DeltaS: Reading the Gated Linear Attention State for KV Cache Eviction in Streaming Video
Taeyoun Kwon, Seungjin Kim, Hyeonyu Kim, Moon Hwan Kim
cs.CV · cs.CL · cs.LG
Quick verdict
Proposes gated-delta state drift as a free KV-eviction signal for streaming video; controlled gains are solid, but single backbone and no significance tests.
Summary
DeltaS addresses KV cache eviction in streaming video-language models built on hybrid architectures that interleave gated-delta linear attention with full attention. Because eviction must happen before the question arrives, retention has to be query-agnostic. The core idea is to read the fixed-size recurrent state of the linear-attention layers: gated-delta updates write only the residual between an incoming value and what the state can already retrieve, so the normalized state change over a chunk ('state drift') measures how much new information that chunk contributes. DeltaS retains high-drift chunks inside a fixed token budget, protecting attention-sink and recent-window tokens and using temporal bucketing (N=4) to counteract early-stream bias. On a frozen Qwen3.5-9B backbone with 8,192/16,384-token per-layer budgets, DeltaS outperforms the strongest query-agnostic bounded-memory baseline (StreamMem) by 2.1/1.7 average points across six long-video benchmarks, and by 5.6 points on LVBench at the tighter budget. A controlled signal comparison (Table 2) shows state drift beating position-, attention-, key-, and value-based scores under a shared policy, and the signal costs only 2.07 ms (~1.9% of a 107 ms chunk forward). Main limitations: a single backbone and scale, no variance or significance testing behind sub-2-point margins, and chunk-level scores that cannot discriminate within a chunk.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The normalized change ('state drift') of the gated-delta linear-attention recurrent state over a chunk measures how much new information that chunk brings and works as a query-agnostic, training-free KV retention signal.
- DeltaS beats the strongest query-agnostic bounded-memory baseline (StreamMem) by 2.1 average points at M=8,192 and 1.7 at M=16,384 across six long-video benchmarks, and by 5.6 points on LVBench at M=8,192.
- Under a fixed budget and retention policy (Table 2), state drift outperforms position-, attention-, key-, and value-based signals by 0.7-2.1 points over the best cache-internal signal (VaN).
- The state-drift signal costs 2.07 ms per chunk (1.9% of a 107 ms chunk forward pass), and 6.4 ms total per eviction event - the cheapest among methods that compute a retention signal.
- Temporal bucketing (N=4) reduces early-stream concentration of retained tokens and improves accuracy over unbucketed selection on all three tested benchmarks.
Strengths
Retention signal read from an existing, fixed-size memory at near-zero cost
DeltaS scores chunks from recurrent states already produced by the forward pass (Sec 3.2): 2.07 ms per chunk, 1.9% of the 107 ms reference forward, 6.4 ms total per eviction event vs InfiniPot-V's 15.4 ms and >107 ms for StreamMem/HERMES (Sec 4.5, Fig 8). No proxy queries or extra attention computation.
Unusually rigorous controlled comparison of signals
Table 2 and Appendix A.2 hold M=8,192, four sink tokens, a 1,024-token recent window, 2-frame chunks, and a shared temporal-bucket policy constant across uniform, SnapKV, L2-norm, KeyDiff, TaR, VaN, and HOLA's beta||e||; Table 4 documents every signal's read/head-axis/granularity. This cleanly isolates the signal from the system.
Honest, self-critical evaluation protocol
Large, pressure-correlated gains at tight budgets
Baseline fidelity and verification effort
Weaknesses
No statistical significance or variance reporting
Table 1 margin conflates signal with retention policy
Single backbone, single scale
Strongest baseline reimplemented from its paper
Chunk-level scoring is coarse
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DeltaS targets KV cache eviction for streaming video understanding on hybrid backbones that interleave gated-delta linear attention with full attention (Qwen3.5-9B: 24 Gated DeltaNet + 8 full-attention layers). The premise: the linear layers' recurrent state is fixed at 48 MiB while the full-attention KV cache grows unboundedly (Fig 1), and since streaming questions arrive after eviction decisions, retention must be query-agnostic. The core claim is that the gated-delta update (Eqs 1-2), which writes only the residual between an incoming value and what the decayed state already retrieves, makes the normalized per-chunk state change ('state drift', Eq 3) a measure of a chunk's new information content - usable as a free retention signal. Three headline claims: (i) +2.1/+1.7 average points over StreamMem at M=8,192/16,384 across six benchmarks (Table 1); (ii) state drift beats position-, attention-, key-, and value-based signals under a shared policy (Table 2); (iii) the signal costs 1.9% of a forward pass (Sec 4.5).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DeltaS in one sentence?
DeltaS is a training-free KV cache eviction method for streaming video-language models that decides which video chunks to keep in the full-attention KV cache by measuring how much each chunk changes the fixed-size recurrent state of gated-delta linear attention - a quantity the authors call state drift.
How much does DeltaS improve accuracy and at what memory budget?
On a frozen Qwen3.5-9B hybrid backbone with per-layer full-attention KV budgets of 8,192 or 16,384 tokens (~256/512 MiB, roughly one to two minutes of video), DeltaS beats the strongest query-agnostic bounded-memory baseline (StreamMem) by 2.1 and 1.7 average accuracy points respectively across six long-video benchmarks, and by 5.6 points on LVBench at the 8,192-token budget. DeltaS at half the budget roughly matches StreamMem at the larger one (58.7 vs 58.8 average).
What does DeltaS cost at inference time?
The state-drift score takes 2.07 ms per chunk, about 1.9% of the 107 ms chunk forward pass measured on an A100; the total added cost including selection, eviction and re-indexing is 6.4 ms per eviction event, the lowest among the evaluated scoring methods (InfiniPot-V: 15.4 ms; StreamMem/HERMES exceed the chunk forward time itself; the beta||e|| state baseline costs 206.1 ms).
Is DeltaS code and data available?
The paper links a public code repository (github.com/MaumAI-Company/DeltaS) and evaluates on six public benchmarks (MLVU, Video-MME, Video-MME v2, LongVideoBench, LVBench, EgoSchema 500-item subset) with a public checkpoint (Qwen3.5-9B). The appendix documents layer indices, KV head configuration, budgets, splits, and baseline adaptations in detail, which supports reproduction - though the repository itself was not verified in this review.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-8 are available only via captions and axis labels in the extracted text; curve shapes and exact plotted values (e.g., Fig 5 margins, Fig 7 per-benchmark gains) could not be inspected directly.
- arXiv v1 preprint, apparently not yet peer-reviewed; several 2026-dated references (venues/pages) could not be verified offline.
- The linked code repository's existence and contents were not verified (offline review).
- Appendix B per-task numbers are cited as printed in the extracted text; minor OCR-level errors in dense tables cannot be ruled out.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.