as a source of supervision and designing these complementary teacher conditions to address progression and retention in streaming generation. This follows the broader OPSD principle of transferring task-relevant privileged knowledge through on-policy supervision, with a distinct supervision design tailored to telic execution.

Distinction from OPSD-V

OPSD-V and State Forcing share the OPSD training paradigm but construct teacher supervision for different long-horizon failure modes. OPSD-V addresses error accumulation and weakened motion dynamics by providing cleaner temporal contexts derived from real long-video sequences. State Forcing addresses telic execution: completing a requested transition without regression and preserving its consequence afterward. The key distinction is the information supplied to the teacher and the corrective behavior it encourages. Motivated by our observation that verified terminal states can guide both progression and retention in a frozen generator, we construct two complementary conditions. ADVANCE supplies terminal-state evidence alongside unfinished execution, while HOLD isolates the achieved outcome under a consequent-state description. On-policy distillation transfers these signals to the student’s own trajectories.

AspectOPSD-VState Forcing
Targeted challengeLong-horizon degradation, including error accumulation and weakened motionTelic execution: completion, non-regressing progression, and consequence retention
Privileged informationReal-video temporal contextVerified terminal chunks from self-generated source rollouts
Teacher-context constructionCleaner temporal context incorporating real-video historyADVANCE: Terminal evidence replaces the initial sink alongside unfinished history. HOLD: Terminal-only initialization under a consequent-state description
InferenceOriginal few-step inference without privileged reference contextOriginal streaming inference without terminal-state inputs or additional computation

Accordingly, we adapt OPSD with terminal-state-privileged supervision, where complementary teacher conditions jointly guide event progression and post-completion retention. This contribution focuses on the supervision design for telic instruction following rather than introducing a fundamentally new distillation framework.

AuthorsOpenReview ↗

Author Response to Reviewer yp4i (Part 2/5)

Matched Experimental Comparisons

Following the reviewer’s suggestion, we retrain OPSD-V from the same Warm-up Only checkpoint using the same accepted telic records and matched training compute. This baseline retains OPSD-V’s temporal-context supervision procedure, adapted to the common telic source rollouts. We also evaluate direct SFT from the same checkpoint under a matched compute budget. The accepted records contain verified prefixes and terminal intervals, but their complete source rollouts may contain temporal violations. We therefore construct reliable SFT demonstrations by retaining fully successful executions from this pool. This SFT baseline uses a filtered subset and is reported separately from the data-matched OPSD-V. All methods use the same TelicBench evaluation protocol. Scores are percentages.

MethodCR ↑TVR ↓TSR ↑
Warm-up Only39.8720.0219.20
Matched SFT33.3315.2516.67
Matched OPSD-V44.6723.0517.33
State Forcing51.337.6037.33

State Forcing exceeds matched OPSD-V by 20.00 percentage points in TSR, while reducing TVR by 15.45 points and achieving higher completion. These results support the benefit of terminal-state-oriented supervision beyond applying existing video OPSD to the same telic records. The tested SFT configuration also does not reproduce these gains.

Taken together, our distinct state-privileged supervision design and matched experimental results establish State Forcing as a substantive methodological contribution beyond existing OPSD-based methods. To the best of our knowledge, State Forcing is the first to formulate telic instruction following as a distinct problem in streaming video generation and develop a dedicated training approach that jointly addresses completion, non-regressing progression, and consequence retention.

Weakness 2 — Insufficient evidence for the proposed mechanism

Counterfactual ablations of terminal-state privilege.

To test whether the gains depend on correctly matched terminal-state evidence, we replace the verified terminal chunk with four alternatives: random future chunks sampled without terminal-state filtering, non-terminal chunks, terminal chunks from mismatched rollouts, and terminal-state text alone. All variants retain the same training settings, differing only in the privileged teacher context. Scores are percentages.

Teacher ContextCR ↑TVR ↓TSR ↑
Random future chunk30.1311.3220.53
Non-terminal chunk31.6010.7919.73
Mismatched terminal chunk29.0716.7918.27
Terminal-state text only39.3317.0219.60
State Forcing51.337.6037.33

Correct terminal chunks achieve the highest completion and strict-success rates and the lowest violation rate. Their TSR exceeds the strongest alternative by 16.80 percentage points. These controls support the importance of verified, correctly matched visual terminal-state evidence: arbitrary future context, incomplete-state evidence, mismatched outcomes, and terminal descriptions alone do not reproduce the gains.

Independent VLM evaluation.

