ProAct-Helper: AI-Driven Proactive Assistant

• ProAct-Helper is an AI-driven proactive agent that minimizes human task cost by synthesizing assistant policies based on user belief updates and formal cost/recognition constraints.
• It employs a two-level Monte Carlo Tree Search combined with conformant planning to optimize joint action strategies under uncertainty.
• Experimental results demonstrate that ProAct-Helper can reduce human solo costs by 35–70% and significantly lower cognitive workload in collaborative settings.

A ProAct-Helper is an AI-driven agent or system designed to provide proactive task assistance, typically in environments where humans and AI agents co-exist with only partial mutual observability and potentially limited shared intent. The core challenge is to synthesize agent behavior that not only reduces the human's cost-to-go for a target task but also makes this benefit both practically effective and cognitively accessible to the human, even when the human may not expect, request, or fully recognize the need for assistance. These helpers are formally motivated and evaluated in frameworks that integrate planning under uncertainty, user-belief modeling, and principled cost/recognition constraints. The term has gained currency in work on collaborative planning, mixed-initiative interaction, and the evaluation of proactive assistant architectures (Kulkarni et al., 2021).

1. Formal Environment, Objective, and Recognition Constraints

A canonical ProAct-Helper operates in a stochastic environment with a set of state variables $S$, human and robot (assistant) action spaces $A_H$ and $A_R$, and a transition function $\Gamma$ encoding the effects of each action. Each human action $a_h \in A_H$ and assistant action $a_a \in A_R$ is described by preconditions, add/del effects, and an immediate cost $C_H(a_h)$ or $C_R(a_a)$. The human perceives a finite observation space $\Omega_H$, with $O_H(a,s)$ mapping action-state pairs to observation symbols (including the null $\omega_\emptyset$ for unperceived events). The assistant is assumed to have full state observability.

The core optimization is to synthesize an assistant policy $\pi_R$ mapping beliefs $b$ to actions so as to minimize the overall human cost-to-go under three core principles (Kulkarni et al., 2021):

1. Cost reduction: The assistant's actions must decrease the human's total cost towards their goal, i.e., for any joint plan $\pi_{\text{joint}}$,

$$C_H^\Delta(\pi_{\text{joint}}) = C_H(\text{human's part of } \pi_{\text{joint}}) - C_H(\pi_H^*) < 0$$

where $\pi_H^*$ is the optimal solo plan.

2. Recognition: The human must be able to recognize this cost reduction, formalized as a recognition constraint for some $\delta > 0$,

$$\mathbb{E}_{s \sim b_k}\left[ C_H(s,\pi_H^*) - C_H(s,\pi_H^{b_k}) \right] \geq \delta$$

at the assistant's hand-off time $k$ and human's belief $b_k$.

3. Overall cost-optimality: The assistant must optimize the human's expected cost (including workload and time), as measured after the human belief update.

In this mixed-observability framework, the human's belief $b_t$ is updated by a Bayesian filter,

$$b_{t+1}(s') = \eta \cdot O_H(a_t,s',o_{t+1}) \sum_{s \in S} T(s'|s,a_t) b_t(s),$$

with identity update on null observations.

2. Proactive Planning Algorithms and Policy Computation

Policy synthesis for a ProAct-Helper employs a two-level Monte Carlo Tree Search (MCTS) architecture combined with conformant planning (Kulkarni et al., 2021). The high-level pseudocode structure incorporates:

• Outer-loop MCTS over assistant actions and beliefs.
• Belief updates based on simulated human perceptions.
• Leaf evaluation by conformant planning (e.g., Conformant-FF), which estimates the human's minimum residual cost from the current belief.
• Only positive rewards propagate for simulated joint plans satisfying $C_H(\pi_H^b) < C_H^*$.

The backpropagated reward at each node is formed from the combined cost surrogate

$$\text{costSim} = \alpha \cdot \text{(assistant steps)} + (1-\alpha) \cdot C_H(\pi_H)$$

with $\alpha \in [0,1]$ a tunable weight. Only policies satisfying a user-specified recognition constraint are considered valid.

Following tree construction, a greedy rollout searches down the top-k utility children per level, selecting the assistant plan that minimizes this cost surrogate subject to the recognition constraint.

