Source-linked AI summary
Clarify the User or Verify the World? Uncertainty Routing for Proactive Agents
Zhaofeng Li, Xuan Zhang, Xiaokui Xiao, Yang Deng
TL;DR
Tool-using agents must determine whether missing information belongs with the user or the external world, not only whether they are uncertain. PROUR decomposes action uncertainty and trains a shared query generator to acquire information from the routed source. On τ-bench it improves average success rate while using fewer interaction steps, and it generalizes without retraining to stronger task agents and τ^3-bench.
Problem
Proactive approaches often specialize in user clarification or environment verification without explicitly determining the appropriate information source for each decision.
Source: §1 · pp. 1–2Method
PROUR decomposes action uncertainty into user-side ambiguity and residual world-side uncertainty, routes among ACT, CLARIFY, and VERIFY, and trains a shared query generator with mode-conditioned information gain.
Source: §1 · p. 2, §3.2 · p. 5Results
4.57% higher average success rate than the strongest prior method, with 2.17 fewer interaction steps on τ-bench.
Source: §1 · p. 3, §5 · p. 10Takeaways & Limitations
The results support source-aligned uncertainty routing as a principled approach for proactive agents, with generalization to stronger task agents and τ^3-bench without retraining.
Source: §1 · p. 3, §5 · p. 10
Abstract
from arXiv · showhide
Tool-using LLM agents must decide not only whether additional information is needed, but also which source can resolve the uncertainty. Existing proactive approaches often specialize in either user clarification or environment verification, without explicitly determining the appropriate information source for each decision. We formulate this problem as uncertainty routing among ACT, CLARIFY, and VERIFY, and propose PROUR, a proactive uncertainty routing framework. PROUR decomposes action uncertainty into disagreement across plausible user-goal interpretations, which signals user-side ambiguity, and the entropy remaining within each interpretation, which signals missing world-side evidence. To acquire information from the routed source, a query generator is trained with a mode-conditioned information-gain reward, targeting user-goal identification under CLARIFY and next-action identification under VERIFY. On $\tau$-bench, PROUR achieves 28.17% average success rate across retail and airline, outperforming the strongest prior method by 4.57% while using 2.17 fewer interaction steps. The learned policy further generalizes to stronger task agents and transactional domains of $\tau^3$-bench without retraining, demonstrating the benefit of source-aligned uncertainty resolution for proactive agents.
1 INTRODUCTION
Tool-using agents must decide both whether information is missing and whether to obtain it from the user or the world. PROUR routes uncertainty between acting, clarifying, and verifying, and reports higher average success with fewer interaction steps on τ-bench.
1 INTRODUCTION
- 89.1% of 19,022 τ-bench events concern world-side information, compared with 10.9% concerning user-side information.Clarification-only approaches can ask users for facts available from tools, while verification alone can miss user preferences.Source: §1 · p. 2, Figure 2 · p. 2
- Uncertainty routing addresses whether missing information should come from the user or the external world, beyond deciding whether to seek information.User preferences require clarification, while system state, policy constraints, and action validity should be verified using tools or other grounded evidence.Source: §1 · pp. 1–2
- PROUR decomposes action uncertainty into user-side and world-side uncertainty to route agents among ACT, CLARIFY, and VERIFY.Its mode-conditioned information-gain objective trains one query generator to acquire information from users or tool-grounded evidence.Source: §1 · p. 2
- On τ-bench, PROUR improves average success rate by 4.57% with 2.17 fewer interaction steps; without further retraining, it achieves 29.00% average success rate on τ 3-bench.Source: §1 · p. 3
2 RELATED WORK
Prior work addresses uncertainty decomposition, proactive clarification, or tool-grounded verification, but typically assumes an information source in advance. PROUR instead routes uncertainty to a source and adapts a shared policy to that route.
Uncertainty Decomposition for LLM Agents
- Unlike classical BALD, which ensembles over model parameters, PROUR ensembles over candidate goal interpretations to distinguish user-side ambiguity from residual world-side uncertainty.The paper treats AU and EU as operational source labels, not universal Bayesian identities.Source: §2 · p. 3
Proactive Clarification and Verification
- Prior clarification and verification methods typically assume a predefined information source, whereas PROUR routes uncertainty first and adapts a shared policy to CLARIFY or VERIFY.Prior approaches include asking users based on expected value or predictive uncertainty and using tool-grounded critique or verification.Source: §2 · p. 3
3 METHODOLOGY
PROUR routes tool-using agents among ACT, CLARIFY, and VERIFY according to whether uncertainty comes from an ambiguous user goal or missing world evidence. A shared query generator then acquires information from the routed source using a mode-conditioned information-gain objective.
