A preprint describes LongVU-TTT, a video-language method that reported higher scores than LLaVA-Video on each of five video-understanding benchmarks: MLVU, LongVideoBench, Video-MME, NExT-QA and LVBench. On a separate evidence-withholding diagnostic, the paper reported that the benefit of carrying state attenuated beyond half the video.
LongVU-TTT inserts a convolutional Test-Time Training (TTT) resampler between the vision encoder and the large language model, or LLM. It recurrently updates grouped-convolutional fast weights before temporal compression. In the paper’s description, those fast weights are the part of the resampler updated from frame to frame.
The paper defines “causal” at the fast-weight update rule: each frame uses only preceding updates. Importance normalization and fixed-budget frame selection occur after the full video has been processed. In the paper’s terminology, the running update is causal, while the later selection step uses the whole clip.
Five benchmarks, with a narrower comparison
In the main setup, LongVU-TTT took 512 input frames and reduced them to 128 frames for the LLM. It reported 71.80 on MLVU, 60.40 on LongVideoBench, 64.43 on Video-MME, 83.50 on NExT-QA and 44.30 on LVBench. The supplied analysis reports no confidence intervals or variability estimates for these benchmark scores.
LLaVA-Video was the primary reference. The authors describe the broader comparison as contextual and report higher performance than that reference on all five benchmarks, with the largest reported differences on LongVideoBench and LVBench. Those wider comparisons involve models with different training data, backbones and frame policies, so the paper does not present them as a universal ranking.
To examine the temporal operator—the component used to process information across time—the study used a fixed 30% subset shared by all variants. It held the base model, insertion point, feature width, frame input, LLM token budget and training setup constant. Under those conditions, the reported TTT-Conv score was up to 2.12 units higher than TTT-MLP on MLVU and up to 3.04 units higher than bidirectional Mamba2.
A test of carried information
Other ablations tested how the model handled state across chunks. The reported benchmark scores placed the cross-chunk state-carry condition above the 1 × 1 and shuffled-grid controls. The paper says that pattern was not explained by adding convolution alone.
Frame-selection ablations showed another score ordering. Among the listed policies, the hybrid selector had the highest reported scores: 71.80 on MLVU and 60.40 on LongVideoBench, compared with 70.63 and 59.72 for uniform sampling.
In a LongVideoBench temporal-subset diagnostic, the state-carried variant scored 56.42, 53.07, 49.84 and 48.66 across the listed gap bins, from the nearest omitted evidence to the farthest. The explicit-evidence upper bound scored 61.35, 58.68, 55.29 and 53.11 over the same bins. The paper reports that the carry benefit attenuated beyond half the video.
The authors interpret this pattern as consistent with the fast-weight state acting as a temporal aggregator rather than reliable long-horizon episodic memory. Explicit evidence remained stronger, and the carry benefit fell with distance in this test. Because the diagnostic covered only a LongVideoBench temporal subset and specified gap bins, it does not answer whether the same behavior would hold across other videos or deployment settings.
The engineering trade-offs
Training was staged. The first stage used LLaVA-CC3M-Pretrain-595K and short-video instruction samples, followed by joint training on 1.5M samples. The joint stage used 32 to 512 frames per video and a global batch size of 128.
System measurements showed separate capacity and throughput results. On a fixed eight-GPU setup, the maximum trainable frame count was 64 without asynchronous CPU activation offloading and 256 with it. Batch-tiled vision-transformer processing was measured at speedups from 1.39 times at eight GPUs to 2.23 times at 64 GPUs, with 87% throughput-scaling efficiency.
Latency separated the methods. At 512 frames on one A100 40GB GPU, reported projector-stage inference latency was 993.15 milliseconds for TTT-Conv, compared with 184.60 milliseconds for TTT-MLP and 108.89 milliseconds for Gated DeltaNet. TTT-Conv was slower than both alternatives but faster than the temporal Transformer at every tested context length.
What the preprint leaves open
These results support a bounded reading. The five benchmark figures are direct reported comparisons; the broader baseline set varies in training data, backbones and frame policies, and the evidence-recovery test is limited to one LongVideoBench temporal subset. The supplied analysis reports no confidence intervals or variability estimates for the benchmark, ablation and diagnostic results, making the size and repeatability of the reported differences difficult to judge from the results alone.
The manuscript is a preprint identified as arXiv:2608.25729v1 and dated 26 August 2026. It says that code and models will be publicly released, but the reviewed text provides no current release link. The work acknowledges support from King Abdullah University of Science and Technology, with KSL Ibex access and JURECA supercomputer access.
Paper data and sources
Original title: LongVU-TTT: Causal Test-Time Training for Visual Resampling in Long Video Understanding
Authors: Mahmoud Ahmed, Sameh Abdulah, Olatunji Ruwase et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text