3. Experimental Evaluation and User Studies

Empirical evaluation covers both simulation-based benchmarks and user studies:

• On empirical domains (e.g., USAR, Driverlog), proactive assistance reduced human solo cost $C_H^*$ by 35–70%, with assistant plans remaining within a fixed step-budget (e.g., $L=15$ steps), and planning times under 20 s with $m=5000$ MCTS iterations.
• User studies (n≈30) found that without explicit legible or obfuscating display actions only $\sim$20% of participants recognized the proactive assist. Introducing a single legible or obfuscating action increased recognition to 80–90%, validating the necessity of making cost reduction “legible."
• Workload and processing-load ratings (NASA-TLX, 1–7 scale) saw substantial drops when ProAct-Helper plans were used: workload increases of +2.7 (baseline to assistant), effect sizes $d>1.6$.

The results support (a) sufficiency of a single explicit action for recognition and (b) large reductions in workload and cognitive load, validating the composite cost and recognition framework (Kulkarni et al., 2021).

4. Integration of ProAct-Helper with Broader Proactivity Architectures

The ProAct-Helper abstraction appears as a recurrent motif in subsequent proactive agent research. The PROPER framework (Kaur et al., 14 Jan 2026) splits proactivity into two agents: a Dimension Generating Agent (DGA) that surfaces implicit knowledge gaps and a Response Generating Agent (RGA) that integrates both explicit and implicitly inferred needs in system responses. In both, proactive assistance is formalized by surfacing unarticulated but contextually salient dimensions and calibrating the initiative (avoiding both underreach and overreach), with explicit scoring on coverage, initiative appropriateness, and user-intent alignment.

These architectures operationalize ProAct-Helper’s principles outside classical planning settings—in LLM-based language or task assistants, proactive programming systems (Chen et al., 2024), and LLM-powered visual analytics (Zhao et al., 24 Jul 2025)—by maintaining explicit representations of observed and inferred user needs, and proactively suggesting actions or modifications that can be directly integrated by the user.

5. Generalizations, Limitations, and Future Research Directions

Current ProAct-Helper frameworks rely on formal task models (fluents, transition functions) and explicit modeling of user beliefs and observability. Limitations acknowledged in (Kulkarni et al., 2021) include:

• The models' reliance on full state enumerability and tractable belief update may limit scalability in high-dimensional, continuous, or partially known domains.
• Recognition constraints presuppose that a single explicit action suffices to induce belief update for the human agent, which can break down if users ignore, misinterpret, or forget such cues.
• Cost models are constructed for single users; generalization to multi-human teams remains an open line.

Research directions include richer user models (memory, persistence, social context), dynamic adaptation of recognition thresholds based on observed human feedback, and integration with large-scale, partially observable LLM-based environments as in (Kaur et al., 14 Jan 2026) and (Lu et al., 2024). Cross-domain extensions (robotics, design, programming, safety) increasingly couple ProAct-Helper mechanisms with multimodal and hybrid architectures (DNN+symbolic, task graph planners, reward models) to maintain proactive, interpretable, and intent-aligned assistive behaviors.

6. Summary Table: ProAct-Helper—Core Model Components

Component	Formalization	Purpose
State Space	$S =$ instantiations of fluents $F$	Describes all possible world states
Actions (H, R)	$A_H$, $A_R$, with $\text{pre}, \text{add}, \text{del}$, $C_H$, $C_R$	Available moves and their costs
Transition Function	$\Gamma$, $T(s'|s,a)$	Dynamics of environment
Observation Model	$\Omega_H$, $O_H$	What the human perceives per action
Belief Updates	$b_{t+1}(s') = \eta\,O_H\,T\,b_t$	Captures human's world belief
Assistant Policy	$\pi_R: b \mapsto a_a$	Maps beliefs to actions
Objective	Minimize human cost, ensure recognition, optimize total step cost	Satisfies proactive principles (Kulkarni et al., 2021)

The ProAct-Helper concept thus formalizes and operationalizes proactive, legible, and cost-optimal AI assistance under uncertainty, combining structured planning, user model updating, and empirical validation in diverse collaborative settings (Kulkarni et al., 2021).