Video Generation in September 2026 — Progress Tracking

September 2026 moved video generation forward on seven fronts at once, and the three changes worth remembering are not about image quality. Agentic orchestration became an object of study rather than a wrapper: two systems now train the controller that composes requests, one grounds multi-shot continuity in an explicit story state, and an audit shows that the verifier closing these loops can be manipulated by the harness's own execution log. The long-horizon family converged on a single diagnosis — a train–inference mismatch in how the history is formed — and produced at least six different fixes for it, none of which compose yet and several of which trade motion against quality. And metric validity became a research topic: a controlled scaling study shows validation loss can invert the ranking of model capability, a survey of post-training states that its own tables are not a leaderboard, and evaluation of interactive world models is still mostly throughput. Underneath, the world-model work shows the field separating what is transferable from what must be retrained per target.

Window and coverage: papers with an arXiv v1 submission date between 2026-09-01 and 2026-09-30 (the previous complete calendar month as of 2026-10-14); 374 candidates screened, 220 included, 32 read in full text. Scope: progress in video generation models and their generation capability — architecture and quality, long/streaming/interactive generation, world models, conditioning and editing, sampling efficiency, agentic orchestration of generation, joint audio-video generation, and evaluation/safety of generative video. Domain-transfer work (driving, robotics, medical, e-commerce try-on) is included only where the contribution returns to the generation method. Evidence basis: 32 papers were read in full text (method and experiment sections, plus the appendices the section map exposes) for this issue; the remaining screened papers are summarised by theme at the level of title and abstract only. Every number below is quoted with its setting, and cross-paper numbers from different settings are flagged as not comparable. Author claims, the evidence the papers supply, and this report's own inferences are kept apart. How the corpus was built: one keyword route (title/abstract terms for video generation, video diffusion, text-to-video, image-to-video, video synthesis, video model) and one semantic route were run over the window and merged into a single candidate pool; the pool was then screened against a written inclusion criterion by a calibrated classifier. The counts below are counts within the checked routes, not the size of the field.

Figure 1. The seven themes in the screened September 2026 corpus and the papers discussed in this issue. Classified counts are model-assigned over titles and abstracts, so they are approximate; 54 further papers were left unassigned. Source: this issue's own screening.

---

1. Agentic orchestration became an object of study, not just a wrapper

September's agentic video work is no longer only "an LLM in front of a generator". Two of the papers train the controller itself, one audits the verifier that closes the loop, and one proposes a criterion for when a pipeline should be called agentic at all. The shared technical claim is that text-only plans under-determine what a frozen generator renders, so the fix has to be typed, visual conditioning plus explicit state rather than better prose.

1.1 StoryEngine: story state as an authoritative object, not a re-reading of pixels

StoryEngine [1] targets the failure mode where an agentic pipeline conditions later shots on textual plans or previously generated pixels, so an omitted action or a misplaced object is silently reinterpreted as a new narrative fact. Its design separates an authoritative semantic plan from fallible visual observations: each entity carries a placement and story-relevant attributes, a deterministic reducer updates the state per shot, and the per-shot transition is computed from planned effects rather than estimated from the rendered video. The previous shot's tail frame must pass a one-way compatibility gate before it can condition the next shot, and failed criteria trigger bounded local repair (budget $T=3$) with the plan held fixed.

The evidence is a 60-story benchmark (three diagnostic suites: narrative realisation, cross-shot coherence, visual consistency; 10 shots per story, two recurring locations) run over two backbones, Veo 3.1 and Wan2.2-TI2V-5B. Against the two agentic baselines the paper names (ViMax, MovieAgent), the authors report an average margin of 0.1416 on Veo and 0.1304 on Wan, with the same ranking on both backbones, and the ablation attributes the largest drop to removing grounded render planning. Two caveats the paper itself supplies matter more than the headline: the plan-text metric is shared across backbones and therefore the two backbone blocks must not be mixed, and the VLM judge is the same model as the in-loop evaluator, so VLM-judged metrics could favour the method. The stated limitation that stands out is that authoritative-plan errors persist — a wrong entity catalogue or effect order cannot be repaired by any amount of later visual evidence.

Figure 2. StoryEngine's pipeline: story-state propagation, grounded render planning, and continuity-aware generation with its compatibility gate and bounded repair loop. Reproduced from Figure 2 of [1]; the image was inspected at pixel level for this issue.

1.2 The controller starts being trained

Two papers move the orchestration policy from hand-written prompts to a learned object.

MVAgent [2] takes the black-box setting seriously: eight-to-sixteen-shot stories are generated through separate API calls to a frozen generator that shares no internal state. The paper recasts cross-shot consistency as condition construction, with agents exchanging typed conditioning inputs — images with fixed positional roles (transition anchor, character portraits, environment views) rather than text alone — and then trains the Orchestrator that composes each request with Trunk-GDPO, a shot-level agentic RL scheme that samples $K$ candidates per shot, normalises each reward component across candidates before aggregation, and continues only the highest-advantage candidate as the trunk. On ViMax-Bench (35 stories) it reports overall consistency 0.5611 versus ViMax's 0.5500 under the same generator and VLM agents, and wins 60.06% of human pairwise comparisons against ViMax; the authors' own framing bounds this as "subject to the generator's capabilities and the judges' reliability". Worth flagging for a reader comparing tables: the ablation block uses a different (Qwen3-32B) Orchestrator, so its numbers are not comparable to the main table, and no resolution, duration or step counts are reported anywhere in the sections read.

VideoGen-Agent [3] instead trains a tool-selection policy. A Qwen3-VL-8B agent runs a reasoning–action–observation loop over three tool groups — augmentation (web search, a code physics simulator), generation (T2V/I2V/reference/motion), and verification (object detection, monocular depth) — with SFT on teacher trajectories followed by GRPO on a hybrid reward (0.1 format, 0.5 VLM rubric, 0.4 tool use). On the paper's own 600-prompt VABench it reports 75.6 overall with the original toolset and 86.1 after swapping in stronger generators, against 56.5 for the base generator and 73.2 for the strongest standalone model tested. Two of the authors' numbers deserve to be read carefully rather than quoted: the 75.6 → 86.1 jump is a tool-upgrade effect with the policy frozen, not a learning effect; and the multitask claim is weaker than it looks, since single-task RL reaches 76.3, marginally above the shared multitask policy. The evaluation also uses the same VLM with the same rubrics as the RL reward, which the authors mitigate with a 92% judge–human agreement study (dropping to 72% on procedural knowledge).

