acceptodds
Under review as a conference paper at ICLR 2027

Clarify the User or Verify the World? Uncertainty Routing for Proactive Agents

Abstract

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 -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 -bench without retraining, demonstrating the benefit of source-aligned uncertainty resolution for proactive agents.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM agents, tool-using agents, proactive agents, uncertainty routing, clarification question generation, information gain

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jrj5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a proactive tool-using agent that must decide not only whether it is uncertain, but also whether that uncertainty should be resolved by the user or by the world. Its core idea is presented in the phrase “A proactive agent must determine not only whether it is uncertain, but also whether that uncertainty should be resolved by the user or by the world. We refer to this problem as uncertainty routing.” The proposed PROUR framework has two parts: The Decomposer samples several plausible interpretations of the user’s goal from the interaction history, evaluates the next-action distribution with the history, and splits action uncertainty into two source-aligned components (AU from user and EU from world). Calibrated thresholds on these two quantities route the agent to ACT, CLARIFY or VERIFY. To make routing determines where to look and what to ask, a mode-conditioned Query Generator is trained using DAPO. On τ -bench, the method improves success over clarification-only and verification-only baselines with fewer interaction steps.

Critical strengths
  • The motivation is convincing. Through the example of Figure 1, authors show that treating clarification and verification as two answers to the same question is a useful way.
  • I also liked the idea that ambiguity can be measured through disagreement between next-action distributions across plausible interpretations of the user’s goal. Thus, high AU indicates userside ambiguity that should be resolved through CLARIFY.
  • Experiments section is well-structured. All baselines are reimplemented on the same backbone. The ablations are also well chosen.
Critical weaknesses
  1. To me, the core insight seems not new. The idea of “Some missing information can only come from the user while other information can be obtained from the environment” is already present in prior work such as Ask-before-Plan (2024). Ambig-SWE (2026) makes the distinction explicit by separating information recoverable from the codebase from information only the user can provide. From Assumptions to Actions (2026) chooses between communicating and acting based on assumptions about both the environment and the partner. What PROUR adds is making this choice at every step from an uncertainty estimate. The related work should clarify the similarities and differences with these approaches, and the claim to “identify uncertainty routing as a fundamental problem for tool-using agents” should be toned down.
  2. The decomposition itself follows the input clarification view of “Decomposing Uncertainty for Large Language Models through Input Clarification Ensembling, 2024”. This means the decomposition itself is not a methodological contribution. What is new in this part is only the reading of each term as an information source.
  3. I don’t agree with the assumption that “high EU should mean missing world-side evidence”. The entropy over next actions under a fixed interpretation can be high for reasons unrelated to the world. For example, when the user ask for “the latest option after 6 pm”, retrieving the reservation first or searching flights first are both correct (so the entropy is high although nothing is missing). Conversely, if the sampled interpretations all make the same assumption, AU stays low even when the goal is ambiguous. Authors must clarify this with appropriate experiments.
  4. The mode-conditioned information-gain objective is essentially that of IG-clarifier. The paper states that “We generalize the belief-update estimator and group-relative DAPO optimization of Deng et al. (2026) to the mode-conditioned source and target”, IG-clarifier also uses the same small Qwen backbone and a strict user simulator. So, calling it mode-conditioned does not change much. The only addition is a second reward target.
  5. I am not sure VERIFY is really a separate source. Paper states that “ACT may itself include ordinary tool calls”, and the verifier uses the same read-only tools the agent already has. So it is hard to tell whether the gains come from routing or simply from extra tool use.
Questions for authors

Q1. I would like to see whether high-EU steps are ones where the verifier returns decision-changing evidence rather than “insufficient evidence”. Also, what happens when the missing piece is a user-held fact that the interpretation does not specify (e.g., which payment method to use for the fare difference)? Such step would raise EU and be routed to VERIFY, although only the user can resolve it. Q2. In Hou et al. (2024), the expected conditional entropy reflects what the model itself does not know. Why should this term be read as evidence retrievable from the world, rather than as the model’s own uncertainty about the task or the presence of several valid actions? Q3. Apart from adding the reference next action as a second target, what is conceptually new in the mode-conditioned information-gain objective compared to IG-clarifier? Q4. How does uncertainty routing differ in approach from prior work that already separates information only the user can provide from information obtainable from the environment, or that chooses between communicating and acting by expected utility?

11 comments
AuthorsOpenReview ↗

Response to Reviewer jrj5 (1/4)

Thank you for the constructive review: we ran the experiments you asked for (W3, W5, Q1, Q2; new Tables R1–R5, with R1 and R4 added to Appendix A.6 of the revised PDF as Tables 13 and 14) and narrowed our novelty claims as you suggested (W1, W2, W4, Q3, Q4).

[W3/Q2/Q1] On reading high EU as missing world-side evidence.

Sorry for the confusion. Indeed, high EU does not only involve missing world-side evidence. To clarify your concern, we measured what high-EU steps need. At 72% of them, world-side evidence is missing: a lookup, a policy check or a calculation decides the next action. Only 13% are cases with several valid actions, and verifying there makes no detectable difference to success. 

We labeled 400 VERIFY steps from PROUR's τ-bench runs by what each step needed. GPT-5.5 labeled every step and an author checked every label, both without seeing the replay results; a second author labeled 100 steps independently (Cohen's κ = 0.79). We then replayed each step with and without the verification. Success is measured on the episodes replayed from each sampled step (3 runs per branch).

Table R1. What high-EU (VERIFY) steps need (n = 400).