To assess the robustness of our evaluation to the choice of VLM judge, we additionally evaluate all methods using Gemini 3.8 Flash as an independent evaluator. We retain the same TelicBench videos, evaluation rubrics, frame-sampling protocol, and scoring criteria, changing only the evaluator.

MethodCR ↑TVR ↓TSR ↑
SkyReels-V219.3318.906.00
MAGI-132.6714.2715.33
Self Forcing44.6722.4416.67
Rolling Forcing48.6716.3320.00
Context Forcing50.6717.3223.33
LongLive49.3315.5725.33
State Forcing52.009.8735.33

Despite differences in absolute scores, Gemini independently identifies State Forcing as the best-performing method across all three metrics. Our blinded human study (Section 5.3) further corroborates this finding: State Forcing achieves the highest completion rate (60%) and strict success rate (34%), as well as the lowest violation extent (1.56). The consistent findings across GPT-5.6, Gemini 3.8 Flash, and human evaluations provide strong evidence for the reliability of our evaluation and the effectiveness of State Forcing.

AuthorsOpenReview ↗

Author Response to Reviewer yp4i (Part 3/5)

Distinguishing consequence retention from post-completion freezing.

TVR measures task-related temporal violations rather than motion magnitude. Importantly, whether inactivity constitutes a failure depends on the instruction semantics: maintaining an achieved terminal state may legitimately require static content, whereas suppressing explicitly required scene dynamics would violate the instruction. We therefore do not penalize inactivity alone. TSR further requires both successful completion and violation-free execution, preventing a model from exploiting reduced motion or inactivity without actually executing the requested transition.

To directly investigate post-completion freezing, we conduct an additional dynamics evaluation on TelicBench instructions that explicitly require non-task-related scene dynamics to continue after task completion. We evaluate videos satisfying the strict-success criterion, assessing whether the specified dynamics persist from the first completion timestamp until the end of the rollout. Importantly, the expected dynamics can range from subtle local motions to large-scale background activities, which cannot be adequately assessed using generic motion-magnitude measurements. We therefore use an independent Gemini 3.8 Flash evaluator to assess the preservation of instruction-specified dynamics. We further validate the findings through blinded human evaluation using the same 0–2 scoring rubric.

MethodVLM score ↑Human score ↑
LongLive1.571.22
Context Forcing1.111.05
State Forcing1.521.24

Scores range from 0 (required dynamics are not preserved) to 2 (required dynamics are fully preserved). VLM evaluation covers all eligible strictly successful videos, while human evaluation uses 30 randomly sampled videos per method, independently rated by five annotators.

Notably, LongLive and Context Forcing serve as general streaming-generation baselines without telic-specific training or post-completion supervision. State Forcing achieves post-completion dynamics scores comparable to LongLive and higher than Context Forcing in both independent evaluations. These results demonstrate that State Forcing can retain the achieved task outcome while preserving explicitly required scene dynamics after completion, rather than relying on post-completion freezing.

Weakness 3 — Visual results and method presentation

Figure 1 is intended to demonstrate telic instruction-following behavior, rather than compare the frame-wise visual quality of individual frames. In the door example, LongLive reverses the opening transition and returns to a closed door. In the character-entry example, it repeats the entry event after completion, producing additional boy instances instead of retaining the requested single-character outcome. State Forcing avoids these failures in the displayed sequences. Importantly, visual consistency and correct telic instruction execution are distinct evaluation dimensions: visually coherent frames do not necessarily imply correct event progression and consequence retention. These side-by-side examples directly illustrate State Forcing's qualitative advantage in telic execution. Table 2 further shows that State Forcing maintains higher subject consistency than LongLive at both 30 seconds (0.9860 vs. 0.9751) and 60 seconds (0.9840 vs. 0.9734), together with higher background consistency at both durations.

Figure 2 presents a controlled empirical study using a frozen LongLive-based generator to motivate State Forcing, rather than qualitative results of the trained State Forcing model. With model parameters and sampling fixed, it isolates the effect of state-privileged context on telic instruction following, demonstrating how verified terminal-state information supports progression and consequence retention. These observations motivate the subsequent ADVANCE and HOLD designs.

