Source-linked AI summary
Anticipate and Learn: Unleashing Idle-Time Compute in Proactive Agents
Haoyi Hu, Qirong Lyu, Xianghan Kong, Weiwen Liu, Jianghao Lin, Zixuan Guo, Yan Xu, Yasheng Wang, Weinan Zhang, Yong Yu
TL;DR
AI agents remain reactive, leaving idle time unused for preparing future support. ProAct predicts likely needs from dialogue history and persistent memory, acquires grounded evidence during idle intervals, and reports lower turns, user effort, and hallucination rates on ProActEval. Its results also include strong reflective memory accuracy on MemBench.
Problem
AI agents initiate computation only after explicit requests, leaving idle time unused for anticipating and preparing for likely future needs.
Method
ProAct combines Future-State Prediction and Idle-Time Acquisition to use dialogue history and persistent memory for predicting needs and preparing grounded evidence.
Results
14.8% fewer required turns, 11.7% lower user effort, and 28.1% lower hallucination rates were reported on ProActEval versus a reactive baseline.
Takeaways & Limitations
ProAct converts idle intervals into grounded preparation and improves proactive efficiency, coverage, factual integrity, and reflective memory accuracy.
Takeaways & Limitations
Results come from a closed-world synthetic benchmark and depend on an LLM judge and value-aware delivery gate, so they do not guarantee open-world deployment performance.
Abstract
from arXiv · showhide
While AI agents demonstrate remarkable capabilities in reasoning and tool use, they remain fundamentally reactive: they compute responses only after explicit user prompts. This paradigm ignores a critical opportunity: the idle time between interactions is largely wasted, leaving agents unable to prepare for future user needs. To bridge this gap, we introduce ProAct, a proactive agent architecture that leverages idle-time compute to anticipate and fulfill likely upcoming user needs. By analyzing evolving dialogue history together with persistent memory, ProAct predicts upcoming needs and iteratively acquires information, allowing the agent to resolve knowledge gaps and prepare evidence before the user initiates a query. To rigorously evaluate proactive capabilities, we also introduce ProActEval, a comprehensive benchmark comprising 200 scenarios across 40 domains, featuring predictable need chains and diverse user cognitive profiles. Empirical results demonstrate significant advantages over reactive baselines. ProAct accelerates task completion by reducing required turns by 14.8%, decreases user effort by 11.7%, and cuts hallucination rates by 28.1% on ProActEval. Furthermore, MemBench evaluations confirm that ProAct achieves state-of-the-art reflective accuracy, underscoring its sustained and robust performance.
1 Introduction
Current agents remain reactive, leaving idle computation unused for anticipating future needs. ProAct uses dialogue history and persistent memory to predict needs, acquire evidence during idle time, and improve subsequent interactions.
- Current agents initiate processing only after explicit requests, then return to dormancy when tasks finish.
- ProAct turns idle time into proactive work by predicting likely future needs from dialogue history and persistent memory.
- Future-State Prediction integrates conversation history, user profiles, prior summaries, stored facts, and unresolved memory gaps to project upcoming intents.
- Idle-Time Acquisition prioritizes predicted needs by relevance, knowledge gaps, incremental value, and timeliness, then retrieves and verifies supporting evidence.
- 14.8% fewer required turns, 11.7% lower user effort, and 28.1% fewer hallucinations were reported on ProActEval versus a reactive baseline.
- ProActEval contributes a 200-scenario, 40-domain framework with predictable need chains and diverse user cognitive profiles.
2 Related Work
Related work spans persistent-memory agents, proactive dialogue and always-on assistants, and inference-time computation. These lines of work differ in memory structure, anticipation, and whether extra computation occurs before or after a user request.
- Persistent-memory agents support reflection, virtual memory hierarchies, or hierarchical summaries, but the cited systems lack some combination of lifecycle management, user profiles, or proactive behavior.
- Proactive-agent research includes need prediction from conversational context, self-reflection triggered by uncertainty, and always-on assistants with scheduled checks or reminders.
- Inference-time compute methods allocate additional reasoning to planning, reflection, or refinement, but remain reactive because computation is triggered after a user request.
3 Method
ProAct forms a closed loop that updates memory, predicts future needs, allocates idle-time computation, acquires grounded evidence, and decides how prepared artifacts should be delivered.