What the step neededProportionVerifier: "insufficient evidence"Agent's decision changedSuccess rate: verify → act directly
Missing world-side evidence72%4%33%34.4 → 23.9
Several valid actions, nothing missing13%2%4%31.9 → 31.4
A fact only the user holds6%63%13%26.7 → 21.1
Other (the model is simply unsure)9%14%22%30.6 → 28.1
All VERIFY steps100%8%27%33.3 → 25.1

Several valid actions (W3). These steps are rare, and verifying there changes success by 0.5 points, which is not significant. The verifier is read-only (L287–289), and in 80% of these steps its lookup is one of the valid actions. In your example ("the latest option after 6 pm"), retrieving the reservation and searching flights are both lookups: the verifier does one, and the agent goes straight to the other. In most of these steps, the verification replaces a step rather than adding one.

The model's own uncertainty (Q2). Our definition of world-side uncertainty already includes most of it. In the taxonomy of Appendix A.1 (Table 5; L718–719), about a third of the epistemic events are tool-argument verification (2,356), calculation verification (1,751) and plan-completion verification (1,243). There the model is unsure about its own arguments, arithmetic or progress. A grounded check resolves this, and the user cannot. Table R5 (part 2/4) shows what VERIFY returns in evaluation. 38% retrieve state not yet in the history. The other 54% re-check evidence already in the history: they quote the policy, recompute a value or check subgoal completion, but it is not free reflection: the verifier must cite a tool result, a policy quote or a computation with its inputs, or answer "insufficient evidence" (L1363–1371). Either way, the user cannot supply it. We route it to VERIFY because a grounded check resolves most of the uncertainty and the user cannot (Tables R1 and R5); “world-side” in the paper means exactly this (L195–197), which is broader than "retrieved by a new tool call".

Decision-changing evidence (Q1). The verifier answers "insufficient evidence" in only 8% of VERIFY steps, consistent with Table 11. Its answer changes the agent's next decision in 27% of steps (33% where evidence is missing). Skipping the verification lowers the success of the replayed episodes by 8.2 points.

We revised §3.2 (L229–233) to say that high EU mostly reflects missing world-side evidence, in the sense defined above, and that part of EU cannot be reduced by any source (several valid actions).

AuthorsOpenReview ↗

Response to Reviewer jrj5 (2/4)

[W3/Q1] On user-side information that AU could miss

A missing user-held fact raises AU, not EU, in most cases, so PROUR asks the user (Q1). Goal interpretations are sampled independently (L200–201). When the user has not said how to pay, different interpretations assume different payment methods, so their next actions differ. That disagreement is AU. We built 160 steps where the next action needs a choice only the user can make: 50 payment-method cases (your example) and 110 other choices and preferences.

Table R2. When only the user holds the missing fact (n = 160).

OutcomeBase agentLLM RouterPROUR
Asked the user right away71.3%62.5%81.9%
Routed to VERIFY first, then asked–21.9%12.5%
Acted on a guessed value28.7%15.6%5.6%

The base agent has no VERIFY route; its own lookups before asking count in the first row. PROUR sends 110 of the 160 cases straight to CLARIFY, and in 21 more the agent's own reply asks the user (81.9%). The case you describe, routing to VERIFY, occurs in 20 cases (12.5%) and corrects itself: the verifier answers "insufficient evidence" or lists the options on file, the interpretations then disagree, and PROUR asks the user within two steps.

Interpretations that share one assumption (W3). We agree that this failure exists, and measured it on 120 annotated ambiguous steps.

Table R3. Ambiguous steps (n = 120).

The 5 goal interpretationsStepsRouted to CLARIFYShared assumption correct
Differ104 (87%)84 (81%)–
All make the same assumption16 (13%)1 (6%)10 of 16

As you predicted, PROUR rarely clarifies when all interpretations agree, but their shared assumption is usually correct, and a wrong assumption occurs in 6 of 120 steps (5%). With K = 10, the share of agreeing steps falls from 13% to 6%.

[W5] On whether the gain is only extra tool use

It is not. With the same verifier, tools and trained query generator, changing only when verification happens removes almost all of the gain, even though the controls make more tool calls. You are right that VERIFY uses the same read-only tools as ACT (L180–183). It differs in when it acts (only when EU is high) and in what it asks (a query trained to identify the next action, answered with a cited tool result, computation or policy quote; L1363–1371). Clarify-only is PROUR with VERIFY branch disabled.

Table R4. Who decides when to verify (τ-bench, 3 seeds).

TriggerSuccess (%)Verify turns / episodeTool calls / episodeGain over Clarify-only
None (Clarify-only)25.1508.9–
Random steps, count matched to PROUR25.614.7912.6+0.46
The task agent (verifier offered as a tool)25.326.9014.8+0.17
PROUR: when EU above its threshold28.174.7910.3+3.02

The controls keep only 15% and 6% of PROUR's gain, and the paired 95% CIs of their difference from PROUR exclude zero ([−5.1, −0.1] and [−5.4, −0.3]). The random control makes as many verifier calls as PROUR, so the gain is not an effect of extra model calls either. More verification does not help either: lowering the EU threshold from 0.20 to 0.15 raises verification from 4.79 to 7.60 turns per episode and lowers success from 28.17% to 25.18% (Fig. 8). VERIFY is also more than tool calls: 54% of VERIFY turns make none (Table R5).

What VERIFY returns (W5/Q2). We classified every VERIFY turn by the evidence its answer cites, and re-ran the agent without that turn.

Table R5. What VERIFY returns.

Evidence cited by the verifierShare of VERIFY turnsNew tool call?Next action changed
State lookup not yet in the history38%Yes41%
Policy check (quotes the domain policy)22%No36%
Computation over existing observations21%No44%
Subgoal-completion check11%No29%
Insufficient evidence8%–3%

54% of VERIFY turns make no tool call, yet they change the next action about as often as new lookups do (29–44% vs 41%). Unlike Table R1, this column also counts read-only lookups, hence 36% on average rather than 27%. §3.1 now states that VERIFY is a separate channel, opened only when EU is high, whose answer must cite a state lookup, a policy check, a computation or a subgoal check.

