betaFeedback

WorksComputer Science

Receiver-Conditioned Clarification for Multi-Agent Food Assistance

Abstract

Food insecurity creates a pressing need for guidance that helps households find usable food resources and assess food choices in light of their circumstances. Such guidance must connect eligibility requirements, practical access constraints, and dietary needs rather than return isolated listings or generic nutritional advice. Multi-agent food assistants can coordinate these tasks, with uncertainty indicating when clarification or verification is needed before a recommendation. However, uncertainty can be lost across agent handoffs, allowing tentative interpretations to propagate as facts and obscuring opportunities for correction. Moreover, confidence that a service requirement has been interpreted correctly is not confidence that a recommendation satisfies the user's needs; local confidence cannot simply be reused as downstream reliability. We introduce an inference-time, receiver-conditioned clarification framework that preserves shared premise identities and replays consistent interpretations across dependent agents. Each receiver exposes admissible decisions and remaining evidence obligations, so question relevance reflects both decision changes and missing support. Information questions and user-confirmable fallback adjustments are evaluated together with their anticipated follow-up interactions under shared budgets. This structure connects clarification to jointly supported plans for resource navigation, nutritional reasoning, and supply-side planning.

CCS Concepts

• Computing methodologies → Multi-agent systems; • Information systems → Question answering.

Introduction

Relevance to the Web and the track. Conversational access to Web service directories hinges on resolving user constraints before any resource is recommended. This paper shows how a multi-agent assistant preserves uncertain premises across agents and selects the clarification that most improves its downstream decisions under an interaction budget. The work falls squarely within interactive information access, domain-specific recommendation, and agentic search in the Search, Recommendation, and Retrieval-Augmented AI track.

Food insecurity is defined as limited or uncertain availability of nutritionally adequate and safe food, or limited or uncertain ability to acquire acceptable food through socially acceptable means [43]. It remains a pressing societal problem in the United States: the U.S. Department of Agriculture (USDA) estimates that 13.7% of households (18.3 million) experienced food insecurity at some point in 2024 [44]. Meeting this need requires more than food supply: people must determine where and how to obtain usable food, while providers must decide what to offer and where to reach unmet demand. A food assistant must connect supply with those in need and support both complementary decisions.

On the demand side, nearby resource lists are insufficient: people experiencing food insecurity, including those without stable housing, need resource discovery linked to personal suitability and feasible access plans. Eligibility, documentation, opening hours, and travel constraints determine usability [22], while food must meet individual nutritional needs [54]. On the supply side, food banks, pantries, and community meal programs must decide what to provide and where, aligning provision with community preferences and dietary needs [26] and balancing accessibility, equitable coverage, and limited operational capacity in distribution [39]. Provider-facing assistance must therefore compare food and site options against unmet demand and operational feasibility, beyond enumerating candidates.

Therefore, effective food assistance must connect individual needs and access constraints with providers’ supply and distribution capacity. To support this coordination, we present a conversational food assistant integrating demand-side resource navigation with supply-side provision and distribution planning; it is deployed here. These functions require heterogeneous evidence and distinct but interdependent eligibility, nutritional, and planning judgments with different constraints and verification needs, making it difficult for a single agent to isolate task-specific reasoning and verify intermediate decisions. We therefore build a role-specialized multi-agent system (MAS) that separates task-specific reasoning and verification and coordinates findings into actionable plans.

Role specialization supports distinct judgments but complicates coordination: downstream decisions rely on potentially uncertain intermediate information [37, 49]. Since users and providers do not fully specify their needs upfront, the assistant must identify which uncertainties to clarify before they affect subsequent decisions, while limiting unnecessary interaction [4, 5]. Value-of-information (VOI) methods provide a decision-theoretic foundation by weighing clarification’s expected benefit against communication cost under a dialogue budget [10]. However, existing formulations primarily consider a single decision-maker, whereas MAS decisions depend on shared, potentially uncertain premises. This distinction introduces three challenges. First, a shared uncertain premise may affect receivers with different evidence requirements, making clarification value depend on their decisions. Second, handoffs may lose alternative interpretations, preventing receivers from evaluating how clarification would change their decisions [34]. Third, multiple agents’ reliance on the same upstream evidence may resemble independent confirmation and inflate confidence [13, 23]. Together, these challenges show that uncertainty propagation or calibration alone is insufficient: clarification must account for shared uncertainty’s effects on downstream decisions and their evidence dependencies.

Food assistant motivation

Figure 1. Food assistant motivation. (a) Linking food-access and dietary needs to provision and site planning. (b) Clarification depends on how requirements affect each receiving agent’s decision and whether fallback acceptance needs follow-up checks.

To address these challenges, we propose a receiver-conditioned clarification framework that evaluates questions by their effects on downstream decisions. We preserve alternative interpretations and evidence provenance across handoffs so receivers can reason over unresolved premises rather than collapsed conclusions. During receiver replay, consistent interpretations of shared premises ensure that affected receivers evaluate dependent decisions under the same hypothesis; each receiver produces a decision signature recording admissible decisions and outstanding evidence obligations, enabling the controller to identify uncertainties affecting

the collective plan. Beyond information gathering, the framework considers user-confirmable fallback adjustments when constraints prevent feasible plans, including the additional interactions needed to implement them. Both clarification types share an interaction budget, allowing the controller to prioritize questions by their expected contribution to an actionable plan.

We integrate this framework into our conversational food assistant to coordinate resource navigation, nutritional reasoning, and provider-side planning, evaluating the system across these complementary tasks on HoosierHelp, NGOA, and FoodGo, respectively. Our approach outperforms the compared baselines across all three tasks, while controlled ablations examine the clarification mechanisms’ contributions.

Our contributions are summarized as follows.

  • We build and deploy a conversational food assistant that, to our knowledge, is the first to span both the demand and supply sides of food assistance, supporting personalized resource navigation, nutritional reasoning, and evidence-grounded food provision and distribution planning in one system.

  • We develop a receiver-conditioned clarification framework that extends single-decision VOI to dependent receiver decisions. It enforces consistent shared-premise assignment during replay, records decision signatures that retain admissible decisions and outstanding evidence obligations, and jointly values information clarification, fallback acceptance, and their anticipated follow-up interactions under one budget, with responses driving revision and replanning.

  • We evaluate the assistant on HoosierHelp, NGQA, and a FoodGo-based task, covering resource navigation, nutritional reasoning, and transfer-site selection. Our approach outperforms all compared baselines in every setting, and atomic-question and component ablations confirm that each clarification mechanism contributes to the gains.

Related Work

Food assistance and personalized nutritional reasoning. HoosierHelp [22] evaluates conversational social-service navigation on real service records with simulated users, covering composite needs, constraint fallback, and non-ideal interactions. NGQA [54] evaluates personalized nutritional reasoning over user, food, and nutrient information. On the provider side, the FoodGo dataset [31, 32] supports food-rescue pickup-and-delivery scheduling. TRACE [12] is the closest system: a retrieval-first pipeline for public-service directories that parses a query into constraints and enforces them during retrieval. All of these take the user’s constraints as given. None asks which constraint is still unknown, which decision it would change, or whether asking is worth the interaction; that is the subject of this paper. We use the three benchmarks as evaluation foundations; TRACE-style retrieval is a drop-in replacement for our retriever, not an alternative to the controller that precedes it.

Uncertainty and dependencies in multi-agent systems. PropUQ-MAS [23] propagates local uncertainty with risk inherited over a communication graph; CAGE-Cal [13] calibrates confidence against a counterfactual non-communicating graph. Both establish that inter-agent agreement is not independent confirmation, and we build on that premise. But a calibrated number on a message does not tell the consuming task whether the uncertainty matters to it or whether resolving it would change what it does: propagation and calibration are inputs to clarification, not a policy. From the execution side, VeriMAP [48] plans subtasks as a DAG with structured I/O and planner-generated verification functions, and ToCoMAS [60] derives agent responsibilities, handoffs, and memory boundaries from the task graph. Both decide how dependent agents execute and assume the premise being passed is already resolved; neither can choose which unresolved premise to ask about or value that question against the decisions downstream. Our framework fills exactly that gap and is complementary to all four: decision signatures and the dependency footprint can be read off any DAG-structured plan.

