Jiang, Yuqiu, Qiao, Xiaozhen, Chen, Yifan, Zheng, Ye, Sun, Zhe, Li, Xuelong · Nov 24, 2025
arXiv cs.AI: Jiang and coauthors propose adaptive memory for rare interaction detection
Markdown · JSON · Original source
Yuqiu Jiang and coauthors argue that cache-based methods for long-tailed human-object interaction (HOI) detection may have limited ability to capture informative features of underrepresented interactions when they allocate equal memory to every class. Their proposed Adaptive Diversity Cache (ADC) assigns more capacity to underrepresented interactions and selects cache entries using prediction confidence and local feature diversity. Cache retrieval then refines predictions from a frozen detector; the framework is described as training-free and plug-and-play. In experiments on HICO-DET and V-COCO with three base detectors, the authors report consistent improvements. On HICO-DET, they report rare-interaction mean average precision (mAP) gains of 4.02–8.57 percentage points, alongside gains for non-rare interactions. They also report rare-interaction gains of 2.87–7.36 percentage points with online adaptation. The authors present these results as evidence that adaptive test- memory can complement existing HOI detectors, though the reported results concern the evaluated datasets and detectors.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 24, 2025 (2025-11-24).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T20:40:06.927Z. First observed: 2026-11-06T05:11:10.537Z. Retrieved: 2026-11-08T20:38:31.488Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported experiments cover HICO-DET and V-COCO and three base detectors; the source excerpt does not identify those detectors or provide further experimental details.
Open question
Would ADC retain its reported gains on long-tailed HOI detection datasets beyond HICO-DET and V-COCO?
Source-stated
Jiang and coauthors say cache-based methods often allocate uniform memory capacity across classes, limiting their ability to capture features of underrepresented interactions.
existing cache-based methods typically assign uniform memory capacity across classes, limiting their ability to capture informative features of underrepresented interactions
Source-stated
The authors propose ADC, a training-free framework that allocates cache capacity according to interaction frequency and selects entries using confidence and local feature diversity.
we propose Adaptive Diversity Cache (ADC), a training-free, plug-and-play framework that adaptively constructs test-time caches through frequency-aware capacity allocation and confidence-diversity joint sample selection.
Source-stated
The authors report that ADC improved rare-interaction mAP on HICO-DET by 4.02–8.57 percentage points, while also improving non-rare interaction performance.
On HICO-DET, ADC improves mean average precision (mAP) for rare interactions by 4.02--8.57 percentage points, while also improving performance on non-rare interactions.
Source-stated
The authors report rare-interaction gains of 2.87–7.36 percentage points with online adaptation.
Online adaptation further achieves gains of 2.87--7.36 percentage points on rare interactions.
Source and contributor context
Jiang, Yuqiu, Qiao, Xiaozhen, Chen, Yifan, Zheng, Ye, Sun, Zhe, Li, Xuelong · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: b8975f2f-9f9d-4b32-88f1-674b357ce2f5.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T20:40:06.927Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: d227e0149d352da5d9a4ce62d989ab1eb30947cc4e05cbc0a0a5a33db0b0291a. Source text hash: cd998e5c5797236062aa73dcaeb3f18ec9b0083475f1bf4eee39667277ccd5bc.
Revision and source-check history
Version 1 · 2026-11-08T20:40:06.927Z
arXiv cs.AI: Jiang and coauthors propose adaptive memory for rare interaction detection
Yuqiu Jiang and coauthors argue that cache-based methods for long-tailed human-object interaction (HOI) detection may have limited ability to capture informative features of underrepresented interactions when they allocate equal memory to every class. Their proposed Adaptive Diversity Cache (ADC) assigns more capacity to underrepresented interactions and selects cache entries using prediction confidence and local feature diversity. Cache retrieval then refines predictions from a frozen detector; the framework is described as training-free and plug-and-play. In experiments on HICO-DET and V-COCO with three base detectors, the authors report consistent improvements. On HICO-DET, they report rare-interaction mean average precision (mAP) gains of 4.02–8.57 percentage points, alongside gains for non-rare interactions. They also report rare-interaction gains of 2.87–7.36 percentage points with online adaptation. The authors present these results as evidence that adaptive test- memory can complement existing HOI detectors, though the reported results concern the evaluated datasets and detectors.