AuthorsOpenReview ↗

Response to Reviewer jrj5 (3/4)

[W1/Q4] On prior work that separates user-side from environment-side information

Thanks for helping us better position our work with three related works (Ask-before-Plan, Ambig-SWE, and From Assumptions to Actions), which separate information only the user has from information in the environment. We removed the "fundamental problem" claim and added all three to Related Work. As you note, PROUR's addition is a per-step choice made from an uncertainty estimate. We reframed our claim as: measuring the source from the agent's own action distribution routes more reliably than judging it.

Ask-before-Plan assigns the two sources to fixed stages: execution agents query the environment at every turn, a fine-tuned clarification agent decides whether to ask the user, and planning starts after the dialogue ends. PROUR chooses at each step among the user, the world, or neither, interleaved with actions that change state. On the same backbone, success rises from 23.60% to 28.17% and steps fall from 23.50 to 21.33 (Table 1).

Ambig-SWE finds that LLMs struggle to judge whether an input is underspecified merely through prompting, as our prompted LLM Router shows. On 240 held-out airline steps, with thresholds frozen on retail (Fig. 7). The accuracy gap is significant (McNemar p < 0.001).

PROURLLM Router
Routing accuracy73.8%53.8%
Macro-F174.0454.24
Unnecessary user questions (ACT or VERIFY steps sent to CLARIFY)927
User-side steps sent to VERIFY618

PCE (Q4) picks between communication and physical actions by expected gain minus cost, so it must enumerate assumptions and estimate a gain and a cost for each action. PROUR needs no utility model: it splits next-action uncertainty into a user part and a world part, and learns what to ask. PCE targets embodied planning with a partner, and τ-bench defines no per-action gain or cost, so we compare with the closest expected-utility method for tool-using agents, SAGE-Agent (expected value of perfect information): 21.13% vs 28.17% on τ-bench and 23.95% vs 29.00% on τ³-bench (Tables 1 and 4). A cost term like PCE's could be added to our thresholds, so the approaches are complementary.

[W2] On the decomposition of Hou et al.

Although both Hou et al. and our work use the conditional mutual-information identity [1] for decomposition, applying it to the tool-using agent is not a trivial and straightfoward adaptation. It requires re-inferring the interpretations from the multi-turn history at every step and scoring schema-validated tool calls on a shared candidate set (Appendix A.2.3), but we do not claim this as conceptual novelty. The difference is how the epistemic term is used. Hou et al. use the aleatoric term to trigger clarification (their §3.6), attribute the epistemic term to knowledge the model lacks, and do not query for it. For a tool-using agent, much of that missing knowledge is environment state that read-only tools can retrieve, so PROUR routes it to verification. Clarify-only routes on AU alone, the closest analogue of Hou et al.'s usage, and reaches 25.15% vs 28.17% on τ-bench and 24.83% vs 29.00% on τ³-bench (Tables 1 and 4). This reading of EU fails when several next actions are equally valid; Table R1 (part 1/4) shows that this case covers 13% of high-EU steps and has no detectable effect on success.

References

[1] Cover, T. M. and Thomas, J. A. Elements of Information Theory. Wiley-Interscience, 2nd edition, 2006.

Reviewer_jUfpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses how tool-using LLM agents should choose between acting, clarifying user intent, and verifying external information. The proposed framework, PROUR, decomposes action uncertainty into disagreement across plausible user-goal interpretations and residual uncertainty within each interpretation to guide this choice. A shared query generator is trained with mode-conditioned information-gain rewards. The reported experiments show that PROUR improves task success while reducing interaction steps compared with prior baselines, with additional gains observed across stronger agents and a separate benchmark without retraining.

Critical strengths

The paper is well motivated by the distinction between unclear user intent and missing external information. This distinction provides a useful perspective on two practical failure modes: burdening users with questions that tools could answer, and taking action before user preferences are sufficiently understood.

Critical weaknesses
  1. PROUR extends the input-clarification decomposition in [1] to next-action distributions for ACT/CLARIFY/VERIFY routing, and extends the information-gain training framework in [2] by conditioning the response source and reward target on the selected mode. These are meaningful but incremental adaptations of existing components. The paper does not sufficiently establish the independent value of its key additions.
  2. The paper’s computational-efficiency comparison with the baselines is incomplete. Table 1 reports interaction steps, which do not capture PROUR’s additional routing computation, while the runtime analysis in Table 8 only compares PROUR configurations. The results do not show lower overall inference cost relative to the baselines.

[1] Hou, Bairu, et al. “Decomposing Uncertainty for Large Language Models through Input Clarification Ensembling.” ICML, 2024.

[2] Deng, Mengyi, et al. “Uncertainty-Aware Clarification in LLM Agents with Information Gain.” 2026.

Questions for authors
  1. Could the authors extend Table 1 with end-to-end runtime and total model-call counts for PROUR and the baselines, including all internal routing and verification calls? Could the authors also report the uncertainty decomposer’s runtime and its share of the total execution time to quantify the routing overhead?
  2. If several next actions are equally valid—for example, when independent subgoals can be completed in any order—could the resulting high EU trigger unnecessary verification and potentially lead to task failure? How does the method handle such cases?
  3. How do the authors validate that the estimated AU reflects ambiguity in the user’s request, rather than biases or sampling variability in LLM-generated goal interpretations? Specifically, (i) is there evidence that the sampled interpretations cover plausible user goals without overrepresenting particular interpretations, and (ii) how stable are AU/EU estimates and routing decisions across repeated sampling of the same decision state and different values of KK, given the default K=5K=5?
7 comments
AuthorsOpenReview ↗

Response to Reviewer jUfp (1/4)

Thank you for the careful review. We tested each of our additions in isolation (W1), added the cost table you asked for (W2, Q1), and measured the two cases you raised (Q2, Q3). Line and table numbers refer to the revised PDF.