3 METHODOLOGY
- PROUR separates user-goal ambiguity from world-side uncertainty, routing the agent among CLARIFY, VERIFY, and ACT before a shared Query Generator acquires information from the selected source.CLARIFY targets the user; VERIFY targets grounded world evidence.Source: §3 · p. 3, §3.2 · p. 4, §3.3 · p. 5
3.1 PROBLEM FORMULATION
- The interaction formulation distinguishes user-side information, such as underspecified intent, from world-side information, such as environment state, policy constraints, or action validity.The agent may act on available information, ask the user, or verify the world.Source: §3.1 · pp. 3–4
3.2 UNCERTAINTY DECOMPOSITION AND ROUTING
- Aleatoric uncertainty measures disagreement in next-action distributions across plausible goal interpretations, while epistemic uncertainty measures residual action entropy under a fixed interpretation.High aleatoric uncertainty routes to CLARIFY; high epistemic uncertainty routes to VERIFY, with clarification prioritized when user ambiguity remains.Source: §3.2 · pp. 4–5
3.3 MODE-CONDITIONED INFORMATION-GAIN REWARD
- A shared query generator is trained with a mode-conditioned information-gain objective: CLARIFY targets identification of the ground-truth user goal, while VERIFY targets identification of the reference next action.The reward estimates the change in target likelihood after incorporating the query–response pair; the responder, agent, and decomposer remain frozen during query-generator training.Source: §3.3 · pp. 5–6
4 EXPERIMENTS
The experiments evaluate PROUR on τ-bench and test its training, calibration, end-to-end performance, generalization, and interaction dynamics. PROUR outperforms the base agent and shows gains across stronger agents and transactional domains, while decomposer-call budgets trade cost against performance.
4 EXPERIMENTS
- The evaluation trains on 500 τ-retail trajectories and tests on held-out retail tasks and airline tasks, with airline excluded from training and calibration.Qwen3-1.7B is the only optimized module; Qwen3-8B implements the default frozen modules.Source: §4 · p. 6
4.2 MAIN RESULTS AND ABLATION STUDIES
- 28.17% average success and 21.33 interaction steps: PROUR leads τ-bench results, versus 17.40% and 28.90 steps for the base agent.Clarification-only methods do not acquire grounded world-side evidence; VerifiAgent requires 43.01 steps on average.Source: Table 1 · p. 7, §4.2 · pp. 7–8
4.1 TRAINING DYNAMICS AND ROUTING CALIBRATION
- The query generator’s posterior target belief exceeds its prior during training, and mean reward rises from near zero to approximately 0.7 after 2,010 optimization steps.CLARIFY- and VERIFY-mode rewards improve alongside the aggregate reward without evidence of mode collapse.Source: §4.1 · p. 7
- Routing thresholds are calibrated on 240 disjoint τ-retail decision steps, with five goal interpretations per step, to maximize three-class macro-F1.The resulting thresholds are (τAU, τEU) = (0.25, 0.20).Source: §4.1 · p. 7
4.3 GENERALIZATION ACROSS AGENTS AND BENCHMARKS
- Across Qwen3-8B, Qwen3-32B, and DeepSeek-V4-Flash task agents, PROUR reaches 28.17%, 40.25%, and 48.20% average success, respectively.The trained query generator and calibrated thresholds are held fixed, and gains persist as the task agent becomes stronger.Source: Table 3 · p. 8, §4.3 · pp. 8–9
- On τ 3-bench, PROUR achieves 29.00% average success and 29.25 total steps without retraining or threshold recalibration, ahead of VerifiAgent’s 27.23% success.The evaluation covers transactional text domains: retail, airline, and telecom; banking and voice are outside the current action space.Source: Table 4 · p. 8, §4.3 · p. 9
4.2 MAIN RESULTS AND ABLATION STUDIES
- The trained query generator reaches 28.17% average success, outperforming prompted generators while the task agent, decomposer, and thresholds remain fixed.The LLM Router, Clarify-only, and Verify-only variants each underperform the full method on average; Verify-only achieves 30.0% on airline but 15.7% on retail.Source: Table 2 · p. 7, §4.2 · p. 8
4.4 FURTHER DISCUSSIONS
- With a decomposer-call budget of 16, interaction steps reach 20.42 as average success rises from 18.37% to 29.42%; larger budgets add cost without further benefit.PROUR’s total uncertainty falls from 0.58 to 0.11 over 16 steps; EU decreases from 0.26 to 0.08, while residual AU is expected from sampling variability.Source: §4.4 · p. 9, Figure 5 · p. 9, Figure 6 · p. 9
5 CONCLUSION
PROUR routes action uncertainty among ACT, CLARIFY, and VERIFY, then uses a mode-conditioned information-gain objective to acquire user-side information or grounded world-side evidence. It improves τ-bench success with fewer interaction steps and generalizes without retraining to stronger task agents and τ^3-bench.