Clarification and information-value control. MAC [1] coordinates clarification between a supervisor and domain experts using a taxonomy of user ambiguities. Dong et al. [10] instantiate inference-time VOI through hypothetical answers, posterior decision utility, communication costs, and a clarification budget. Zhang et al. [51] separate failure prediction from intervention value and learn action-sensitive oversight from alternative continuations. All three value a question against a single decision-maker’s posterior, and that is where they stop short in a multi-agent system: one premise is consumed by several receivers needing different evidence, alternatives are collapsed at handoffs, and a fallback is worth nothing until the follow-ups that make it usable are counted. We keep their decision-theoretic core and change what is valued: dependent receiver decisions replayed under a shared premise assignment, decision signatures that retain admissible decisions and outstanding obligations, and fallback acceptance valued with its bounded continuation, all under one budget. No prior method evaluates clarification at this level; doing so is the methodological contribution of this paper, and the ablation in Section 4.3.1 shows it carries the gains.

Methodology

Our framework operates entirely at inference time with frozen language models. Agents preserve candidate interpretations and their dependencies; a planner connects unresolved premises to receiving decisions; and a cost-aware controller selects clarification based on complete-task consequences. When constraints block a plan, counterfactual relaxation diagnosis identifies user-confirmable adjustments that could recover it. Actual responses then revise the evidence and trigger replanning.

Problem Formulation

In a role-specialized multi-agent workflow, agents exchange intermediate artifacts that become premises for subsequent decisions. The same uncertain premise may affect several receivers differently, making its clarification value depend on the tasks that consume it. We represent these dependencies with a time-unrolled graph Gt\mathcal{G}_t. Nodes represent agent invocations, evidence events, and scoped premises, while edges record evidential support and downstream consumption. Each premise identifies its subject, condition, and applicability scope.

At interaction step tt, the public state is

St=(Ht,Et,Mt,Gt,Bt,It),(1)S_t = (H_t, E_t, \mathcal{M}_t, \mathcal{G}_t, B_t, I_t), \tag*{(1)}

where HtH_t is the interaction history, EtE_t is observed evidence, Mt\mathcal{M}_t contains candidate messages, and BtB_t and ItI_t are the remaining shared interaction and inference budgets. Active user needs and constraints are extracted from HtH_t and EtE_t. The controller chooses whether to proceed, clarify an unresolved premise, or request acceptance of a preference adjustment, aiming to complete the task with supported decisions and limited interaction.

Evidence-Grounded Candidate States

Preserving alternatives and their support. A single committed artifact can conceal alternatives that matter to a downstream decision. Each agent therefore generates a bounded set of candidate artifacts using the frozen backbone. A role-specific canonicalizer groups equivalent proposals while retaining the greedy artifact and alternative interpretations. Agent ii communicates

Mi={(aik,wik,Pik,Fik)}k,(2)M_i = \{(a_{ik}, w_{ik}, P_{ik}, F_{ik})\}_k, \tag*{(2)}

where kk indexes retained candidates, aika_{ik} is a candidate artifact, wikw_{ik} is its empirical proposal support, PikP_{ik} records its sources and consumed message versions, and FikF_{ik} contains its scoped assumptions and unresolved requirements. These premise-linked alternatives let receivers assess how clarification changes their decisions.

Shared-premise dependencies. The graph links each candidate to its premises and each premise to its originating evidence events.

Receiver-conditioned clarification framework

Figure 2. Receiver-conditioned clarification framework. (a) Candidate states retain alternatives, evidence provenance, and shared premises. (b) The planner links these premises to receiving decisions and unmet evidence obligations, and diagnoses minimal useful fallback adjustments for blocked plans. (c) The controller values information and fallback questions together with their follow-up costs; actual replies update evidence or accepted preferences and trigger replanning.

Shared premises retain a common identity across handoffs, and forwarding a claim adds a consumer relationship while preserving its evidence origin. This makes repeated uses of the same information traceable and allows one clarified premise to update all relevant receivers.

Let Ωt\Omega_t contain joint interpretations compatible with the observed evidence. Each ω∈Ωt\omega\in\Omega_t assigns one value ω(z)\omega(z) to an unresolved premise zz, which is reused by every dependent receiver during replay. Consequently, hypothetical replies are evaluated through coherent joint states, preventing different agents from constructing a seemingly feasible plan using incompatible assumptions about the same condition.

Receiver-Conditioned Decision Relevance

Planning and decision signatures. The planner constructs candidate plans for active needs and identifies the conditions needed to support them. For receiver jj and joint interpretation ω\omega, it returns

σj(ω;St)=(Ajadm(ω),Oj(ω)),(3)\sigma_j(\omega; S_t) = \left(\mathcal{A}^{\mathrm{adm}}_j(\omega), O_j(\omega)\right), \tag*{(3)}

where Ajadm\mathcal{A}^{\mathrm{adm}}_j contains decisions not ruled out by the instantiated conditions, and OjO_j records the remaining obligations required to support those decisions. The planner generates information questions from unresolved premises affecting either component.

Interpretations yielding the same supported outcome are immaterial to that receiving decision. Conversely, a decision may remain unchanged while requiring missing evidence. Including OjO_j retains such questions instead of judging relevance solely by whether the preferred decision changes.

Each question qq has a dependency footprint

Rt(q)={j∣v(q)⇝j in Gt},(4)\mathcal{R}_t(q) = \{j \mid v(q) \leadsto j \text{ in } G_t\}, \tag*{(4)}

where v(q)v(q) is the targeted premise or preference node and ⇝\leadsto denotes a directed dependency path. Replaying these receivers reveals how alternative replies affect their decisions and support obligations. Questions targeting the same scoped premise can therefore be shared across receivers, avoiding repeated requests while retaining their combined downstream effects.

Counterfactual relaxation diagnosis. When an identified constraint blocks a plan, asking for additional facts may not resolve the incompatibility. The planner instead evaluates eligible user-confirmable preference adjustments. Each adjustment is applied consistently across affected receivers while holding evidence and external requirements fixed. Replanning returns recovered candidate plans and their remaining obligation.

Let Dt\mathcal{D}_t contain the evaluated adjustments, let rt(δ)=1r_t(\delta)=1 indicate that adjustment δ\delta provides the required recovery, and let δ′≺δ\delta' \prec\delta denote a strictly smaller adjustment to the original preferences. The retained alternatives are

Dtmin⁡=Min⁡≺{δ∈Dt∣rt(δ)=1},(5)\mathcal{D}^{\min}_t = \operatorname{Min}_{\prec}\{\delta\in\mathcal{D}_t \mid r_t(\delta)=1\}, \tag*{(5)}

where Min⁡≺\operatorname{Min}_{\prec} selects minimal elements under ≺\prec. Retaining minimal useful adjustments avoids requesting larger changes when a smaller evaluated alternative provides the required recovery. The planner converts these alternatives into questions requesting user acceptance and retains any follow-up obligations for subsequent value evaluation.

Together, the generated questions form

Qt=Qtinfo∪Qtfallback.(6)Q_t = Q_t^{\mathrm{info}} \cup Q_t^{\mathrm{fallback}}. \tag*{(6)}

Here, QtinfoQ_t^{\mathrm{info}} resolves missing or ambiguous information, and QtfallbackQ_t^{\mathrm{fallback}} requests acceptance of a planner-supported adjustment.

Dependency-Aware Budgeted VOI

Complete-task consequence evaluation. For each question, the controller simulates possible replies, updates the targeted premise once per branch, and replays the affected receivers. An acceptance branch activates the proposed adjustment. The resulting outcome is scored against the active needs and their evidence requirements as one complete task. This captures the benefit of resolving a shared premise without counting the same improvement repeatedly across agents, while exposing consequences that help one decision but conflict with another.

A question is evaluated together with a bounded, response-contingent continuation that can ask follow-up questions or resume retrieval and repair. Evaluating the continuation captures complementary questions and the additional interactions needed to make an accepted fallback usable, rather than rewarding an intermediate change that leaves the plan blocked.