[W1] On the independent value of what PROUR adds to [1] and [2]

Thanks for your valuable suggestions. To reveal the independent value of what PROUR adds to [1] and [2], we evaluated each PROUR component through ablation study (Table R13). Replacing EU-triggered verification or the mode-conditioned target lowers success by 2.6 to 8.5 points. Replacing the decomposition with a prompted router lowers routing accuracy from 73.8% to 53.8%; its success gap is 2.42 points on τ-bench and 2.67 points on τ³-bench.

Table R13. Contribution of each PROUR component. Each variant replaces one component with the stated alternative; everything else is PROUR. Δ is the variant's success minus PROUR's, with its paired 95% CI where computed. Rows 2 and 4 were already in Table 1 in the original submission.

#Component → alternativeSuccess (%)Δ (95% CI)StepsSource
1Full PROUR28.17–21.33Table 1
2EU-triggered verification → route on AU only (Clarify-only)25.15−3.02 [−5.6, −0.4]23.20Table 1
3EU trigger → random steps (same verifier, count matched)25.61−2.56 [−5.1, −0.1]29.10Table R4
4AU/EU decomposition → prompted LLM Router25.75−2.42 [−6.3, 1.4]23.39Table 1
5Mode-conditioned target → goal target of [2] in both modes22.57−5.6030.18Table 6
6Mode-conditioned target → next-action target in both modes19.70−8.4734.31Table 6

The addition to [1] is routing on the epistemic term. [1] uses the aleatoric term to trigger clarification. It attributes the epistemic term to knowledge the model lacks, and does not query for it. PROUR sends that term (EU: the next-action uncertainty that remains once the goal interpretation is fixed) to tool-grounded verification, at every step. Routing on AU alone loses 3.02 points (row 2). The gain comes from the EU signal, not from having a verifier: with the same verifier, tools and number of verifications, verifying at random steps recovers only 0.46 of those 3.02 points (row 3).

Table R4. Who decides when to verify (τ-bench, 3 seeds; same verifier, tools and trained query generator in every row).

TriggerSuccess (%)Verify turns / episodeTool calls / episodeGain over Clarify-only
None (Clarify-only)25.1508.9–
Random steps, count matched to PROUR25.614.7912.6+0.46
The task agent (verifier offered as a tool)25.326.9014.8+0.17
PROUR: EU above its threshold28.174.7910.3+3.02

An agent that calls the verifier itself does no better than random verification (+0.17 vs +0.46), though both controls make more tool calls.

Against a prompted router that sees the same history and tools (row 4), the decomposition routes more accurately on 240 held-out airline steps: 73.8% vs 53.8% accuracy (macro-F1 74.04 vs 54.24; Fig. 7, L898–900; McNemar p < 0.001). It sends 9 unambiguous steps to the user, where the prompted router sends 27. On τ-bench the 2.42-point success gap has a paired CI that includes 0, but episodes are 2.06 steps shorter and contain 1.10 fewer user questions (both CIs exclude 0). On τ³-bench, which has 278 tasks against 165, we ran two more seeds for both methods (5 in total); there the success gap is significant: +2.67 points (95% CI [0.4, 4.9], p = 0.02).

The addition to [2] is that the routed mode selects the reward target. CLARIFY is rewarded for identifying the user's goal and VERIFY for identifying the next action, so each mode is trained to reduce the uncertainty that triggered it. In Table R13, replacing mode-conditioned target with the goal target of [2] in both modes loses 5.60 points and adds 8.85 steps. The reason is that the goal contains order IDs and zip codes, so a goal target also rewards asking the user for facts that tools can retrieve (Table R12).

Table R12. Share of user questions that ask for tool-retrievable facts (τ-bench; questions per episode from Table 1).

MethodUser questions per episodeTool-retrievableSuccess (%)
IG-clarifier (goal target only)2.7034%22.85
PROUR (CLARIFY)2.208%28.17

We rewrote the openings of §3.2 and §3.3 to state this lineage and exactly what is added.

References

[3] Cover, T. M. and Thomas, J. A. Elements of Information Theory. Wiley-Interscience, 2nd edition, 2006.

[4] Chaloner, K. and Verdinelli, I. Bayesian experimental design: A review. Statistical Science, pp. 273–304, 1995.

AuthorsOpenReview ↗

Response to Reviewer jUfp (2/4)

[W2/Q1] On end-to-end inference cost

As you asked, Table 10 (Appendix A.2.5; §4.4 points to it at L494–495) now reports end-to-end time and LLM calls per episode for every method. PROUR makes 668 model calls per episode. With a budget of at most 16 routing decisions per episode, PROUR is the fastest method in Table 10 (559.16 s per episode, 17% below the LLM Router and 17% below IG-clarifier) and also the most accurate (29.42%). Even without the budget, its wall-clock time is within 4% of the fastest baseline.

"LLM calls" counts every model call: the task agent, the user simulator, every decomposer call (goal proposals, action samples and likelihood scoring; 30 per routing decision, L863–864), the query generator, and every verifier round. "Time" is wall-clock time per episode and includes the verifier's tool calls. All methods ran on the same hardware and serving configuration.

Table 10. Success and cost per episode on τ-bench.