Figure 3 separates our designed state-privileged teacher-context construction in Panel (a) from its integration into on-policy distillation in Panel (b). Panel (a) explicitly shows ADVANCE's terminal-sink replacement and HOLD's terminal-only initialization under a consequent-state description. These branch-specific designs are distinct from OPSD-V's use of real-video history as privileged context. Panel (b) illustrates how these contexts are integrated into the shared OPSD training process. Its left side shows the student and privileged teacher contexts, while its right side depicts the student–teacher rollout and prediction-matching pipeline. We follow the schematic presentation in D-OPSD to illustrate this general OPSD procedure. Any visual similarity in this shared component reflects the underlying OPSD mechanics, rather than the method-specific teacher-context construction. The methodological distinction is further supported by the matched comparisons presented in our response to Weakness 1. We expand the Fig.3 caption for clarification.

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

This paper addresses telic instruction following, which requires models to preserve the resulting state after an action has been completed. The authors introduce a benchmark for evaluating this capability and propose an OPD framework to improve it.

Critical strengths

Strength 1: A complete problem-to-solution pipeline. The paper identifies a concrete failure mode, introduces an evaluation metric, and proposes a framework to address it.

Critical weaknesses

Weakness 1 Narrow scope and unclear distinction from general instruction following. The paper focuses on a specific failure mode, but it is unclear why this should be treated separately from the broader problem of instruction following or prompt faithfulness. These failures could potentially be addressed through better training data and supervision. For example, successful trajectories in which a door opens and remains open could be paired with explicit captions such as “The door opens and remains open for the rest of the video,” then used for direct fine-tuning with dense caption supervision. For distilled models, an RL framework for streaming video generation, such as Astroble, could potentially address the issue using a VLM-based instruction-following judge as the reward signal. Comparisons with these alternatives would help establish the need for a specialized framework.

Weakness 2 The proposed method requires three forward passes, which is complex. The authors argue that appending the correct terminal state helps guide the student, and attribute the failure partly to conflicts between the initial frame and the current generation history. This explanation suggests a simpler alternative: recache at inference time to remove stale context and place greater emphasis on recent history.

Questions for authors

Question 1: Could you evaluate an inference-time recache ? Question 2: Could you directly fine-tune the base model on the 3,000 verified successful records using a training budget comparable to that of the proposed method? Question 3: Table 2 reports only six VBench dimensions. Could you explain the rationale for selecting these dimensions?

8 comments
AuthorsOpenReview ↗

Author Response to Reviewer Dyji (Part 1/2)

Response to Reviewer Dyji

Thank you for the constructive feedback. We address each concern and question in detail below.

Weakness 1 / Question 2 — Scope and alternative training methods

Distinction from general instruction following.

We view telic instruction following as a temporally structured subproblem of general instruction following. It highlights a key challenge in streaming video generation: local visual coherence does not guarantee correct execution of temporally extended events, which requires completing the requested transition, avoiding progress regression, and retaining the resulting consequence. As generation extends beyond the native clip horizon, this gap can lead to systematic failures in event completion and consequence retention.

Existing streaming generators primarily rely on chunk-level generation and distillation objectives, which do not explicitly supervise event-level progression or post-completion retention. This limitation is particularly relevant to distillation-based methods. Although students can be rolled out beyond the teacher's native clip horizon, standard distillation supervision does not explicitly enforce globally consistent event progression toward the desired terminal state or its retention over long horizons. This motivates State Forcing, which uses verified terminal chunks as privileged teacher context to supervise progression and consequence retention along student-generated trajectories.

Meanwhile, general video-quality metrics, such as VBench, do not assess event completion, progress regression, or consequence retention throughout a rollout. To address this evaluation gap, we introduce TelicBench, which evaluates full rollouts through Completion Rate (CR), Telic Violation Rate (TVR), and Telic Success Rate (TSR). In particular, TSR requires both successful completion and violation-free execution, ensuring that a rollout that never completes the requested transition cannot be counted as successful.

Taken together, telic instruction following is not a separate alternative to general instruction following, but an underexplored temporally structured subproblem that becomes particularly exposed in streaming generation. Its specific requirements are not adequately addressed by existing training objectives or general video-quality metrics, motivating dedicated supervision and evaluation through State Forcing and TelicBench.

Question 2: Dense-caption SFT with matched training budget.

The 3,000 accepted records provide compatible execution prefixes and verified terminal-state intervals, but acceptance does not guarantee that the entire source rollout is successful and violation-free. Since standard SFT requires reliable target demonstrations, we further filter the associated source rollouts to retain only fully successful executions. We fine-tune the same Warm-up Only checkpoint using these selected successful records, paired with dense telic captions. Unlike State Forcing, this baseline uses standard supervised fine-tuning without privileged teacher contexts or OPSD. We evaluate both methods using identical settings.