Original source
cs.AI · Article
Source details
Original source date: Nov 24, 2025 (2025-11-24).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T20:40:06.927Z. First observed: 2026-11-06T05:11:10.537Z. Retrieved: 2026-11-08T20:38:31.488Z.
Limits
The reported experiments cover HICO-DET and V-COCO and three base detectors; the source excerpt does not identify those detectors or provide further experimental details.
Open question
Would ADC retain its reported gains on long-tailed HOI detection datasets beyond HICO-DET and V-COCO?
Source-stated
Jiang and coauthors say cache-based methods often allocate uniform memory capacity across classes, limiting their ability to capture features of underrepresented interactions.
existing cache-based methods typically assign uniform memory capacity across classes, limiting their ability to capture informative features of underrepresented interactions
Source-stated
The authors propose ADC, a training-free framework that allocates cache capacity according to interaction frequency and selects entries using confidence and local feature diversity.
we propose Adaptive Diversity Cache (ADC), a training-free, plug-and-play framework that adaptively constructs test-time caches through frequency-aware capacity allocation and confidence-diversity joint sample selection.
Source-stated
The authors report that ADC improved rare-interaction mAP on HICO-DET by 4.02–8.57 percentage points, while also improving non-rare interaction performance.
On HICO-DET, ADC improves mean average precision (mAP) for rare interactions by 4.02--8.57 percentage points, while also improving performance on non-rare interactions.
Source-stated
The authors report rare-interaction gains of 2.87–7.36 percentage points with online adaptation.
Online adaptation further achieves gains of 2.87--7.36 percentage points on rare interactions.
Source and contributor context
Jiang, Yuqiu, Qiao, Xiaozhen, Chen, Yifan, Zheng, Ye, Sun, Zhe, Li, Xuelong · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: b8975f2f-9f9d-4b32-88f1-674b357ce2f5.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T20:40:06.927Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: d227e0149d352da5d9a4ce62d989ab1eb30947cc4e05cbc0a0a5a33db0b0291a. Source text hash: cd998e5c5797236062aa73dcaeb3f18ec9b0083475f1bf4eee39667277ccd5bc.
uncertain · 2026-11-08T20:40:06.927Z · anthropic/claude-opus-5.5
Publication: b8975f2f-9f9d-4b32-88f1-674b357ce2f5
attribution — pass: The paper is correctly attributed to Yuqiu Jiang and coauthors on arXiv. Claims remain framed as the authors' arguments and reported results, not established facts.
Authors:Yuqiu Jiang, Xiaozhen Qiao, Yifan Chen, Ye Zheng, Zhe Sun, Xuelong Li
summary — uncertain: The reading is faithful but hard for a non-specialist to follow. It never states the source's core premise: rare interactions are hard because they appear sparsely in training data. Central terms such as long-tailed, test-time cache, frozen detector and mAP are not explained in plain words. Smallest repair: add one sentence restating the source's sparse-training-data rationale. Briefly gloss the cache as memory built from test observations and used to refine a fixed, unretrained detector. Otherwise use term tooltips. The question is fair and follows from the stated evaluation scope.
Long-tailed human-object interaction (HOI) detection remains challenging because rare interactions are sparsely represented in training data.
date — pass: The displayed original date, 2025-11-24, matches the v1 submission date. It is not the later v3 revision (5 Nov 2026) or the feed date.
[Submitted on 24 Nov 2025 (v1), last revised 5 Nov 2026 (this version, v3)]
support — pass: All claims and figures match the abstract, including the 4.02–8.57 and 2.87–7.36 point gains, the non-rare improvements and the training-free, plug-and-play framing. The scope caveat and question add no unsupported benefits or costs.
On HICO-DET, ADC improves mean average precision (mAP) for rare interactions by 4.02--8.57 percentage points, while also improving performance on non-rare interactions.
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv