Awesome Proactive Agents

A curated collection of proactive AI research and applications, bringing together papers, agents, models, systems, projects, products, benchmarks, and supporting infrastructure.

If this list is useful, a ⭐ helps others find it.

Companion Guides

• Agents & Applications: personal assistance, coding, collaboration, wearables, robotics, and domain-specific settings.

• Proactive Personal Agents: assistants, research methods, interaction models, and evaluations across desktop, mobile, and physical-world interfaces.

• Supporting Infrastructure: observation, memory, activation, and serving roles.

• Proactive Atlas Website: searchable papers, projects, models, datasets, selected overview figures, and benchmarks under LowEntropyAI.

• Website Development & Deployment: shared content sources, local preview, validation, and automatic GitHub Pages publishing.

• Projects & Products: agent runtimes, ambient assistants, products, and supporting components, with concrete activation mechanisms.

• Streaming Proactivity: visual triggers, evidence-conditioned response timing, full-duplex interaction, model resources, and evaluation routes.

• Research Map: question-driven clusters for quickly locating papers by intervention timing, inference, long-term intent, personalization, evaluation, and safety.

• Benchmark Matrix: side-by-side benchmark comparison by domain, input stream, proactive target, user model, data type, and metrics.

---

Contents

• Scope
• Must Read
• Agents & Applications
• Projects & Products
• Streaming Proactivity
• Supporting Infrastructure
• Papers
  • Foundations, Surveys and Human Factors
  • Proactive Interaction and Planning
  • GUI, Mobile, OS and Coding Agents
  • Multimodal, Wearable and Embodied Agents
  • Benchmarks, Personalization and Optimization
• Benchmarks
• Research Map
• Benchmark Matrix
• Tag Vocabulary
• Contributing

---

Scope

This collection covers work related to proactive AI: anticipating needs, inferring latent intent, taking initiative, and deciding when to ask, suggest, remind, act, or remain silent. It includes foundations and human factors, agent methods, interaction models, implemented systems and products, benchmarks and datasets, and supporting infrastructure.

Application settings include personal assistance, dialogue, information seeking, coding and digital workflows, streaming multimodal interaction, wearables, robotics, collaboration, and domain-specific assistance. These are starting points for browsing; the collection can grow as new proactive capabilities and applications emerge.

Proactivity can appear in an agent runtime (event, schedule, or heartbeat activation), a model policy (response timing, speak / wait decisions, or intent prediction), or a product workflow (background research and contextual delivery). A paper's main contribution must directly study proactive behavior, active clarification, or the evaluation/governance of those decisions. Merely being usable inside a proactive system is not sufficient.

Application, device/modality, decision mechanism, and resource type are separate axes. Personal assistance, streaming interaction, and embodied collaboration are overlapping routes through the collection. Papers retain canonical bibliography entries; projects, products, models, and evaluations are connected through companion guides.

Typical inclusion signals:

• Studies definitions, mechanisms, or human factors of initiative, anticipation, and proactive assistance.
• Builds an agent, model, system, or product that infers needs or decides when and how to intervene.
• Evaluates initiative, need inference, event-conditioned response or silence, or agent-selected clarification and stopping; user utility, consent, and interruption cost must be tied to those decisions.
• Provides a benchmark, dataset, or analysis with a direct proactive decision target, or studies threats specifically affecting proactive assistance.
• A documented component can appear in an infrastructure guide, but this does not automatically qualify its paper for the bibliography.

For each resource, describe whether it studies, implements, evaluates, or supports proactivity, and keep evidence and limitations visible. Generic memory/QA, user simulation, full-duplex turn-taking, compute scheduling, internal exploration, and agent reliability are not sufficient standalone paper-inclusion signals. A title containing "proactive" is not sufficient either. Active clarification after an incomplete request and authorization governance are adjacent routes, not proof of discovering an unsolicited user need.

Resource labels:

• Paper: arXiv, ACL Anthology, DOI, OpenReview, ACM, Springer, or official proceedings page.
• Website: project page, conference page, lab page, or documentation.
• Code / Dataset: GitHub, released code, released benchmark, or released dataset.
• Model: named checkpoint/model card; linked and publicly accessible are tracked separately.
• Product / Docs: official product documentation; product claims are distinct from research evaluation.
• Component: memory, sensing, or serving infrastructure with an explicit role in a proactive system.
• Notes: short English decision card with why the paper matters, proactivity signal, evaluation setup, limitations, and use cases.

---

Must Read

A layered starting path, selected for coverage of different mechanisms and evidence, rather than publication date or a shared leaderboard. Read Tier 1 first, choose an application or mechanism in Tier 2, then use Tier 3 to evaluate it. Tiers are a reading order, not a quality ranking. Products and code are included alongside papers.

Route	Representative resources	Why start here
Tier 1 · Foundations — define proactivity, build a baseline, understand interruption
Human-centered definition	Towards Human-centered Proactive Conversational Agents SIGIR 2024 · conceptual foundation	Intelligence, adaptivity, and civility provide a starting vocabulary for useful, respectful initiative.
Canonical agent baseline	Proactive Agent · Code · Notes ICLR 2025 · method + benchmark	Connect desktop event streams to task anticipation and ProactiveBench evaluation.
User experience and timing	Need Help? · Notes CHI 2025 · human factors	Understand when proactive IDE help is useful and how users retain control.
Tier 2 · Mechanisms and systems — choose the kind of initiative you want to build
Sensory context → assistance	ContextAgent · Code NeurIPS 2025 · context-aware agent	Move from desktop logs to open-world sensory context and tool-mediated help.
Persistent agent activation	OpenClaw · Heartbeat docs Proactivity SDK Runtime projects · see Projects & Products	Compare periodic attention checks with durable goals and model-selected wake cadence.
Product-level initiative	OpenAI dot MineContext Always-on product / ambient desktop project	Compare background research and connected-app context with proactive desktop summaries and tips.
Latent needs and persistent memory	PASK / IntentFlow · Code / project Notes Technical report · need detection and assistance routing	Connect silent / fast / full assistance decisions to ongoing conversations and hierarchical personal memory.
Personal routines across devices	PersonalAlign / HIM-Agent Personal Agents guide Mobile user history · preference and routine memory	Separate completing omitted preferences from anticipating an unrequested routine; personal assistance extends beyond desktop interfaces.
Wearable and physical-world help	Satori ProMemAssist AR guidance · working-memory-aware intervention	Ground assistance in inferred user/task state and weigh the benefit of help against interruption.
Embodied ask / act decisions	PACT PACE Continual robot assistance · action-completion timing	Compare asking versus acting from interaction history with timing physical help from observed task progress.
Visual events → speak / silence	JoyAI-VL-Interaction · Code MOSS-VL-Realtime · Code Streaming interaction models	Study internally controlled response timing and vision-triggered intervention, including JoyAI's delegation decision.
Evidence and memory → timely answer	OneStreamer · Code · Notes StreamReady Streaming memory and answer readiness	Contrast proactive evidence recording with waiting until a standing question has enough evidence.
Full-duplex interaction → agent work	MiniCPM-o 4.5 · Code Gander · Code Simultaneous perception, speech, and asynchronous work	Study speech/listening control, scene-driven comments, and the bridge to a background reasoning agent.
Tier 3 · Evaluation and boundaries — test the route you chose
Real workflows and personalization	ProAgentBench KnowU-Bench · Code	Pair real desktop traces with personalized, consent-aware mobile evaluation.
GUI intent recommendation	PIRA-Bench / PIRF · Project Data Benchmark with a memory-aware recommendation baseline	Evaluate latent future intents, user profiles, interleaved tasks, and rejection of noisy visual context before instruction-following execution.
Long-horizon initiative	π-Bench · Code VibeLifeBench	Evaluate hidden intents and act / ask / stay-silent decisions over evolving personal workflows.
Streaming timing and intent	OmniMMI StreamGaze · Code Streaming evaluation routes	Compare multimodal alert/turn timing with gaze-conditioned intent; inspect proactive subsets separately from general streaming QA.
Personalized physical-world assistance	EgoPro-Bench Vinci2 / EgoServe Egocentric timing, user context, and silence	Connect streaming to personalized assistance while keeping recorded-video evaluation separate from live wearable deployment.
Silence and oversight	Why2Speak · Notes Benchmark Matrix	Audit act-versus-abstain decisions, class imbalance, and whether exposed reasoning changes the policy being inspected.