MethodCR ↑TVR ↓TSR ↑
Warm-up Only39.8720.0219.20
Direct SFT33.3315.2516.67
State Forcing51.337.6037.33

Under matched training budgets, State Forcing achieves higher completion and strict success with fewer temporal violations than successful-rollout SFT. These results suggest that directly imitating successful demonstrations is less effective than state-privileged on-policy self-distillation.

VLM-reward RL comparison.

To evaluate VLM-reward RL as an alternative, we implement an RL baseline based on the Astrolabe framework, using Qwen3-VL-8B-Instruct as the reward judge. The judge scores generated rollouts against the original instructions using task-specific rubrics for completion and temporal validity. We initialize the RL policy from the same Warm-up Only checkpoint and use the same telic training prompts. Both methods are evaluated under the same TelicBench protocol with comparable training budgets. VLM-reward RL optimizes sampled rollouts through evaluator feedback. In contrast, State Forcing provides denoising-level supervision on student-generated trajectories using verified terminal-state contexts. ADVANCE and HOLD further provide complementary supervision for event progression and consequence retention.

MethodCR ↑TVR ↓TSR ↑
Astrolabe-based VLM-reward RL38.6713.8121.33
State Forcing51.337.6037.33

Under comparable training budgets, State Forcing outperforms the VLM-reward RL baseline by 12.66 percentage points in CR and 16.00 points in TSR, while reducing TVR by 6.21 points. These results show that state-privileged on-policy distillation is more effective than rollout-level VLM-reward optimization for telic instruction following in our setting.

AuthorsOpenReview ↗

Author Response to Reviewer Dyji (Part 2/2)

Weakness 2 / Question 1 — Training complexity and inference-time recaching

Training and inference computation.

Training computation. State Forcing follows the standard OPSD rollout-and-matching structure. The student generates on-policy rollouts using the original few-step sampler. Selected noisy states visited during these rollouts are then reused for student–teacher prediction matching. Only the student prediction used for matching requires gradients; rollout sampling and frozen-teacher prediction are performed without gradient tracking. We sample only one denoising timestep per rollout and reuse it across all generated chunks. Unlike standard DMD-style distillation, which typically requires both a frozen real-score model and a trainable fake-score model, OPSD uses a frozen copy of the base model as its teacher without an additional trainable score estimator. State Forcing retains this structure. ADVANCE and HOLD are alternative teacher-context initializations sampled with equal probability, introducing no additional denoising passes. Source rollouts and VLM-verified terminal-state records are constructed offline, requiring no online VLM calls or state verification during training.

Inference computation. At inference, only the trained student runs with the original streaming pipeline. The teacher and privileged contexts are entirely removed, requiring no additional forward passes, online VLM calls, or state detection.

Therefore, State Forcing introduces no additional denoising passes beyond the standard OPSD training structure and no additional inference-time cost.

Question 1: Inference-time recaching.

We implement the suggested inference-time recaching on the Warm-up Only checkpoint by periodically replacing stale sink context with the latest generated chunk, emphasizing recent history while keeping the cache budget unchanged. All methods use identical evaluation settings.

Method / Cache PolicyCR ↑TVR ↓TSR ↑
Warm-up, standard cache39.8720.0219.20
Warm-up, inference-time recaching31.4711.9820.13
State Forcing, standard cache51.337.6037.33

Inference-time recaching reduces TVR from 20.02% to 11.98%, but also decreases CR from 39.87% to 31.47%, yielding only a marginal TSR improvement (19.20% to 20.13%). In contrast, State Forcing simultaneously improves completion and reduces temporal violations, achieving substantially higher TSR. These results suggest that refreshing historical context alone is insufficient to ensure both event completion and temporal validity, whereas state-privileged supervision directly addresses both.

Question 3 — Rationale for selecting the six VBench dimensions

Table 2 evaluates generated videos using MovieGenBench prompts. We report all six dimensions explicitly supported by VBench's official custom-video evaluation protocol: subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality, and imaging quality. This choice is consistent with existing streaming video generation evaluations, such as Rolling Forcing. We apply all six dimensions uniformly across the compared methods to ensure a consistent evaluation. We clarify the custom-input evaluation protocol in the revision.

We thank the reviewer once again for the constructive suggestions. We hope these responses address your concerns and would be happy to discuss any remaining questions.

Dear Reviewer Dyji,

Thank you again for your constructive feedback. With the discussion window closing very soon, we wanted to gently check whether our response has addressed your concerns. If so, we would be grateful if you would consider raising your score accordingly. Please let us know if anything needs further clarification.