- Foreground interactions update persistent memory, which conditions prediction, acquisition, and delivery decisions during the next idle interval.
- Proactive policy: The policy jointly selects candidates, allocates budget, generates artifacts, and assigns delivery decisions under interruption, computation, and hallucination constraints.
- Future-State Prediction: The predictor generates candidate needs from dialogue history and memory, including user profiles, summaries, stored artifacts, unresolved goals, and memory gaps.
- Future-State Prediction: Each candidate is represented with an anticipated need, grounding rationale, prediction confidence, and retrieval plan.
- Idle-Time Acquisition: Candidates receive a value score combining relevance, knowledge gap, incremental value, and timeliness, with acquisition restricted to candidates meeting the value threshold.
- Idle-Time Acquisition: Accepted candidates reuse sufficient memory coverage, search for missing subtopics when coverage is partial, or undergo iterative decomposition and evidence checking when coverage is low.
- Delivery and memory: Generated artifacts include provenance and are pushed, queued, or stored silently according to expected utility and interruption cost before being written back to memory.
4 ProActEval
ProActEval is a controlled benchmark for memory-grounded anticipation, using fact sheets and ordered, predictable user needs across diverse domains and cognitive archetypes. Its protocol measures efficiency, factual integrity, and coverage.
- ProActEval contains 200 scenarios across 40 domains, with fictional entities, scenario-specific fact sheets, and predictable need chains.
- Evaluation protocol: The benchmark spans five cognitive archetypes covering memory, gap resolution, dependency reasoning, handoff consistency, and readiness follow-through.
- Scenario construction: Scenarios are generated from manually designed seeds by separating fact-sheet construction from ordered need-sequence generation for auditable grounding and predictability.
- Scenario construction: Automatic checks enforce identifiers, fact references, predictability links, turn order, and reveal-group consistency before manual review.
- Evaluation protocol: Each scenario uses an ordered need sequence and a fact sheet, while proactive coverage can cause the simulator to skip already addressed future needs.
- Metrics: T80 and T100 measure turns to 80% and 100% must-have coverage, while User Effort, Fact Accuracy, Hallucination Rate, and coverage metrics assess complementary outcomes.
5 Experiments
The experiments compare reactive assistance, unguided idle-time acquisition, and prediction-guided ProAct across proactive assistance, memory, and search-budget settings. Results indicate that predictive direction improves proactive coverage and efficiency beyond idle-time computation alone, while larger search budgets incur diminishing returns.
- Main Proactivity Evaluation: The evaluation compares Reactive, Undirected Idle, and Directed Idle across 200 ProActEval scenarios to measure proactive assistance and predictive direction.Reactive disables both proactive modules; Undirected Idle removes predictive direction; Directed Idle is the full ProAct configuration.
- Main Proactivity Evaluation: Directed Idle improves all non-cost metrics over both baselines, including turn efficiency, proactive coverage, and factual grounding.The reported comparison covers coverage, anticipation, factual integrity, and compute cost.
- Ablation Study: Adding predictive direction reduces T100 by 14.1% and User Effort by 10.7% relative to Undirected Idle, while reaching 0.428 Anticipation Recall.Undirected Idle spends 69.8k active tokens per scenario but only slightly improves over Reactive, reducing T100 by 0.9% and User Effort by 1.1%.
- Comparison with ProactiveAgent: ProAct anticipates 703 of 1,572 predictable needs, reaching 0.447 judge-labeled Anticipation Recall versus ProactiveAgent’s 32 needs and 0.020.The comparison uses judge-labeled Anticipation Recall because ProactiveAgent does not expose structured delivered fact IDs.
- Memory Evaluation: ProAct achieves the strongest MemBench reflective accuracy at both context lengths, improving from 0.742 to 0.843 at 10k tokens and from 0.833 to 0.863 at 100k tokens.The evaluation uses the reflective participation setting with Qwen2.5-7B-Instruct.
- Search Budget Analysis: On a matched 50-scenario subset, increasing k raises Directed Idle’s Anticipation Recall from 0.253 at k = 4 to 0.432 at k = 16, but creates a cost–efficiency trade-off.Directed Idle has lower T100 and User Effort than Undirected Idle at every matched budget; Figure 3 also reports active-token cost.