Agents & Applications

Browse by the assistance setting, then follow the model and evaluation links. Full application map.

Application route	Representative starting points	What makes it proactive
Personal assistance	Proactive Agent, PersonalAlign, dot, GAIA, MineContext	Infer needs or routines and initiate useful help using ongoing personal context.
Coding and knowledge work	Need Help?, CodingGenie, Ask or Assume?	Choose when to offer help or request missing requirements.
Meetings and collaboration	InsightToast, Speak for Me, DocuTeam	Surface context or contribute when conversation or artifact state warrants it.
Wearable assistance	Satori, ProMemAssist, WatchGuardian, Vinci2	Use physical-world observations, user state, and memory to select an intervention.
Embodied and robot assistance	PACE, PACT, ProRobo	Anticipate physical help and decide whether to ask or act.
Domain-specific assistance	See, Infer, Intervene; AI Assistants Overassist; Proact-VL	Choose appropriate initiative under task-specific benefits and interruption costs.

Projects & Products

Implementation resources are grouped by where initiative comes from. Full entries, release boundaries, and official sources: PROJECTS.md.

Route	Representative projects	Activation mechanism
Persistent runtimes and SDKs	OpenClaw, Hermes Agent, Letta Code, Proactivity SDK	Heartbeat, scheduled/event activation, or durable goal-driven wake cadence.
Ambient and personal assistants	MineContext, GAIA, memUBot	Desktop context delivery, service events, and recurring personal workflows.
Research implementations	Proactive Agent, ContextAgent	Learned task anticipation and sensory-context-driven assistance.
Product reference	OpenAI dot	Ongoing responsibility, background research, proactive app-context review, and scheduled work.
Supporting components	memU, Screenpipe	Memory and observation layers; intervention policy belongs to the integrated assistant.

Streaming Proactivity

Streaming proactivity has its own module because it crosses models, inference frameworks, datasets, and benchmarks. Paper records also remain in the bibliography. Full resource matrix: STREAMING.md.

Capability route	Representative work	Proactive decision
Scene-triggered interaction	JoyAI-VL-Interaction, MOSS-VL-Realtime, Proact-VL	Speak, stay silent, or intervene as visual events unfold.
Evidence-conditioned response and memory	OneStreamer, StreamReady, StreamOV, Response-G1, Eyes Wide Open	Wait for sufficient evidence, retain useful history, then respond. Often conditioned on a standing query.
Full-duplex interaction and delegation	MiniCPM-o 4.5, Gander	Coordinate perception and speech; Gander also connects interaction to asynchronous agent work.
Intent, timing, and silence evaluation	OmniMMI, StreamGaze, ESTP-Bench, SOVBench, ProReady-QA	Measure the proactive task/subset and its timing policy, alongside content quality.

---

Supporting Infrastructure

Infrastructure guide: observation and context capture → persistent memory → activation and goal continuity → inference interfaces. These support proactive systems; their own contribution is recorded separately from a user-facing intervention policy.

Papers

The complete bibliography retains five research areas for archival browsing. Use Applications, Personal Agents, Streaming, and the Research Map as overlapping reading routes. A work can appear in several routes while keeping one bibliographic record and one evidence card.

Foundations, Surveys and Human Factors

Date	Title	Venue / Source	Tags	Resources
2024-04	Towards Human-centered Proactive Conversational Agents	SIGIR 2024	Definition · Human Factors · Dialogue
2024-10	Redefining Proactivity for Information Seeking Dialogue	SICON 2024	Definition · Dialogue · Intent Inference
2025-01	When AI-Based Agents Are Proactive: Implications for Competence and System Satisfaction in Human-AI Collaboration	BISE 2026	Human Factors · Intervention Timing · Trust
2025-02	Assistance or Disruption? Exploring and Evaluating the Design and Trade-offs of Proactive AI Programming Support	CHI 2025	Human Factors · Intervention Timing · IDE
2025-03	Proactive Conversational AI: A Comprehensive Survey of Advancements and Opportunities	ACM TOIS 2025	Survey · Definition · Dialogue
2025-08	When not to help: planning for lasting human-AI collaboration	arXiv 2508	Human Factors · Intervention Timing · Trust
2025-10	Check Yourself Before You Wreck Yourself: Selectively Quitting Improves LLM Agent Safety	RegML 2025 Workshop	Safety & Consent · Tool Use · Intervention Timing
2026-01	Developer Interaction Patterns with Proactive AI: A Five-Day Field Study	IUI 2026	Human Factors · Real-world Data · IDE
2026-02	From Fragmentation to Integration: Exploring the Design Space of AI Agents for Human-as-the-Unit Privacy Management	CHI 2026	Safety & Consent · Privacy · Human Factors
2026-02	Exploring The Impact of Proactive Generative AI Agent Roles in Time-Sensitive Collaborative Problem-Solving Tasks	CHI 2026	Human Factors · Collaboration · Intervention Timing
2026-06	Designing for Doubt: The Case for Informed Abstention in Autonomous Agents	AIES 2026	Safety & Consent · Benchmark · Tool Use
2026-07	AI Assistants Overassist	arXiv 2607	Human Factors · Intervention Timing · Education
2026-08	VeriForge: Mitigating Latent Knowledge Gaps in Narrative Drafting via Mixed-Initiative Scaffolding	UIST 2026	Human Factors · Intervention Timing · GUI
2026-08	AI Watchdog: Agent Interfaces for Detecting and Defending Against Manipulative Dark Patterns in AI Conversations	arXiv 2608	Human Factors · Safety & Consent · Intervention Timing
2026-09	Designing Proactive Thought Partners for Writing	arXiv 2609	Human Factors · Intervention Timing · Personalization
2026-09	Proactive Service Agents: A Unified Decision Framework, Methods, and Evaluation	arXiv 2609	Survey · Definition · Safety & Consent
2026-09	When Intelligence Becomes Agency: A Theory of Governed, Proactive Agency for Symbiotic AI Systems	arXiv 2609	Definition · Safety & Consent · Intervention Timing
2026-09	Oops, Not Now: PEARL, a RAG-Based Support Agent for Gameplay and What Players Want from AI Help	IEEE CoG 2026	Human Factors · Intervention Timing · Education
2026-09	Breaking Up is Hard to Do: AI Companions that Won't Let Their Users Go	arXiv 2609	Human Factors · Safety & Consent · Dialogue
2026-09	Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness	arXiv 2609	Human Factors · Safety & Consent · Intervention Timing
2026-09	Working with Agentic 'Teammates': When a New Organizational Actor Collides with the Human Ecosystem of Work	arXiv 2609	Human Factors · Collaboration · Real-world Data
2026-09	Foundations of Proactive Agents: Principles, Technical Layers, and Proactivity-Gym	arXiv 2609	Definition · Human Factors · Benchmark
2026-09	Referential Uncertainty in Human--AI Collaboration	arXiv 2609	Human Factors · Clarification · Collaboration
2026-10	Vision: "What Don't I Know About User's Preferences?" How Uncertainty Estimation Improves Proactive Assistant Alignment	UbiComp 2026 Companion · PEACE workshop (arXiv record)	Definition · Personalization · Human Factors	DOI