Let πt,q\pi_{t,q} denote the evaluated continuation beginning with qq, τ\tau a simulated trajectory, and SτS_\tau its terminal state. Let U(Sτ)U(S_\tau) measure supported task completion and U^tproceed\widehat{U}_t^{\mathrm{proceed}} be the estimated utility of proceeding without further clarification. Then

ΔU^t(q;Bt,It)=E^τ∼πt,q[U(Sτ)]−U^tproceed,C^t(q;Bt,It)=E^τ∼πt,q[∑ℓ∈L(τ)c(qℓ,oℓ)],(7)\begin{aligned} \widehat{\Delta U}_t(q; B_t, I_t) &= \widehat{\mathbb{E}}_{\tau\sim\pi_{t,q}}[U(S_\tau)] - \widehat{U}_t^{\mathrm{proceed}},\\ \widehat{C}_t(q; B_t, I_t) &= \widehat{\mathbb{E}}_{\tau\sim\pi_{t,q}}\left[\sum_{\ell\in L(\tau)} c(q_\ell,o_\ell)\right], \tag*{(7)} \end{aligned}

where E^\widehat{\mathbb{E}} denotes the replay-based expectation estimate, L(τ)L(\tau) indexes clarification interactions along τ\tau, and c(qℓ,oℓ)c(q_\ell,o_\ell) is the interaction cost of question qℓq_\ell and reply oℓo_\ell, expressed on the utility scale. Thus, ΔU^t\widehat{\Delta U}_t measures improvement over proceeding now, while C^t\widehat{C}_t includes both the initial question and anticipated follow-up costs.

Question selection. The controller selects

qt∗=arg⁡max⁡q∈Qtfeas[ΔU^t(q;Bt,It)−C^t(q;Bt,It)].(8)q_t^* = \underset{q \in Q_t^{\mathrm{feas}}}{\arg\max}\left[\widehat{\Delta U}_t(q; B_t, I_t) - \widehat{C}_t(q; B_t, I_t)\right]. \tag*{(8)}

where QtfeasQ_t^{\mathrm{feas}} contains questions whose evaluated continuations fit the remaining shared interaction and inference budgets. Accounting for continuation feasibility prevents the controller from initiating an apparently inexpensive question while ignoring the resources needed to complete its anticipated follow-ups.

The selected question is issued when its estimated net value is positive; otherwise, the assistant proceeds with verification and repair. Only the first question of the continuation is executed. Its actual reply determines the revised state and the next value comparison.

Selection error bound. Let Ct=Qtfeas∪{proceed}C_t = Q_t^{\mathrm{feas}} \cup\{\mathrm{proceed}\}, and let Vt(a)V_t(a) and V^t(a)\widehat{V}_t(a) denote the expected and estimated net values of action aa for its evaluated continuation. Proceeding has value zero, while a question has estimated value ΔU^t−C^t\widehat{\Delta U}_t - \widehat{C}_t. Let at∗=arg⁡max⁡a∈CtVt(a)a_t^* = \arg\max_{a \in C_t} V_t(a) and a^t=arg⁡max⁡a∈CtV^t(a)\widehat{a}_t = \arg\max_{a \in C_t} \widehat{V}_t(a). If the uniform estimation error is at most ϵt\epsilon_t, then

max⁡a∈Ct∣V^t(a)−Vt(a)∣≤ϵt(9)\max_{a \in C_t} \left|\widehat{V}_t(a) - V_t(a)\right| \le\epsilon_t \tag*{(9)}
⟹Vt(at∗)−Vt(a^t)≤2ϵt.\Longrightarrow\quad V_t(a_t^*) - V_t(\widehat{a}_t) \le2\epsilon_t.

Proof. Estimated-value maximization gives V^t(a^t)≥V^t(at∗)\widehat{V}_t(\widehat{a}_t) \ge\widehat{V}_t(a_t^*). Applying the error bound to both actions yields

Vt(at∗)≤V^t(at∗)+ϵt≤V^t(a^t)+ϵt≤Vt(a^t)+2ϵt.\begin{aligned} V_t(a_t^*) &\le\widehat{V}_t(a_t^*) + \epsilon_t\\ &\le\widehat{V}_t(\widehat{a}_t) + \epsilon_t\\ &\le V_t(\widehat{a}_t) + 2\epsilon_t. \end{aligned}

Evidence Revision and Closed-Loop Control

Scoped response integration. An actual reply is parsed into an update with a subject, condition, applicability scope, and provenance. Information responses revise the relevant evidence, while explicitly accepted fallback responses update the specified preference. The resulting state is

St+1=Revise⁡(St,qt,ot),(10)S_{t+1} = \operatorname{Revise}(S_t,q_t,o_t), \tag*{(10)}

where qtq_t is the issued question and oto_t is the actual response. Scope determines which premises are revised and which receivers can consume the update. This lets a shared fact benefit all applicable tasks while preserving requirements unrelated to the response.

Dependency-guided replanning. Each update creates a new evidence or message version. The framework invalidates candidate artifacts, decision signatures, and question values that depended on the superseded version, then reruns unaffected roles in dependency order. Independent evidence and unrelated artifacts remain available. Replies introducing new interpretations also trigger candidate regeneration, allowing the plan to incorporate information beyond its earlier alternatives.

After replanning, the controller refreshes information questions and fallback alternatives, removes resolved questions, and reevaluates the remaining options using the updated state and budgets. The loop continues until a complete response is supported, no feasible clarification has positive estimated net value, or a shared budget is exhausted. The assistant then completes verification and repair to produce the final response.

Experimental Evaluation

We conduct one main experiment on HoosierHelp and two supplementary experiments on NGQA and a FoodGo-based transfer-site selection task. We then examine clarification policies and system components through two controlled ablation studies and assess controller hyperparameter sensitivity.

Experimental Setting

Backbone. We use Qwen3-8B [41] for our assistant and all baselines in the main experiment, two supplementary experiments, and both ablations, except the GPT-4.1 [28] and gpt-oss-120b [27] single-agent baselines identified in the main results table. Our assistant freezes the backbone weights; candidate generation, planning, clarification, and response revision occur entirely at inference time.

Benchmarks and evaluation construction. HoosierHelp [22] evaluates conversational social-service navigation with real service records, structured search tools, and simulated users, covering composite needs, constraint fallback, and non-ideal interactions. We evaluate 240 instances per run under the official interaction and evaluation protocol, with GPT-4.1-mini as the user simulator. The assistant interacts with the user, retrieves service information through the provided tools, and recommends resources matching the user’s requirements. The official protocol allows questions requesting multiple information items; only our controlled ablations impose atomic questions (§ 4.3), so results under the two protocols are not directly comparable.

NGQA [54] evaluates personalized nutritional reasoning from user, food, and nutrient information. Our interactive episodes initially withhold selected profile fields, reveal them through clarification, and score final answers against reference annotations. A 20% food-name hash split yields 2,652 test instances; each of three seeds shuffles this pool and evaluates 300 instances. We report the benchmark’s profile-complexity groups as sparse, standard, and complex, with per-seed counts in Appendix § 9.

FoodGo [31, 32] provides food-rescue pickup-and-delivery data. As no existing benchmark evaluates interactive requirement elicitation for provider-side planning, we construct a transfer-site selection task from these records, with 12 candidate sites and partially disclosed requirement profiles. The assistant clarifies relevant requirements and selects one or two sites. A deterministic planner generates reference decisions from complete profiles, isolating elicitation of fields that change the site set; success thus measures this skill without confounding it with planning quality. The task evaluates requirement elicitation and site selection rather than the original pickup-and-delivery scheduling problem.

Baselines. HoosierHelp baselines include single-agent free questioning with Qwen3-8B, GPT-4.1, and gpt-oss-120b; our MAS backbone; majority vote; text-confidence handoff; and four clarification methods: ProCoT [6], ATD [17], SAGE-Clarify [40], and LLM-VoI [10]. NGQA baselines include Role-chain MAS, nine-way majority vote, LLM-Blender [16], HybridLLM [7], and NG-Router [36].

FoodGo-based baselines include single-agent direct answering, single-agent free questioning, and the same four clarification methods.

The atomic-question study compares the four clarification methods under the controlled protocol, with ask-on-disagreement, fixed-order questioning, and myopic VOI without interaction cost as atomic controls. Component ablations are named by the removed component.