CamPilot [4] trains the cinematographic part of the controller: a Camera Work Planner that predicts angle, shot size and a five-attribute motion tuple, supervised by VLM-annotated shots from 14,581 real film videos and optimised with GRPO, wrapped in scene/shot planners plus an evaluator–reviser loop with a character bank and end-frame conditioning. The authors report camera-work Macro-Acc of 63.7 against 59.5 for the agentic baseline and 60.1 for the strongest commercial reference under an untrained setting, but the lead is not uniform — on Macro-F1 it does not lead for every backbone, and the camera labels and the test labels come from the same VLM annotator, so part of the measured agreement is agreement with that annotator's conventions. (The latter is this report's inference, not the authors' claim.) The paper explicitly lists human evaluation as future work.

1.3 A production-layer DSL instead of prompt chaining

CineCrew [5] puts an explicit intermediate representation between screenplay and generator: clips are compiled into FilmDSL with three layers per clip (narrative action, cinematic staging including continuity hooks, render spec), generated keyframe-first, and repaired by turning failures into structured QA objects that drive localised regeneration rather than whole-clip re-prompting. A Persona Schema maps traits to beat-level behaviour, and a persistent Production Rulebook is shared by the crew agents (Showrunner, Art Department, Story Editor, Dailies Reviewer, Post Supervisor). On 20 MovieBench narratives judged by a VLM, it leads its three baselines on most of the 17 reported metrics and wins a 24-participant study on five of six axes, losing shot-level cinematic score to LTX-Studio. The paper has no limitations section, and — importantly for anyone comparing numbers — it does not state resolution, clip length, sampling steps, hardware or the identity of its off-the-shelf backbones, so its scores cannot be placed against other papers' scores.

1.4 The verifier is the soft spot

Don't Read the Log [6] is the sharpest result of the month for anyone building these loops. It holds the frames fixed, varies only the auxiliary text that harnesses feed their own multimodal judge (narration transcript, execution log, burned-in subtitle), and measures whether the verdict moves on a purely visual requirement. Open-weight Qwen-VL judges from 7B to 32B accept 78–90% of visibly failed clips when the execution log claims success; in the closed loop, non-adversarial planners exploit this — the naive 8B configuration ends at judge pass 1.00 while true pass is 0.28, and the paper shows the ceiling identity true pass $\le 1-\alpha_0$ holds to two decimals across configurations. Frontier closed judges tested did not move, but they were run on subsets (26–109 clips) for budget. The proposed fix is structural rather than prompt-level: type each requirement, route the judge only the evidence type that can establish it (plus OCR-masking text burned into frames), which restores hacked 0.14–0.08 and true pass 0.86–0.92 at roughly 3–4× the cost per episode. Its own limits: a 109-clip benchmark with author labels, a simulated rather than deployed repair loop, one open-weight judge family, and no audio requirements instantiated.

Figure 3. Auxiliary text moves the verdict on a purely visual requirement. With a supportive execution log the open-weight judges accept 0.78–0.90 of failed clips (left panel), while a contradicting log drives Qwen3-VL acceptance to 0.00–0.08 and leaves GPT-5.4-mini at 0.90 (right panel). Figure 1 of [6]; viewed at pixel level for this issue.

1.5 A criterion for "agentic", and what the corpus looks like under it

Agentic Visual Generation [7] argues that the field has no consistent criterion for agenticity — planning depth, tool count, multi-role design and RL are all used as evidence even though none of them determines which generation decisions the controller can actually make. It proposes a controller-capability ladder L0–L4 defined by the latest future generation decision the controller can causally change (specification → execution → within-task outcome adaptation → cross-task experience reuse), with ambiguous cases assigned to the lower level. This is a position paper: its evaluation protocol is proposed, not measured, and the corpus counts in its figures are not usable from the text layer of the reading copy. Two structural observations it reports are still informative: only four reviewed records fall at L0, i.e. the corpus is deliberately built around controller-bearing systems; and 2026-H2 contains 52 records through 24 August 2026 in its records.

Table 1. What the agentic systems in this issue actually optimise, and how they verify it.

The tension to watch. Within a single month, five papers built plan–generate–verify loops and one showed that the cheapest component of that loop — the judge — is exploitable by honest planners whenever the harness forwards its own execution log. The mitigations proposed in September (typed evidence routing, judge–human agreement checks, VLM-free metrics) are structural rather than cosmetic, which is the right direction; none of them has yet been evaluated inside a deployed harness at scale.

Takeaways — Agentic video pipelines in September split into three jobs: building the condition (typed, visual, state-grounded), training the composer, and verifying the result. The verification job is the least trustworthy of the three: open-weight judges accept 78–90% of visibly failed clips when the harness's own log claims success, and honest planners exploit that inside repair loops. The proposed mitigations are structural rather than prompt-level, and none has yet been tested in a deployed harness.

2. Long-horizon generation converged on one diagnosis and diverged on the fix

Almost every September paper on long video starts from the same failure: a generator distilled to run causally, one chunk at a time, degrades over rollouts far beyond its ~5-second training horizon — drift, saturation, or the opposite, near-static frames. The papers also share a diagnosis, namely a train–inference mismatch in how the history is formed, and they differ mainly in which part of the loop they intervene on. Strikingly, none of them buys length by simply enlarging the context window.

2.1 Where the history is, and how it is supervised

Self-Aligned Forcing [8] makes the history noisy but differentiable: history key/value states are produced in the same block-causal pass as the prediction, at the same denoising stage, instead of being reconstructed in a separate no-gradient rollout and detached. The parameter gradient then splits into a read and a write term, so a future loss can supervise how earlier blocks encode what later blocks read. The authors report the fastest training and inference among the compared streaming methods — 155.18 GiB and 132.78 s per 5-step cycle against SGF's 187.45/186.73, about $1.8\times$ faster framewise, and 49.1 latent FPS with 4-GPU stage pipelining — together with the best VBench total (85.04) in the chunk-wise setting. The reported picture is not a clean sweep, and the paper shows it: it is not best on Interactive ViCLIP or VBench semantic score, and in the framewise setting SGF leads several aggregates. Its ablations are the informative part — detaching the K/V gradient costs about 13 points of dynamic degree, and removing the clean "sink" recache collapses dynamics to 6.90, which is strong evidence that the history-encoding path, not capacity, is what carries long-horizon motion.