Proactive Interaction and Planning

Date	Title	Venue / Source	Tags	Resources
2024-03	ProMISe: A Proactive Multi-turn Dialogue Dataset for Information-seeking Intent Resolution	Findings of EACL 2024	Clarification · Dialogue · Benchmark
2024-03	Asking the Right Question at the Right Time: Human and Model Uncertainty Guidance to Ask Clarification Questions	EACL 2024	Clarification · Dialogue · Intervention Timing
2024-06	Ask-before-Plan: Proactive Language Agents for Real-World Planning	Findings of EMNLP 2024	Clarification · Planning · Intent Inference
2024-10	Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance	ICLR 2025	Intent Inference · Benchmark · Desktop
2024-11	PCQPR: Proactive Conversational Question Planning with Reflection	EMNLP 2024	Dialogue · Planning · Intent Inference
2025-01	Proactive Conversational Agents with Inner Thoughts	CHI 2025	Dialogue · Intent Inference · Intervention Timing
2025-01	ProTOD: Proactive Task-oriented Dialogue System Based on LLMs	COLING 2025	Dialogue · Planning · Tool Use
2025-07	Tunable LLM-based Proactive Recommendation Agent	ACL 2025	Recommendation · Personalization · Intent Inference
2025-09	PRINCIPLES: Synthetic Strategy Memory for Proactive Dialogue Agents	Findings of EMNLP 2025	Dialogue · Memory · Simulation
2025-10	ProMediate: A Simulation Testbed for Evaluating Proactive Mediation in Multi-Party Negotiation	Findings of ACL 2026	Dialogue · Collaboration · Benchmark
2026-01	Proactivity-driven Personalized Agents for Advancing Human Learning through Engagement, Reflection, and Self-Efficacy	ACM CHIIR 2026 Workshop	Personalization · Intent Inference · Education
2026-01	Long-term Task-oriented Agent: Proactive Long-term Intent Maintenance in Dynamic Environments	arXiv 2601	Long-horizon · Intent Inference · Benchmark
2026-01	Value of Information: A Framework for Human–Agent Communication	ACL 2026	Clarification · Intervention Timing · Safety & Consent
2026-03	PsyProbe: Proactive and Interpretable Dialogue through User State Modeling for Exploratory Counseling	Findings of EACL 2026	Dialogue · Intent Inference · Human Factors
2026-05	Anticipate and Learn: Unleashing Idle-Time Compute in Proactive Agents	arXiv 2605	Intent Inference · Memory · Benchmark
2026-07	PROPER Agents: Proactivity Driven Personalized Agents for Advancing Knowledge Gap Navigation	Findings of ACL 2026	Personalization · Intent Inference · Dialogue
2026-07	ProACT: Towards Breakdown-Aware Proactive Agent in Multi-User Collaboration	arXiv 2607	Collaboration · Intervention Timing · Benchmark
2026-07	Reasoning While Asking: Transforming Reasoning Large Language Models from Passive Solvers to Proactive Inquirers	ACL 2026	Clarification · Dialogue · Optimization
2026-07	ProMed: Shapley Information Gain Guided Reinforcement Learning for Proactive Medical LLMs	ACL 2026	Clarification · Dialogue · Optimization
2026-07	"Excuse me, may I say something..." CoLabScience, A Proactive AI Assistant for Biomedical Discovery and LLM-Expert Collaborations	ACL 2026	Collaboration · Intervention Timing · Benchmark
2026-07	Let LLM Tutors Ask First: Proactive LLM-Based Tutoring at Scale in a 1,500-Student Online Classroom	ACL 2026 Industry Track	Education · Intent Inference · Real-world Data
2026-07	When and What to Ask: AskBench and Rubric-Guided RLVR for LLM Clarification	Findings of ACL 2026	Clarification · Benchmark · Optimization
2026-07	ProActor: Timing-Aware Reinforcement Learning for Proactive Task Scheduling Agents	ACL 2026	Intervention Timing · Planning · Optimization
2026-08	A Self-Triggered Agentic Push Recommendation System	RecSys 2026	Recommendation · Intervention Timing · Real-world Data
2026-08	CLAIM: Leading Open-domain Active Clarification of Large Language Models with Uncertainty Measurement	arXiv 2608	Clarification · Dialogue · Optimization
2026-08	JarvisBench: Always-on Intelligence Between Humans and Agents	arXiv 2608	Intervention Timing · Collaboration · Benchmark
2026-08	Bridging Search and CRM: Productionizing AI Product Research Agents for Customer Re-Engagement	arXiv 2608	Recommendation · Personalization · Real-world Data
2026-08	Why2Speak: Faithful Reasoning for Abstaining Action Policies	arXiv 2608	Intervention Timing · Collaboration · Safety & Consent
2026-08	Clarify User Expertise: Towards Proactive Conversational Agents Tailoring Responses to User Proficiency	Findings of EMNLP 2026	Clarification · Personalization · Dialogue
2026-08	Beyond Information Seeking: Severity-Aware Question Supervision for Proactive Medical Dialogue	arXiv 2608 · ICASSP 2027 submission	Clarification · Dialogue · Optimization
2026-08	From Passive Response to Proactive Correction: Enhancing LLM Robustness Against Input Fact Perturbations	EMNLP 2026	Intent Inference · Safety & Consent · Dialogue
2026-09	Speak for Me: Giving LLMs the Situational Awareness to Participate in a Meeting	EMNLP 2026	Collaboration · Intervention Timing · Dialogue
2026-09	Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization	arXiv 2609	Clarification · Planning · Benchmark
2026-09	CC-Mediation: Evaluating Large Language Models for Cross-Cultural Conflict Mediation	arXiv 2609	Dialogue · Intervention Timing · Benchmark
2026-09	IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications	arXiv 2609	Clarification · Benchmark · Tool Use
2026-09	ProMediConv: Benchmarking Proactive Conversational Agents in Legal Dispute Mediation	Findings of EMNLP 2026	Dialogue · Collaboration · Benchmark
2026-09	Asking for What Was Never Requested: Horizontal and Vertical Proactivity in Agents	arXiv 2609	Definition · Intent Inference · Optimization
2026-09	When and How Should an Agent Clarify? CIGAsk: Teaching LLMs to Clarify via Counterfactual Information Gain	Findings of EMNLP 2026	Clarification · Optimization · Dialogue
2026-09	DocuTeam: Mixed-Initiative Multi-Agent Discussions around Evolving Documents	arXiv 2609	Collaboration · Intervention Timing · GUI
2026-09	Clarify the User or Verify the World? Uncertainty Routing for Proactive Agents	arXiv 2609	Clarification · Tool Use · Planning
2026-09	When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents	arXiv 2609	Intent Inference · Dialogue · Benchmark
2026-09	Proactive Dialogue Policy Optimization via Cognitive-State Transition	arXiv 2609	Dialogue · Optimization · Personalization
2026-09	Rational Clarification by Assistive Agents via Value-of-Information Reasoning	arXiv 2609 · Under review	Clarification · Planning · Human Factors
2026-09	Beyond Oracle Communication: Benchmarking Interactive Intent Alignment Under Miscommunication and Evolving User Intent	arXiv 2609	Clarification · Intent Inference · Benchmark
2026-10	Knowing When to Yield: Grounded Arbitration of User Corrections in Text-Based Embodied Agents	arXiv 2610	Clarification · Planning · Embodied
2026-10	Beyond Correctness: Resolving Underspecification in Agentic Text-to-SQL	arXiv 2610	Clarification · Planning · Tool Use
2026-10	Ask, Relax, or Act? Evaluating Actionable Indeterminacy in LLM Preference Reasoning	arXiv 2610	Clarification · Planning · Benchmark
2026-10	Learning to Clarify Underspecified Intents Under Limited Interaction	arXiv 2610	Clarification · Optimization · Human Factors
2026-10	Proactive AI: From Turns to Replannable Dialogue Timelines	arXiv 2610	Definition · Runtime Activation · Planning
2026-10	CLEAT: Learning Value-Guided Clarification for Interactive Agents	arXiv 2610	Clarification · Optimization · Planning
2026-10	Let Me Clarify: Interactive Modeling of Soft Attributes	arXiv 2610	Clarification · Human Factors · Personalization
2026-10	Stop Training Against Soft Users: Mentalized Adversarial Self-Play for Proactive Dialogue	EMNLP 2026 Main (per arXiv record)	Dialogue · Planning · Simulation	Code
2026-10	When Reasoning Is Not Enough: Proactive Inquiry for Cross-Modal Safety in Multimodal Large Language Models	arXiv 2610	Clarification · Safety & Consent · Multimodal / Wearable
2026-10	ProactGen: Learning Proactive Clarification for Video Generation	arXiv 2610	Clarification · Optimization · Multimodal / Wearable