Metrics and reporting. We report the following metrics: for HoosierHelp, episode success, Tool EM, Resource EM, and total tokens (thousands per episode); for NGQA, overall and profile-complexity-specific binary accuracy, label F1, and explanation ROUGE-L; for the FoodGo-based task, exact-site-set success against the deterministic planner’s reference decision under complete requirements. HoosierHelp success follows the official end-to-end criterion: an explicit final recommendation plus exact tool-call-set and resource-ID-set matches. Total tokens sum input and output across all backbone calls per episode, including every controller call for candidate sampling, receiver replay, and simulated users (Appendix C). All experiments, including both ablations, use three runs; means are computed after rounding each seed’s result to three decimal places. All tables and error bars report mean ± standard deviation. NGQA overall accuracy is computed over the 300 instances in each run before averaging across runs; ablations also report mean questions per episode. Score differences are in percentage points (pp). Complete results and evaluation details appear in Appendices A–C.

Main Results

HoosierHelp

Table 1 compares our assistant with single-agent, multi-agent coordination, and clarification baselines on end-to-end resource navigation under the official HoosierHelp interaction protocol [22]. Our system leads on all three metrics: 57.8% success, 59.9% Tool EM, and 72.1% Resource EM. It beats the strongest baseline, LLM-VoI, by 3.7 pp on success while spending 5.1% fewer tokens (28.2k per episode), the lowest consumption of any clarification method. Higher accuracy at lower cost rules out the explanation that the gain is bought with extra interaction.

Supplementary Experiments

We further evaluate nutritional question answering and provider-side planning, using Qwen3-8B for all systems.

Nutritional question answering on NGQA. Figure 3 shows that our system leads on all three NGQA metrics: binary accuracy 0.778, label F1 0.603, and explanation ROUGE-L 0.655. The accuracy margin over the best baseline, HybridLLM, is 15.5 pp, with F1 also ahead (+0.017). Accuracy is highest in every profile-complexity group, so the advantage does not depend on profile difficulty. These scores evaluate the complete profile-disclosure pipeline. Table 2 and Appendix § 9 provide full results and sampling details.

Figure 3. NGQA results. Our assistant achieves the highest accuracy overall and across all profile-complexity groups. It also leads in label F1 and explanation ROUGE-L. Accuracy axes use a restricted range.

MethodOverallSparseStandardComplexLabel F1 ↑\uparrowROUGE-L ↑\uparrow
Role-chain MAS0.571 ±\pm 0.0200.580 ±\pm 0.0350.529 ±\pm 0.0290.615 ±\pm 0.0760.447 ±\pm 0.0030.473 ±\pm 0.005
Majority vote (9-way)0.591 ±\pm 0.0210.618 ±\pm 0.0140.507 ±\pm 0.0320.623 ±\pm 0.0910.455 ±\pm 0.0060.476 ±\pm 0.004
LLM-Blender0.617 ±\pm 0.0150.614 ±\pm 0.0130.599 ±\pm 0.0550.666 ±\pm 0.0560.491 ±\pm 0.0020.491 ±\pm 0.002
HybridLLM0.623 ±\pm 0.0150.623 ±\pm 0.0180.598 ±\pm 0.0330.673 ±\pm 0.0670.586 ±\pm 0.0060.516 ±\pm 0.008
NG-Router0.613 ±\pm 0.0120.597 ±\pm 0.0130.621 ±\pm 0.0590.672 ±\pm 0.0900.549 ±\pm 0.0120.511 ±\pm 0.002
Ours0.778 ±\pm 0.0210.750 ±\pm 0.0270.894 ±\pm 0.0450.700 ±\pm 0.0780.603 ±\pm 0.0020.655 ±\pm 0.019

Table 2. NGQA results.

Provider-side transfer-site selection. We construct an interactive transfer-site selection task from FoodGo data [31, 32] to evaluate provider-side planning. Success requires an explicit final recommendation that exactly matches the deterministic planner’s site set under the complete provider profile. The reference is produced by the same deterministic planner the agent calls, and this is by design: once the complete profile is known the planner has exactly one answer, so the only way to fail is to leave a decision-changing requirement unasked or to stop before asking it. The task therefore measures requirement elicitation and nothing else. Direct scenarios, half of the set, are solvable without any question and establish a 50% floor that single-agent direct answering reproduces exactly; every point above that floor is earned by eliciting the right fields in the sensitive half. Figure 4 shows that our system reaches 97.2% success, 6.6 pp above the strongest baseline, ProCoT, under an exact-match criterion that gives no credit for partially correct site sets. Full results and task details appear in Table 5 and Appendix B.

Figure 4. FoodGo transfer-site selection results show that our assistant achieves the highest success rate in recovering the reference site set from partially disclosed provider requirements.

SystemSuccess ↑\uparrow
Single agent: direct answer0.508 ±\pm 0.000
Single agent: free questioning0.655 ±\pm 0.005
ProCoT0.906 ±\pm 0.008
ATD0.739 ±\pm 0.019
SAGE-Clarify0.778 ±\pm 0.021
LLM-VoI0.892 ±\pm 0.025
Ours0.972 ±\pm 0.013

Table 5. FoodGo transfer-site selection results.

Ablation Studies

Controlled protocol. Both ablations use atomic questions, each requesting one information item or one fallback-acceptance decision, under a cap of eight rounds; policies may stop earlier, and we report actual question counts. The first study removes system components, the second compares clarification policies. This protocol is deliberately different from the main experiment, and numbers are not comparable across the two. The official protocol lets one question request several items, which confounds question selection with question packing; the atomic protocol removes that freedom and charges every item one round. Every policy scores lower under it, and clarification gains less, because less information fits under the cap. The ablations therefore answer one question the main experiment cannot: which component or policy makes the best use of a fixed, item-level budget. The without-clarification variant is the full system with questions disabled under this protocol, not the MAS backbone of Table 1.

CategorySystemSuccess ↑Tool EM ↑Resource EM ↑Tokens (k/episode) ↓
Single-agent methodsQwen3-8B: free questioning0.392 ± 0.0160.407 ± 0.0130.623 ± 0.02615.0 ± 0.15
Single-agent methodsGPT-4.1: free questioning0.454 ± 0.0250.489 ± 0.0270.653 ± 0.02917.1 ± 0.5
Single-agent methodsgpt-oss-120b: free questioning0.389 ± 0.0060.406 ± 0.0020.629 ± 0.01110.2 ± 1.5
Multi-agent coordination baselinesOur MAS backbone0.336 ± 0.0020.350 ± 0.0030.510 ± 0.01117.1 ± 0.1
Multi-agent coordination baselinesMajority vote0.350 ± 0.0100.364 ± 0.0080.525 ± 0.00935.4 ± 0.2
Multi-agent coordination baselinesText-confidence handoff0.324 ± 0.0040.343 ± 0.0050.509 ± 0.00838.8 ± 7.13
Clarification methodsProCoT0.539 ± 0.0080.553 ± 0.0020.678 ± 0.01030.2 ± 0.1
Clarification methodsATD0.533 ± 0.0030.566 ± 0.008<u>0.708 ± 0.008</u>30.2 ± 0.2
Clarification methodsSAGE-Clarify0.539 ± 0.007<u>0.581 ± 0.009</u>0.706 ± 0.01129.0 ± 0.1
Clarification methodsLLM-VoI<u>0.541 ± 0.015</u>0.574 ± 0.0150.696 ± 0.01029.7 ± 0.2
Ours0.578 ± 0.0080.599 ± 0.0080.721 ± 0.00928.2 ± 0.2

Table 1. HoosierHelp main results. Bold and underlined means mark the best and second-best quality scores.

System Component Ablations

Without receiver conditioning. The decisive ablation removes receiver conditioning: questions are selected without conditioning on the downstream receiving decisions. In the full system, a premise is relevant exactly when its possible values change a receiver’s admissible decisions Gjadm\mathcal{G}^{\mathrm{adm}}_j or its remaining above obligations OjO_j (§ 3.3), so a question is worth asking both when the answer can flip a decision and when it supplies the support a fixed decision still lacks. Without this conditioning, the selection no longer sees which premises any receiver depends on.