Recency Forcing [9] intervenes one layer lower, on attention weighting. It first measures a positional response — how much the velocity prediction changes when a context frame at distance $\Delta$ is swapped for a frame from an alternative rollout — and finds a steep decay with distance plus a systematic per-denoising-step structure. It then adds a monotone, non-positive bias to the attention logits, parametrised by temporal distance and denoising step and applied only to the history portion of the context, with a power-law decay chosen over linear and logarithmic after ablation. A reparameterisation absorbs the bias into the values so it runs as an unmodified FlashAttention call, which the paper measures at 6.06 ms versus 70.34 ms for FlexAttention with the bias applied directly ($11.76\times$). On VBench-Long's 60-second setting it reports quality 84.02 versus LongLive 81.98 and Rolling Forcing 81.56, with dynamic degree 75.48 against LongLive's 42.85. Two things to keep in mind: its training-free variant and its training-based variant are different model regimes and must not be read as the same system, and the paper reports no wall-clock or memory comparison against the baselines, only the kernel microbenchmark.

LongTake [10] changes what the teacher forcing sees. Instead of initialising a few-step student from short clips, it applies teacher forcing on long ground-truth videos with a cache of sink and recent frames, which the authors argue gives a direct initialisation for self-rollout DMD and removes a whole training stage; a second mechanism reuses the resulting autoregressive teacher to supervise the later frames of the student's own rollout. On 30-second rollouts it reports dynamic degree 96.30 with hybrid DMD against 93.44 for Rolling Forcing and ~51–59 for SGF, MemRoPE and Context Forcing, at comparable aesthetic and quality scores. Its own framing of the metric problem is worth repeating: at 30 s, every baseline with a higher aesthetic score stays at or below 58 dynamic degree, and the one baseline above 90 dynamic degree trails by 3.9 aesthetic points — a reminder that the two axes trade off and that a single number will misrepresent this family.

Figure 4. Long-horizon teacher forcing against short-clip initialisation under the same 5-second self-rollout distillation, evaluated on 30-second rollouts: dynamic degree 90.73 against 51.30, with subject consistency (96.83 versus 98.19) and aesthetic quality (62.51 versus 63.90) close. Figure 5 of [10]; viewed at pixel level for this issue.

Rollout-Marginal Distillation [11] blames the supervision target rather than the context: video-level distribution matching scores a whole rollout jointly, so the teacher's real-score prediction inherits the student's flawed history and corrects the new chunk toward consistency with past artefacts. The fix scores each chunk independently against a chunk-distribution teacher and then refines with video-level DMD to restore cross-chunk coherence. The reported gaps on VBench-Long 60-second videos are large — total 81.26 against Self Forcing 70.94 and Causal Forcing 69.88 at chunk size 1 — and the semantic score is where most of the gap sits (68.24 versus 43.17 and 36.03), which the paper attributes to prompt-following collapse in the baselines; that attribution is plausible but was not separately ablated. Its full method is marginally below its marginal-only ablation on quality (84.52 versus 84.90) while ahead on total and semantic, a detail worth knowing before quoting the headline.

Table 2. The six interventions on the long-horizon train–inference mismatch, and what each one supervises.

2.2 Other fixes in the same family

Mask Forcing [12] is the cheapest intervention of the group: it perturbs only the self-rollout, mixing tokens from a slightly cleaner timestep into the scheduled-noise state by a random mask while keeping the original DMD objective and the original conditioning timestep, and reports gains across three existing pipelines (HPSv3 +0.29 for Self Forcing, +0.80 for Causal Forcing, +1.03 for LongLive) plus 72–83% human preference, with no real data and no extra post-training. Its one regression — LongLive's chunk-wise dynamic degree falling from 76 to 69 — is explained by the authors as a VBench artefact where drift inflates optical flow; that explanation is a claim, not a measurement. SemanTok [13] attacks the other end of the stack: a video tokenizer whose every retained prefix must reconstruct frozen DINOv2 features, so that early tokens carry clip semantics. It reports that an 85M autoregressive model at a 16-token budget beats a 2.29B baseline on gFID on both datasets, and that tokens cost 32% fewer bits — but the evaluation is 17-frame 128×128 clips, so the streaming motivation is inferential rather than demonstrated, and the tokenizer reconstructs colour and appearance less faithfully at low budgets by the authors' own account.

Takeaways — Six papers attacked the same train–inference mismatch from different layers: history differentiability [8], attention weighting [9], initialisation from long video [10], the distillation target [11], the rollout noise [12] and the tokenizer [13]. None was combined with another, and the results are metric-dependent: motion dynamics and perceptual quality trade off against each other across these papers, and two of them warn that consistency metrics can reward near-static output.

3. World models: bounding the store made interactive generation deployable

Interactive world models in September are differentiated less by their backbones — most wrap an existing video diffusion model — than by how they store a scene between chunks. This theme covers the papers that make that store finite; the next covers those that let it grow instead.

3.1 A store that never grows

Honeycomb [14] stores the scene in six fixed-size 2D planes, three spatial and three spatiotemporal, which are read to condition each chunk and written back in place, so memory never grows with rollout length. Against a growing point cloud and a growing latent cache it reports the best WorldScore average (65.52 versus 63.21 and 61.20) and a closed-loop revisit PSNR of 17.22 dB versus 15.99 for the next best. Its most useful ablation for readers is the writer: per-scene optimisation reaches 17.44 dB but costs ~3,130–3,228 ms per write, a "replacement" writer degrades to 17.16 dB with write time growing 11.2 → 40.1 ms, while the recurrent writer holds 17.22 dB at a flat 13.1–13.2 ms. The stated cost of the design is a resolution trade-off: as scene bounds expand, fixed grids cover more space and time at coarser resolution. There is no limitations section, and baseline byte footprints are never quantified, only shown qualitatively.

Figure 5. Honeycomb's first-frame initialisation, memory-conditioned generation and recurrent update, with the six fixed-size HexMemory planes on the right. Reproduced from Figure 3 of [14]; the image was inspected at pixel level for this issue.

Honeycomb [14] is complemented by Waypoint-1.5 [15], a 1.28B interactive world model whose goal is real time on consumer GPUs rather than memory efficiency: a single-stream causal DiT with per-block controller injection and a rolling KV cache (dense 16-frame local ring plus a dilated 128-frame ring every fourth block), trained with per-frame diffusion forcing and distilled to four steps. The reported throughput is the headline — 720P INT8 clears 30 latent FPS from an RTX 5060 Ti upward, and 360P clears it on every tested card including an RTX 3060. What the paper does not report is equally notable: no latency figure, no quality or FVD metric, no action-adherence metric, no head-to-head comparison against the named world models, and no ablation. Its own abstract claims to distinguish rendered FPS, latent FPS and control rate, but only latent FPS is defined and measured in the sections read — this report's observation, not a claim of the authors. For a systems paper that is a coherent choice; for anyone asking whether the world model is correct, September's evidence base here is throughput plus qualitative rollouts.