GUI, Mobile, OS and Coding Agents

Date	Title	Venue / Source	Tags	Resources
2024-10	Need Help? Designing Proactive AI Assistants for Programming	CHI 2025	IDE · Intervention Timing · Human Factors
2025-03	CodingGenie: A Proactive LLM-Powered Programming Assistant	FSE 2025 Companion	IDE · Intent Inference · Tool Use
2025-07	FingerTip 20K: A Benchmark for Proactive and Personalized Mobile LLM Agents	ICLR 2026	Mobile · Personalization · Benchmark
2025-07	ProactiveVA: Proactive Visual Analytics with LLM-Based UI Agent	IEEE VIS 2025 / TVCG 2026	GUI · Intervention Timing · Human Factors
2025-08	AppAgent-Pro: A Proactive GUI Agent System for Multidomain Information Integration and User Assistance	CIKM 2025	GUI · Intent Inference · Tool Use
2025-09	VeriOS: Query-Driven Proactive Human-Agent-GUI Interaction for Trustworthy OS Agents	arXiv 2509	OS · Safety & Consent · Clarification
2026-02	ProAgentBench: Evaluating LLM Agents for Proactive Assistance with Real-World Data	arXiv 2602	Real-world Data · Intervention Timing · Benchmark
2026-02	ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence on Mobile Devices	CVPR 2026	Mobile · Intent Inference · Benchmark
2026-03	PIRA-Bench: A Transition from Reactive GUI Agents to GUI-based Proactive Intent Recommendation Agents	arXiv 2603	GUI · Intent Inference · Benchmark
2026-03	GUIDE: A Benchmark for Understanding and Assisting Users in Open-Ended GUI Tasks	CVPR 2026	GUI · Intent Inference · Benchmark
2026-03	Ask or Assume? Uncertainty-Aware Clarification-Seeking in Coding Agents	EMNLP 2026	IDE · Clarification · Intervention Timing
2026-04	Ambig-SWE: Interactive Agents to Overcome Underspecificity in Software Engineering	ICLR 2026	IDE · Clarification · Benchmark
2026-04	Help Without Being Asked: A Deployed Proactive Agent System for On-Call Support with Continuous Self-Improvement	arXiv 2604	Real-world Data · Intervention Timing · Skill Learning
2026-04	Proactive Agent Research Environment: Simulating Active Users to Evaluate Proactive Assistants	arXiv 2604	Simulation · Intervention Timing · Benchmark
2026-04	KnowU-Bench: Towards Interactive, Proactive, and Personalized Mobile Agent Evaluation	arXiv 2604	Mobile · Personalization · Safety & Consent
2026-05	An Empirical Study of Proactive Coding Assistants in Real-World Software Development	arXiv 2605	IDE · Real-world Data · Benchmark
2026-05	ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents	arXiv 2605	GUI · Tool Use · Optimization
2026-04	From Reactive to Proactive: Assessing the Proactivity of Voice Agents via ProVoice-Bench	Interspeech 2026	Multimodal / Wearable · Intervention Timing · Benchmark
2026-06	Perceive Before Reasoning: A Pre-Reasoning Perception Framework for Efficient and Reliable Proactive Mobile Agents	arXiv 2606	Mobile · Intervention Timing · Tool Use
2026-07	InquireMobile: Teaching VLM-based Mobile Agent to Request Human Assistance via Reinforcement Fine-Tuning	ACL 2026	Mobile · Safety & Consent · Clarification
2026-07	PersonalAlign: Hierarchical Implicit Intent Alignment for Personalized GUI Agent with Long-Term User-Centric Records	ACL 2026	GUI · Personalization · Intent Inference
2026-08	Active-SWE: Benchmarking Coding Agents for Proactive Bug Fixing without Issue Reports	arXiv 2608	IDE · Benchmark · Tool Use
2026-08	ITO: Real-time Browser Tab Orchestration Through Intent Detection	UIST 2026	GUI · Intent Inference · Human Factors
2026-08	Act2Intention: A Benchmark For Developing Active Mobile Agents Through Inferring User Intention from GUI Actions	arXiv 2608	Mobile · Intent Inference · Benchmark
2026-08	On the Prospects of Dynamic LLM Conversations in Software Development	arXiv 2608	IDE · Human Factors · Intervention Timing
2026-08	Towards Cognitive Process-Aware Proactive Writing Support	arXiv 2608	GUI · Intent Inference · Human Factors
2026-08	InsightToast: Proactive Information Retrieval & Glanceable Visualization in the Side Channel of Data-Rich Meetings	UIST 2026	GUI · Collaboration · Intervention Timing
2026-09	Beyond Instruction-Driven Editing: Source-Grounded Problem Discovery with User-Governed Repair for Scientific Posters	arXiv 2609	GUI · Human Factors · Safety & Consent
2026-09	Do GUI Agents Know When Not to Act? Enabling Conflict-Aware Termination for Multimodal GUI Agents	arXiv 2609	GUI · Safety & Consent · Intervention Timing
2026-09	Relevance Does Not Imply Applicability: Experience Activation for Personal GUI Agents	arXiv 2609	GUI · Personalization · Intervention Timing
2026-09	PrecogUI: Proactive GUI Agents via Pre-cognitive Simulation and Experience Retrieval	arXiv 2609	GUI · Planning · Tool Use
2026-09	Learning When and How to Intervene: A Hindsight-Distilled Sentinel for Coding Agents	arXiv 2609	IDE · Intervention Timing · Safety & Consent
2026-10	OverAct: Measuring and Mitigating Proactive Over-Authorization in LLM Tool-Calling Agents	arXiv 2610	Tool Use · Safety & Consent · Benchmark
2026-10	Who Is Your Agent Serving? Provider-Side Indirect Prompt Injection in Proactive Agents	arXiv 2610	Safety & Consent · Privacy · Benchmark
2026-10	Don't Guess, Just Ask: Evaluating Interactive Requirement Clarification for Repository-Level Software Development	arXiv 2610	Clarification · IDE · Benchmark

