00 / Start
One continuous video.
Scrolling is the animation timeline. No click is required to understand the method.
Full video evidence
Watch the entire video.The same temporal object stays on screen while the method transforms around it.
prompt-conditioned temporal prior
V = log(Repeat(h,P) + ε)
Original logitsA
Steered logitsà = A + gβV
Generated summary
The model keeps the full temporal evidence, while allocating more attention to salient interactions and retaining lower-scored context for coherence.
01 / Background
Hard selection deletes evidence.
Importance becomes a binary keep/drop decision. Bridging context can disappear before the MLLM ever sees it.
02 / Rewind
HAS restores every frame.
Low highlight means less emphasis, not deletion. The full stream remains available.
03 / Highlight prior
Scores become a temporal prior.
H(F,q) → Interp → MinMaxNorm. The curve aligns with the entire video timeline.
04 / Token microscope
A frame opens into visual tokens.
The scalar hₜ is copied over the P visual tokens that originate from that frame.
05 / Steering vector
Move the prior into log-space.
V = log(Repeat(h,P)+ε) creates an additive bias aligned to visual-token logits.
06 / Attention steering
Inject gβV before softmax.
Selected cross-attention heads redistribute mass while the backbone MLLM stays frozen.
07 / Output
Summarize with evidence intact.
The model generates with soft temporal guidance rather than a hard frame bottleneck.