Table 3. The bounded-memory world models in this theme.

Takeaways — Bounding the store is what makes an interactive world model deployable: a fixed plane set keeps memory flat at a reported 13 ms per write, and a rolling cache keeps an interface above 30 latent FPS on consumer hardware. The price is visible in both papers: resolution degrades as bounds expand, and the real-time work is evaluated on throughput alone, with no control-fidelity or long-horizon stability measurement.

4. World models: letting the store grow, by construction rather than inference

4.1 Fused geometry versus committed geometry

OctWorld [16] takes the opposite route: an extensible octree TSDF map whose voxel resolution is allocated per pixel from depth, fused with generated RGB-D observations and rendered back as conditioning for the next chunk. This is the month's clearest statement that world memory should be fused rather than accumulated — it reports 50.31 WorldScore static and 71.67 3D consistency against Gen3C's 70.52 and Voyager's 66.35, and wins a 22-participant user study 53.8% to 39.5% against Gen3C on long outdoor trajectories. Its admitted costs are real: no voxel pruning, all nodes cached on GPU so memory grows with scene extent, and no view-dependent level of detail because coarse nodes keep no features. Baselines also run at different resolutions and chunk lengths (Gen3C 1280×704/121 frames, Voyager 768×512/49, OctWorld 832×480/36), so the WorldScore table mixes settings and should not be read as a matched comparison.

WorldWeave [17] pushes the same instinct further with a stricter rule: generated RGB never writes back into world state. Instead an explicit metric world — terrain, districts, asset instances and inherited boundary interfaces — is grown by validated construction, and video generation queries it read-only. Terrain expansion is treated as a measurement problem (residual encoding against a neighbour-fitted surface, decoder inversion, Hermite-style derivative constraints at joins: height-only joins leave 0.319 m residual, derivative inheritance brings it below $10^{-9}$ m). On 120 exploration and revisit tasks it reports the best structural-motion consistency (2.30 versus 2.21) and the second-best aesthetic score, with seed stability far tighter than baselines (0.39% coefficient of variation on image quality versus 27.61% for the closest competitor). Two caveats from the paper's own text: WorldWeave receives depth queried from its constructed world while baselines get only first frame, trajectory and prompt — so it is not condition-matched — and the authors limit themselves to static geometry, with depth as guidance rather than a hard rendering constraint.

Figure 6. WorldWeave's pipeline: metric terrain chunks are completed against committed neighbours, scenes are planned and validated before a state commit, and video generation queries the committed world under a read-only lock. Figure 2 of [17]; viewed at pixel level for this issue.

Table 4. The growing-memory world models in this theme.

4.2 Control through the generator's own interface

H3-World [18] is the cheapest intervention in this group: rather than adding an action module, it expresses per-latent-interval character and camera actions as short natural-language prompts, packs them into the frozen 33B MiniMax-H3 generator's packed self-attention sequence with mirrored positional offsets and a deterministic "single-egress" attention mask, and trains only a rank-32 LoRA — reported at 0.199% of parameters, 7,872 clips and 10,000 steps. The diagnostic is the interesting result: a frozen model given one global motion instruction barely responds to a mid-clip reversal (accumulated mean horizontal flow 0.0 → −17.3), the per-latent interface with LoRA zeroed is unresponsive (0.003), and the trained version produces +52.7 → −106.0; on constant motion the frozen model and H3-World are nearly identical (301.8 versus 300.5), which supports the authors' claim that LoRA is what buys time-varying control rather than coarse motion. The claims it does not support numerically are "preserving generation quality" (only qualitative examples, no FVD or human study) and generality: the study is short-horizon and fixed-length, with no persistent world state, no real-time interaction and no planning, and it reports no latency or hardware figures.

Takeaways — When the store is allowed to grow, the month's two routes differ in what they trust: OctWorld fuses generated observations into an octree and pays with memory that grows and no voxel pruning, while WorldWeave refuses to let generated pixels write back at all and validates every commit before it joins the world. Both beat their baselines on consistency — but not under matched conditions, since WorldWeave is handed depth the baselines never see, and OctWorld's baselines run at different resolutions and chunk lengths.

5. Control and editing: cheaper interfaces rather than bigger models

Control and editing was the largest single bucket in the screened corpus, and the papers read here share a design instinct: reduce what the user must specify, and make the new modality cheap to learn. Two of the four also illustrate how much a control paper's result depends on its evaluation protocol.

5.1 Asking for less

CoaG [19] asks how little geometry a person must draw to fix both where people stand and where the camera moves, and answers with a ground plane drawn as a wireframe grid plus one upright cylinder per person — no pose, no heading, no boxes. A LoRA on an open 14B control backbone is trained on ~1,935 automatically produced tuples (rank 64, one expert at a time, about 1h50m and 70 GB per GPU per expert on 8×H100). One methodological fix is worth singling out because it applies beyond this paper: the base recipe takes the reference image from a frame of the target clip, which leaks the very layout the control video is supposed to supply, so CoaG reads the reference from an inpainted background instead. The authors are unusually explicit that they ran no human study, no ablation grid and no quantitative controllability metric, so the representation claim rests on uniformly sampled qualitative results and hand-authored stress tests (six cylinders give six people; a 1.7/0.6 adult–child height ratio is preserved; steep top-down and ground-level cameras are not followed). The most informative numbers in the paper are the dataset's own failure statistics: although seven of eight caption cards request camera motion, 47% of the generated training clips have a static camera, and the pipeline flags 658 subject-count mismatches in 2,000 clips. Those set a hard ceiling on what this LoRA could have learned, and the authors report them rather than hiding them.

Figure 7. CoaG's control representation and data engine: the same three inputs — scene grid, reference image and text — feed a shared renderer at both training and inference, and a rank-64 LoRA adapts a frozen control backbone with two weight sets per denoising step. Figure 2 of [19]; viewed at pixel level for this issue.