Figure 5 shows the cost of that change. Success drops from 39.2% to 35.4% and tool-call-set matching from 40.7% to 36.2%. Resource matching drops from 62.3% to 46.4% (15.9 pp), a direct loss in the correctness of the recommended resource set. At the same time, mean questions rise from 3.75 to 5.25 (+40%). The ablated system asks 40% more questions and makes worse decisions on every metric. The pattern is exactly what the mechanism predicts. Without receiver conditioning, the most uncertain premise is asked first whether or not any receiver’s decision depends on it, so rounds go to premises whose answers change nothing, and the rounds left for retrieval and repair shrink accordingly. The loss in resource matching is the direct consequence of asking the wrong questions under a fixed cap, and it cannot be recovered by asking more: the ablation already asks more and does worse. Additional questions cannot substitute for the link between missing information and the decisions that consume it; uncertainty alone does not tell the controller which premise is worth the budget. Receiver conditioning is what turns uncertainty representation into useful information gathering under the shared budget.

Figure 5. System component ablations show that removing receiver conditioning reduces decision quality while requiring more clarification questions.

Other components. Every remaining component contributes. Removing fallback clarification is the largest single loss (−12.0 pp): information questions whose answers cannot repair a blocked plan still consume rounds, which is why removing all clarification (34.5%) scores higher than removing fallback alone (27.2%). Asking without the ability to act on the answer is worse than not asking. Removing the deterministic planner costs 7.2 pp and counterfactual diagnosis 2.0 pp. No ablation matches the full system, and none reduces questions without also reducing success. Every variant that severs the link between questions and receiving decisions converges to the same operating point. Without receiver conditioning, without interaction-cost awareness, without the clarification policy, and the ask-on-disagreement and fixed-order controls of Section 4.3.2 all settle at about five questions per episode and resource matching near 46%. This is the signature of the removed mechanism, not an artifact. Once selection ignores what the receivers need, the rule used to pick the next premise no longer matters; every such policy spends the cap on decision-irrelevant questions and lands in the same place. The full system is the only configuration that escapes it. Appendix D defines each intervention.

Atomic-Question Policy Comparison

Figure 6. Atomic-question policy comparison shows that our controller achieves the highest decision quality while asking fewer questions than most baselines.

Decision quality. We compare our controller with ProCoT, ATD, SAGE-Clarify, and LLM-VoI, and with ask-on-disagreement, fixed-order questioning, and myopic VOI without interaction cost, all under the same atomic-question protocol. Figure 6 shows that our controller is the only policy that leads on success, tool matching, and resource matching simultaneously; every baseline trails on all three. Myopic VOI without interaction cost, the closest decision-theoretic baseline, confirms that the gain does not come from VOI scoring alone but from valuing receiver decisions under a cost-aware budget.

Interaction costs. We measure mean questions per episode. Figure 6 shows that our controller asks fewer questions than SAGE-Clarify and LLM-VoI while beating both on success. ProCoT and ATD ask fewer questions but trail on all three quality metrics.

Controller hyperparameter sensitivity. Appendix G.8 sweeps ∣Ω∣|\Omega|, MM, HH, citemc_{\mathrm{item}}, and pacceptp_{\mathrm{accept}} one at a time. Decision quality is flat across the entire grid, within seed variance of the default, while tokens rise monotonically with retained interpretations, branches, and depth. The defaults were fixed outside the evaluation set and sit at the cost knee: no setting buys quality, and larger settings only buy cost. The results do not depend on these choices.

Conclusion

We presented a conversational food assistant for demand-side resource navigation, nutritional reasoning, and supply-side provision planning. Its receiver-conditioned clarification framework extends single-decision VOI to dependent receiver decisions using shared-premise replay and decision signatures, and jointly values clarification and fallback under a shared budget. Across all tasks, the system outperforms every baseline; under the official HoosierHelp protocol, it also uses fewer tokens than any other clarification method. The ablation supports the central claim: questions should be chosen for their effect on receiver decisions, not premise uncertainty alone. Removing receiver conditioning leads to more questions and worse performance on every decision metric. Jointly evaluating fallback acceptance and its required follow-up questions further connects clarification to the recovery of feasible, evidence-supported plans. These results show the value of grounding multi-agent clarification in the decisions it is meant to improve.

References

A Complete Numerical Results

Table 2–Table 4 report mean ± standard deviation over three runs; quality scores in [0, 1], questions per episode. Bold marks the best mean, including ties. Role-chain MAS omits the proposed clarification and recovery mechanisms and is distinct from the without-clarification intervention.

VariantSuccess ↑\uparrowTool EM ↑\uparrowResource EM ↑\uparrowQ/ep. ↓\downarrow
Full system0.392±\pm0.0160.407±\pm0.0130.623±\pm0.0263.75±\pm0.02
Without receiver conditioning0.354±\pm0.0060.362±\pm0.0100.464±\pm0.0035.25±\pm0.01
w/o fallback clarification0.272±\pm0.0050.278±\pm0.0070.410±\pm0.0113.17±\pm0.00
w/o counterfactual diagnosis0.372±\pm0.0210.383±\pm0.0190.601±\pm0.0063.76±\pm0.03
w/o clarification0.345±\pm0.0190.353±\pm0.0190.557±\pm0.0120.00±\pm0.00
w/o deterministic planner0.320±\pm0.0150.333±\pm0.0130.554±\pm0.0243.71±\pm0.03
w/o interaction-cost awareness0.344±\pm0.0110.352±\pm0.0110.464±\pm0.0095.13±\pm0.01
w/o clarification policy0.350±\pm0.0060.359±\pm0.0070.461±\pm0.0145.27±\pm0.01

Table 4. Component ablations (w/o: without).

B FoodGo Interactive Transfer-Site Selection

Source data and constructed task. No public benchmark evaluates interactive requirement elicitation for provider-side food planning. FoodGo [31, 32] is the only released food-rescue dataset we know of with the required pickup, delivery, and volunteer records. Its synthetic test release contains 500 rescue orders (donor/recipient locations, one to three units, perishable flag) and 50 drivers (origin, service interval). Our task adds candidate sites, hidden profiles, clarification, and exact-site-set evaluation; kk-medoids clustering of available locations yields 12 candidate transfer sites, the complete action space.

Hidden requirements and disclosure. An organization representative seeks one or two sites; its six-field profile specifies (1) donation category: all, perishable only, or bulky only; (2) objective: shorter donor pickup routes, more pickup locations within a radius, or proximity to recipients; (3) coverage radius; (4) operating interval; (5) whether nearby volunteers are required; (6) volunteer distance threshold. The opening message states the site count and part of the profile; replies disclose only the requested item.

Tools and interaction limit. A deterministic planner scores and ranks all feasible site combinations under supplied constraints; a final-recommendation tool accepts site identifiers and ends the dialogue. Dialogues allow eight agent turns; each question consumes one.

Reference solution and success criterion. The same planner enumerates all combinations under the complete hidden profile to obtain the optimal site set. This deterministic reference introduces no planning error, isolating requirement elicitation: agents succeed by eliciting every decision-changing field, and fail only by omitting one, stopping early, or omitting the final call. Success requires the final-recommendation tool call and exact reference-set matching; missing, extra, or wrong sites fail.

Scenario stratification. The 120 scenarios balance site count (one/two) and type (direct/sensitive), 30 per cell. Undisclosed requirements change the optimum in sensitive scenarios; defaults already recover it in direct scenarios. The direct half gives a 50% floor any policy can reach without asking, reproduced by single-agent direct answering; method differences arise from the sensitive half.

Reported results. Table 5: three Qwen3-8B runs over 120 scenarios; highest mean in bold.

Evaluation Reporting Details

Reporting convention. All experiments use three runs; reported means average per-seed results rounded to three decimals; tables and error bars report mean and standard deviation across runs.

HoosierHelp. Each run evaluates 240 instances using Qwen3-8B (except identified GPT-4.1/gpt-oss-120b baselines), with GPT-4.1-mini simulating users under the official setting. Success requires an explicit final recommendation, exact normalized tool-call-set matching, and exact resource-ID-set matching; tool matching covers service, location, schedule, intake, documents, and eligibility [22].