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

This paper studies telic instruction following in streaming generation problems requiring 1) completion, 2) non-regressing progression, and 3) consequence retention. Using GPT 5.6 as a source, State Forcing feeds verified terminal chunks from generated videos as privileged teacher context for OPSD (on policy self-distillation). Authors introduce two modes 1) ADVANCE where teacher’s attention sink is replaced with terminal evidence and 2) HOLD where retention is encouraged by isolating the completed state under a state description. This paper also introduces TelicBench, where the suggested method outperforms most baselines with a Telic Success Rate of 37.33% (others typically lie between 10~25).

Critical strengths
  • For me, this paper introduces an interesting problem focused on “Telic Instructions” which less recent papers focus on. Most recent papers focus on anti-drifting and infinite length generation but in most cases, those infinite generations without persistent states are not useful. Personally, I don’t find technical components (OPSD) to be highly novel or benchmark qualities to be of substantial quality, but this problem setup (+ how authors tackle it) was the biggest component that made me lean towards weak acceptance.

  • Methods are not breakingly new ideas, but are well/reasonably motivated. I also think experiments are quite useful (especially the ablations in Table 4 and the automatic human judgements in Table 3). w/o Hold and w/o Advance results sound interesting for me. Also unlike many other papers (those in lines of coding world models) that adopt offline LLMs in the inference to preserve states, the fact that State Forcing doesn’t rely on external models during inference could make it practical for downstream uses.

Critical weaknesses
  • Despite the name being State Forcing, I think this method is only partially applicable to very limited scenarios. Because the entire “state” is limited to the KV-cache, whatever that leaves that KV Cache, has no way to be maintained. For example, in longer rollout scenarios where people go out of camera and come back (long enough so that they leave KV cache), or whatever that cannot be directly told based on simple observation (such as knocking the door twice / turn off lights and turn it back on) wouldn’t work. Essentially, this work only covers simple states that are visible directly just from the current frame. TelicBench also seems to be designed specifically for static scenes with scenario aligned cases starting with things like “A fixed daylight…”. While I like the idea of Telic, I don’t know if framing it to “states” is a right thing (especially where we have a bunch of world action models that actually aim at learning states, or coding world models that really do have persistent states).

  • State Forcing trains on data the GPT 5.6 has considered as successful, and primarily evaluates using the GPT 5.6. Also other major baselines do not receive “telic warm-up” and GPT 5.6-verified training samples which would probably be great for increasing the telic benchmark scores. Authors do have ablations but they do not fully address concerns for the data part.

Questions for authors
  • I would appreciate it if authors could provide a comparison with other baselines that also train on the telic warmup + same 3K accepted records with similar training budget. One meaningful baseline with matched compute+data would help resolving the concerns. As mentioned above, using different VLMs as judge would also be helpful.

  • I would reconsider the claims/scope of this paper to be more targeted to telic in simple/static scenes and less relate to generic concepts of state. When I look at the supplementary videos, I feel like this method is very targeted to specific type of video samples where you have easily verifiable conditions in the frame without much camera / object movements / dynamics.

  • This is very minor, and can be done after rebuttal, but I think Wan 1.3B backbones are quite outdated. Given that this paper targets very application specific practical cases, I would also be curious about its performance on larger models (Wan 2.2 / Cosmos 3 / or Minimax H3).

5 comments
AuthorsOpenReview ↗

Author Response to Reviewer FHDX (Part 1/2)

Response to Reviewer FHDX

Thank you for your supportive assessment and for recognizing the value of our problem formulation. We address your concerns and questions in detail below.

Weakness 1 / Question 2 — Scope of state and scene dynamics

Meaning of state and scope of the task.

In our paper, state refers to the condition relevant to the instructed event. A telic instruction specifies an initial condition and an identifiable target outcome, which we call the terminal state. The generated video should depict the transition between them. For example, "open a closed door fully and keep it open" requires showing the door transition from closed to fully open and remain open. Our evaluation therefore considers whether the video completes the transition without regression and consistently retains the resulting consequence throughout the rollout. It goes beyond simply checking whether the terminal state appears in a single frame.

Instructions such as counting repeated knocks or turning a light off and then back on involve tracking event history or composing multiple state transitions. Our benchmark focuses on individual telic events under fixed instructions. We envision composing such events through successive prompts to support more complex interactions. For example, the model can first be prompted to turn a light off stably and then to turn it back on.

Consequence retention and the KV cache.