Multimodal, Wearable and Embodied Agents

Date	Title	Venue / Source	Tags	Resources
2024-09	AssistantX: An LLM-Powered Proactive Assistant in Collaborative Human-Populated Environments	IROS 2025	Embodied · Collaboration · Planning
2024-10	Satori: Towards Proactive AR Assistant with Belief-Desire-Intention User Modeling	CHI 2025	Multimodal / Wearable · Intent Inference · Human Factors
2025-01	YETI: Proactive Interventions by Multimodal AI Agents in Augmented Reality Tasks	arXiv 2501	Multimodal / Wearable · Intervention Timing · Human Factors
2025-01	AiGet: Transforming Everyday Moments into Hidden Knowledge Discovery with AI Assistance on Smart Glasses	CHI 2025	Multimodal / Wearable · Intent Inference · Personalization
2025-02	WatchGuardian: Enabling User-Defined Personalized Just-in-Time Intervention on Smartwatch	arXiv 2502	Multimodal / Wearable · Personalization · Intervention Timing
2025-02	Mirai: A Wearable Proactive AI Inner-Voice for Contextual Nudging	CHI EA 2025	Multimodal / Wearable · Intervention Timing · Human Factors
2025-03	OmniMMI: A Comprehensive Multi-modal Interaction Benchmark in Streaming Video Contexts	CVPR 2025	Multimodal / Wearable · Intervention Timing · Benchmark
2025-05	ContextAgent: Context-Aware Proactive LLM Agents with Open-World Sensory Perceptions	NeurIPS 2025	Multimodal / Wearable · Personalization · Tool Use
2025-05	PACE: Proactive Assistance in Human-Robot Collaboration Through Action-Completion Estimation	ICRA 2025	Embodied · Intervention Timing · Optimization
2025-06	Proactive Assistant Dialogue Generation from Streaming Egocentric Videos	EMNLP 2025	Multimodal / Wearable · Dialogue · Intervention Timing
2025-07	ProMemAssist: Exploring Timely Proactive Assistance Through Working Memory Modeling in Multi-Modal Wearable Devices	UIST 2025	Multimodal / Wearable · Intervention Timing · Human Factors
2025-10	Eyes Wide Open: Ego Proactive Video-LLM for Streaming Video	NeurIPS 2025	Streaming · Intervention Timing · Multimodal / Wearable	Paper Code Notes
2025-12	ProAgent: Harnessing On-Demand Sensory Contexts for Proactive LLM Agent Systems	arXiv 2512	Multimodal / Wearable · Sensing · Intervention Timing
2025-12	StreamGaze: Gaze-Guided Temporal Reasoning and Proactive Understanding in Streaming Videos	arXiv 2512	Streaming · Intervention Timing · Benchmark	Paper Code Notes
2026-01	ROMA: Real-time Omni-Multimodal Assistant with Interactive Streaming Understanding	Findings of ACL 2026	Multimodal / Wearable · Intervention Timing · Optimization
2026-01	Clarify or Answer: Reinforcement Learning for Agentic VQA with Context Under-specification	arXiv 2601	Multimodal / Wearable · Clarification · Optimization
2026-03	ProactiveBench: Benchmarking Proactiveness in Multimodal Large Language Models	ICLR 2026	Multimodal / Wearable · Intervention Timing · Benchmark
2026-03	StreamReady: Learning What to Answer and When in Long Streaming Videos	CVPR 2026	Multimodal / Wearable · Intervention Timing · Benchmark
2026-03	Em-Garde: A Propose-Match Framework for Proactive Streaming Video Understanding	ECCV 2026	Multimodal / Wearable · Intervention Timing · Optimization
2026-03	Proact-VL: A Proactive VideoLLM for Real-Time AI Companions	ICML 2026	Streaming · Intervention Timing · Multimodal / Wearable	Paper Code Notes
2026-04	AURA: Always-On Understanding and Real-Time Assistance via Video Streams	arXiv 2604	Multimodal / Wearable · Intervention Timing · Optimization
2026-04	MiniCPM-o 4.5: Towards Real-Time Full-Duplex Omni-Modal Interaction	arXiv 2604	Streaming · Intervention Timing · Full-duplex	Paper Code Notes
2026-05	IPIBench: Evaluating Interactive Proactive Intelligence of MLLMs under Continuous Streams	arXiv 2605	Multimodal / Wearable · Intervention Timing · Benchmark
2026-05	OmniPro: A Comprehensive Benchmark for Omni-Proactive Streaming Video Understanding	arXiv 2605	Multimodal / Wearable · Intervention Timing · Benchmark
2026-05	EgoPro-Bench: Benchmarking Personalized Proactive Interaction in Egocentric Video Streams	arXiv 2605	Multimodal / Wearable · Personalization · Benchmark
2026-05	MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory	arXiv 2605	Memory · Multimodal / Wearable · Benchmark
2026-05	PACT: Proactive Asking for Continual Task Assistance in Human-Robot Collaboration	arXiv 2605	Embodied · Clarification · Long-horizon
2026-05	StreamOV: Streaming Omni-Video Understanding via Evidence-Guided Memory and Response Triggering	arXiv 2605	Streaming · Intervention Timing · Memory	Paper Notes
2026-05	Response-G1: Explicit Scene Graph Modeling for Proactive Streaming Video Understanding	arXiv 2605	Streaming · Intervention Timing · Memory	Paper Code Notes
2026-06	JoyAI-VL-Interaction: Real-Time Vision-Language Interaction Intelligence	arXiv 2606	Streaming · Intervention Timing · Multimodal / Wearable	Paper Code Notes
2026-07	Vinci2: Providing Proactive Assistance in Continuous Egocentric Videos	ECCV 2026	Multimodal / Wearable · Intervention Timing · Memory
2026-07	ChatMuse: Supporting In-Person Small-Group Conversation Experience with a Proactive Assistive AI Agent in Mixed Reality	UIST 2026	Multimodal / Wearable · Collaboration · Human Factors
2026-08	GROVE: Growing and Reasoning over Temporally Stratified Memory from Streaming Video Experience	arXiv 2608	Multimodal / Wearable · Memory · Intervention Timing
2026-08	From Sports to Safety: Benchmarking Proactive Risk Inference in MLLMs	arXiv 2608	Multimodal / Wearable · Safety & Consent · Benchmark
2026-08	StreamArena: Toward Continuous, Interactive, and Long-Horizon Agentic Streaming Video Understanding	arXiv 2608	Multimodal / Wearable · Long-horizon · Benchmark
2026-08	MOSS-VL Technical Report	arXiv 2608	Multimodal / Wearable · Intervention Timing · Optimization	Code Model
2026-08	StreamSoccer: Event-Driven Memory for Streaming Soccer Commentary	arXiv 2608	Multimodal / Wearable · Intervention Timing · Memory
2026-08	OmniAssistBench: Assistant-style Interaction Benchmark for Omni-LLMs	arXiv 2608	Multimodal / Wearable · Dialogue · Benchmark
2026-08	When Seeing Is Not Enough: Benchmarking Interactive Visual Grounding in LVLMs	EMNLP 2026	Multimodal / Wearable · Clarification · Benchmark
2026-08	MMPCBench: Benchmarking Multimodal Large Language Models on Proactive Critique of Flawed Inputs	arXiv 2608	Multimodal / Wearable · Safety & Consent · Benchmark
2026-09	Time-Aware Assistive Navigation	ICRA 2026	Multimodal / Wearable · Intervention Timing · Benchmark
2026-09	Ambient @ EgoProactive 2026: Proactive Egocentric Assistance with Visually Grounded Supervision	ECCV 2026 Wearable AI Challenge	Multimodal / Wearable · Intervention Timing · Optimization
2026-09	Omni Interaction Agent Technical Report	arXiv 2609	Multimodal / Wearable · Dialogue · Intervention Timing
2026-09	Realtime-Venus: A Full-Duplex Interaction System with Asynchronous Delegation	arXiv 2609	Multimodal / Wearable · Dialogue · Tool Use
2026-09	Full-Duplex Speech Models Take the Floor When Asked, Not When Needed	arXiv 2609	Multimodal / Wearable · Intervention Timing · Benchmark
2026-09	Cognitive Action Reasoning for Proactive Robots from Human-Centered Multimodal Observations	arXiv 2609	Embodied · Intent Inference · Benchmark
2026-09	Generative Tutorial: Towards Live Contextualized Visual Instructions for Physical Tasks	arXiv 2609	Multimodal / Wearable · Intervention Timing · Human Factors
2026-09	Live Assistant: Learning Whether, When, and Whom to Assist in Real-World Live Social Streams	arXiv 2609 · Under review	Multimodal / Wearable · Intervention Timing · Real-world Data
2026-09	Robots That Take Initiative: A Framework for Building and Evaluating Proactive Robots	arXiv 2609	Embodied · Definition · Intervention Timing
2026-09	TRACE: Temporal Audit and Condition-aware Evaluation of Streaming Video Understanding	arXiv 2609	Multimodal / Wearable · Intervention Timing · Benchmark
2026-09	HiThink Turn: An Intent-Aware Turn-Taking Control Module for Full-Duplex Dialogue	arXiv 2609	Multimodal / Wearable · Dialogue · Intervention Timing
2026-09	Bayesian Active Learning for Intent Disambiguation in Interactive Robot Planning	CoRL 2026	Embodied · Clarification · Planning
2026-09	RobotEQ 3.0: Towards Personalized Social Proactive Intelligence in Embodied Agents	arXiv 2609	Embodied · Personalization · Benchmark
2026-09	APM-Bench: Benchmarking Cross-session Persistent Memory for Egocentric Streaming Video Assistants	arXiv 2609	Multimodal / Wearable · Memory · Benchmark
2026-09	DuplexAct-Bench: Broadening Full-Duplex Speech Evaluation toward Proactive Interaction across Diverse Behavioral Requirements	arXiv 2609	Multimodal / Wearable · Intervention Timing · Benchmark
2026-10	OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction	arXiv 2610	Multimodal / Wearable · Memory · Intervention Timing	Code Model Dataset
2026-10	From Social Reasoning to Embodied Interaction: An Agentic Framework for Social Robots	arXiv 2610	Embodied · Memory · Intervention Timing	Project
2026-10	Improving Proactive AI Assistance with Hierarchical Procedural Understanding	arXiv 2610	Streaming · Personalization · Benchmark	Project
2026-10	Jarvis: A Proactive Speech Agent for Multi-Party Conversations	arXiv 2610	Dialogue · Collaboration · Intervention Timing
2026-10	Proactive Robot Assistance via Event-Driven Reasoning over State Changes	arXiv 2610	Embodied · Runtime Activation · Intent Inference
2026-10	EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams	EMNLP 2026 Main (per arXiv record)	Streaming · Multimodal / Wearable · Intervention Timing	Project
2026-10	Proactive Physical AI through Continuous Observation and Timely Intervention	arXiv 2610	Streaming · Embodied · Intervention Timing
2026-10	An Agentic Framework for Proactive Robots with Procedural Assistance Memory	CoRL 2026 (per arXiv record)	Embodied · Memory · Intervention Timing	Project
2026-10	From Perception to Prediction and Intervention: Benchmarking Proactive Video Assistants in Real-Life Scenarios	arXiv 2610	Streaming · Intervention Timing · Benchmark	Project Code Dataset
2026-10	SurgStreamer: proactive, request-driven monitoring of surgical video streams with structured state and a reusable reasoning layer	arXiv 2610	Streaming · Response Timing · Multimodal / Wearable
2026-10	ProPerAgent: Toward Memory-Driven Proactive Assistance	arXiv 2610	Streaming · Memory · Personalization