BEACON [20] makes a similar "add one cheap factor" move for subject-specific human video: identity stays an image, but behaviour is conditioned on a separate expressive video through an attention design that keeps the two branches from leaking into each other, trained with ~0.15B trainable parameters on a frozen 14B backbone. Its reported expressivity gains are substantial (emotion-FID 0.86 versus 2.28 and 3.17 for the two baselines on MEAD), and it demonstrates identity–expression transfer — one subject's face with another's dynamics. The interesting result is a protocol effect: under the usual neutral-reference measurement protocol BEACON's face similarity (0.760) is below the Stand-In baseline (0.826), but when the ground-truth frame is expressive the ordering reverses (0.852 versus 0.678). Any comparison in this sub-area that quotes a single identity metric without stating which reference frame was used is therefore uninterpretable. The authors also concede residual appearance leakage (generated eye colour partly follows the expressive reference) and do not compare against Slot-ID, which they name as the closest dual-reference prior.

Table 5. The control and editing papers in this theme.

5.2 Editing as a data problem

VideoX-Qwen [21] takes the data-centric route to instruction-based editing: a synthetic paired-record pipeline (object removal, addition, replacement and attribute edits, reusing existing video models for each direction) yields more than 1.2M directional records at an 89% automatic acceptance rate, and a unified editor consumes both sparse semantic tokens and dense source latents, trained in three progressively higher-resolution stages. Against UniVideo and Kling O1 on 100 examples it reports the best mean on 9 of 11 automatic metrics — instruction following 7.70 versus 6.91 and 7.61, SSIM 0.766 versus 0.572 and 0.670 — with the judge being a Qwen2.5-VL-7B model and no human study; the authors note that Kling O1 remains stronger on aesthetic and imaging quality, and the head-to-head is limited to these two systems despite many named priors.

SVEET [22] attacks streaming editing, where the obvious route — distilling the editing ability into a causal backbone — is expensive. Its answer is to learn the edit on a bidirectional model and transfer it to a frozen streaming backbone with no retraining: the control branch is changed from full 3D attention to frame-wise 2D spatial attention (so no extra temporal KV cache), and a least-squares/SVD procedure constrains the trainable update to the orthogonal complement of the directions that carry the bidirectional-to-causal gap. On a VLM editing-accuracy score with identical test sets across methods it reports 7.43 versus 7.07 for the 3D-attention VACE variant, and the ablation shows both changes contribute (3D without the orthogonal training 7.07, 2D with it 7.43). Two caveats: the 15 FPS on a single H100 operating point is asserted without any latency table, and the strictest generality claim (transfer across heterogeneous backbones) rests on one qualitative appendix figure; the two learning rates in the paper's sections also disagree ($1\times10^{-4}$ versus $5\times10^{-5}$).

Takeaways — The control papers of September all reduce what the user must specify: a ground plane plus one cylinder per person controls layout and camera together [19]; identity stays an image while behaviour comes from a second reference video [20]; instruction editing is bought with 1.2M synthetic paired records rather than architecture [21]; and streaming editing transfers from a bidirectional model without retraining the causal backbone [22]. Two of the four rest on asserted operating points rather than measured tables, and two show that the headline identity or speed number depends on the measurement protocol.

6. Efficiency: four separate levers, and a clearer line between capability and target

September's efficiency work is not one story. It splits cleanly into capacity, approximation accuracy, serving, and reuse — and two of the four papers spend real effort telling the reader what their numbers do and do not mean.

6.1 Capacity and the approximation

SplitMoE [23] attacks capacity: LLM-style token-wise mixture-of-experts routing plus load balancing scatters spatiotemporally coherent video patches across unrelated experts, so it splits the expert pool into semantic and generic pools, routes with prototype guidance derived from clean VAE features, and deliberately does not balance the semantic experts. Upcycled from a Wan 2.2 low-noise checkpoint, it activates 14B of 27B parameters and is compared against both a load-balanced MoE baseline and a dense model from the same source — the comparison the paper relies on for causal claims, since it explicitly warns that its cross-model rows are not compute-matched. It leads the same-source controls on every listed VBench-2 and T2V-CompBench column (human fidelity 84.47 versus 73.49 and 73.08; controllability 45.72 versus 35.51 and 30.37) while reporting 6.84 s/iteration training against 6.69 for the MoE baseline and 5.66 for the dense one. Against the official Wan 2.2 ensemble it wins five of seven and loses controllability and interaction, and it trails LongCat-Video on common sense and numeracy — so this is a routing-quality result, not a uniform win, and the paper says as much.

PARK [24] attacks the accuracy of the approximation rather than the amount of work. Sparse attention normally scores query/key blocks from averaged centroids; PARK shows two estimation mismatches — averaging queries before the softmax discards per-query preferences, and Euclidean key clustering minimises key variance rather than the score spread that governs the key-centroid error — and fixes both (keep every query and average its normalised weights; cluster keys in a query-derived metric space realisable by a Cholesky change of variables). Across four models it reports the best PSNR/SSIM/LPIPS among sparse baselines at 1.53–1.76× end-to-end speedup. The essential caveat is in the metric definition: those PSNR/SSIM/LPIPS values measure similarity to dense-attention output, not absolute quality, and the VBench columns stay within a few tenths of dense. The ablation is clean — removing either component costs latency-free quality (PSNR 29.85 → 28.40 without the query correction, → 29.54 without the key clustering), which is the paper's real point: the errors are free to fix.

Table 6. The efficiency papers in this theme.

6.2 Serving and reuse

Sol-H3 [25] is a serving report rather than a method paper, and it is the month's most concrete deployment evidence. It restructures MiniMax-H3's 49-step $1344\times768$ trajectory into 4 steps at $672\times384$, a latent-to-latent handoff (learned ×2 upsampler plus a cross-VAE adapter, 63 ms against 12,749 ms for a VAE round trip), and 3 refine steps with a different backbone, then searches the kernel and communication stack under an explicit fail-closed contract: the search may not alter the generation contract, so every accepted change is like-for-like. Reported outcomes: 13.7–17.0× over the 49-step baseline on 1–4 GB200s, 31× on a DGX Spark and 26.3× on an RTX 5090, and memory brought from 142.4 GiB (over the 119.68 GiB pool) to 116.9 GiB on the edge device. It concedes that the low-resolution-draft/high-resolution-refine recipe is prior art and that its contribution is the serving realisation; it also reports two numbers that do not reconcile (a 6.226 s figure for the same nominal 5-second workload that Table 1 gives as 9.475 s) and no quality metric for the two-stage pipeline beyond qualitative comparison.