Token accounting. Reported tokens cover all episode backbone calls, including controller candidate sampling, receiver replay, and simulated replies. Replay reruns only receivers in Rt(q)R_t(q) over structured candidates and their evidence, excluding dialogue history; at most six prefiltered questions are evaluated per step; planning and candidate comparison use the deterministic planner. External dialogue dominates cost: each user turn reruns full-history intake, receivers, and retrieval. Scoped replay enables fewer targeted turns: lowest token cost among clarification methods under the official protocol (Table 1); fewer questions than SAGE-Clarify and LLM-VoI under the atomic protocol (Table 3).

PolicySuccess ↑\uparrowTool EM ↑\uparrowResource EM ↑\uparrowQ/ep. ↓\downarrow
Ours0.392±\pm0.0160.407±\pm0.0130.623±\pm0.0263.75±\pm0.02
ProCoT0.324±\pm0.0120.333±\pm0.0140.592±\pm0.0263.51±\pm0.05
ATD0.192±\pm0.0110.200±\pm0.0120.499±\pm0.0353.40±\pm0.09
SAGE-Clarify0.369±\pm0.0050.372±\pm0.0020.507±\pm0.0094.96±\pm0.00
LLM-VoI0.356±\pm0.0120.364±\pm0.0120.510±\pm0.0134.55±\pm0.02
Ask on disagreement0.357±\pm0.0080.366±\pm0.0080.465±\pm0.0105.26±\pm0.01
Fixed order0.350±\pm0.0060.359±\pm0.0070.461±\pm0.0145.27±\pm0.01
Myopic VOI (no cost)0.344±\pm0.0110.352±\pm0.0110.464±\pm0.0095.13±\pm0.01

Table 3. Atomic-question policy results. Myopic VOI (no cost) and Fixed order report the same runs as w/o interaction-cost awareness and w/o clarification policy, respectively, in Table 4.

Protocols. Official HoosierHelp permits multiple items per question; both ablations permit one item or fallback acceptance per question within eight rounds. We compare no absolute numbers across protocols. Atomic questions isolate selection from packing; fewer items fit under the cap, lowering all policies’ scores and compressing clarification gains.

NGQA sampling. A food-name hash split reserves 20% for testing: 2,652 instances with sparse, standard, or complex profiles. Each seed shuffles the pool and takes the first 300; group sizes vary (Table 6), and samples may overlap.

ScopeSparseStandardComplexTotal
Test pool16406843282652
Seed 01946838300
Seed 11827444300
Seed 21788240300

Table 6. NGQA test-pool and per-seed sample counts.

NGQA aggregation and scoring. For group gg, seed ss, sample count ns,gn_{s,g}, and accuracy as,ga_{s,g},

as=1300∑gns,gas,g,aˉ=13∑s=02round⁡3(as).(11)a_s = \frac{1}{300}\sum_g n_{s,g}a_{s,g}, \qquad\bar{a} = \frac{1}{3}\sum_{s=0}^{2}\operatorname{round}_3(a_s). \tag*{(11)}

Group columns average rounded as,ga_{s,g} across seeds. All systems use Qwen3-8B and staged disclosure: designated profile fields remain hidden until clarification; scoring follows the final answer. Binary accuracy, label F1, and explanation ROUGE-L measure distinct output components. Of 328 complex-profile instances, 114 require weighted rules; 214 also admit simpler rules.

FoodGo and controlled ablations. FoodGo uses the 120 scenarios in Appendix B. Atomic question counts reflect actual use, including early stopping; quality-metric definitions are unchanged. Token costs are reported for the official protocol (Table 1) and controller sensitivity (Table 7).

SettingSuccess ↑\uparrowTool set ↑\uparrowResource ↑\uparrowTokens (k) ↓\downarrow
Default39.2±1.639.2 \pm1.640.7±1.340.7 \pm1.362.3±2.662.3 \pm2.628.2
∣Ω∣=4|\Omega|=437.6±1.537.6 \pm1.539.2±1.539.2 \pm1.562.2±0.662.2 \pm0.624.7
∣Ω∣=16|\Omega|=1639.0±1.639.0 \pm1.641.4±1.041.4 \pm1.063.8±1.763.8 \pm1.736.7
∣Ω∣=32|\Omega|=3239.3±0.639.3 \pm0.640.8±0.640.8 \pm0.663.0±0.563.0 \pm0.543.4
M=2M=238.8±1.138.8 \pm1.140.1±1.140.1 \pm1.164.6±1.964.6 \pm1.927.3
M=4M=439.0±2.139.0 \pm2.140.6±2.740.6 \pm2.763.3±1.863.3 \pm1.828.9
M=5M=539.2±0.839.2 \pm0.840.6±0.940.6 \pm0.963.2±1.063.2 \pm1.029.6
H=1H=138.9±1.638.9 \pm1.640.6±1.740.6 \pm1.763.5±0.263.5 \pm0.227.5
H=3H=339.4±0.539.4 \pm0.540.9±0.540.9 \pm0.564.0±0.764.0 \pm0.730.5
citem=0c_{\mathrm{item}}=038.5±0.638.5 \pm0.640.0±0.540.0 \pm0.562.3±0.662.3 \pm0.627.5
citem=0.02c_{\mathrm{item}}=0.0238.3±0.838.3 \pm0.839.8±1.239.8 \pm1.263.0±0.363.0 \pm0.327.4
citem=0.08c_{\mathrm{item}}=0.0839.2±0.739.2 \pm0.740.8±0.540.8 \pm0.563.4±1.163.4 \pm1.127.4
paccept=0.3p_{\mathrm{accept}}=0.340.4±1.440.4 \pm1.441.4±1.341.4 \pm1.364.7±1.064.7 \pm1.028.4
paccept=0.7p_{\mathrm{accept}}=0.739.6±0.439.6 \pm0.441.2±0.541.2 \pm0.564.1±0.864.1 \pm0.827.4

Table 7. Controller sensitivity (quality scores in %).

Component Ablation Definitions

Variants in Figure 5 and Table 4 name the removed component. Without receiver conditioning: selection is unconditioned on downstream receiving decisions. Without fallback clarification: disables fallback-acceptance questions; retains information questions. Without counterfactual diagnosis: removes counterfactual relaxation diagnosis. Without clarification: disables all questions. Without the deterministic planner: substitutes LLM planning. Without interaction-cost awareness: myopic VOI omits interaction cost, removing cost-aware continuation evaluation and changing lookahead. Without the clarification policy: fixed ordering replaces state-dependent whether/what-to-ask decisions.

Selection Error Bound

Let Ct=Qtfeas∪{proceed}C_t = Q_t^{\mathrm{feas}} \cup\{\mathrm{proceed}\}, with Vt(a)V_t(a) and V^t(a)\widehat{V}_t(a) the expected and estimated net values of action aa; proceeding has value zero and a question has estimated value ΔU^t−C^t\widehat{\Delta U}_t-\widehat{C}_t. Let at⋆=arg⁡max⁡aVt(a)a_t^\star= \arg\max_a V_t(a) and a^t=arg⁡max⁡aV^t(a)\hat{a}_t = \arg\max_a \widehat{V}_t(a). Then

max⁡a∈Ct∣V^t(a)−Vt(a)∣≤ϵt⟹Vt(at⋆)−Vt(a^t)≤2ϵt,(12)\max_{a\in C_t}\left|\widehat{V}_t(a)-V_t(a)\right|\le\epsilon_t \quad\Longrightarrow\quad V_t(a_t^\star)-V_t(\hat{a}_t)\le2\epsilon_t, \tag*{(12)}

since V^t(a^t)≥V^t(at⋆)\widehat{V}_t(\hat{a}_t)\ge\widehat{V}_t(a_t^\star) gives Vt(at⋆)≤V^t(at⋆)+ϵt≤V^t(a^t)+ϵt≤Vt(a^t)+2ϵtV_t(a_t^\star)\le\widehat{V}_t(a_t^\star)+\epsilon_t\le\widehat{V}_t(\hat{a}_t)+\epsilon_t\le V_t(\hat{a}_t)+2\epsilon_t.

Ethical Considerations

Benchmark and simulated-interaction results do not establish deployment safety or fairness; service or nutritional errors could harm users. Deployment would require verification of current information, protection of sensitive disclosures, user consent for fallback adjustments, and qualified human oversight.

