What happens inside a streaming video model when the prompt changes mid-rollout, which cache tricks do the papers use, and which failure modes do they report?
Corrected 2026-10-24 - see C3-B1086-01; the original statements below are kept for history and are superseded where the correction says so. Four papers handle a mid-rollout prompt change by editing the attention cache. LongLive recomputes cached key-value states with the new prompt (KV recache) and reports the trade-off: with no cache the new prompt is followed but transitions are abrupt; with a plain cache visuals are smooth but the new prompt is delayed or ignored. Zing-0.5 refreshes only the text K/V and pins the first frame after the switch; GameCraft-2 uses a prompt recaching mechanism with a sink token; WorldSonus replaces its cross-attention cache at the next 100 ms chunk boundary. MaineCoon gates the next prompt on whether the current spoken line has finished. Reported failure modes: prompt inertia, abrupt transitions, adherence decay across switches, drift beyond about 500 frames and no persistent entity state.
Four papers handle a mid-rollout prompt change by editing the attention cache. LongLive recomputes cached key-value states with the new prompt (KV recache) and reports the trade-off: with no cache the new prompt is followed but transitions are abrupt; with a plain cache visuals are smooth but the new prompt is delayed or ignored. Zing-0.5 refreshes only the text K/V and pins the first frame after the switch; GameCraft-2 uses a prompt recaching mechanism with a sink token; WorldSonus replaces its cross-attention cache at the next 100 ms chunk boundary. MaineCoon gates the next prompt on whether the current spoken line has finished. Reported failure modes: prompt inertia, abrupt transitions, adherence decay across switches, drift beyond about 500 frames and no persistent entity state.
Curator correction 2026-10-24 (discovery audit record C3-B1086-01, curator-reviewed C3R-0693; excerpts re-located by independent curator GET 2026-10-24): Correction. Finding 2 of C3-B074-W15-08 says 'plain cache gives the best consistency (94.77/93.69)' and that recache gives 'similar consistency (94.81/94.04)'. In LongLive Table 4 (arXiv 2509.22622v2, 10-second video, one prompt switch at 5 s) KV recache has the HIGHER background and subject consistency (94.81 and 94.04 versus 94.77 and 93.69 for plain KV cache), so the plain cache does not give the best consistency; the margin is small (0.04 and 0.35 points). The CLIP ordering stated in the record is right: no cache 28.95, recache 27.87, plain cache 25.92. The paper's own caption says 'KV recache achieves the best consistency score and CLIP score', which conflicts with its table on CLIP (no cache 28.95 is higher than recache 27.87); both are preserved here. The record's numbers were accurate when written (v2 read 2026-10-22; v2 is the current arXiv version of the paper at 2026-10-24); the error is the ranking sentence, not the digits. Table values are reported in the evidence rows, not as excerpts.
Curator qualifications: All excerpts located; KV-recache trade-offs for mid-rollout prompt changes.
Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.
Structured fields
{
"citations": [
{
"url": "https://arxiv.org/abs/2509.22622v2",
"note": "candidate evidence excerpts"
},
{
"url": "https://arxiv.org/abs/2609.17909v1",
"note": "candidate evidence excerpts"
},
{
"url": "https://arxiv.org/abs/2610.08760v1",
"note": "candidate evidence excerpts"
},
{
"url": "https://arxiv.org/abs/2511.23429v2",
"note": "candidate evidence excerpts"
},
{
"url": "https://arxiv.org/abs/2606.17800v1",
"note": "candidate evidence excerpts"
},
{
"url": "https://arxiv.org/abs/2605.18601v2",
"note": "candidate evidence excerpts"
}
]
}Evidence claims
- [sourced_fact] Curator correction 2026-10-24 (discovery audit record C3-B1086-01, curator-reviewed C3R-0693; excerpts re-located by independent curator GET 2026-10-24): Correction. Finding 2 of C3-B074-W15-08 says 'plain cache gives the best consistency (94.77/93.69)' and that recache gives 'similar consistency (94.81/94.04)'. In LongLive Table 4 (arXiv 2509.22622v2, 10-second video, one prompt switch at 5 s) KV recache has the HIGHER background and subject consistency (94.81 and 94.04 versus 94.77 and 93.69 for plain KV cache), so the plain cache does not give the best consistency; the margin is small (0.04 and 0.35 points). The CLIP ordering stated in the record is right: no cache 28.95, recache 27.87, plain cache 25.92. The paper's own caption says 'KV recache achieves the best consistency score and CLIP score', which conflicts with its table on CLIP (no cache 28.95 is higher than recache 27.87); both are preserved here. The record's numbers were accurate when written (v2 read 2026-10-22; v2 is the current arXiv version of the paper at 2026-10-24); the error is the ranking sentence, not the digits. Table values are reported in the evidence rows, not as excerpts. · locator: curator correction paragraph in this record body (C3-B1086-01) source
- [sourced_fact] SUPERSEDED IN PART 2026-10-24 by curator correction C3-B1086-01 (see the correction claim above); original kept for history: Four papers handle a mid-rollout prompt change by editing the attention cache. LongLive recomputes cached key-value states with the new prompt (KV recache) and reports the trade-off: with no cache the new prompt is followed but transitions are abrupt; with a plain cache visuals are smooth but the new prompt is delayed or ignored. Zing-0.5 refreshes only the text K/V and pins the first frame after the switch; GameCraft-2 uses a prompt recaching mechanism with a sink token; WorldSonus replaces its cross-attention cache at the next 100 ms chunk boundary. MaineCoon gates the next prompt on whether the current spoken line has finished. Reported failure modes: prompt inertia, abrupt transitions, adherence decay across switches, drift beyond about 500 frames and no persistent entity state. · locator: candidate evidence excerpts source
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | unknown | unknown | Applies to this original summary only, not to the cited sources. |