Evicting the frames depicting the original transition does not necessarily remove evidence of its consequence. Once a door opens, subsequent frames can continue to depict the open door and refresh that evidence in the recent cache. State Forcing trains the generator to produce such consequence-consistent continuations under the original instruction. Retention can therefore extend beyond the cache window containing the original transition. This mechanism relies on the consequence remaining represented in the evolving visual history. We agree that reliable recall after the relevant evidence has disappeared from the available context requires additional memory mechanisms. Our current experiments do not establish that capability; dedicated memory mechanisms would complement the behavior studied here. We clarify these distinctions and the intended scope of our formulation in the revised paper.

Controlled viewpoints and scene dynamics.

Many TelicBench prompts use controlled viewpoints to keep the task-relevant subject observable. This design isolates event execution and consequence retention from additional challenges such as occlusion, viewpoint changes, and subject re-identification. It also provides clearer visual evidence for VLM-based evaluation.

Importantly, controlled viewpoints do not imply static scenes. Our benchmark includes multi-stage object interactions (tray placement and putting on sunglasses), gradual task-relevant transformations (strawberry ripening, cactus blooming, and iron-bar cooling), and concurrent scene dynamics (moving reflections, passing insects, and heat shimmer), while requiring the requested consequence to persist.

This controlled setting is not trivial: even under these controlled visibility conditions, the strongest baseline (Context Forcing) achieves only 41.47% CR and 24.93% TSR (Table 1). This indicates that making the target outcome visually observable does not resolve the execution challenge: the generator must still complete the requested transition without regression and retain its consequence throughout the rollout.

For evaluation, the VLM examines chronologically ordered frames spanning the full 30-second rollout, together with task-specific rubrics. This controlled setting provides temporally ordered evidence for tracking the same primary subject, assessing completion, and identifying subsequent regression, reset, and re-execution. State Forcing imposes no fixed-camera constraint. We additionally provide qualitative examples with camera motion and substantial subject displacement (Appendix D.1.1), illustrating the model's ability to complete instructed transitions and retain their consequences under camera motion and subject movement. The revision distinguishes these qualitative examples from the quantitative evaluation on TelicBench.

AuthorsOpenReview ↗

Author Response to Reviewer FHDX (Part 2/2)

Weakness 2 / Question 1 — Training-data controls and evaluator dependence

Role of the accepted training records.

The 3,000 accepted records provide verified state evidence for constructing the ADVANCE and HOLD teacher contexts, rather than serving as complete successful-video demonstrations. GPT-5.6 verifies the initial condition and visible execution, identifies terminal-state intervals, and checks compatibility between unfinished prefixes and terminal chunks from the same source rollout. Acceptance applies to these selected portions; the entire source rollout need not be violation-free. During distillation, the student generates subsequent chunks, and the frozen teacher provides denoising-prediction supervision along these student-generated trajectories. Thus, the records initialize state-privileged supervision rather than supplying complete successful-video continuations for the student to reproduce.

Comparisons with matched training data and compute

Following the reviewer’s suggestion, we further train all four distilled baselines in Table 1 using the same telic warm-up procedure, the same 3,000 verified records, and matched training compute. All methods are evaluated using the same TelicBench protocol. Scores are percentages.

MethodCR ↑TVR ↓TSR ↑
Self Forcing (retrained)42.1323.6016.93
Rolling Forcing (retrained)42.2717.5317.33
Context Forcing (retrained)36.8011.3225.60
LongLive (retrained)39.0718.5120.67
State Forcing51.337.6037.33

State Forcing retains the highest completion and strict-success rates and the lowest violation rate under these matched conditions. Its TSR exceeds the strongest retrained baseline, Context Forcing, by 11.73 percentage points. These results support the contribution of the State Forcing training procedure beyond access to telic warm-up and the verified records.

Independent VLM evaluation.

To examine dependence on GPT-5.6 as the evaluator, we re-evaluate all original Table 1 videos using Gemini 3.8 Flash, which was not involved in training-data construction. We keep the generated videos, evaluation rubrics, chronological visual evidence, and anonymized evaluation protocol unchanged.

MethodCR ↑TVR ↓TSR ↑
SkyReels-V219.3318.906.00
MAGI-132.6714.2715.33
Self Forcing44.6722.4416.67
Rolling Forcing48.6716.3320.00
Context Forcing50.6717.3223.33
LongLive49.3315.5725.33
State Forcing52.009.8735.33

Under Gemini evaluation, State Forcing maintains the highest CR (52.00%) and TSR (35.33%), as well as the lowest TVR (9.87%). Our human evaluation on the paired subset provides further support. State Forcing achieves the highest completion rate (60%) and strict-success rate (34%), as well as the lowest violation extent (1.56). Together, these independent evaluations provide complementary evidence that the observed improvements are not specific to the GPT-5.6 evaluator.