MethodSuccess (%)Interaction stepsLLM callsTime (s)
Base Agent17.4028.9071.52728.71
Ask-before-Plan23.6023.50169.78724.54
VerifiAgent22.5743.01298.781206.90
SABER20.8734.43165.45767.13
SAGE-Agent21.1334.54159.57950.66
IG-clarifier22.8523.55104.67670.23
LLM Router25.7523.39139.04674.09
PROUR28.1721.33668.13696.51
PROUR (budget 16 routing decisions)29.4220.42471.49559.16

Decomposer runtime and share (Q1). The decomposer takes 282 s per episode, 40.5% of PROUR's runtime. It issues most of PROUR's calls but takes less than half of its time, because each call is short: it writes a one-sentence goal or a short JSON action, or scores existing text without generating anything. The calls share the dialogue history, so the server reuses the cached prefix (95% hit rate). A decomposer call takes about 0.5 s, against several seconds for an agent or verifier call.

Why the time stays comparable. Routing shortens the episode. PROUR takes 14.34 ACT steps per episode, against 28.90 for the base agent (Table 1), and each of those is a long reasoning call. PROUR's time (696.51 s) is therefore below the base agent's (728.71 s) and within 4% of the two fastest baselines, IG-clarifier (670.23 s) and the LLM Router (674.09 s), with 5.32 and 2.42 points higher success.

A budget on routing removes the overhead. The original submission already studied a budget on routing decisions per episode (Fig. 5 and Table 9; L486–495). Once the budget is reached, the base agent continues alone (L1013–1014). With a budget of 16, PROUR reaches 29.42% success in 559.16 s: 17% faster than the LLM Router, with 3.67 points higher success. We chose 16 after seeing that sweep, but the conclusion does not depend on the choice: when budget sweeps from 8 to 32, Prour is both faster and more accurate than every baseline (Table 9: 27.15% in 538.01 s at 8, 28.19% in 669.24 s at 32). On VitaBench we fixed the budget at 16 before running, and PROUR was again the fastest method.

The revised Appendix A.2.5 says that PROUR reduces interaction steps and user questions (2.20 per episode, against 3.30 for the LLM Router and 2.70 for IG-clarifier; Table 1) at comparable wall-clock time, and that this takes more model calls.

AuthorsOpenReview ↗

Response to Reviewer jUfp (3/4)

[Q2] On several equally valid next actions

This case accounts for 13% of VERIFY steps (51 of 400), and we find no measurable harm from verifying there. A verification cannot change the task state, it usually replaces one of the agent's own steps, and replays give 31.9% success with the verification against 31.4% without it.

We labeled 400 VERIFY steps from PROUR's τ-bench runs by what each step actually needed (Cohen's κ = 0.79). We then replayed each step twice, once with the verification and once with the agent acting directly. Success is measured on episodes replayed from each sampled step.

Table R1. What VERIFY steps need (n = 400).

What the step neededShare (steps)Agent's decision changedSuccess (%): verify → act directly
Missing world-side evidence72% (289)33%34.4 → 23.9
Several valid actions, nothing missing13% (51)4%31.9 → 31.4
A fact only the user holds6% (24)13%26.7 → 21.1
Other (the model is simply unsure)9% (36)22%30.6 → 28.1

Why verifying there is harmless. The verifier has read-only tools, so a verification cannot change the task state (L287–289). In 80% of these steps the verifier's lookup is itself one of the valid actions, so it replaces a step the agent would take anyway. In the other 20% it costs one extra read-only step. With 4.79 VERIFY turns per episode (Table 1), that is about 0.6 such turns per episode and about 0.1 extra steps. In your example of independent subgoals, the verifier cannot execute either subgoal. It returns read-only evidence, and the agent then proceeds with one of them. By contrast, where world-side evidence is missing (72% of VERIFY steps), skipping the verification lowers success by 10.5 points.

How PROUR handles the case. It does not detect it. It limits how often the case arises and what it costs. Sampled actions are canonicalized and deduplicated before scoring, so one action written in two ways adds no entropy (L858–859). The EU threshold is calibrated on annotated steps (L884–895). A verification can only read. The paper also never assumes that the reference action is the only valid one (L269, L1079).

We revised §3.2 (L229–233) to say that high EU mostly reflects missing world-side evidence, and that the part caused by several valid actions cannot be reduced by any source. We list this in Appendix A.6 (Table 13).

Reviewer_GzcyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how tool-using agents should choose between acting, clarifying user intent, and acquiring additional world-side evidence. It proposes Prour, which estimates action distributions under sampled goal interpretations and decomposes their uncertainty into between-interpretation disagreement and within-interpretation entropy. These quantities determine the routing decision. A shared query generator is trained with a mode-conditioned belief-update reward, targeting the user goal under CLARIFY and a reference next action under VERIFY. Experiments on τ-bench report improvements in task success and interaction steps, with additional evaluations across task-agent backbones and on τ³-bench.

Critical strengths

The paper distinguishes between ambiguity in user intent and missing world information, then uses this distinction to decide whether to ask the user or gather evidence through tools. This is a practically useful distinction for interactive agents: not every uncertainty should be passed to the user, and tools cannot resolve preferences the user has not expressed. The motivation is clear, and the overall approach is easy to follow.

The experiments cover routing strategies, query generators, and different task-agent models, helping establish the role of each component. The trained 1.7B query generator outperforms its untrained counterpart, and the full method achieves competitive average success rates on both τ-bench and τ³-bench. The cross-agent results also suggest that the learned query-generation strategy is useful beyond a single task model.