Benchmarks, Personalization and Optimization

Date	Title	Venue / Source	Tags	Resources
2025-08	ProactiveEval: A Unified Evaluation Framework for Proactive Dialogue Agents	ACL 2026	Benchmark · Dialogue · Intent Inference
2025-09	ProPerSim: Developing Proactive and Personalized AI Assistants through User-Assistant Simulation	ICLR 2026	Personalization · Simulation · Benchmark
2025-10	Beyond Reactivity: Measuring Proactive Problem Solving in LLM Agents	arXiv 2510	Benchmark · Intent Inference · Tool Use
2025-10	AgentFold: Long-Horizon Web Agents with Proactive Context Folding	ICLR 2026	Long-horizon · Memory · Tool Use
2025-11	Training Proactive and Personalized LLM Agents	COLM 2026	Personalization · Optimization · Simulation
2025-11	Structured Uncertainty guided Clarification for LLM Agents	Findings of ACL 2026	Clarification · Tool Use · Benchmark
2026-02	Pushing Forward Pareto Frontiers of Proactive Agents with Behavioral Agentic Optimization	ICML 2026	Optimization · Human Factors · Safety & Consent
2026-03	ProEvent: An Event-centric Benchmark for Proactive Agents	arXiv 2607 / ACL ARR 2026	Benchmark · Long-horizon · Intervention Timing
2026-04	SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization	arXiv 2604	Optimization · Skill Learning · Memory
2026-04	PASK: Toward Intent-Aware Proactive Agents with Long-Term Memory	arXiv 2604	Long-horizon · Memory · Intervention Timing
2026-05	CogniFold: Always-On Proactive Memory via Cognitive Folding	arXiv 2605	Memory · Intent Inference · Benchmark
2026-05	MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation	arXiv 2605	Optimization · Skill Learning · Memory
2026-05	Do Proactive Agents Need an LLM to Decide When to Act?	arXiv 2605	Intervention Timing · Optimization · Tool Use
2026-05	ProactBench: Beyond What The User Asked For	arXiv 2605	Benchmark · Dialogue · Intent Inference
2026-05	Claw-Anything: Benchmarking Always-On Personal Assistants with Broader Access to User's Digital World	arXiv 2605	Long-horizon · Benchmark · Tool Use
2026-05	π-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows	arXiv 2605	Long-horizon · Intent Inference · Benchmark
2026-05	VitaBench 2.0: Evaluating Personalized and Proactive Agents in Long-Term User Interactions	arXiv 2605	Long-horizon · Personalization · Memory
2026-06	Ψ-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues	arXiv 2606	Dialogue · Personalization · Benchmark
2026-06	Communication Policy Evolution for Proactive LLM Agents	arXiv 2606	Dialogue · Intervention Timing · Optimization
2026-06	See, Infer, Intervene: Proactive World Modeling for Goal-Oriented Social Intelligence	arXiv 2606	Multimodal / Wearable · Intervention Timing · Benchmark
2026-07	Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents	arXiv 2607	Long-horizon · Memory · Intervention Timing
2026-07	ENPMR-Bench: Benchmarking Proactive Memory Retrieval for Emotional Support Agents	Findings of ACL 2026	Memory · Dialogue · Benchmark
2026-08	Preference-Driven Online Adaptation for Personalized Interaction Initiation in Proactive AI Assistants	arXiv 2608	Personalization · Intervention Timing · Optimization
2026-08	VibeLifeBench: Can Your Life Agent Be Proactive and Persistent in a Living World?	arXiv 2608	Long-horizon · Benchmark · Safety & Consent
2026-08	Not Just Reason, Not Just Scan: Reinforcement Learning for Proactive Scientific Error Verification over Academic Paper	Findings of EMNLP 2026	Multimodal / Wearable · Optimization · Safety & Consent
2026-09	RPCBench: A Benchmark for Proactive Premise Critique in LLM-based Recommendation	arXiv 2609	Recommendation · Benchmark · Safety & Consent
2026-09	Propose to Learn, Learn to Propose: Evaluability-Aware Assistance under Bounded Rationality	arXiv 2609 · Under review	Personalization · Planning · Human Factors
2026-09	New Evidence, Same Choice: Testing Physical Experiment Selection in Vision Language Models	arXiv 2609	Benchmark · Planning · Tool Use
2026-09	FinInteract: Benchmarking Clarification and Intent Integration in Ambiguous Financial Question Answering	arXiv 2609	Clarification · Benchmark · Dialogue
2026-09	PASTABench: Proactive Assessment of Sequential Trajectories for Agent Safety	EMNLP 2026	Benchmark · Safety & Consent · Intervention Timing
2026-09	TWIST: A Proposed Benchmark for Intervention Quality in Conversational Memory, with a Human-Validated Draft-Alignment	arXiv 2609	Memory · Benchmark · Safety & Consent
2026-10	DelegationBench: Measuring When AI Agents Should Ask Before Acting	arXiv 2610	Safety & Consent · Clarification · Benchmark	Code / Data
2026-10	InteractionBench: A Real-Time Interaction Benchmark for Streaming Video Systems	arXiv 2610	Streaming · Intervention Timing · Benchmark	Project Code Dataset
2026-10	DuplexAct: How Interactive Are Frontier Full-Duplex Spoken Language Models?	arXiv 2610 · Under review	Full-duplex · Intervention Timing · Benchmark	Code / Data
2026-10	LexPilot: Benchmarking the Proactive Interaction Ability of Legal LLMs with Laypeople in Chinese Civil Case Consultation	arXiv 2610	Clarification · Dialogue · Benchmark	Dataset

---

Benchmarks

For detailed comparison, see BENCHMARKS.md.