Question 3 — Model backbone

We adopt Wan2.1-T2V-1.3B to match the backbone used by the primary distilled streaming baselines: Self Forcing, Rolling Forcing, Context Forcing, and LongLive. This enables controlled comparisons that isolate the contribution of telic supervision without confounding backbone differences. We agree that evaluating newer video-generation backbones would strengthen the evidence for generality, and we are exploring this extension beyond the current study.

We thank the reviewer once again for the constructive suggestions. We hope these responses address your concerns and would be happy to discuss any remaining questions.

Dear Reviewer FHDX,

Thank you again for your thoughtful review. We wanted to gently follow up and ask whether our response has addressed your concerns. Please let us know if anything needs further clarification.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Fully Revised Draft

Summary of Revisions

Dear Reviewers and Area Chair,

Thank you for your constructive feedback. We have revised the manuscript accordingly, with the main changes highlighted in blue.

In the main paper:

  1. Clarified the distinction between State Forcing and OPSD-V, emphasizing their different target problems and supervision designs within the shared OPSD framework (Section 2).
  2. Added independent VLM evaluation using Gemini 3.8 Flash (Section 5.4; Table 5).
  3. Added counterfactual terminal-context ablations to examine the role of verified, task-matched terminal evidence (Section 5.5; Table 6).
  4. Clarified the six evaluation dimensions supported by VBench’s custom-video protocol (Section 5.1).
  5. Expanded the captions of Figures 1–3 to clarify the illustrated failure modes and teacher-context construction.

In the supplementary material:

  1. Added training-data related statistics, and clarification of what state verification establishes (Appendix A.1).
  2. Clarified RoPE positions, cache initialization, and training computation within OPSD (Appendices A.4 and A.6).
  3. Explained the controlled viewpoints related settings in TelicBench (Appendix B.1).
  4. Added qualitative examples with camera motion and substantial subject displacement (Appendix D.1.1).
  5. Added comparisons with OPSD-V and four retrained streaming baselines under matched data and training compute (Appendices D.2.2 and D.2.4).
  6. Added dense-caption SFT and Astrolabe-based VLM-reward RL comparisons (Appendices D.2.5–D.2.6).
  7. Added an inference-time recaching experiment (Appendix D.2.7).
  8. Clarified the meaning of state and the scope of consequence retention, and discussed potential combinations with multi-prompt generation and memory mechanisms (Appendix E).

Thank you again for your time and consideration.

Best regards, Paper 2922 Authors

AuthorsOpenReview ↗

Author Response to Reviewer yp4i (Part 1/2)

Follow-up to Reviewer yp4i

Thank you for the follow-up. We clarify the three remaining concerns below.

1. Backbone choice and relevance to larger streaming systems

Our experimental setting follows the streaming video generation baselines such as Self Forcing (NeurIPS 2025), Rolling Forcing (ICLR 2026), and Context Forcing (ICML 2026), all of which evaluate their proposed methods using Wan2.1-T2V-1.3B as the sole pretrained generator backbone. Trajectory Forcing (NeurIPS 2026) and OmniMem (NeurIPS 2026), both recently accepted, explicitly evaluate their proposed methods using Wan2.1-T2V-1.3B as the sole generator backbone. Recent work such as OPSD-V (arXiv, July 2026) also evaluates its method on Self Forcing and LongLive, using Wan2.1-T2V-1.3B as the sole pretrained generator backbone across both implementations. These studies support the continued relevance of Wan2.1-T2V-1.3B as an experimental backbone for evaluating methodological contributions in streaming video generation, and our choice of generator backbone follows this established practice.

Our central contribution is to formulate telic instruction following and develop verified terminal-state supervision to advance this capability in streaming video generation. State Forcing achieves state-of-the-art telic instruction-following performance on TelicBench among the evaluated streaming generators (Table 1).

To assess its competitiveness against larger streaming systems, we additionally evaluated LongLive 2.0 and TaoMate-H3 using their official implementations and released model weights with the matched evaluation setting. Scores are percentages.