Method Implementation Parameters

System Architecture

Frozen Qwen3-8B; non-thinking mode; inference only. Workflow: intake →\rightarrow three parallel receivers →\rightarrow joint planner →\rightarrow retriever. Intake maps the full dialogue to candidates with premise identities; receivers read only intake candidates and domain evidence, emitting candidates and decision signatures; the deterministic planner combines signatures and external hard constraints into a jointly feasible plan with consistent fallback adjustments; the retriever maps the planner’s retrieval program and results to recommendations. Receiver roles: navigation—access/eligibility, schedule, logistics; nutrition—profile constraint, food-nutrient, rule applicability; site selection—category/objective, coverage, operations.

Premise fields: id, subject, condition, scope, provenance, version, consumers; forwarding adds consumers without duplicating premises.

Candidates and Joint Interpretations

Sampling: one greedy plus four random draws (temperature 0.7, top_p 0.9); retain K=3K = 3 candidates including the greedy one. Canonicalization merges equal normalized sets of structured constraint fields; support ww is cluster size over total samples. Messages Mi={(a,w,P,F)}M_i = \{(a,w,P,F)\} contain artifacts, support, provenance including consumed message versions, and scoped assumptions and unresolved requirements. Ω\Omega is the Cartesian product of unresolved-premise candidate values, weighted by products of ww; retain at most 8 highest-probability interpretations and renormalize. Each shared premise appears once; each ω\omega assigns one value used by all dependent receivers during replay.

Decisions and Questions

Signature σj(ω)=(Aj(ω),Oj(ω))\sigma_j(\omega) = (A_j^{(\omega)}, O_j(\omega)): decisions not ruled out and remaining obligations. QinfoQ_{\mathrm{info}} targets unresolved premises affecting either, one field per question; unchanged decisions retain questions if OjO_j is nonempty. Rt(q)R_t(q) comprises receivers reachable from v(q)v(q) along dependencies; each scoped premise generates one cross-receiver question. QfallbackQ_{\mathrm{fallback}} requests acceptance of Dmin⁡D_{\min} adjustments. Prefilter Qt=Qinfo∪QfallbackQ_t = Q_{\mathrm{info}} \cup Q_{\mathrm{fallback}} by ∣Rt(q)∣|R_t(q)| and nonempty OjO_j; retain at most 6 feasible questions per step.

Counterfactual Relaxation Diagnosis

Apply adjustments consistently across affected receivers, holding evidence and external requirements fixed. Dmin⁡=Min⁡≺{δ∈Dt:rt(δ)=1}D_{\min} = \operatorname{Min}_{\prec}\{\delta\in D_t : r_t(\delta) = 1\} excludes larger adjustments whenever a smaller one succeeds. Navigation order: time-window relaxation < distance expansion (×1.5\times1.5, ×2\times2) < intake-method relaxation < document-requirement relaxation. Site-selection order: radius expansion (×1.5\times1.5, ×2\times2) < operating-hours relaxation < volunteer-distance requirement removal. Per step: at most 6 adjustments evaluated, 3 fallback questions proposed.

VOI and Selection

Each simulated-response branch updates its target premise once, replays affected receivers, activates accepted adjustments, and scores the complete task. Information branches: M=3M = 3 (top two candidate values and other/unknown); fallback: accept/reject with paccept=0.5p_{\mathrm{accept}} = 0.5. Depth H=2H = 2: initial question plus at most two follow-ups or retries; one greedy rollout per branch. Simulated user: same backbone, conditioned on sampled ω\omega, answering only the requested field. Select q∗=arg⁡max⁡q∈QfeasΔU(q)q^* = \arg\max_{q \in Q_{\mathrm{feas}}} \Delta U(q), where ΔU(q)=E[U(ST)]−Uproceed\Delta U(q) = \mathbb{E}[U(S_T)] - U_{\mathrm{proceed}} and C(q)=E[∑ℓc(qℓ,oℓ)]C(q) = \mathbb{E}[\sum_{\ell} c(q_{\ell},o_{\ell})] includes every continuation interaction. Require positive net value; execute only the first question; exclude continuations exceeding remaining budgets from QfeasQ_{\mathrm{feas}}.

Utility and Cost

U(S)∈[0,1]U(S) \in[0,1]: fraction of active requirements supported by evidence with all OjO_j cleared; count each premise’s improvement once. Inputs: confirmed evidence and retrieval results, never reference answers or hidden profiles. UproceedU_{\mathrm{proceed}}: same criterion after verification and repair without further questions. Costs: citem=0.04c_{\mathrm{item}} = 0.04 per information field; cfallback=0.06c_{\mathrm{fallback}} = 0.06 per fallback question, including confirmation. Tune only citemc_{\mathrm{item}}, cfallbackc_{\mathrm{fallback}}, and pacceptp_{\mathrm{accept}}, exclusively outside evaluation; sensitivity in Appendix G.8.

Budgets, Revision, and Termination

BtB_t: external rounds remaining under the protocol limit; inference caps: 60 backbone calls per step, 180 per episode. Response fields: subject, condition, scope, provenance. St+1=Revise⁡(St,qt,ot)S_{t+1} = \operatorname{Revise}(S_t,q_t,o_t) updates evidence for information responses and preferences for accepted fallbacks; invalidates candidates, signatures, and question values dependent on old versions; reruns affected receivers and the planner in dependency order. New interpretations regenerate candidates; refreshed question sets discard resolved questions. Stop when a complete answer is supported, no feasible question has positive net value, or a budget is exhausted. Verification/repair: at most 2 rounds; unsupported recommendations trigger renewed retrieval.

Sensitivity to Controller Hyperparameters

Table 7: one-at-a-time sweeps around defaults (∣Ω∣|\Omega|=8, MM=3, HH=2, citemc_{\mathrm{item}}=0.04, pacceptp_{\mathrm{accept}}=0.5), atomic protocol, three seeds per setting. Quality is flat: success 0.383–0.404 except ∣Ω∣|\Omega|=4 (0.376); resource matching 0.622–0.647. Tokens increase monotonically with interpretations (24.7k–43.4k), branches (27.3k–29.6k), and depth (27.5k–30.5k), without consistent quality gains; defaults sit at the cost knee. Setting citemc_{\mathrm{item}}=0 lowers success 0.7 pp; this scalar change is distinct from removing cost-aware continuation evaluation altogether (Table 4). For paccept∈{0.3,0.5,0.7}p_{\mathrm{accept}} \in\{0.3,0.5,0.7\}, success is 0.404, 0.392, 0.396; retain 0.5, fixed outside evaluation. Online acceptance estimates from confirmed responses, shrunk toward this prior for sparse histories, are a natural extension; this fixed-prior sweep does not establish their robustness. Mechanism contributions follow from removal (Table 4), not parameter magnitudes: small ∣Ω∣|\Omega| still preserves alternatives; shallow HH still evaluates continuations.

Food Assistant Interface

Figure 7 shows the Food Assistant interface. The chat view clarifies user requirements, while the plan view presents recommended resources and service details. Users can switch between both views within the same conversation.

Food Assistant interface showing clarification and resource plan

Figure 7. Food Assistant interface: clarification (top) and resource plan (bottom). The plan includes service locations, hours, and access details.

Additional Related Work

Large language models (LLMs) have advanced rapidly in recent years [46, 47, 49]. Building on this progress, LLM-driven agents and agentic reinforcement learning have gained increasing attention for their ability to plan, interact with external environments, and solve complex tasks with limited human oversight [8, 9, 55]. In this section, we provide a boarder review of our related work.

Prior Works in Nutrition Personalization. With growing awareness of the importance of dietary health, various studies have sought to incorporate health metrics into personalized food recommendation systems to tackle the challenge of food insecurity [14, 54, 59], aiming at not only provide sufficient but healthy food options. These approaches can be grouped into three primary stages. First, some research emphasizes single indicators like calorie or fat content, as highlighted in works by Ge et al. [11] and Shirai et al. [38], though such metrics often fail to represent the multifaceted nature of a balanced diet. Second, simulated health data has been utilized, as demonstrated by Wang et al. [45], but these methods often diverge from real-world data distributions. However, foods deemed healthy by general standards can still negatively affect certain individuals [50, 58], highlighting the importance of personalized nutrition in-taking.