5 CONCLUSION
- 28.17% average success rate, up from 23.60%, with fewer interaction steps on τ-bench.The framework generalizes without retraining to stronger task agents and τ^3-bench.Source: §5 · p. 10
- PROUR decomposes action uncertainty into aleatoric and epistemic uncertainty to route agents among ACT, CLARIFY, and VERIFY.Uncertainty is resolved by acquiring information from the routed source.Source: §5 · p. 10
- A mode-conditioned information-gain objective trains a query generator to acquire user-side information under CLARIFY and grounded world-side evidence under VERIFY.Source: §5 · p. 10
AI USE STATEMENT
The authors used generative AI tools for manuscript editing and assistance with trajectory annotations and initial uncertainty labels. They manually reviewed and corrected those annotations and labels.
AI USE STATEMENT
- Generative AI tools assisted with drafting and editing parts of the manuscript for clarity, readability, and presentation.Source: AI USE STATEMENT · p. 10
- AI-assisted annotation covered τ-bench trajectory case taxonomy and initial uncertainty labels for development-set routing analysis.The authors manually reviewed and corrected all such annotations before calibration and analysis.Source: AI USE STATEMENT · p. 10
- The authors state that generative AI was not used to develop the conceptual framework, propose hypotheses, formulate mathematical claims, or implement the method.They reviewed all AI-assisted work and take responsibility for the final content, analyses, claims, and results.Source: AI USE STATEMENT · p. 10
A.1 UNCERTAINTY CASE TAXONOMY
The taxonomy classifies uncertainty at individual decision steps in historical retail and airline τ-bench trajectories, distinguishing user-side from world-side sources. Most recorded events are world-side.
A.1 UNCERTAINTY CASE TAXONOMY
- The taxonomy assigns decision-step events to nine source-relevant cases after uncertainty annotations are manually verified, corrected, normalized, and consolidated.The annotations distinguish uncertainty types and record the uncertain object, available evidence, resolution outcome, and a provisional case label.Source: §A.1 · p. 14, Table 5 · p. 14
- The 19,022 reported events are step-level rather than trajectory-level, so a trajectory can contribute multiple events.Source: §A.1 · p. 14, Table 5 · p. 14
- 16,947 of 19,022 decision-step events (89.1%) were world-side, while 2,075 (10.9%) were user-side.The authors conclude that clarification alone addresses only a small portion of observed uncertainty; most cases require world-information retrieval or verification.Source: §A.1 · p. 14
A.2.1 IMPLEMENTATION DETAILS
The implementation details describe trajectory collection, training-instance construction, model configuration, optimization, baselines, and evaluation. The held-out airline routing figure compares calibrated PROUR with a prompted router.
A.2.1 IMPLEMENTATION DETAILS
- 500 retail-training trajectories were collected with a Qwen3-8B task agent, Qwen3-32B teacher query generator, and Qwen3-8B user simulator.Teacher queries elicited feedback and enriched histories but were not retained as SFT targets; optimization queries came from the Qwen3-1.7B policy’s on-policy rollouts.Source: §A.2.1 · pp. 14–15
- Step-level extraction yielded 3,520 CLARIFY or VERIFY instances: 2,816 training and 704 validation examples.CLARIFY targets the normalized user goal; VERIFY targets the reference next action, while ACT states are omitted.Source: §A.2.1 · p. 15
- The query generator was the only optimized component; the task agent, decomposer, user simulator, and verifier were frozen Qwen3-8B models.Training used DAPO under VERL, with two 80 GiB NVIDIA H100 GPUs.Source: §A.2.1 · p. 15
- Optimization used 2,010 steps, with 128 rollouts per training step and six retained candidates per prompt.The setup used four PPO epochs and the final checkpoint for evaluation.Source: §A.2.1 · p. 15
- Baselines included five information-seeking approaches, and controlled ablations replaced the decomposer or disabled either routing branch.Additional comparisons replaced the trained query generator with prompted Qwen3-8B, Qwen3-32B, or DeepSeek-V4-Flash.Source: §A.2.1 · p. 15
- Evaluation used 115 held-out retail tasks and 50 out-of-distribution airline tasks; Average is the unweighted mean across domains.Figure 7 compares calibrated PROUR with a prompted Qwen3-8B router on the same disjoint airline test set; rows are ground-truth modes, columns predicted modes, and cells give counts and row-normalized percentages. τ^3-bench uses release v1.0.1.Source: §A.2.1 · pp. 15–16, Figure 7 · p. 16
A.2.2 UNCERTAINTY ESTIMATION AND ROUTING CALIBRATION
PROUR estimates uncertainty from sampled goal interpretations and candidate actions, then calibrates routing thresholds on a development set. Its routing performance and downstream effects depend on matching calibration to the model.