Date	Benchmark	Paper	Environment	What it tests	Resources
2024-03	ProMISe	ProMISe	information-seeking dialogue	proactive clarification for intent resolution
2024-10	RealHumanEval	Need Help?	programming tasks	proactive IDE assistance with human users
2024-10	ProactiveBench	Proactive Agent	desktop activity events	proactive task prediction and acceptance
2025-03	OmniMMI	OmniMMI	continuous audio-visual streams	proactive alerts, turn-taking, and streaming interaction
2025-05	ContextAgentBench	ContextAgent	wearable sensory contexts	proactive service prediction and tool calling
2025-07	FingerTip 20K	FingerTip 20K	Android trajectories	proactive task suggestion and personalized execution
2025-08	ProactiveEval	ProactiveEval	proactive dialogue	target planning and dialogue guidance across six domains
2025-10	PROBE	Beyond Reactivity	web problem-solving tasks	bottleneck discovery and autonomous resolution
2025-10	ESTP-Bench	Eyes Wide Open	egocentric video and evolving queries	coherent just-in-time streaming answers	Paper Code
2025-11	UserVille	Training Proactive and Personalized LLM Agents	SWE and deep-research user simulation	productivity, proactivity and personalization under vague prompts
2025-11	ClarifyBench	Structured Uncertainty Guided Clarification	dynamic tool-calling tasks	decide which tool argument to clarify and when to stop asking
2025-12	StreamGaze	StreamGaze	causal video with human gaze	gaze-conditioned anticipation and alerts; past/present tasks tracked separately	Paper Code
2026-01	ChronosBench	Long-term Task-oriented Agent	dynamic task environments	proactive long-term intent maintenance
2026-02	ProAgentBench	ProAgentBench	real workflow logs	when-to-assist and how-to-assist
2026-02	ProactiveMobile	ProactiveMobile	mobile device context	latent intent to executable API sequence
2026-03	ProEvent	ProEvent	future event tracking	proactive event maintenance and reminders
2026-03	PIRA-Bench	PIRA-Bench	continuous GUI screenshots	proactive GUI intent recommendation
2026-03	ProReady-QA	StreamReady	long streaming videos with annotated evidence windows	answer only after sufficient evidence appears and avoid both premature and delayed responses
2026-03	GUIDE	GUIDE	open-ended GUI workflow videos	behavior state, intent, and help prediction
2026-03	ProactiveBench (MLLM)	ProactiveBench / Trento	visual difficulty scenarios	MLLM proactive help-seeking from visual context
2026-03	Live Gaming Benchmark	Proact-VL	streaming game video and companion interaction	timely solo/co-commentary and guidance	Paper Code
2026-04	ProVoice-Bench	ProVoice-Bench	voice interaction streams	proactive voice intervention timing, over-triggering, monitoring
2026-04	Pare-Bench	Pare	multi-app FSM environment	active user simulation, intervention timing, multi-app execution
2026-04	KnowU-Bench	KnowU-Bench	Android emulator	personalization, proactive tasks, consent and rejection handling
2026-04	LatentNeeds-Bench	PASK	real speech transcripts over multi-turn sessions	detect latent demand, assist, or stay silent under latency constraints
2026-05	IPIBench	IPIBench	streaming video, multi-turn	interactive proactive monitoring, task management, reactive-proactive coordination
2026-05	CogEval-Bench	CogniFold	streaming event memory	proactive concept emergence and cognitive-structure formation
2026-05	MemEye	MemEye	multimodal long-term memory	visual evidence granularity and temporal state reasoning
2026-05	ProCodeBench	Proactive Coding Assistants	real IDE traces	proactive coding intent prediction and sim-to-real evaluation
2026-05	ProactBench	Beyond What The User Asked For	synthetic multi-turn conversations	grounded unstated-need inference at emergent, critical, and recovery triggers
2026-05	OmniPro	OmniPro	continuous omni-modal video streams	autonomously decide when and what to report across multiple triggers
2026-05	EgoPro-Bench	EgoPro-Bench	egocentric streaming videos and user profiles	personalized interaction timing with explicit attention-or-silence labels
2026-05	Claw-Anything	Claw-Anything	months of cross-service and multi-device activity	anticipate needs and complete proactive personal-assistant tasks amid noise
2026-05	π-Bench	π-Bench	persistent personal workspaces	proactive hidden-intent resolution and checklist completion in long-horizon workflows
2026-05	ProActEval	Anticipate and Learn	proactive assistant scenarios	idle-time anticipation, evidence acquisition, user effort and hallucination reduction
2026-05	VitaBench 2.0	VitaBench 2.0	long-term user interaction sequences	preference extraction, memory use, updates, and proactive missing-information acquisition
2026-05	SOVBench-O / SOVBench-T	StreamOV	streaming audio/video and multi-turn context	answer timing and absent-evidence silence	Paper / protocol
2026-06	Ψ-Bench	Ψ-Bench	persuasive dialogue	persona-sensitive influencing with simulated clients and user profiles
2026-06	GuidanceSalesBench	See, Infer, Intervene	pre-interaction retail videos and customer-state manifests	select greet, elicit, inform, recommend, or hold
2026-07	ProACT Collaboration Bench	ProACT	multi-user collaboration	breakdown detection, intervention timing, and non-interruptiveness
2026-07	InquireBench	InquireMobile	mobile GUI	proactive confirmation before risky or uncertain actions
2026-07	AndroidIntent	PersonalAlign	long-term Android records	vague-intent resolution and instruction-free routine suggestions
2026-07	BSDD	CoLabScience	streaming biomedical discussions	when and how to contribute without disrupting collaborators
2026-07	ENPMR-Bench	ENPMR-Bench	emotional-support dialogue	latent emotional-need inference and proactive supportive-memory retrieval
2026-07	EgoServe	Vinci2	continuous egocentric video	when-and-whether assistance across four memory horizons
2026-07	Int-Bench	AI Assistants Overassist	simulated learners solving 1,500 problems	whether, when, and how much a tutor should intervene
2026-07	AskBench	When and What to Ask	interactive intent-deficient and false-premise QA	decide when and what to clarify without sacrificing task performance
2026-08	Active-SWE	Active-SWE	repository snapshots without issue reports	autonomous discovery and repair of single, multiple, and potential bugs
2026-08	SPRINT	From Sports to Safety	real sports videos	early hazard warning, cause grounding, and false-alarm control
2026-08	StreamArena	StreamArena	hour-scale audio-visual streams	proactive monitoring alongside perception, retrospection, and tool use
2026-08	VibeLifeBench	VibeLifeBench	simulated multi-week life tasks	autonomous act, ask, and stay-silent decisions under silent world changes and implicit constraints
2026-08	JarvisBench	JarvisBench	ongoing single- and multi-agent work	recognize consequential user-owned decisions, request judgment, and route it back to working agents
2026-08	Act2Intention Bench	Act2Intention	continuous mobile GUI trajectories	understand, predict, suggest, and execute the next user intention from action history
2026-08	StreamSoccer	StreamSoccer	continuous soccer video	causal commentary across current, recent, and historical scopes with rule-assisted mode-or-silence scheduling
2026-08	OmniAssistBench	OmniAssistBench	continuous assistant-style video interactions	guide a user through a goal, incorporate visual prompts, and delay responses until the target event
2026-08	Interactive Visual Grounding	When Seeing Is Not Enough	ambiguous visual references under multi-turn protocols	ask follow-up questions and synthesize acquired evidence to ground the target
2026-08	MMPCBench	MMPCBench	flawed multimodal user inputs	autonomously detect, diagnose, and repair errors without an extra prompt
2026-09	PROS-Bench	Beyond Instruction-Driven Editing	source papers paired with editable scientific posters	discover source-grounded problems, obtain user acceptance, repair native objects, and validate outcomes
2026-09	RPCBench	RPCBench	evidence-grounded recommendation requests with corrupted premises	proactively detect, localize, and handle premise failures without fabricating unsupported recommendations
2026-09	CONFLICTGUI	Do GUI Agents Know When Not to Act?	feasible and conflicting instructions grounded in GUI states	verify feasibility and terminate rather than over-comply when instructions or GUI evidence conflict
2026-09	CC-Mediation	CC-Mediation	ten-turn cross-cultural conflict dialogues	decide when to intervene and select a mediation strategy with persistent downstream benefit
2026-09	OR-Clarify	Ask Before You Optimize	partially specified operations-research requests with hidden formulation slots	ask only formulation-critical questions, avoid silent assumptions, and stop when ready to model
2026-09	TIMELI	Time-Aware Assistive Navigation	egocentric navigation videos and closed-loop urban simulation	decide when to give concise guidance or remain silent for safety
2026-09	IdeaAMBIG	IdeaAMBIG	incomplete research-method specifications grounded in papers and code	detect implementation-critical gaps and issue clarification actions instead of assuming
2026-09	Physical Experiment Selection	New Evidence, Same Choice	controlled visual-physics measurements with known experiment costs	answer when evidence is sufficient or select the cheapest resolving experiment
2026-09	ProMediConv	ProMediConv	multi-stage legal dispute dialogues reconstructed from real cases	select mediation strategies proactively and improve party behavior over the trajectory
2026-09	ProAction	Cognitive Action Reasoning for Proactive Robots	instruction-free multimodal observations across daily-life scenes	infer a cognitively grounded high-level robot action from human and environmental cues
2026-09	FinInteract	FinInteract	bilingual ambiguous financial questions grounded in regulatory filings	elicit the intended interpretation and integrate the answer instead of guessing the default
2026-09	Live Assistant	Live Assistant	real livestream audio-video, comments, gifts, viewer dynamics, and room metadata	choose silence, private memory, or grounded assistance for the right recipient
2026-09	PASTABench	PASTABench	multi-step agent trajectories with accumulating safety risk	decide whether and when to intervene inside an annotated optimal window
2026-09	TWIST	TWIST	conversational memory, belief changes, drafts, and matched hard negatives	intervene on contradictions or sensitive recall without over-flagging safe cases
2026-09	TRACE	TRACE	streaming video with temporally audited evidence and trigger windows	separate answer quality, delay, false alarms, missed windows, and processing workload
2026-09	IntentFlux	When Users Change Their Minds	executable multi-turn tasks with controlled intent changes	recover the active request while ignoring superseded or withdrawn requirements
2026-09	RobotEQ 3.0	RobotEQ 3.0	embodied social situations paired with user profiles and preferred actions	predict which proactive action a specific person prefers rather than the population average
2026-09	PROACTIVITY-GYM	Foundations of Proactive Agents	multi-day stateful scenarios with persona-conditioned simulated users	jointly evaluate task capability, temporal allocation, and evolving trust
2026-09	APM-Bench	APM-Bench	cross-session egocentric life trajectories with persistent-memory constraints	answer and assist across interruptions while recognizing missing evidence
2026-09	Drift-Bench++	Beyond Oracle Communication	executable tasks with miscommunication, finite patience, and silent intent shifts	recover and continuously track current intent under imperfect communication
2026-09	DuplexAct-Bench	DuplexAct-Bench	bilingual streaming speech across pre-session, in-session, and no-explicit conditions	interruption, yielding, proactive initiation, active silence, and backchannel timing/content
2026-09	SWE-Intervene	Learning When and How to Intervene	software-engineering action trajectories before tool execution	allow, autonomously redirect, or pause for human assistance with actionable feedback
2026-10	OverAct	OverAct	controlled tool-calling tasks across eight privacy-sensitive domains	measure and filter data access beyond the user's authorized scope
2026-10	Actionable Indeterminacy	Ask, Relax, or Act?	preference reasoning	Act on a shared acceptable action, clarify otherwise, or propose a permitted repair	arXiv
2026-10	DelegationBench	DelegationBench	tool-agent consent	Act, ask permission, ask for information, or refuse	arXiv Code / Data
2026-10	InteractionBench	InteractionBench	streaming video interaction	Respond with correct content at the right time and remain silent on negatives	arXiv Project Code Dataset
2026-10	ProactiveCoach-Bench	ProactiveCoach	procedural video assistance	Select phase, step, or action guidance, or stay silent	arXiv Project
2026-10	CHI-180-proactive	Jarvis	multi-party speech assistance	Contribute missing or corrected evidence only when the group does not self-resolve	arXiv
2026-10	EgoVoice benchmark	EgoVoice	egocentric spoken assistance	Choose guidance or silence without a new question	arXiv Project
2026-10	ReqClarifyBench	Don't Guess, Just Ask	repository-level requirements	Agent-selected clarification and faithful requirements	arXiv
2026-10	DuplexAct — Help/Safety Interrupt subset	DuplexAct	spoken proactive intervention	Take the floor and deliver situation-appropriate assistance or warning	arXiv Code / Data
2026-10	P3IBench	P3I-Assistant / P3IBench	streaming video assistance	Monitor events, anticipate hazards, or issue grounded guidance with restraint	arXiv Project Code Dataset
2026-10	ProPerMem	ProPerAgent	query-free personalized assistance	Retrieve evidence that justifies one timely reminder, or remain silent	arXiv
2026-10	SSUI-Dialog	SAFE-PRO	cross-modal safety clarification	Ask a safety-grounded clarification instead of unsafe commitment or over-refusal	arXiv
2026-10	LexPilot	LexPilot	chinese civil consultation	Elicit missing case facts and improve claim-level prediction	arXiv Dataset