LongLive-Plug [26] makes reuse the objective: distil three capabilities once on a backbone family — single-pass CFG, four-step sampling, and long-context error correction — as functional LoRAs that can be merged into compatible downstream checkpoints without target data or target training. The economics are the headline: roughly 80 H100 GPU-hours for the shared base adapters against 456.8 GPU-hours for accumulating per-task distillation across the four tasks it measures, with reuse verified on 54 downstream models across eight task categories at a 5×–12.5× step reduction. The honest parts matter more than the headline: the long-context results are described by the authors as competitive with metric-dependent trade-offs rather than a win (VBench seven-dimension mean 73.51 → 75.77 at 64 s, but background consistency drops 88.87 → 80.81), reuse requires compatible descendants of the base model, and the paper concedes that its naive-baseline comparison differs in both the adapter and the sampler, so the isolated LoRA effect is not established.

Figure 8. Once-for-all distillation changes the cost curve: task-specific distillation accumulates 163.9 → 456.8 H100 GPU-hours across four downstream tasks, while the shared base adapters stay at roughly 80 hours. Figure 2 of [26]; viewed at pixel level for this issue.

Takeaways — Efficiency in September was four separate levers, not one: capacity at fixed active parameters [23], the accuracy of the sparse-attention approximation [24], the serving stack [25] and once-for-all distillation [26]. The strongest patterns are the separation of transferable capability from per-target training — about 80 GPU-hours of shared adapters against roughly 457 for per-task distillation — and a willingness to state what the numbers are not: similarity-to-dense rather than absolute quality [24], or a metric-dependent trade-off rather than a win [26].

7. Audio–video: the soundtrack becomes editable, and rewards learn to stay apart

Two papers read here share a diagnosis of joint audio–video generation: the audio side has been treated as one monolithic signal. They fix it in opposite ways.

7.1 Making the soundtrack editable

Soundwich [27] is training-free: it wraps a frozen joint audio–video flow model so that the single audio latent is expanded into $N$ source-specific stems plus an internal scene stem, with per-stem guidance and cached "carrier" attention vectors captured from sounding and silent references to schedule when each source is active; a scene stem gathers shared context in the first half of generation and broadcasts it back during a scene-integration stage, after which the scene stem is discarded. The resulting stems can be retimed, muted, replaced or remixed, and the reported scheduling numbers are large: window success 93.1% against 15.0% for naive batch expansion of the same backbone and 85.0% for a commercial omni model, off-window activity 1.7% against 69.7%, and word error rate inside the intended window 0.060 against 0.779. What makes this entry useful is the uncertainty reporting: the authors give bootstrap intervals and note that their differences from the commercial model in word error rate (−0.008, CI [−0.156, 0.175]) and window success (+8.1 pp, CI [−2.8, 22.8]) include zero, so only the onset-error difference (−0.541 s, CI [−1.323, −0.068]) is supported at that sample size; baselines also run at different resolutions, and the human study covers only three of the compared systems.

Figure 9. Soundwich's layered generation: a scene prompt with per-source time windows drives stem formation over one shared video, entity routing binds each stem to its visual source, and the output is a video with separately editable audio stems. Figure 2 of [27]; viewed at pixel level for this issue.

7.2 Keeping the rewards apart

AV-GRPO [28] instead post-trains the joint model with reinforcement learning, and its key idea is credit assignment: rather than treating the paired audio–video sample as one policy output, each phase conditions on a fully realised trajectory of the anchor modality and updates only the other tower, alternating every $n$ steps — so within a group the anchor reward is constant and cancels after centring, leaving the synchronisation comparison controlled. On the LTX-2.3 22B backbone with a difficulty-stratified prompt set it reports audio quality 5.097 → 5.798 and desynchronisation 0.757 → 0.542 over the base, beating the multi-reward baseline on most columns. It also concedes two regressions (video quality 5.816 in the LoRA variant, below the 5.855 base; visual realism 4.395, marginally below 4.399), and its cost-reduction claim — the stated motivation for freezing towers — comes with no wall-clock or memory measurement.

Audio is also a first-class constraint in two papers read elsewhere: [25] keeps the generated audio dense and exact while sparsifying video attention and routes it past refinement unchanged, and [18] preserves the native audio stream while adding control but never evaluates or scores it.

Table 7. The joint audio–video papers in this theme.

Takeaways — Two opposite routes to the same complaint: Soundwich makes a frozen model's single mixed track editable by expanding it into scheduled stems without any training [27], while AV-GRPO post-trains a joint model so that each modality's reward updates only its own tower, with the anchor trajectory held fixed [28]. The first reports bootstrap intervals that include zero for its most quotable gaps; the second concedes two quality regressions while claiming a training-cost reduction it never measures.

8. Evaluating and governing generation: metric validity became a research topic

The month's evaluation and safety papers are unusually self-critical, and three of them make claims that are mostly about how the field measures.

8.1 What the metrics actually measure

Video Generation Models: A Survey of Post-Training and Alignment [29] functions here as an inventory: 117 method entries across supervised fine-tuning, self-training/distillation, preference and reward methods, and inference-time steering, plus 20 training datasets and 28 benchmarks with sizes. Its most useful content for a reader is a warning the authors place on their own tables — the compiled scores are not a leaderboard, because the protocols differ in base model, size, sampling budget and prompt set, and resolution, clip duration and sampling steps are simply not reported for most evaluations. Its substantive reading of the compiled evidence is that smoothness and consistency scores do not track dynamic degree, and that optimising one reward can leave other dimensions flat or worse; it also states plainly that safety post-training specific to video models "still lacks a strong body of work" and that autoregressive alignment is under-represented because the literature is overwhelmingly diffusion-based.

Scaling Video Generation for Reasoning [30] turns that genre of doubt into a controlled experiment: a $2\times2\times2$ Rubik's cube with exact simulator ground truth, nine prescribed turns, and models trained from scratch across seven sizes (19.7M–938.4M). Its central result is a metric-validity finding — validation flow MSE does not order models by capability, and can invert them: at 1.5M training videos a 1B model has lower MSE than a 70M model but 3.67% average frame accuracy against 41.00%. The compute story is equally concrete: at about 0.1 PF-days the 70M model reaches 44.6% post-action frame accuracy while the 1B model is at 0.3%, and only reaches 83.7% at 3.14 PF-days. The third finding is the constructive one: adding a symbolic state-prediction pass lifts a 20M model from 31.1% to 67.3% frame accuracy at 0.0445 PF-days, roughly matching a 550M video-only model at 1.33 PF-days — about 30× the compute. Two honest caveats: the sticker representation is handcrafted and may not generalise, and the power-law compute extrapolations are labelled illustrative, not budgets, with strong fitting-window sensitivity. It also reports a diagnostic worth remembering — models can follow 80% of actions correctly while getting 0.3% of frames exactly right, i.e. turning the right faces of the wrong cube.