A.2.2 UNCERTAINTY ESTIMATION AND ROUTING CALIBRATION
- AU and EU scales are model-specific, so changing the goal-proposal or action-scoring backbone can require recalibrating both routing thresholds.Replacing the model can change action distributions and resulting uncertainty scores.Source: §A.2.2 · p. 16
- Five sampled goal interpretations and four structured actions per interpretation form candidate action distributions used to estimate AU and EU.Invalid actions are discarded without resampling, and valid actions are canonicalized and deduplicated before scoring.Source: §A.2.2 · p. 16
- Thresholds are calibrated by grid search for three-class macro-F1 on 240 manually reviewed retail development decisions, then frozen for airline evaluation.The development labels comprise 60 CLARIFY, 90 VERIFY, and 90 ACT decisions; a separate 240-step airline set is used for held-out evaluation.Source: §A.2.2 · p. 16
- PROUR reaches 74.04% macro-F1 on held-out airline routing, versus 54.24% for a prompted router that directly predicts the three modes.The thresholds were selected on retail development data and applied unchanged to airline; Figure 7 reports the held-out comparison.Source: §A.2.2 · p. 17
- Model-specific calibration yields similar macro-F1 across GLM-4-32B and Qwen3-8B, but their class-level strengths differ by routing mode.GLM-4-32B scores higher on ACT and CLARIFY, while Qwen3-8B scores higher on VERIFY.Source: §A.2.2 · p. 17, Table 6 · p. 17
- Model-matched thresholds raise average success for both query generators, while the trained generator's gain comes with more intervention steps.With the trained generator, success rises from 33.26% to 37.00%; with the prompted generator, it rises from 29.65% to 31.95%.Source: §A.2.2 · p. 17, Table 7 · p. 17
A.2.3 PERFORMANCE–COST TRADE-OFF
The appendix examines the routing frequency trade-off and details PROUR's training and prompt protocols. It describes mode-conditioned query optimization, grounded responders, and the prompts and data procedures used to implement and evaluate the system.
A.2.3 PERFORMANCE–COST TRADE-OFF
- 28.17% average success rate and 21.33 interaction steps make routing every decision the best overall balance among the evaluated frequencies.Routing every second step yields 27.32% success and 25.26 steps; every-fourth and every-eighth-step routing reduces runtime but lowers success to 23.00% and 19.55%.Source: §A.2.3 · pp. 17–18, Figure 8 · p. 18, Table 8 · p. 18
A.3 MODE-CONDITIONED INFORMATION-GAIN TRAINING FRAMEWORK
- Mode-conditioned information-gain training uses one query generator to improve identification of the user goal in CLARIFY and the reference next action in VERIFY.The generator also scores its own belief before and after the query–response exchange; strict user and verifier responders provide mode-specific responses, with the verifier limited to grounded evidence.Source: §A.3 · pp. 18–19
A.4.1 UNCERTAINTY DECOMPOSER PROMPT TEMPLATES
- The decomposer samples plausible goal interpretations, samples structured next actions conditioned on them, and teacher-forces each canonical action for scoring.The prompts require one goal interpretation and one strict-JSON action at a time.Source: §A.4.1 · pp. 19–20
A.4.2 QUERY GENERATOR PROMPT TEMPLATES
- The query generator uses separate CLARIFY and VERIFY templates, with belief scoring targeting normalized user intent or the reference next tool action.VERIFY queries request answerable world evidence from a separate verifier; CLARIFY asks one concise question about user intent.Source: §A.4.2 · pp. 21–22
A.4.3 USER SIMULATOR AND VERIFIER PROMPT TEMPLATES
- The verifier is restricted to read-only tools and must ground its response in retrieved, computed, or policy evidence, returning insufficient evidence otherwise.Evaluation uses the standard τ-bench user simulator rather than reusing the strict user prompt used for training.Source: §A.4.3 · pp. 22–23
A.4.4 DATA PREPARATION AND CALIBRATION PROMPT TEMPLATES
- Data preparation normalizes raw instructions into first-person, flat natural-language goals, while calibration labels distinguish factual lookup, user preference, and sufficient information to act.GPT-5.5 labels decisions under a fixed rubric before grid search.Source: §A.4.4 · pp. 24–25
A.4.5 ABLATION PROMPT TEMPLATES
- The prompted-router ablation replaces goal-conditioned decomposition with a direct three-way mode decision while keeping the downstream query generator unchanged.The prompt distinguishes ACT, CLARIFY for user-owned ambiguity, and VERIFY for evidence obtainable from the world.Source: §A.4.5 · p. 26