Related-Work Table (Verified References)

May 26, 2026 · View on GitHub

Date: 2026-05-10 Related prose map: literature-map-2026-04-16.md

Canonical table of methods we cite or compare against. The literature map is prose; this file is the machine-friendlier state.

verified means we have read the abstract and at least one headline table in the primary source as of the listed date. Seed-derived or unverified methods do not belong in this table; keep them in private scratch notes until a primary source check is complete.

Benchmark and runtime references used by the manuscript but not listed as comparison methods: TOMATO is ICLR 2025 Poster on OpenReview and arXiv 2410.23266; Qwen2.5-VL is arXiv 2502.13923; Gemma 4 E4B is cited through the Google Hugging Face model card plus the MLX-community 4-bit conversion; MLX-VLM is cited through the project repository.

Columns

columnmeaning
methodpaper name or nickname
traintrain-free or trained
stagereduction stage on the VLM pipeline
signalsignal the method uses for its decision
familyencoder-side-temporal, intra-frame-token, kv-memory, systems, trained-representation, or historical
headline_metricrepresentative result in a short form
portability_blockerwhat prevents this running on our MLX stack
verifiedYYYY-MM-DD of primary-source check
urlprimary URL

Table

methodtrainstagesignalfamilyheadline_metricportability_blockerverifiedurl
CodecSighttrain-freestreaming inference: codec-guided pre-ViT patch pruning + selective LLM KV refreshMV + residuals (α=0 default)systemsup to 3× throughput, up to 87% GPU compute reduction, 0-8% F1 dropNVDEC dependency; no MLX path2026-04-29arXiv 2604.06036
CoPE-VideoLMtrainedlearned codec-native encoderMV + residuals via trained transformertrained-representationTTFT 86% reduction, token 93% reduction, TOMATO Ours-7B 28.3% in its open-source-model tablerequires learned Δ-encoder + alignment2026-04-16arXiv 2602.13191
SimpleStreamtrain-freestreaming-memory baselinerecent-frame recency windowsystemssimple recency baseline matches or beats heavier streaming-memory methods on OVO-Bench and StreamingBenchbenchmark / protocol baseline, not an MLX drop-in method2026-04-22arXiv 2604.02317
FastVtrain-freedecoder-internal layer Kintra-modal attentionintra-frame-token45% FLOPs reduction (LLaVA-1.5-13B)mlx-vlm fused SDPA hides attention scores; fork required2026-04-29arXiv 2403.06764
FastVIDtrain-freepost-encoder + spans temporal groupsdensity clusteringencoder-side-temporal + intra-frame-token90.3% token pruning, FLOPs to 8.3%, 7.1× LLM prefill speedup, 98.0% original accuracy on LLaVA-OneVision-7Btorch-only; hardwired into lmms-eval fork2026-04-29arXiv 2503.11187; OpenReview NeurIPS 2025
EVStrain-free / optional uptrainingpost-encoder video-token pruningtemporally static patch pruning with positional identity preservedencoder-side-temporal + intra-frame-tokenup to 4× TTFT reduction with minimal accuracy loss; optional uptraining improves robustness under aggressive pruningclosest plug-and-play static-patch baseline; MLX C-VISION analog should be tested before claiming typed static-detail wins2026-05-15arXiv 2510.14624
FrameFusiontrain-freepost-encoder, video-nativesimilarity + importanceencoder-side-temporal + intra-frame-token70% vision-token reduction, 1.6–3.6× end-to-endtorch; MLX port TBD2026-04-29CVF ICCV 2025; arXiv 2501.01986
VisionZiptrain-freepost-encoderencoder-dependent attention score: CLS-token attention for CLS encoders, average received attention otherwiseintra-frame-token≥5% SOTA gain over prior art (their setting)torch-only (HF hooks)2026-04-29CVF CVPR 2025; arXiv 2412.04467
EvoPrunetrain-freevisual encoder, early-stage token pruninglayer-wise token similarity, diversity, and attention importanceintra-frame-token2× VideoMME inference speedup with less than 1% performance degradationmodifies visual-encoding internals; no MLX port2026-04-30arXiv 2603.03681
SparseVLMtrain-freedecoder-internal every layercross-modal (text-visual attention)intra-frame-token54% FLOPs reduction, 37% CUDA latency decrease, 97% original accuracy retainedtorch + per-layer HF hooks2026-04-29PMLR ICML 2025
Token Pruning: Are We Solving the Right Problem?analysisevaluation critiquefixed/random coverage, attention-score critique, duplication-vs-importance tradeoffintra-frame-tokenshows many pruning methods underperform naive random token selection; makes random/fixed controls mandatorycritique paper, not a method2026-05-10ACL Findings 2025
When Token Pruning is Worse than Random: Understanding Visual Token Information in VLLMsanalysisdecoder-layer token-pruning critiqueinformation-horizon analysis; structured-vs-random pruning by layer depthintra-frame-tokenargues training-free pruning methods become indistinguishable from random beyond deep-layer information horizons; DivPrune+random keeps 96.9% Qwen2.5-VL-7B performance at 50% pruningarXiv comments report CVPR 2026 acceptance; not a deployable MLX method2026-05-26arXiv 2512.07580