6 Conclusion
ProAct converts idle intervals into grounded preparation using persistent memory, Future-State Prediction, and Idle-Time Acquisition. Across ProActEval and MemBench, it improves proactive efficiency, coverage, factual integrity, and reflective memory accuracy, while evaluation remains bounded by synthetic closed-world conditions and other stated limitations.
- Conclusion: ProAct uses persistent memory, Future-State Prediction, and Idle-Time Acquisition to prepare for likely future needs during idle intervals.The architecture turns idle time into a structured cycle of anticipation and learning.
- Conclusion: Across ProActEval and MemBench, ProAct improves proactive efficiency, coverage, factual integrity, and reflective memory accuracy.The conclusion also reports diminishing returns as larger acquisition budgets increase active-token cost.
- Limitations: The evidence comes from a closed-world synthetic benchmark and depends on an LLM judge and a value-aware delivery gate, limiting deployment guarantees.The authors note that real deployments would need user controls, rate limits, and ongoing monitoring.
- Evaluation Scope: ProActEval keeps ground-truth user-needs graphs and related annotations hidden from systems under test, using them for evaluation instead.Runtime systems may use the user profile, fact sheet, and conversation history but not gold needs, grounding IDs, predictability links, or reveal-group annotations.
B Scenario Data Example
The finance_basic_01 scenario models predictable financial-planning needs with fictional facts, user-specific context, and an operational pipeline that anticipates, acquires, and delivers information. Its trace shows Directed Idle completing all needs in fewer turns than Reactive without coverage loss.
- Scenario structure: The scenario contains 28 facts and 12 user needs across 8 reveal groups, with fictional entities throughout.
- Scenario structure: The simulated user is a 23-year-old entry-level analyst seeking simple explanations about saving and investing.
- Scenario structure: Predictable_after dependencies link earlier needs to later anticipation targets, such as 401k-match information preceding enrollment and vesting needs.
- Operational pipeline: Future-State Prediction generates candidates from current dialogue and persistent memory, while memory-gap candidates address stale or weakly supported information.
- Operational pipeline: Candidates are scored for idle-time acquisition, with immediate acquisition eligible at S(z) ≥60; lower-scoring candidates are queued or stored.
- Trace outcome: Directed Idle completes all 12 needs in 6 turns versus 9 under Reactive, a 33% reduction with no coverage loss.
H Compute and Resource Accounting
The evaluation combines compute disclosures, benchmark adaptations, memory-layer details, and scenario traces to assess proactive behavior. Results show substantial gains in representative cases, while domain-level regressions demonstrate that delivery gates remain important.
- Resource accounting: The reported MemBench 100k run used 2,323 local model calls and 5.38M total model tokens on an Apple-silicon laptop.
- Evaluation setup: The ProactiveAgent comparison uses a GPT-4o framework adapter across all 200 scenarios and 1,685 turns, exposing scenario facts at every turn.
- Evaluation setup: ProAct anticipates 703 of 1,572 predictable needs, compared with 32 for ProactiveAgent under the judge-labeled cross-system metric.
- Memory accounting: The memory layer maintains profiles, facts, summaries, and artifacts, while incremental extraction records updates that support persistent grounding for prediction.
- Domain-level outcomes: The few domains with non-negative ∆T100 illustrate that additional proactive content can change the closed-loop trajectory even when final coverage remains high.
- Scenario traces: In insure_denial_language_05, Directed Idle reduces T100 from 11 to 5 and user effort from 10 to 7 while increasing total coverage from 0.769 to 1.000.
- Scenario traces: In museum_conservation_priority_queue_05, Directed Idle worsens T100 from 6 to 11 and total coverage from 1.000 to 0.500 despite anticipation recall of 0.300.
N MemBench Detailed Results
MemBench results are reported through reflective-accuracy, scenario-level accuracy, and memory-operation efficiency tables. The supplied passages identify these result categories but do not provide their numerical table entries.
- Reflective accuracy: Table 14 reports overall reflective accuracy on MemBench alongside aggregate reflective-accuracy baselines.
- Reflective accuracy: Table 15 reports ProAct’s reflective memory accuracy by scenario type.
- Operation efficiency: Table 16 reports memory-operation efficiency in seconds per operation, including read and write latency.
O Failure Mode Analysis
ProActEval exposes several failure modes: proactive context can interfere with reactive answers, anticipation does not always reduce effort, and unguided or over-budget search can create low-value outputs. Gains also depend on scenario structure, while deployment raises privacy concerns.