Figure 10. Fitted and extrapolated compute–accuracy frontiers: the same measurements support widely different projections depending on which part of the frontier is fitted — 23.55 PF-days from the last third against 422,700 PF-days from the middle third for the average-frame criterion. Figure 7 of [30]; viewed at pixel level for this issue.

Table 8. The evaluation and safety papers in this theme.

8.2 Erasure and new attack surfaces

EraseSAE [31] brings sparse autoencoders to concept erasure in text-to-video diffusion: a partitioned convolutional SAE separates context from concept features, contrastive log-ratio attribution selects concept kernels offline, and a timestep-resolved mask gates a spatially modulated guidance pass at inference. On nudity erasure it reports 2.62 on the generator split for CogVideoX-5B against 2.88 for the strongest baseline, with far better preservation of unrelated content (SSIM 74.99 versus 42.39), and large residual-detection reductions for the five celebrity identities. Three cautions: the two headline margins are computed against different baselines, so they are not one margin; the ablation shows an 18.32-point SSIM contribution from a single loss term, which is a large effect to rest on one component; and the introduction's headline percentage gains could not be reproduced from any table in the experiments section as read — and the paper has no limitations section at all.

BadAction [32] studies a backdoor attack on a channel nobody had attacked: not the prompt and not the pixels, but the sequential action stream of an interactive video model. A fixed four-step motion pattern spliced into the action sequence, paired with a frozen target video, teaches the model to stop responding to further user actions once triggered, while benign sequences behave normally. It reports 91.0% attack success by SSIM-based measurement and 89.6% by ten human judges, against 69.4% and 61.3% for the strongest image-channel baseline, with benign utility close to the clean model (CLIPSIM 82.3 versus 85.1). The defence results are the practical finding: two diffusion backdoor detectors reach F1 66.2 and 29.3 on the action-channel attack, because they inspect image and text channels rather than the action control. The scope limits are real — one backbone, one egocentric dataset, 100 clips, and no resolution, clip length or hardware reported — but the attack surface is genuinely new relative to the literature the paper names.

Cross-cutting. Together with [6]'s audit of verifier contamination, this section suggests that in September the field's weakest measurement links were the ones closest to the loop: the judge inside an agentic harness, the loss curve used to compare model sizes, the compiled leaderboard, and the detector that never sees the action channel.

Takeaways — Measuring generation, not just generating it, became a research topic this month. Validation loss can point the wrong way about capability [30]; a survey of 117 post-training methods warns in its own tables that the compiled scores are not a leaderboard [29]; and new attack surfaces are appearing where defences are not looking — the sequential action channel of interactive models [32] and the auxiliary text of agent harnesses [6]. Safety work specific to video post-training is still thin, which the survey states plainly.

Also noted

Beyond the papers read in full, 220 screened papers were classified into seven themes (single-label; 54 remained unassigned where the primary contribution spanned themes or confidence was low). The counts below are counts within the checked routes for this month, not field shares, and they say nothing about how many papers exist on these topics overall.

Table 9. Themes in the screened corpus and representative papers not read in full (single-label classification over titles and abstracts; counts are within the checked routes only).

Scope and method

Retrieval and screening

The corpus is a merged keyword and semantic retrieval over v1 submission dates 2026-09-01 to 2026-09-30, screened against a written inclusion criterion: contributions to the video-generation model or its generation capability, with domain-transfer work admitted only where the contribution returns to the generation method. 374 candidates were screened and 220 passed; the screened working set is 视频生成 2026-09 · AI 精筛.

Evidence depth and gaps

32 of the 220 were read in full text — method and experiment sections, plus the appendices the section map exposes — and the rest are summarised from title and abstract only. 20 of the 154 exclusions were re-inspected; one borderline exclusion, a benchmark for the procedural correctness of image-to-video generation, was screened out just below the decision threshold and is recorded as a near-miss. Three numbers in this issue come from figures rather than text (LongTake's dynamic degree, Recency Forcing's user study, and the CamPilot camera subtasks) and are marked as such in the underlying notes.

Boundary cases excluded

Domain-flavoured generative work — driving-scene and embodied world models, video try-on, sign-language and dubbing systems — was treated as out of scope except where the contribution is the generative method itself. A few entries carry arXiv identifiers from the October series (2610.xxxxx) because their v1 was submitted on 2026-09-30 and announced in the next batch; the window was applied to v1 submission dates, not to identifiers. Thematic counts come from a single-label classifier over titles and abstracts, with 54 of the 220 left unassigned, so they are approximate and a paper appears under one theme only.

Featured papers at a glance

Table 10. The papers discussed in this issue, the lever each pulls, and the caveat that most limits its headline number.

References

[1] Yingrui Wang, Zeqing Wang, Yeying Jin. StoryEngine: A State-Grounded Agentic Framework for Video Storytelling. arXiv:2609.33627, 2026. https://arxiv.org/abs/2609.33627

[2] Xiangyu Kong, Wenjie Zhou, Fengping Tian, et al. MVAgent: Multi-Agent Video Generation via Consistent Condition Construction and Shot-Level Policy Optimization. arXiv:2609.30609, 2026. https://arxiv.org/abs/2609.30609

[3] Binxu Li, Haoyi Duan, Yuhui Zhang, et al. VideoGen-Agent: Reinforcing Video Generation Agents. arXiv:2609.24997, 2026. https://arxiv.org/abs/2609.24997

[4] Jian Xu. Don't Read the Log: Execution Traces Contaminate Verifiers in Video-Generation Agents. arXiv:2609.28564, 2026. https://arxiv.org/abs/2609.28564

[5] Yang Wu, Stefano Petrangeli, Ishita Dasgupta, Yu Shen. CamPilot: A Multi-Agent Cinematic Assistant for Camera-Controlled Movie Generation. arXiv:2609.10943, 2026. https://arxiv.org/abs/2609.10943

[6] Jiaben Chen, Sixun Dong, Qinhong Zhou, et al. Better Call CineCrew: Consistent Ultra-Long Narrative-to-Film Generation. arXiv:2609.07720, 2026. https://arxiv.org/abs/2609.07720