DARTtrain-freetoken pruningduplication-aware token retention instead of importance rankingintra-frame-tokenreports 88.9% visual-token pruning with comparable performance and 1.99× total-time speeduptorch / attention-operator compatibility; MLX port TBD2026-05-10EMNLP 2025; arXiv 2502.11494
VScantrain-freemulti-stage (ViT + LLM mid-layer)intra-modal + cross-modalintra-frame-token2.91× prefill, 10× FLOPs reduction, 95.4% original performance retained (LLaVA-NeXT-7B)torch; modifies both forward passes2026-04-29arXiv 2505.22654; OpenReview TMLR
VLCachetrain-freeencoder-cache + KV-cachecross-request similaritykv-memory1.2–16× TTFT, 2–5% tokens computedSGLang only2026-04-16arXiv 2512.12977
STTMtrain-freepost-ViT (inside LLM layers)spatio-temporal redundancy (directed pairwise matching)intra-frame-token2× speedup at 0.5% acc drop, 50% tokens (LLaVA-Video-7B, 6 benchmarks)torch (GitHub: HYUNJS/STTM); MLX port TBD2026-04-16arXiv 2507.07990
T3Strain-freeinference wrapper (temporal sampling)diverse subsequence packingencoder-side-temporal+3.1% acc, 2.04× TTFT reduction (long-video MLLMs)model-agnostic wrapper; MLX compat likely2026-04-16 (abstract)arXiv 2511.17945
FlashVIDtrain-freepost-ViT: tree-based spatiotemporal mergeattention/diversity selection + tree-based spatiotemporal token mergingencoder-side-temporal + intra-frame-tokenAt 10% visual-token retention, preserves 99.1% of LLaVA-OneVision performance; on the LLaVA-OneVision VideoMME efficiency comparison, reports 6.3× prefill and 2.1× TTFT speedups versus FastVID; separately enables Qwen2.5-VL to process 10× more frames with +8.6% relative performance under matched budgettorch (ICLR 2026 Oral); MLX port possible but non-trivial2026-04-29arXiv 2602.08024; OpenReview ICLR 2026
Déjà VutrainedQKV + FFN reuseMV + attention (Gumbel-Softmax)encoder-side-temporal1.81×/2.64×/2.54× (retrieval/QA/grounding)requires training2026-04-16arXiv 2506.14107
Eventful Transformerstrain-freeper-block token gating at each transformer blockframe-to-frame delta magnitudeencoder-side-temporal2–4× computational savings (ImageNet VID, EPIC-Kitchens)port to MLX not attempted2026-04-29arXiv 2308.13494
RLTtrain-free or fine-tunedpre-model video tokenization / redundancy scoringrepeated patch runs over time, inspired by run-length encodingencoder-side-temporal30% fine-tuning wall-clock reduction while matching baseline performance; no-training inference throughput +35% with 0.1% accuracy drop in the paper settingfull length-encoding/training stack not reproduced; current paper uses RLT-style redundancy as a cheap frozen-VLM scorer2026-05-25arXiv 2411.05222
CoViARtrainedclassical CNN per compressed streamcodec I / MV / residualhistorical4.6× faster than Res3D (UCF-101, HMDB-51)CNN, not a VLM method2026-04-29arXiv 1712.00636
FastVLMtrainedvision encoder architectureFastViTHD hybrid encoder, fewer visual tokens, lower encoding latencytrained-representation3.2× TTFT improvement in LLaVA-1.5 setup; 85× faster TTFT than LLaVA-OneVision at 1152² in the paper settingrequires trained FastVLM/FastViTHD weights, not drop-in pruning2026-04-29CVPR 2025
LLaVA-PruMergetrain-freepost-vision-encoder token pruning/mergingCLS-token attention sparsity + key-similarity clusteringintra-frame-token4× visual-token reduction on LLaVA-1.5 and Video-LLaVA with comparable or better performanceCLIP/LLaVA-specific token plumbing; MLX port TBD2026-04-29ICCV 2025
SparseVILAtrain-freeprefill pruning + decode-time retrievalquery-agnostic pruning + query-aware visual-token retrievalkv-memory + intra-frame-token1.4× image-benchmark speedup; +5.9% average image accuracy; 3.6× prefill and 1.7× decoding speedup on long-context/generation tasksAWQ/CUDA pipeline and retrieval path; no MLX path2026-04-29ICCV 2025
HERMEStrain-freehierarchical KV memory for streaming videocache hierarchy over streaming visual contextkv-memory10× faster TTFT than prior SOTA, up to 68% fewer video tokens than uniform sampling, and superior or comparable accuracy with up to 11.4% gains on streaming datasetsstreaming/KV runtime design; not a drop-in frame-domain sparse-vision method2026-04-30ACL 2026 Main / arXiv 2601.14724
QuickVideotrain-freesystem pipeline: decoding + LLM prefillQuickCodec keyframe-aligned parallel CPU decoding + QuickPrefill KV-cache pruning + CPU/GPU overlapsystemsQuickCodec gives 2–3× decoding speedup; the full pipeline reduces long-video input processing time by about one minuteCPU/GPU pipeline integration and QuickPrefill; not MLX drop-in2026-04-29arXiv 2505.16175; OpenReview TMLR