Critical weaknesses
  1. The uncertainty estimates depend substantially on the sampled action set. Candidate coverage, action length, and different representations of semantically equivalent actions can all affect the entropy. For example, a singleton candidate set yields zero estimated uncertainty even when the action is wrong. Conversely, multiple valid actions can produce high entropy without requiring additional information. It remains unclear whether these scores reliably distinguish when to clarify from when to verify.

  2. The main efficiency comparisons use interaction steps, which do not capture substantial internal computation. Each decomposition requires 30 calls, including batched likelihood scoring. Although the appendix reports runtime for different Prour configurations, comparable full-system costs for the main baselines are missing. The success-rate improvement over the direct LLM Router is relatively small, and the statistical uncertainty of this difference is not reported.

Questions for authors
  1. How are the reference next actions used in the VERIFY reward constructed and checked? Are they taken directly from the next actions in the collected trajectories?

  2. How are empty or singleton candidate sets handled? How sensitive are the results to the number of sampled candidates?

  3. Does the reported runtime include all candidate generation, likelihood scoring, and verifier execution? Please clarify the scoring-interface requirements for API-only models/closed-source models and provide comparable costs for the main baselines.

As optional suggestions, testing with a stronger or different-family user simulator would help assess sensitivity to the choice of Qwen3-8B, whose capacity may be limited for faithfully simulating users on τ³-bench. VitaBench, which also uses LLM-based user simulators and includes challenging cross-scenario tasks, would be a useful additional benchmark.

12 comments
AuthorsOpenReview ↗

Response to Reviewer Gzcy (1/5)

Thank you for the careful review, and for noting that the user-versus-world distinction is practically useful and that the experiments establish the role of each component. We ran every analysis you asked for, including both optional suggestions: candidate-set sensitivity (W1, Q2), full-system cost and significance tests (W2, Q3), a decomposer that needs no log-probabilities (Q3), stronger user simulators, and VitaBench. Line and table numbers refer to the revised PDF.

[W1/Q2] On the dependence of AU and EU on the sampled candidate set

You are right that the estimates are computed over a sampled set, so we measured each effect you list. Success saturates at the default (K = 5, L = 4), and the single-candidate case you describe hides a wrong action in only 3 of 240 steps.

At each step the decomposer samples K = 5 goal interpretations and L = 4 actions per interpretation (L855–863). We ran the full pipeline on τ-bench with only K or L changed. We also measured the candidate set on the 240 held-out airline steps (L884–891).

Table R7. Sensitivity to K and L. Gold-in-set is the share of steps whose candidate set contains the next annotated action, read-only lookups included, after both are put in the same canonical JSON form. The right tool with a different argument value does not count. This column measures how well the set covers the agent's actual next step. The first two columns are measured on the 240 held-out airline steps, and success is the τ-bench average of the full pipeline.

(K, L)Single-candidate stepsGold-in-setτ-bench success (%)
(1, 4)31%71%23.1
(3, 4)14%79%26.4
(5, 4), default9%82%28.17
(10, 4)4%85%28.6
(5, 1)28%63%24.8
(5, 2)16%74%26.9
(5, 8)5%86%28.3

Sensitivity (Q2). With K ≥ 3 and L ≥ 2, success stays within 2 points of the default. Doubling K or L raises gold-in-set by 3 to 4 points and success by at most 0.4. Small sets hurt, as you expected. L = 1 loses 3.4 points. K = 1 makes AU zero by construction, so success falls to the Verify-only level (23.1% vs 22.85%, Table 1).

Coverage. Gold-in-set requires an exact argument match. When the annotated action needs an argument that is not yet in the dialogue, such as an ID the user has not given or a value not yet looked up, no candidate can match, because the sampler is told not to invent such values (L1245–1247). Such misses reflect missing information, not a sampling failure. Of the 43 steps without the annotated action, PROUR routed 74.4% to CLARIFY or VERIFY.

Empty and single-candidate sets (Q2). Both give AU = EU = 0, so PROUR routes to ACT. The task agent then chooses its own action; the sampled candidate is never executed. The set was empty in 1 of 240 steps, when every sample failed schema validation (L856–858). It held a single action in 22 steps (9%). A single action means that all 20 samples, drawn under 5 different interpretations, agreed. In 19 of these 22 steps that action was the reference action. The case you describe, zero uncertainty with a wrong action, therefore occurs in 3 of 240 steps (1.3%), and its cost is one missed intervention.

Equivalent representations and action length. Every sampled action is validated against the tool schema, rewritten in one canonical JSON form with sorted keys, and deduplicated before scoring (L856–859). The same call written in two ways therefore counts once. Different calls that return the same information are not merged; that is the several-valid-actions case, measured in part 2/5 (13% of high-EU steps, no loss in success). Actions are scored by their summed token log-likelihood (L860), which could favor short actions. Within candidate sets, the Spearman correlation between an action's length and its probability is −0.08. Scoring by the per-token average instead changes the route in 7% of steps and leaves routing macro-F1 almost unchanged (73.2 vs 74.0). Table R7 is Table 17 in Appendix A.6.

AuthorsOpenReview ↗

Response to Reviewer Gzcy (2/5)

[W1] On whether the scores separate CLARIFY from VERIFY

They do. Of 150 held-out steps that need information, PROUR sends 9 to the wrong source, against 27 for a prompted router. The case you raise, several valid actions with nothing missing, is 13% of high-EU steps, and verifying there leaves success unchanged.

We use the 240 airline steps with human-checked labels, which we used only after freezing the thresholds on retail (L884–891). 60 need the user, 90 need world-side evidence, and 90 need neither. The prompted LLM Router sees the same history and tools.

Routing on the 240 held-out steps (Fig. 7).

