38 lines
7.8 KiB
JSON
38 lines
7.8 KiB
JSON
{"id": "C01", "kind": "stat", "field": "papers", "expected": 134, "claim": "The archive holds 134 unique arXiv papers after de-duplicating 150 raw query hits across 5 queries."}
|
|
{"id": "C02", "kind": "stat", "field": "corpus_chars", "expected": 228277, "claim": "The concatenated abstract pack CORPUS.txt is 228,277 characters, large enough to force real long-context handling rather than skim-reading."}
|
|
{"id": "C03", "kind": "stat", "field": "date_min", "expected": "2025-11-28", "claim": "The oldest paper in the archive was submitted 2025-11-28."}
|
|
{"id": "C04", "kind": "stat", "field": "date_max", "expected": "2026-10-23", "claim": "The newest paper in the archive was submitted 2026-10-23."}
|
|
{"id": "C05", "kind": "category_count", "category": "cs.CV", "expected": 96, "claim": "96 of 134 papers (71.6%) list cs.CV as primary category: long-context multimodal work is still filed mostly under computer vision."}
|
|
{"id": "C06", "kind": "term_count", "term": "video", "expected": 71, "claim": "71 abstracts mention video: video is the dominant long-context modality in this slice."}
|
|
{"id": "C07", "kind": "term_count", "term": "memory", "expected": 44, "claim": "44 abstracts use the word memory, making memory a first-class design axis rather than a side note."}
|
|
{"id": "C08", "kind": "term_count", "term": "retrieval", "expected": 45, "claim": "45 abstracts use retrieval, slightly more than use memory."}
|
|
{"id": "C09", "kind": "term_count", "term": "benchmark", "expected": 32, "claim": "32 abstracts mention a benchmark, i.e. roughly one in four papers ships an evaluation artifact."}
|
|
{"id": "C10", "kind": "term_count", "term": "long-context", "expected": 42, "claim": "42 abstracts spell the hyphenated term long-context explicitly."}
|
|
{"id": "C11", "kind": "term_count", "term": "document", "expected": 30, "claim": "30 abstracts mention documents, the second major long-context modality after video."}
|
|
{"id": "C12", "kind": "term_count", "term": "agent", "expected": 21, "claim": "21 abstracts mention agents: long-context handling is increasingly framed as an agentic loop over evidence."}
|
|
{"id": "C13", "kind": "term_count", "term": "KV cache", "expected": 7, "claim": "Only 7 abstracts discuss KV cache directly, so the serving-level bottleneck is under-discussed relative to algorithmic ones."}
|
|
{"id": "C14", "kind": "term_count", "term": "RAG", "expected": 14, "claim": "14 abstracts name RAG explicitly."}
|
|
{"id": "C15", "kind": "term_count", "term": "frame selection", "expected": 4, "claim": "Only 4 abstracts use the exact phrase frame selection, despite frame selection being a visible research thread in titles."}
|
|
{"id": "C16", "kind": "cooccur", "terms": ["memory", "retrieval"], "expected": 15, "claim": "15 abstracts mention both memory and retrieval, the write-then-read pattern that defines the streaming branch."}
|
|
{"id": "C17", "kind": "cooccur", "terms": ["video", "memory"], "expected": 28, "claim": "28 abstracts mention both video and memory: video memory is the single densest theme in the archive."}
|
|
{"id": "C18", "kind": "cooccur", "terms": ["token", "compression"], "expected": 11, "claim": "11 abstracts pair token with compression, the budget-cutting branch of the field."}
|
|
{"id": "C19", "kind": "month_count", "month": "2026-09", "expected": 33, "claim": "33 archived papers are from September 2026."}
|
|
{"id": "C20", "kind": "month_count", "month": "2026-10", "expected": 38, "claim": "38 archived papers are from October 2026, the archive's densest month."}
|
|
{"id": "C21", "kind": "stat", "field": "multi_query_papers", "expected": 15, "claim": "15 papers were returned by more than one of the 5 queries, evidence that the query set overlaps on a shared core."}
|
|
{"id": "C22", "kind": "quote", "paper_id": "2610.16118", "quote": "the standard recipe of encoding each frame into visual tokens yields sequences that grow linearly with video length", "claim": "FramePack-U states the cost model plainly: per-frame tokenisation grows linearly with video length."}
