Paper search

Can Edge-Deployable Vision-Language Models Identify Species?

Problem space

Evaluate model, prompting, and image-processing configurations for fine-grained species identification on edge hardware while providing verifiable grounds to control risks from open-ended species queries.

Claim

The authors argue that practical 2–8B general-purpose VLMs possess species-classification knowledge above chance but trail smaller expert BioCLIP models, while clean-to-camera-trap degradation more likely reflects image discernibility and domain shift than a VLM-specific fine-grained classification defect.

Contribution

Provides a joint comparison of general-purpose VLMs and BioCLIP with independent resampling, identifying empirical patterns involving model scale, image domain, prompting, and open-ended query hallucination risk.

Contribution type: 实证发现Contribution tags: 经验规律

Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact

Problem space

Evaluate the causal total effect of generative-media exposure on business responses such as sales and conversions at the market-period scale.

Claim

GMMM proposes converting unrecorded GEO and GEM media exposure into expected noticed answer-feature or sponsored-display counts, incorporating them into marketing mix modeling, and comparing business responses across complete treatment sequences.

Contribution

Introduces a method design that adds generative-media measurement and counterfactual treatment sequences to standard marketing mix modeling for evaluating business responses.

Contribution type: 方法设计Contribution tags: 算法设计 · 表示重构 · 测量方法改进 · 组件组合

WSPolypNet: Weakly Supervised Polyp Localization in Colonoscopy Videos

Problem space

Enable semi-automated annotation assistance for temporally consistent presence judgment, frame-level spatial localization, and boundary refinement of polyps in colonoscopy videos.

Claim

WSPolypNet is proposed as a video-level-label-only framework that uses 3D CNN and CAM cues to prompt MedSAM2, temporal propagation, track selection, and frame-wise fallback for polyp localization and segmentation in videos.

Contribution

Provides a video-polyp localization workflow that connects multi-view CAM coarse localization with MedSAM2 point prompting, temporal propagation, track selection, and frame-wise fallback.

Contribution type: 方法设计Contribution tags: 算法设计 · 系统架构 · 组件组合 · 推理策略 · 控制与反馈 · 经验规律

Trust-Aware Adaptive Disclosure for Inference Privacy Preservation in Multi-Agent Networks

Problem space

Trust-aware privacy-preserving communication strategy design for multi-agent systems, using receiver trust to reduce passive inference of private latent goals while maintaining task utility and communication efficiency.

Claim

The paper proposes trust-aware adaptive disclosure that adjusts link-level message masking according to dynamic receiver trust to reduce private-target inference while retaining task-relevant information.

Contribution

The study introduces a method-level trust–disclosure–message-generation structure that adds continuous, link-specific control to privacy-preserving multi-agent communication.

Testing Interchangeability in LLM Agent Teams

Problem space

This research addresses member-swap assessment in two-agent collaboration and aims to determine how exchange affects task outcomes, coordination efficiency, and post-rotation recovery.

Claim

The authors propose that role-matched, similarly experienced LLM teammates are broadly interchangeable for task outcomes but not for coordination efficiency after partner exchange.

Contribution

The study adds a reusable controlled assessment structure for separating general experience from partner-specific coordination residuals across task performance, communication cost, and recovery.