[7] Yinming Huang, Shuyuan Tu, Xi Yan, et al. Agentic Visual Generation: From Generative Models to Agentic Control. arXiv:2609.06758, 2026. https://arxiv.org/abs/2609.06758

[8] Weiqiang Wang, Zhuokun Chen, Yusheng Dai, et al. Self-Aligned Forcing: Streaming Video Diffusion with Differentiable Noisy History. arXiv:2609.38114, 2026. https://arxiv.org/abs/2609.38114

[9] Tri Cao, Hung Nguyen, Phong Nguyen, Khoi Nguyen. Recency Forcing: Bridging the Long-Horizon Gap in Autoregressive Video Generation. arXiv:2609.19729, 2026. https://arxiv.org/abs/2609.19729

[10] Byoungwoo Park, Jaemoo Choi, Juho Lee, Yongxin Chen. LongTake: Learning to Sustain Dynamics in Long-Horizon Video Generation. arXiv:2609.38562, 2026. https://arxiv.org/abs/2609.38562

[11] Chenjian Gao, Zhihao Hu, Jianqi Ma, et al. Rollout-Marginal Distillation for Long-Horizon Autoregressive Video Generation. arXiv:2609.37925, 2026. https://arxiv.org/abs/2609.37925

[12] Zhuoran Zhao, Shengju Qian, Tongtong Liang, et al. Mask Forcing: Improving Autoregressive Video Diffusion Distillation via Dual-Noise Masking Rollout. arXiv:2609.09123, 2026. https://arxiv.org/abs/2609.09123

[13] Mikhail Dereviannykh, Vikram Voleti, Simon Donne, et al. SemanTok: Predictable Semantic Tokens for Efficient Autoregressive Video Generation. arXiv:2610.00686, 2026. https://arxiv.org/abs/2610.00686

[14] Jack Wei Lun Shi, Kaichen Zhou, Haoyu Chen, et al. Honeycomb: Constant-Size Scene Memory Representation for Video World Models. arXiv:2609.37690, 2026. https://arxiv.org/abs/2609.37690

[15] Rajit Rajpal, Shahbuland Matiana, Liew Wei Pyn, et al. Waypoint-1.5: A Real-Time Video World Model for Consumer Hardware. arXiv:2609.37107, 2026. https://arxiv.org/abs/2609.37107

[16] Zelong Lv, Sicheng Xu, Jianfeng Xiang, et al. OctWorld: Long-Range World-Consistent Video Generation with Octree-Based 3D Mapping. arXiv:2609.03919, 2026. https://arxiv.org/abs/2609.03919

[17] Yifan Huang, Lifan Jiang, Qingyue Hao, et al. WorldWeave: Growing Persistent Geometric Worlds for Video Generation. arXiv:2609.34221, 2026. https://arxiv.org/abs/2609.34221

[18] Danze Chen, Zeqing Wang, Ziyue Lin, Xingyi Yang, Yeying Jin. H3-World: Turning Language Understanding into World Control. arXiv:2609.01560, 2026. https://arxiv.org/abs/2609.01560

[19] Zhangsihao Yang, Mengyi Shan. CoaG: Cylinders on a Grid for Coarse 3D Layout Control in Video Generation. arXiv:2609.24208, 2026. https://arxiv.org/abs/2609.24208

[20] Pokrzywa Baptiste, Nabyl Quignon, Yara Bahram, et al. BEACON: Behavior and Appearance Control for Subject-Specific Video Generation. arXiv:2609.13264, 2026. https://arxiv.org/abs/2609.13264

[21] JJiahang Li, Dingbao Shao, Xinyu Chen, et al. VideoX-Qwen: Data-Centric Instruction-Based Video Editing. arXiv:2609.26015, 2026. https://arxiv.org/abs/2609.26015

[22] Yujia Hu, Jiajun Li, Zihao He, Songhua Liu. Streaming Video Editing with Easy Adaptation (SVEET). arXiv:2609.24788, 2026. https://arxiv.org/abs/2609.24788

[23] Yu Xu, Yuxin Zhang, Xiao Yang, et al. Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE. arXiv:2609.38140, 2026. https://arxiv.org/abs/2609.38140

[24] Yun Dai, Jiarui Wen, Huiping Zhuang, Cen Chen, Ziqian Zeng. PARK: Accurate Block Retrieval for Sparse Attention in Video Diffusion Transformers. arXiv:2609.38978, 2026. https://arxiv.org/abs/2609.38978

[25] Yitong Li, Jincheng Yu, Junsong Chen, et al. Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge. arXiv:2609.35110, 2026. https://arxiv.org/abs/2609.35110

[26] Shuai Yang, Luozhou Wang, Wei Huang, et al. LongLive-Plug: Once-for-All Distillation for Video Generation. arXiv:2609.38154, 2026. https://arxiv.org/abs/2609.38154

[27] Zhuo Ning, AmirHossein Naghi Razlighi, Sagi Polaczek, et al. Soundwich: Video Generation with Layered and Controllable Audio. arXiv:2610.00691, 2026. https://arxiv.org/abs/2610.00691

[28] Zhiyu Xu, Weilong Yan, Yufei Shi, et al. AV-GRPO: Modality-Anchored Decoupling Diffusion Reinforcement Learning for Joint Audio-Video Generation. arXiv:2609.29816, 2026. https://arxiv.org/abs/2609.29816

[29] Chaoyu Li, Xiaoyi Gu, Yogesh Kulkarni, et al. Video Generation Models: A Survey of Post-Training and Alignment. arXiv:2610.00812, 2026. https://arxiv.org/abs/2610.00812

[30] Weihang Guo, Xiaoyu Wu, Yifei Wang, Niloofar Mireshghallah, Lydia E. Kavraki. Scaling Video Generation for Reasoning: At What Cost? arXiv:2609.36599, 2026. https://arxiv.org/abs/2609.36599

[31] Xinghao Wang, Dong Li, Wei Yu, et al. EraseSAE: Surgical Concept Erasure in Text-to-Video Diffusion Models via Sparse Autoencoders. arXiv:2609.03629, 2026. https://arxiv.org/abs/2609.03629

[32] Zhihang Wu, Zhongqi Wang, Jie Zhang, et al. BadAction: Backdoor Attacks on Interactive Video Generation via Action-Guided Triggers. arXiv:2609.39047, 2026. https://arxiv.org/abs/2609.39047