Autoregressive video diffusion is already the default route for long-video generation: generate chunk by chunk, with causal attention holding the chunks together. But it carries an unstated flaw. Under teacher forcing, adjacent frames are so similar that a model can smear the previous chunk's pixels across the current one and finish denoising, never needing to understand how objects are about to move.

On October 21, a team from Nanjing University, ARC Lab at Tencent IEG, and Zhejiang University published Looking Ahead (arXiv:2610.29989) on arXiv, treating the problem as a training objective in its own right.

Turning future prediction into a training target

The authors' core diagnosis: temporal redundancy in video is strong enough that local continuity becomes a shortcut, letting the model bypass modeling visual evolution entirely. FCP (Future Chunk Prediction) adds an auxiliary objective that, jointly optimized with the teacher-forced denoising loss, predicts temporally separated chunks from intermediate features of the current chunk, rather than only the next adjacent one.

Three design choices stand out. Source representations come only from the current chunk, not the full history. Each temporal offset gets its own lightweight predictor of three DiT blocks, running in parallel rather than chained. And the auxiliary predictors drop text cross-attention entirely, since text prompts can themselves leak cues about the future.

The numbers, under matched settings

The paper's claims rest on controlled comparison. The authors start from official Wan checkpoints and train both a 1.3B and a 5B variant with fully matched configurations and data.

On the 5B model, the teacher-forcing baseline scores 30.4 IQ on image-to-video and 40.6 on video-to-video; with FCP those become 33.8 and 44.4. On the 1.3B model, 27.6 rises to 30.9. The ablation table breaks the contribution down: clean source representations alone reach 28.8, noisy source representations alone 29.9, and the two combined 30.9.

Two sets of negative results matter here. The chained design reaches only 30.1 on the 1.3B model, below the parallel 30.9. On the 5B, both chained variants barely clear the baseline on image-to-video and fail to improve video-to-video at all; the paper reads this as larger models being better at exploiting the information that leaks along the chained path. Adding explicit text conditioning to the auxiliary predictors, meanwhile, drops IQ by 1.1.

Zero inference cost

Every FCP-specific module is discarded at inference, leaving the rollout identical to the baseline with no added overhead. The paper also compares against Next Forcing-style approaches that condition on the full history: at matched training steps the two score comparably, but the current-chunk-only design reaches the same performance with roughly 40% to 50% less training compute.

On WorldModelBench, PAI-Bench-G, and PhyGround, FCP beats teacher forcing across the board, with gains ranging from 0.03 to 2.39, so the improvement is not confined to a single benchmark.

In the model-ranking table, the FCP 5B reaches IQ 33.8, above Wan2.2-14B (32.2) but below Grok Imagine Video (34.8). Those baseline scores come primarily from the benchmark itself rather than being re-run under one identical setup, so the comparison indicates magnitude rather than a conclusion.

So what

On the world-model track, the competition is over who generates longer video, on how many cards, at what resolution. This paper takes a much smaller cut: no architecture change, no added parameters, zero cost at inference, just a change in what the model is asked to predict during training.

Code and weights for both the 1.3B and 5B variants are on GitHub at TencentARC/Looking-Ahead; the authors validated their environment on an NVIDIA H200, and single-GPU inference is supported. The real takeaway is not the 3.4-point IQ gain but the underlying claim, that local continuity gets treated as a shortcut, which holds for any highly redundant sequence data.

Paper: https://arxiv.org/abs/2610.29989
Code: https://github.com/TencentARC/Looking-Ahead