Streaming generatorBackboneCR ↑TVR ↓TSR ↑
LongLive 2.0Wan2.2-TI2V-5B41.8716.2921.20
TaoMate-H3MiniMax-H343.7320.0720.93
State ForcingWan2.1-T2V-1.3B51.337.6037.33

State Forcing outperforms both systems on all three telic metrics, with TSR gains of 16.13 and 16.40 percentage points over LongLive 2.0 and TaoMate-H3, respectively. These comparisons show that telic failures persist in the evaluated larger streaming systems, while State Forcing achieves stronger telic performance with its smaller Wan2.1-T2V-1.3B backbone. Our central empirical claims are well supported by the completed experiments and analyses, including the additional comparisons with larger streaming systems. These results substantiate the relevance of the problem, the effectiveness of the supervision design, and the practical competitiveness of State Forcing.

We are also training larger-backbone variants of State Forcing to assess transferability. Given the required backbone-specific data construction, verification, and training, the full evaluation will extend beyond the rebuttal period. We will report the completed evaluation in the revision.

AuthorsOpenReview ↗

Author Response to Reviewer yp4i (Part 2/2)

2. Novelty of the problem formulation and terminal-state supervision

Our methodological contribution is the design of verified terminal-state teacher supervision for telic execution. Our use of RoPE for temporal positioning is an implementation detail, rather than a claimed methodological contribution. In OPSD, privileged teacher context determines the predictions that supervise the student. D-OPSD (NeurIPS 2026, Sec. 2.2) explicitly identifies the construction of stronger teacher context as a central challenge. Learning from the Self-future (NeurIPS 2026, Sec. 4.3) likewise treats self-teacher construction as a core design choice. Our core design uses verified terminal-state evidence to construct complementary teacher contexts for supervising event progression and post-completion consequence retention.

To the best of our knowledge, we are the first to explicitly formulate telic instruction following in streaming video generation and develop a dedicated training approach jointly targeting completion, non-regressing progress, and consequence retention. Our controlled study identifies how terminal-state conditioning supports progression and retention in a frozen generator (Fig. 2). We translate these findings into complementary teacher conditions: ADVANCE uses verified terminal evidence alongside unfinished execution to guide progression, while HOLD isolates the achieved outcome under a consequent-state description to supervise its retention. This design translates our empirical findings into complementary teacher supervision for telic progression and consequence retention, rather than merely adapting existing context-handling operations.

The controlled experiments directly test this supervision design. Under matched training data, initialization, and compute, State Forcing achieves 37.33% TSR versus 17.33% for OPSD-V (Table 10). With training settings held fixed, replacing the verified, task-matched terminal context with any of the tested alternatives reduces TSR from 37.33% to 20.53% or lower (Table 6). Removing either ADVANCE or HOLD also reduces performance (Table 4). These controlled comparisons demonstrate the substantive contribution of the proposed terminal-state supervision, rather than merely reporting engineering gains.

Our methodological novelty lies in the telic problem formulation, the empirical findings on terminal-state conditioning, and the complementary supervision design that translates these findings into improved telic instruction following. Together, the identified supervision mechanism and its controlled validation support a substantive methodological contribution.

3. Telic improvements and post-completion dynamics

Our central result is improved telic instruction following while maintaining scene dynamics comparable to baselines such as LongLive (Tables 1 and 2), rather than an improvement in generic video dynamics for streaming video generation. The additional analysis specifically addresses the concern that consequence retention might be achieved through post-completion freezing. For each method, we evaluate strictly successful TelicBench videos whose instructions explicitly require continued scene dynamics from the first completion timestamp to the end of the rollout. State Forcing and LongLive obtain similar mean scores on the 0–2 scale: 1.52 vs. 1.57 under VLM evaluation and 1.24 vs. 1.22 under human evaluation. These results provide evidence of consequence retention with continued scene dynamics, rather than retention achieved merely through post-completion freezing.

The larger-system comparisons in Section 1 show that pronounced telic failures persist in the evaluated larger streaming systems. LongLive 2.0 and TaoMate-H3 achieve 21.20% and 20.93% TSR, respectively, compared with 37.33% for State Forcing, which also outperforms both systems on CR and TVR. Larger-backbone State Forcing training is underway, as described in Section 1. The completed experiments substantiate the relevance of the telic problem and the effectiveness and practical competitiveness of State Forcing.

We thank the reviewer once again for the suggestions. We hope these responses address your concerns.

Dear Reviewer yp4i,

Thank you again for following up with us. With the discussion window closing very soon, we wanted to gently check whether our latest response has addressed your remaining concerns. If so, we would be grateful if you would consider raising your score accordingly. Please let us know if anything needs further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.