---

Tag Vocabulary

Tags are intentionally compact and reusable. They describe the paper's main contribution, not every detail.

Tag	Meaning
Definition	Defines or reframes proactive agents, proactive dialogue, or design-space boundaries.
Survey	Synthesizes a broad proactive-agent subfield or taxonomy.
Human Factors	Studies interruption, control, satisfaction, workload, adoption, or developer experience.
Trust	Focuses on competence perception, calibrated reliance, or trustworthy interaction.
Safety & Consent	Covers confirmation, autonomy boundaries, reversibility, rejection, or risk control.
Privacy	Centers privacy management, data minimization, or personal-context governance.
Intervention Timing	Focuses on when an agent should act, ask, suggest, or remain silent.
Intent Inference	Infers latent goals, hidden constraints, future tasks, or missing information.
Clarification	Proactively asks questions before planning, execution, or recommendation.
Dialogue	Proactive behavior in conversational, persuasive, or task-oriented interaction.
Planning	Proactive decomposition, task planning, scheduling, or future-state reasoning.
Tool Use	Tool calling, API execution, GUI operation, or action orchestration.
Recommendation	Proactive recommendation or suggestion ranking.
Collaboration	Multi-party or human-agent collaborative problem solving.
Education	Learning, tutoring, reflection, or student engagement contexts.
Long-horizon	Multi-session, dynamic, future-event, or long-running task maintenance.
Personalization	User preferences, personas, profiles, long-term user history, or user-specific adaptation.
Memory	Persistent memory, episodic memory, visual memory, skill memory, or cognitive memory structures.
Simulation	User simulation, environment simulation, synthetic users, or synthetic workflows.
Optimization	RL, reward modeling, multi-objective optimization, self-evolution, or behavior tuning.
Skill Learning	Skill creation, skill internalization, skill memory, or reusable procedure learning.
Benchmark	Introduces a dataset, evaluation suite, benchmark, simulator, or diagnostic protocol.
Real-world Data	Uses real user traces, field-study data, or deployment-like logs.
Desktop	Desktop activity streams, workstation context, or event logs.
GUI	Graphical interface agents, browser/app screens, or visual UI interaction.
Mobile	Mobile GUI, Android/iOS workflows, phone sensors, or mobile user context.
OS	Operating-system agents, cross-app workflows, or OS-level verification.
IDE	Programming assistants, code editors, or developer tooling.
Streaming	Continuous causal video/audio input; specify whether perception continues during generation.
Response Timing	Evidence-conditioned wait/respond or trigger decisions for a standing query.
Full-duplex	Overlapping perception and output with explicit interaction/floor control.
Runtime Activation	Event, heartbeat, scheduled, or model-selected wake mechanisms.
Product / Component	Product documentation or an explicitly labeled sensing/memory/serving component.
Multimodal / Wearable	Video, audio, AR, smart glasses, egocentric streams, or open-world sensory context.
Sensing	Active context acquisition, sensor selection, or on-demand sensory capture.
Embodied	Robots, physical environments, or human-populated embodied settings.

---

Contributing

Pull requests are welcome.

Before adding a resource, identify its type (paper, model, framework, product, benchmark, or component), the evidence source, and the actual activation mechanism. Add projects to PROJECTS.md, streaming resources to STREAMING.md, and paper records to the bibliography. Supporting components need a stated role; a proactive keyword or continuous input alone is insufficient.

For projects/models, record an official repository or documentation link, what is publicly released, any access or licensing boundary, and a checked date. Do not treat a README promise, product demo, or star count as measured proactive capability.

Before adding a paper, check that it satisfies at least one of:

• It predicts latent user intent before a complete explicit instruction.
• It decides when to intervene, ask, suggest, execute, remind, or stay silent.
• It evaluates proactive assistance, interruption cost, user control, consent, or personalization.
• It contributes a benchmark or dataset where proactivity is the primary task.

Suggested note template:

# Paper Title

## Why It Matters

...

## Proactivity Signal

...

## Evaluation Setup

...

## Key Limitations

...

## Use For

...

---

Maintained by Low Entropy AI.