PROURLLM Router
User-side steps sent to VERIFY (of 60)618
World-side steps sent to CLARIFY (of 90)39
Right source when intervening on these 150 steps108 of 117 (92%)84 of 111 (76%)
Routing accuracy / macro-F173.8% / 74.0453.8% / 54.24

The accuracy gap is significant (McNemar p < 0.001, a new test in Appendix A.6, Table 15; macro-F1 at L898–900). Most of PROUR's remaining errors (54 of 63) concern whether to intervene, not which source to use.

Several valid actions. We labeled 400 VERIFY steps from PROUR's τ-bench runs by what each step needed (Cohen's κ = 0.79). We then replayed each step twice, once with the verification and once with the agent acting directly.

Table R1. What high-EU (VERIFY) steps need (n = 400).

What the step neededShareSuccess (%) with / without verification
Missing world-side evidence72%34.4 / 23.9
Several valid actions, nothing missing13%31.9 / 31.4
A fact only the user holds6%26.7 / 21.1
Other (the model is simply unsure)9%30.6 / 28.1
All VERIFY steps100%33.3 / 25.1

Verifying when several actions are valid shows no detectable difference (31.9% vs 31.4%). The verifier has read-only tools (L287–289), and in 80% of these steps its lookup is itself one of the valid actions, so it replaces a step rather than adding one. Where evidence is missing, skipping the verification costs 10.5 points.

Which samples are drawn. To measure how much the scores depend on the particular sampled set, we ran the decomposer 5 times with different seeds on each held-out step, so each run draws a different candidate set.

Table R6. Stability over 5 repeated samplings.

KStd of AU / EU across runsSame route in all 5 runsRouting macro-F1
30.06 / 0.0674%69.8 ± 2.4
5 (default)0.04 / 0.0581%74.0 ± 1.6
100.03 / 0.0387%75.1 ± 1.1

At K = 5 the spread is small next to the thresholds of 0.25 and 0.20 (L376). Of the steps whose route varies, 78% have AU or EU within 0.05 of a threshold, and 70% switch only between ACT and VERIFY, which costs one read-only lookup and no user question.

Tables R1 and R6 are Tables 13 and 16 in Appendix A.6.

AuthorsOpenReview ↗

Response to Reviewer Gzcy (3/5)

[W2/Q3] On full-system cost

Yes, the reported runtime includes candidate generation, likelihood scoring and verifier execution, and Table 10 (Appendix A.2.5) now gives the full-system cost of every baseline. PROUR makes more model calls than any baseline. Its wall-clock time is nevertheless within 4% of the fastest baseline, and with budget control (at most 16 routing decisions) it is the fastest method on both τ-bench and VitaBench.

"LLM calls" counts every model call: the task agent, the user simulator, every decomposer call (30 per routing decision, as you note; L863–864), the query generator, and every verifier round. "Time" is end-to-end wall-clock time per episode, including the verifier's tool calls. All methods ran on the same hardware and serving configuration.

Table 10 (revised). Success and cost per episode on τ-bench.

MethodSuccess (%)Interaction stepsLLM callsTime (s)
Base Agent17.4028.9071.52728.71
Ask-before-Plan23.6023.50169.78724.54
VerifiAgent22.5743.01298.781206.90
SABER20.8734.43165.45767.13
SAGE-Agent21.1334.54159.57950.66
IG-clarifier22.8523.55104.67670.23
LLM Router25.7523.39139.04674.09
PROUR28.1721.33668.13696.51
PROUR (budget 16 routing decisions)29.4220.42471.49559.16

The decomposer takes 282 s per episode, 40.5% of PROUR's runtime. Its calls are short (a one-sentence goal, a short JSON action, or scoring existing text) and share the dialogue history, so the server reuses the cached prefix (95% hit rate). A decomposer call takes about 0.5 s, against several seconds for an agent or verifier call. Routing also shortens the episode: PROUR takes 14.34 agent steps per episode, against 28.90 for the base agent (Table 1). Its time is therefore below the base agent's and within 4% of IG-clarifier and the LLM Router, with 5.32 and 2.42 points higher success. With a cap of 16 routing decisions (Fig. 5 and Table 9; L486–495), PROUR is 17% faster than the LLM Router with 3.67 points higher success. On VitaBench we fixed the budget before running, and it is again the fastest method (part 5/5).

We do not claim a lower overall inference cost. Wall-clock time benefits from prefix caching on our own server. Under per-call or per-token API pricing, cost roughly follows the call count, because each decomposer call re-sends the dialogue history, and PROUR makes 4.8 times as many calls as the LLM Router. The decomposer without log-probabilities (Table R9, part 4/5) needs 10 calls per routing decision instead of 30. The revised Appendix A.2.5 says that PROUR reduces interaction steps and user questions (2.20 per episode, against 3.30 for the LLM Router and 2.70 for IG-clarifier; Table 1) at comparable wall-clock time, and that this takes more model calls.

[W2] On the size and significance of the gain over the LLM Router

You are right about τ-bench: the 2.42-point gap has a paired 95% CI that includes 0. It is significant on τ³-bench and on VitaBench, and has the same sign and a similar size in every setting we ran.

Tables 1 and 4 now report success as mean ± std over three seeds. Average success is 28.17 ± 0.98 for PROUR and 25.75 ± 1.87 for the LLM Router on τ-bench, and 29.00 ± 1.11 and 26.33 ± 1.59 on τ³-bench. Appendix A.6 (Table 18) adds the paired tests below: a bootstrap over tasks (95% CI) and a permutation test (p). On τ³-bench over 5 seeds the means stay at 29.00 and 26.33 (± 1.21 and ± 1.84), as in Table 4.

Table R8. Δ = PROUR's success minus the comparator's, in percentage points (positive favors PROUR).

