💬 NLP

ProACT: Towards Breakdown-Aware Proactive Agent in Multi-User Collaboration

This paper introduces ProACT, a breakdown-aware framework that enables conversational agents to proactively identify and address collaboration issues in multi-user settings, alongside a new benchmark demonstrating its superior performance in intervention quality and appropriateness compared to reactive chat models.

Original authors: Shu Yang, Difei Xu, Jiaxin Pei, Di Wang

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shu Yang, Difei Xu, Jiaxin Pei, Di Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are in a busy group chat with friends trying to plan a surprise party. Everyone is talking, throwing out ideas, and occasionally getting confused.

Right now, most AI assistants in these chats are like silent librarians. They sit on the shelf, waiting for someone to specifically ask, "Hey, can you summarize what we just said?" or "What time is the party?" If no one asks, they say nothing, even if the group is going in circles or forgetting important details like the budget.

The paper introduces a new kind of AI called ProACT. Think of ProACT not as a librarian, but as a skilled, observant team captain or a referee who knows exactly when to step in and when to stay quiet.

Here is how ProACT works, broken down into simple concepts:

1. The Problem: The "Reactive" Robot

Current AI is reactive. It only moves when pushed.

The Scenario: Your group chat starts arguing about whether to rent a hall or have a potluck. Then, someone forgets the budget limit. Then, two people start arguing about the same point they argued about an hour ago.
The Old AI: It watches all of this happen but says nothing because no one asked it a question. It misses the chance to fix the mess.

2. The Solution: The "Breakdown-Aware" Captain

ProACT is proactive. It watches the conversation for specific "breakdowns"—moments where the team is getting stuck. It looks for things like:

The Loop: People arguing the same point over and over without deciding.
The Fog: Everyone is confused about the rules or the goal.
The Forgotten Rule: Someone suggests something that breaks a rule the group agreed on earlier (like the budget).
The Imbalance: Only one person is talking while the quiet experts are ignored.

3. How ProACT Decides to Speak (The "Referee" Instinct)

This is the most important part. ProACT doesn't just talk whenever it sees a problem. It asks itself two questions:

Is the team stuck? (Is there a "breakdown"?)
Is the team already fixing it? (Are the humans solving the problem on their own?)

The Analogy of the Soccer Referee:
Imagine a soccer game.

If a player is injured and the game stops, the referee (ProACT) steps in.
But if two players are arguing but then they quickly laugh it off and keep playing, the referee does not blow the whistle. Blowing the whistle would be annoying and interrupt the flow.

ProACT does the same. If the humans are already fixing the confusion, ProACT stays silent. If they are stuck in a loop or forgetting a rule, ProACT steps in with a short, helpful nudge.

4. The Toolkit: "Skills" instead of Random Chat

When ProACT decides to speak, it doesn't just guess what to say. It uses a specific toolkit of skills, like a mechanic choosing the right wrench:

The "Loop Breaker": "Hey, we've discussed this three times. Let's pick a winner."
The "Constraint Reminder": "Wait, we agreed the budget was $500, and this plan costs $600."
The "Clarifier": "I'm not sure what 'Option B' means. Can someone explain?"

5. The Results: Better Timing, Less Noise

The researchers tested ProACT against standard AI chatbots using thousands of real and simulated group conversations (from coding teams to planning committees).

The Old Way: The AI often interrupted when it shouldn't, or stayed silent when it should have helped.
ProACT: It was much better at knowing when to speak.
- It interrupted less often when the humans were doing fine.
- When it did speak, its comments were shorter, more helpful, and less annoying.
- It improved the quality of the group's decision-making, especially in long, complex conversations where people tend to get lost.

Summary

ProACT changes the AI from a passive tool that waits to be used, into an active participant that helps the team stay on track. It acts like a wise friend who knows exactly when to whisper a reminder and when to let the conversation flow naturally, ensuring the group doesn't get stuck in confusion or arguments.

Technical Summary: ProACT: Towards Breakdown-Aware Proactive Agent in Multi-User Collaboration

Problem Statement

Conversational agents are increasingly integrated into human collaborative workflows (e.g., meeting summarization, code drafting, deliberation). However, current deployments remain fundamentally passive and reactive: they respond only to explicit user requests or predefined prompts. This design fails to address the dynamic nature of multi-user collaboration, where progress is often undermined by subtle "breakdowns" such as disagreements, ambiguous goals, forgotten constraints, underspecified plans, discussion loops, and imbalanced participation.

Unlike single-user assistance, where the agent's goal is to fulfill a specific request, multi-user collaboration requires an agent to navigate complex social dynamics. The core challenge is determining when to intervene in a group interaction where users may have conflicting or uncertain intents, and how to intervene without disrupting productive discussion. Existing proactive agent frameworks (e.g., Lu et al., 2025) focus on predicting user intent from single-user inputs and do not adequately address the social and normative challenges of group coordination.

Methodology: The ProACT Framework

The authors introduce ProACT, a breakdown-aware proactive agent framework grounded in theories of common ground, collaborative planning, and coordination work. ProACT operates as a skill-guided system that observes speaker-attributed conversation history ( $H_t$ ) and decides between two actions: remaining silent ( $\emptyset$ ) or producing a concise, group-facing intervention ( $u_t$ ).