Query-Aware Visual Routing Priors

These rows are primarily for the follow-on query-aware visual-routing paper, not for the current VLMaxxing/RLT update. They define the novelty boundary: query-aware selection already exists; our future contribution must be operator-level visual evidence planning with full cost accounting. Do not claim that query-aware admission or token routing is new. The defensible angle is that token-reduction and FLOP ratios are insufficient: only tokens admitted into the expensive path can reduce that path's measured cost, and E2E speedup is bounded by the affected-stage share, selection overhead, decode share, and kernel compatibility.

methodtrainstagesignalfamilyheadline_metricportability_blockerverifiedurl
Static or Dynamictrain-freevideo token allocationquery-aware static-keyframe vs delta-frame allocationquery-aware-video-routingup to +5.8% on video-QA benchmarks by selecting static/dynamic token mix without model updatesmethod overlap; must compare conceptually and avoid claiming first static/dynamic query allocation2026-05-10ACL Anthology EMNLP 2025
Active Video Perceptionagentic / training-free at inferenceiterative evidence seekingplanner-observer-reflector loop over what/when/where to observequery-aware-video-routingoutperforms the best agentic method by 5.7% average accuracy while using 18.4% inference time and 12.4% input tokensagentic loop and long-video protocol; closest high-level prior to evidence planning2026-05-10arXiv 2512.05774
QuoTAtrain-freeante-hoc visual token assignmentCoT query decomposition + query-oriented frame importancequery-aware-video-routing+3.2% average performance across six benchmarks at identical visual token budgetuses LVLM/CoT query decomposition; scalar/budget assignment baseline, not typed operator menu2026-05-10arXiv 2503.08689
Q-Frametrain-freeframe selection + resolution adaptationCLIP-like text-image matching + Gumbel-Max; multi-resolution scalingquery-aware-video-routingimproves Video-LLM performance on MLVU, LongVideoBench, and Video-MME while fitting compute budgetsframe-level, not token-level C-VISION; CLIP dependency if reproduced2026-05-10ICCV 2025; arXiv 2506.22139
VideoRoutertrainedlong-video token routingquery-adaptive semantic router plus early-LLM image routerquery-aware-video-routingup to 67.9% token reduction on InternVL-family long-video settingsupervised routers; close comparator for coverage-vs-detail allocation2026-05-19arXiv 2605.05848
Frame-Voyagertrainedframe-combination selectionquery-conditioned learned policy supervised by Video-LLM loss rankingquery-aware-video-routingimproves four Video-QA benchmarks when plugged into two Video-LLMstrained selector + frame-level policy; not training-free2026-05-10arXiv 2410.03226
M-LLM frame selectiontrainedframe selectionspatial and temporal pseudo-labels from M-LLM/LLM promptsquery-aware-video-routingimproves medium/long-context video-QA benchmarks including ActivityNet, NExT-QA, EgoSchema, and LongVideoBenchtrained frame selector and external pseudo-label pipeline2026-05-10CVPR 2025; arXiv 2502.19680
PruneVidtrain-freevideo visual-token pruningtemporal/static redundancy merge + query-irrelevant token pruningquery-aware-video-routing + intra-frame-tokenprunes over 80% tokens while maintaining competitive performance across video benchmarkspost-encoding/LLM relevance plumbing; not a pre-vision C-VISION scorer2026-05-10ACL Findings 2025
QTSplustrainedvisual token selector between encoder and LLMcross-attention scoring + instance-specific retention budget + temporal re-encoderquery-aware-video-routingreports up to 89% vision-stream compression and 28% E2E latency reduction on long videostrained module; arXiv/preprint status2026-05-10arXiv 2511.11910
LongVUtrained / model-integratedlong-video spatiotemporal compressionDINOv2 frame redundancy + text-guided cross-modal query + inter-frame spatial reductionquery-aware-video-routingprocesses long videos within context limits and reports strong VideoMME/MLVU long-video performancemodel-integrated compression; not a drop-in C-VISION scorer2026-05-10ICML 2025 / PMLR; arXiv 2410.17434
FlashVLMtrain-freevisual-token selectiontext-guided cross-modal similarity + intrinsic saliency + diversity floorquery-aware-video-routingreports strong compression/accuracy tradeoffs across image/video benchmarksarXiv preprint; scalar token scorer, not structured evidence plan2026-05-10arXiv 2512.20561
Adaptive-VoCotrained predictorvisual-token compression-rate selectionpatch-token entropy + attention-map variancecomplexity-aware-routingimproves fixed-rate VoCo-LLaMA tradeoffs across multimodal tasksimage-level complexity predictor, not video codec/query routing2026-05-19arXiv 2512.18496
Principles of Visual Tokensanalysis + trained lightweight modelvideo token-value analysisoracle token value, random baseline, Pareto token-value distributionquery-aware-video-routing + intra-frame-tokenargues most token-selection methods do not beat random discarding; proposes LITE for GFLOPs/accuracy tradeoffbenchmark family differs; primarily token-value theory and controls2026-05-10arXiv 2411.13626
Inference Compute-Optimal Video VLMsanalysis / trained sweepscompute-allocation frontiermodel size, frame count, visual tokens per frame under fixed inference computequery-aware-video-routingstudies compute-optimal allocation across frame count and tokens per framefrontier paper, not a drop-in method; required context for compute-allocation claims2026-05-10ACL 2025; arXiv 2505.18855
Gatekeepertrained calibrationmodel cascade deferralconfidence tuning / selective predictionconfidence-routingimproves deferral performance across image, language, and vision-language taskscross-model cascade and fine-tuned confidence; not within-model pruned-to-dense repair2026-05-19OpenReview ICML 2025 Workshop
ViMaRtrained value modelcaption candidate refinementvalue-guided inference + margin-based rewardconfidence-routingreports over 4× speedup over existing value-guided methods while improving caption fidelitycaptioning/refinement method with trained value model; not MCQA sparse-vision repair2026-05-19arXiv 2506.15649
SpecVLMtrained / speculative decodingtoken generation with video-token pruningdraft-token verification plus verifier-guided token pruningspeculative-decodingreports lossless speculative-decoding acceleration for video LLMstoken-level speculative decoding, not admission-level rollback; forbids calling our policy "lossless speculative decoding"2026-05-19EMNLP 2025; arXiv 2509.11815
SpecPrune-VLAtrain-freeVLA visual-token pruningaction-history-aware self-speculative pruning + action-aware controllerspeculative-pruningreports up to 1.57× simulated and 1.70× real-world VLA speedup with negligible success-rate degradationVLA/action domain and token pruning; weakens novelty of the term "self-speculative pruning" but not our admission-stage cost audit2026-05-19arXiv 2509.05614
Bao / bandit optimizertrained policy over existing optimizer hintsdatabase query planning analogyper-query learned hints over a conventional optimizerquery-optimizer-analogyshows a learned bandit can steer an existing query optimizer without replacing itanalogy only; not a VLM method and needs execution feedback/calibration data2026-05-19arXiv 2004.03814; SIGMOD 2021
Conformal risk controlcalibration methodselective decision/risk controlcalibrated risk sets under exchangeability assumptionsconfidence-routingprovides finite-sample risk-control framework for deciding when a policy can actfuture calibration tool only; needs held-out paired cheap-vs-dense labels and explicit risk definition2026-05-19arXiv 2208.02814

Notes on verification

  • "verified 2026-04-16" next to a row means a research subagent or direct read confirmed the headline metric we cite.
  • "verified 2026-04-16 (abstract only)" means we've confirmed the paper exists and the headline claim from the abstract, but NOT every mechanism detail we might summarize. Before citing a method's internal mechanism, re-read the paper's method section.
  • Add a new row only after checking the primary source and recording the date.

Composition matrix (speculative)

us × themstage overlapsignal overlapcomposes?
us × CodecSightpartial (both pre-ViT)different signal (pixel vs MV)policy stacking, not multiplicative
us × CoPE-VideoLMnone (different representation layer)different (we use pixel diff, they use learned Δ-tokens)clean in principle; future work
us × FastVnone (encoder vs decoder)nonemultiplicative in cost model, measured TBD
us × EVSpartial (both exploit temporal patch redundancy)partial (static-patch pruning vs raw-patch run-length redundancy)comparator first; composition likely overlaps
us × FastVIDpartial (we encoder-side, they post-encoder-temporal)partial (both address temporal redundancy)likely interacts; not cleanly multiplicative
us × VisionZipnone (encoder vs post-encoder-intra-frame)noneclean composition
us × SparseVLMnone (encoder vs decoder)noneclean composition
us × VLCachecomplementary (we encoder-side, they KV-reuse)noneclean composition
us × STTMpartial (both address temporal redundancy post-ViT)partial (directed matching vs pixel-diff)worth measuring; STTM handles remaining temporal redundancy we don't eliminate
us × T3Scomplementary (we reduce per-frame cost, T3S packs more subsequences)noneclean; T3S is a wrapping strategy, not a model change

"Composes" claims are first-order hypotheses; only phase 1.32 FastV measurement will turn one row from speculation into evidence.