ComparatorBenchmark (runs per task)Δ95% CIp
LLM Routerτ-bench (3 seeds)+2.42[−1.4, 6.3]0.22
LLM Routerτ³-bench (5 seeds)+2.67[0.4, 4.9]0.02
LLM RouterVitaBench (4 trials)+2.56[0.5, 4.6]–
Ask-before-Planτ-bench (3 seeds)+4.57[0.7, 8.4]0.04
IG-clarifierτ-bench (3 seeds)+5.32[1.5, 9.2]0.01

PROUR leads the LLM Router by 2.4 to 3.3 points in all six settings we ran (τ-bench and τ³-bench, each also with a GPT-4.1 user; τ³-bench with a Qwen3-32B user; VitaBench), and in each seed of Tables 1 and 4. On τ-bench itself, episodes are 2.06 steps shorter (95% CI [−3.3, −0.8]) and contain 1.10 fewer user questions, a third fewer (95% CI [−1.42, −0.78]). The LLM Router is our own ablation: it keeps our routing modes, verifier and trained query generator, and replaces only the routing signal (Appendix A.5.5). Against the strongest prior method, Ask-before-Plan, the gain is 4.57 points and significant.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response to Reviewers and Area Chairs

We sincerely thank all reviewers and Area Chairs for their careful assessments and constructive discussions. We are encouraged that all three reviewers recognized the practical value of our framework. We have substantially strengthened the manuscript through additional experiments, statistical analyses, stronger-model evaluations, and clearer positioning relative to prior work.

Strengths Noted by Reviewers

  • The value of separating unclear user intent from world-side uncertainty, avoiding both unnecessary user questions and premature actions (Gzcy, jUfp, jrj5).
  • Convincing motivation and intuitive formulation that treats clarification and verification as source-aligned responses to decomposed action uncertainty (Gzcy, jrj5).
  • Well-structured experiments with same-backbone baselines and targeted ablations, alongside cross-agent and tau3-bench generalization of the 1.7B query generator (Gzcy, jUfp, jrj5).

Key Clarifications and Conclusions

1. Independent Contributions Beyond Prior Work (jUfp, jrj5, Gzcy)

We have clarified that our contribution is a unified step-level framework that routes uncertainty by source, triggers grounded verification at the right moments, and aligns query-generation supervision with the selected mode:

  • Uncertainty routing outperforms zero-shot and few-shot prompted routers given the same labeled development steps and call budget (Table 15, Table 24 / R16).
  • VERIFY gains stem from EU-based timing rather than extra tool calls: random or agent-initiated verification with the same verifier and tools recovers almost none of the gain (Table 14 / R4).
  • Mode-conditioned supervision is essential beyond routing: single-target rewards degrade performance, and IG-clarifier's generator behind PROUR's router still asks many tool-answerable questions (Table 6, Table 25 / R17).

2. Grounded VERIFY vs. Self-Reflection and AU/EU Scope (jrj5, jUfp, Gzcy)

High EU does not exclusively indicate missing external tool outputs. Even on re-check turns using evidence already in history, grounded verification substantially outperforms task-agent self-reflection in replay success and correctness, and this advantage persists with a stronger backbone throughout (Table 22 / R14, Table 23 / R15).

We also qualified our uncertainty claims: AU and EU are estimates from finite sampled interpretations rather than guarantees that all uncertainty is removed. Most high-EU steps require missing world-side evidence (multiple-valid-action cases cause no measurable harm; Table 13), while shared wrong assumptions rarely hide user-side ambiguity (Table 26).

3. Statistical Robustness and Generalization (Gzcy, jUfp)

Evaluations confirm statistically significant gains over the LLM Router on tau-bench and tau3-bench, with PROUR leading on every seed (Table 18). Using our log-prob-free sampling estimator, PROUR also transfers to a closed-model setting with Gemini 3.8 Flash as both agent and decomposer (Table R22), and leads under stronger user simulators and on VitaBench without retraining (Tables 19–20). We clarified that the per-seed results in Table R20 come from a fresh five-seed evaluation rather than the original three-seed runs.

4. Efficiency Scope and Cost Accounting (Gzcy, jUfp)

We report full-system runtime and LLM calls across all baselines (Table 10), clarifying that input-token counts in Table R23 are estimated from average prompt length. Our efficiency claim is specifically about reduced user-facing interaction burden: fewer user questions (2.20 versus 3.30 per episode) and fewer interaction steps (21.33 versus 23.39), with comparable wall-clock time. With a 16-decision routing budget, PROUR is fastest (559 s vs. 674 s) at 29.42% success. We do not claim lower total input-token volume or uncached API cost.

Summary of Manuscript Revisions

  • Main Text (Sections 3.2, Related Work): Clarified lineage to Hou et al. (2024) and Deng et al. (2026); added Ask-before-Plan, Ambig-SWE, and PCE; qualified world-side VERIFY claims and AU/EU as estimates (L195–197, L228–233, L479–485).
  • Appendices (A.2.2, A.2.5, A.3, A.6): Added target ablations (Table 6), full-system cost (Table 10, L1026–1060), reference-action filtering (A.3, L1065–1093), and Tables 13–26 (VERIFY breakdown, trigger ablations, McNemar test, K/L stability, 5-seed CIs, stronger simulators, VitaBench, sampling estimator, self-reflection, shared assumptions).

We believe the additional experiments and revisions have resolved key concerns, and provide a substantially stronger basis for evaluating PROUR's scientific contribution and practical value.

We respectfully invite reviewers to reconsider their ratings in light of the complete discussion and revised manuscript. We hope this summary also assists the Area Chairs in assessing final recommendations.

Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.