1. Decision Pipeline

The framework follows a three-step pipeline:

Breakdown-Aware Diagnosis: The agent analyzes the current conversation turn to detect evidence of a collaboration breakdown. It outputs a diagnosis $(b_t, z_t)$ , where $b_t$ is the detected issue type and $z_t \in \{0, 1\}$ indicates if intervention is warranted. The diagnosis considers local context, such as whether the group is already making progress or self-repairing.
Skill-Guided Action: If intervention is warranted ( $z_t=1$ ), the agent routes the detected issue to a specific Collaboration Skill ( $k_t$ ). If not, it outputs silence.
Execution: The selected skill guides the generation of a visible, concise intervention ( $u_t$ ) that addresses the specific breakdown while adhering to social constraints (neutrality, non-interruption, preserving human agency).

2. Collaboration Breakdown Taxonomy

ProACT defines eight specific categories of collaboration breakdowns (Table 1) based on established human collaboration literature:

Conflict: Incompatible goals or constraints.
Uncertainty: Lack of shared clarity on requirements or intent.
Underspecified Plan: Missing execution details (owner, deadline, scope).
Forgotten Constraint: Overlooking prior commitments or policies.
Goal Drift: Moving away from the shared objective.
Discussion Loop: Repeating unresolved arguments without new information.
Participation Imbalance: Dominance by one perspective or missing stakeholders.
Risky Commitment: Moving toward decisions before resolving risks or dependencies.

3. Skill Library

ProACT utilizes a library of lightweight, reusable skills (Table 6). Each skill is defined by:

Trigger Condition: When the skill is applicable.
Diagnostic Cues: Evidence to look for in the conversation.
Social Constraints: Guidelines to remain neutral and avoid interruption.
Response Format: A concise, group-facing utterance.
Examples include Loop Breaking (suggesting a decision rule), Conflict Mediation (naming incompatible goals neutrally), and Constraint Reminders.

4. Structured Interface

The framework separates internal reasoning from external output. The model first determines the action (Silence vs. Intervention) and the skill, then generates the final response. Crucially, the visible output contains only the group-facing message, excluding diagnostic labels, skill names, or meta-commentary, ensuring the agent behaves naturally within the chat.

Benchmark and Evaluation

To evaluate proactive collaboration, the authors constructed a new benchmark containing 3,244 turn-level examples derived from:

Real Data: Public GitHub issue discussions and the QMSum dataset (covering software, product design, research, and education).
Synthetic Data: 800 examples derived from the BEAM dataset, injecting controlled breakdowns to ensure coverage of rare failure modes.

The dataset is balanced between intervention and silence cases. Evaluation is performed using an LLM judge (GPT-5.4-Nano) assessing four criteria:

Appropriateness: Does the action fit the collaboration state (intervening only when a breakdown exists)?
Non-Interruptiveness: Does the agent avoid disrupting productive flow or duplicating human repair?
Conciseness: Is the response short and free of unnecessary explanation?
Intervention Quality: Is the visible response useful, grounded, neutral, and actionable (1–5 scale)?

Experimental Results

The authors evaluated ProACT against a Direct Chat baseline (where the agent responds to every turn) across five LLM backbones: GPT-5.4, Kimi K2.5, Claude Sonnet 4.6, Gemini 3.1 Pro Preview, and GPT-OSS-120B.

Consistent Improvement: ProACT outperformed Direct Chat across all models and all four metrics.
- Kimi K2.5 showed the most dramatic gains: Appropriateness increased from 0.222 to 0.870; Non-interruptiveness from 0.323 to 0.942; and Conciseness from 0.129 to 0.964.
- Similar trends were observed for all other backbones, demonstrating that breakdown-aware skill routing helps agents decide both when to speak and how to intervene.
Participation Timing: Analysis revealed that ProACT significantly reduces unnecessary responses. On examples labeled for silence, Direct Chat responded 100% of the time (with 34.8% judged inappropriate), whereas ProACT responded only 39.3% of the time, increasing overall appropriateness to 89.6%.
Context Sensitivity: Quality gains were strongest in social planning tasks (e.g., committee discussions, academic collaboration) and in later conversation turns (after turn 500), where accumulated context and unresolved coordination issues make direct responses less reliable.
Conservative Abstention: The authors note that many "missed" interventions (where ProACT stayed silent on a reference-labeled intervention turn) were actually judged as acceptable abstentions by the external judge. In these cases, humans were already resolving the issue, and an agent message would have been duplicative.

Significance and Claims

The paper claims that ProACT represents a shift from passive assistance to active participation in multi-user collaboration. Its primary contributions are:

A New Formulation: Defining proactive collaboration not just as "helping a user," but as "repairing group coordination" while respecting social norms.
A Breakdown-Aware Framework: Demonstrating that grounding agent decisions in a structured taxonomy of collaboration failures (conflict, loops, drift) leads to more appropriate and less intrusive behavior.
A New Benchmark: Providing the first evaluation suite for proactive agents across diverse domains (planning, design, logistics, etc.), highlighting the gap between current reactive agents and the needs of collaborative workflows.

The authors conclude that while ProACT improves collaborative appropriateness and reduces interruptions, it operates in a controlled offline setting. They acknowledge that real-world deployment involves risks regarding trust, agency, and manipulation, suggesting that future work must include live studies and mechanisms for transparency and contestability. The work serves as a starting point for building agents that support human collaboration through timely, breakdown-aware participation.