|
|
{"id": "C23", "kind": "quote", "paper_id": "2610.16118", "quote": "Existing approaches, such as token reduction and keyframe selection, shorten the token sequence but leave the O(n) scaling intact", "claim": "The paper's critique of prior work is that token reduction and keyframe selection keep the O(n) asymptote."}
|
|
{"id": "C24", "kind": "quote", "paper_id": "2610.05940", "quote": "most existing works tend to progressively lose long context information as length of input stream increases", "claim": "ReMem reports that streaming VLMs progressively lose long-context information as the stream grows."}
|
|
{"id": "C25", "kind": "quote", "paper_id": "2609.32182", "quote": "During query-agnostic writing, KeyRec preserves fine-grained recent observations in a visual cache and organizes historical evidence into a structured event bank.", "claim": "KeyRec separates query-agnostic writing (recent cache plus event bank) from query-time reading."}
|
|
{"id": "C26", "kind": "quote", "paper_id": "2610.30089", "quote": "even the strongest models fail to capture memory both timely and accurately", "claim": "ProactEgoMem finds that even the strongest baselines fail proactive memory on both timeliness and accuracy."}
|
|
{"id": "C27", "kind": "quote", "paper_id": "2610.26910", "quote": "writes each long video into compact, video-specific adaptive parameters for video-free, text-only question answering", "claim": "MIT^3 internalises a video into per-video weights, then answers from text alone."}
|
|
{"id": "C28", "kind": "quote", "paper_id": "2610.20964", "quote": "Retrieval-based methods substantially underperform the oracle setting with ground-truth supporting pages", "claim": "VELD-QA measures a large gap between retrieval-based systems and the oracle-pages setting."}
|
|
{"id": "C29", "kind": "quote", "paper_id": "2610.20964", "quote": "full-document access does not close the gap", "claim": "VELD-QA reports that giving the model the whole document does not close the retrieval gap."}
|
|
{"id": "C30", "kind": "quote", "paper_id": "2605.08888", "quote": "answer accuracy cannot substitute for trajectory-level evaluation", "claim": "DocScope argues answer accuracy is not a substitute for auditing the evidence trajectory."}
|
|
{"id": "C31", "kind": "quote", "paper_id": "2605.08888", "quote": "region grounding remains the weakest trajectory stage", "claim": "DocScope finds region grounding to be the weakest stage across all evaluated models."}
|
|
{"id": "C32", "kind": "quote", "paper_id": "2609.34314", "quote": "frontier judges reach only 75.4% pairwise accuracy", "claim": "PlaylistEval reports frontier video-language judges at 75.4% pairwise accuracy at day scale."}
|
|
{"id": "C33", "kind": "quote", "paper_id": "2609.34314", "quote": "judge accuracy degrades as the playlist set grows", "claim": "PlaylistEval shows judge accuracy degrades as the evidence collection grows."}
|
|
{"id": "C34", "kind": "quote", "paper_id": "2609.19969", "quote": "reduce its global KV cache footprint (always in HBM) to 890 bytes per token", "claim": "DeepSeek-V4.1-Flash reports an 890-bytes-per-token global KV cache footprint via CSA2 plus FP4 caching."}
|
|
{"id": "C35", "kind": "quote", "paper_id": "2609.17248", "quote": "even strong proprietary models achieve below 60% accuracy", "claim": "Video-HolmesV2 reports strong proprietary models below 60% on spatio-temporal audio-visual reasoning."}
|
|
{"id": "C36", "kind": "quote", "paper_id": "2610.31385", "quote": "Our temporal analysis traces expanding priorities from multimodal content understanding to sufficient, high-fidelity evidence acquisition and sustained evidence management.", "claim": "The long-document VQA survey sees priorities shifting from content understanding to evidence acquisition and management."}
|
|
{"id": "C37", "kind": "quote", "paper_id": "2610.23975", "quote": "Long video understanding requires VideoLLMs to answer queries about videos spanning tens of minutes to hours", "claim": "FrameScout frames the task as queries over videos spanning tens of minutes to hours."}
|