- Reactive compatibility regression: In 6 of 200 scenarios (3.0%), Directed Idle has smaller final must-have coverage than Reactive because proactive context can compete with the reactive answer.The regression can also shift generation toward less relevant facts.
- Precision–recall decoupling: Directed Idle records nonzero anticipation recall in 192 of 200 scenarios, but user effort does not decrease in 82 of those scenarios.Anticipation can arrive too late or duplicate information the user would have requested in the same turn.
- Low-value push pressure: At k = 16, Directed Idle has greater anticipation recall than at k = 4 but does not monotonically improve T100.Additional searches can introduce low-value pushes, change memory state, and alter the closed-loop conversation trajectory.
- Search direction failure: Undirected Idle spends substantial active-token budget for only small gains over Reactive because exploration lacks direction about which future need the search should serve.The failure reflects missing predictive guidance rather than insufficient search alone.
- Opportunity and fragmentation sensitivity: High-opportunity scenarios reduce user effort by 1.48 turns on average, compared with 0.86 turns for medium-opportunity and 1.10 turns for low-opportunity scenarios.Medium- and low-fragmentation scenarios show larger average gains than high-fragmentation scenarios.
- Privacy and surveillance risk: Persistent memory and future-need prediction can create privacy risks when systems infer sensitive needs or monitor behavior without clear consent.The paper uses synthetic scenarios and stores provenance so retrieved evidence can be audited; real deployment requires additional controls.
Q Structural Analysis Details
The structural analysis stratifies scenarios by predictable-need opportunity and topic fragmentation, then compares Directed Idle with Reactive using user-effort deltas. Results favor prediction-guided exploration, especially in medium-fragmentation scenarios, but the strongest cell is small and should be interpreted cautiously.
- Stratification criteria: Proactive opportunity is high when at least 70% of needs have a predictable_after link, medium at 55–70%, and low otherwise.Topic fragmentation is high with at least 10 reveal groups, medium with 8–9, and low with at most 7.
- Comparison measure: Table 17 reports Directed Idle minus Reactive user-effort deltas, where negative values indicate fewer explicit user turns.The table is organized by proactive opportunity and topic fragmentation, with scenario counts included.
- Structural results: The strongest cell is high opportunity with high fragmentation (∆UE = −2.33), but it contains only three scenarios and requires cautious interpretation.The more stable pattern appears in medium-fragmentation scenarios, where prediction-guided exploration reduces user effort across all opportunity levels.
- Structural results: High-fragmentation scenarios still improve on average, but the gain is smaller at −0.76 user turns, consistent with disconnected topic clusters limiting next-need prediction.This contrasts with the broader cross-opportunity improvement observed for medium-fragmentation scenarios.
R Per-Archetype Detailed Results
Per-archetype analyses compare Reactive, Undirected Idle, and Directed Idle across ProActEval, while the paper also documents benchmark composition and knowledge-lifecycle processing. The detailed results emphasize archetype-specific benefits and prediction-guided search.
- Per-archetype comparison: Table 18 reports Reactive-to-Directed Idle deltas for T100, user effort, and hallucination across archetypes.Negative deltas indicate improvement for each of these metrics.
- Archetype results: Trace and Dependency Reasoning benefits most in convergence speed, while Handoff and Consistency Control shows the largest user-effort reduction.The paper associates these patterns with explicit causal or temporal chains and useful preparation windows during topic transitions.
- Experimental conditions: Reactive disables both Future-State Prediction and Idle-Time Acquisition; Undirected Idle enables acquisition without predictive direction; Directed Idle enables both modules.These conditions isolate the contribution of predictive direction and idle-time acquisition.
- Budget-scaling experiment: At matched search budgets k ∈{4, 8, 12, 16}, the difference between Directed Idle and Undirected Idle estimates the value of predictive direction.Matching budgets controls search volume when comparing the two idle-compute conditions.
- Knowledge lifecycle: The knowledge lifecycle applies exact hash matching, vector near-duplicate search, and LLM arbitration before adding, skipping, replacing, or merging knowledge.Merged records preserve provenance through merged_into and merged_from pointers.
- Benchmark composition: ProActEval summarizes needs per scenario, facts per scenario, macro-domain coverage, and predictability structure.Figure 4 presents these four composition properties across its panels.