Intent uncertainty and clarification timing. An assistant can recognize several plausible interpretations of a request before deciding whether to ask the user which one applies. AmbigQA represents the alternatives by pairing plausible answers with disambiguated rewrites [25]. CAmbigNQ uses a follow-up question and the user’s reply to resolve the ambiguity, while CLAM selectively decides when to ask [18, 20]. The timing decision depends on what the uncertainty signal measures. Semantic entropy measures disagreement among sampled answer meanings. SAGE uses sampled answers to calibrate verbal uncertainty, and GCPO calibrates an answer-level uncertainty signal for policy optimization [19, 35, 56]. IntentSim estimates entropy over plausible user intents and evaluates when to ask under different trade-offs between speed and carefulness [53]. Human question timing provides a separate signal. In the collaborative drawing task annotated by CoDraw-iCR [24], Testoni and Fernández [42] find that human question timing diverges from model uncertainty and that questions guided by model uncertainty improve task success.

Learning from subsequent dialogue and incomplete replies. Later turns provide supervision for learning which questions help an assistant proceed. STaR-GATE trains question asking in dialogues with simulated users whose preferences are hidden, using oracle-generated task responses as a reference [3]. Modeling Future Conversation Turns assigns preference labels based on the outcomes of simulated continuations [52]. Even after the user replies, the assistant may still lack a fact needed to act. Not Every User Says It All simulates incomplete replies by masking sentences from user turns. Its StaGPO method assigns credit to individual turns by tracking what the task requires, what remains missing, and how the agent is filling the gap [5]. Drift-Bench tests whether agents recover from input faults such as missing parameters, false presuppositions, and ambiguous wording before execution [4]. Food4All traces failures further downstream, separating failures to translate user requirements into search parameters from failures to carry valid search results into the final referral [21].

Information gain versus decision value. A question can narrow the set of plausible interpretations while leaving the preferred action unchanged. Active Task Disambiguation selects questions by their expected information gain over viable solutions [17]. Qulac evaluates clarification by the quality of the resulting document retrieval [2]. Expected-value question ranking scores the usefulness of anticipated answers, and inference-time value of information weighs the expected utility gain against the cost of asking [10, 29]. Decision-region determination formalizes when an agent has enough information to act. Testing can stop once a single decision succeeds under every hypothesis still consistent with the observations [15]. In a multi-receiver system, this stopping condition can hold for one decision while another still requires clarification. For a food referral, for example, confirming whether the user lives within a pantry’s service area can change the eligibility decision while leaving the nutritional assessment unchanged.

Choosing between assistance and additional computation. Once a gap matters to a decision, its source determines which intervention can close it. KnowNo uses conformal sets over candidate robot actions to decide when to request assistance. Its sequence-level task-completion guarantee depends on its calibration and scenario-distribution assumptions and on receiving correct assistance [30].

When the system needs more capable reasoning, HybridLLM routes requests by predicted difficulty and target quality, and AgentRouter learns allocations from agent performance [7, 57]. Interpreting a routing score as a probability of correctness requires separate calibration. When the gap concerns external evidence, NG-Router couples routing with evidence retrieval, while CiteAudit verifies the existence and metadata of scientific references [33, 36]. When the decision turns on an undisclosed constraint known only to the user, a targeted question can supply the missing information.

References

  1. [1]Emre Can Acikgoz, Jinoh Oh, Joo Hyuk Jeon, Jie Hao, Heng Ji, Dilek Hakkani-Tur, Gokhan Tur, Xiang Li, Chengyuan Ma, and Xing Fan. 2026. MAC: A Multi-Agent Framework for Interactive User Clarification in Multi-turn Conversations. In Proceedings of the 16th International Workshop on Spoken Dialogue System Technology. 1–17.arxiv.org/abs/2512.13154
  2. [2]Mohammad Aliannejadi, Hamed Zamani, Fabio Crestani, and W Bruce Croft. 2019. Asking clarifying questions in open-domain information-seeking conversations. In Proceedings of the 42nd international acm sigir conference on research and development in information retrieval. 475–484.DOI
  3. [3]Chinmaya Andukuri, Jan-Philipp Fränken, Tobias Gerstenberg, and Noah D Goodman. 2024. Star-gate: Teaching language models to ask clarifying questions. arXiv preprint arXiv:2403.19154 (2024).arxiv.org/abs/2403.19154
  4. [4]Han Bao, Zheyuan Zhang, Pengcheng Jing, Zhenqiong Yuan, Kaiwen Shi, and Yanfang Ye. 2026. Drift-bench: Diagnosing cooperative breakdowns in LLM agents under input faults via multi-turn interaction. arXiv preprint arXiv:2602.02455 (2026).arxiv.org/abs/2602.02455
  5. [5]Han Bao, Zheyuan Zhang, Kaiwen Shi, and Yanfang Ye. 2026. Not Every User Says It All: Training Tool-Using Agents Beyond Ideal Simulated Users. arXiv preprint arXiv:2610.24294 (2026).arxiv.org/abs/2610.24294
  6. [6]Yang Deng, Lizi Liao, Liang Chen, Hongru Wang, Wenqiang Lei, and Tat-Seng Chua. 2023. Prompting and evaluating large language models for proactive dialogues: Clarification, target-guided, and non-collaboration. In Findings of the Association for Computational Linguistics: EMNLP 2023. 10602–10621.arxiv.org/abs/2305.13626
  7. [7]Dujian Ding, Ankur Mallick, Chi Wang, Robert Sim, Subhabrata Mukherjee, Victor Rühle, Laks Lakshmanan, and Ahmed H Awadallah. 2024. Hybrid llm: Cost-efficient and quality-aware query routing. In International Conference on Learning Representations, Vol. 2024. 41348–41366.arxiv.org/abs/2404.14618
  8. [8]Guanting Dong, Licheng Bao, Zhongyuan Wang, Kangzhi Zhao, Xiaoxi Li, Jiajie Jin, Jinghan Yang, Hangyu Mao, Fuzheng Zhang, Kun Gai, et al. 2025. Agentic entropy-balanced policy optimization. arXiv preprint arXiv:2510.14545 (2025).arxiv.org/abs/2510.14545
  9. [9]Guanting Dong, Hangyu Mao, Kai Ma, Licheng Bao, Yifei Chen, Zhongyuan Wang, Zhongxia Chen, Jiazhen Du, Huiyang Wang, Fuzheng Zhang, et al. 2025. Agentic reinforced policy optimization. arXiv preprint arXiv:2507.19849 (2025).arxiv.org/abs/2507.19849
  10. [10]Yijiang River Dong, Tiancheng Hu, Zheng Hui, Caiqi Zhang, Ivan Vulic, Andrea Boba, and Nigel Collier. 2026. Value of Information: A Framework for Human–Agent Communication. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 42879–42896.
  11. [11]Mouzhi Ge, Francesco Ricci, and David Massimo. 2015. Health-aware food recommender system. In RecSys.DOI
  12. [12]Touseef Hasan, Laila Cure, and Souvika Sarkar. 2026. TRACE: Trustworthy Retrieval-Augmented Conversational Engine. arXiv preprint arXiv:2608.10176 (2026).arxiv.org/abs/2608.10176
  13. [13]Jiatan Huang, Mingchen Li, Ziming Li, Sunjae Kwon, Hong Yu, and Chuxu Zhang. 2026. Counterfactual Graph for Multi-Agent LLM Calibration. arXiv preprint arXiv:2605.30653 (2026).arxiv.org/abs/2605.30653
  14. [14]Jiatan Huang, Zheyuan Zhang, Tianyi Ma, Mingchen Li, Yaning Zheng, Yanfang Ye, and Chuxu Zhang. 2026. Glen-bench: A graph-language based benchmark for nutritional health. arXiv preprint arXiv:2601.18106 (2026).arxiv.org/abs/2601.18106
  15. [15]Shervin Javdani, Yuxin Chen, Amin Karbasi, Andreas Krause, Drew Bagnell, and Siddhartha Srinivasa. 2014. Near optimal bayesian active learning for decision making. In Artificial Intelligence and Statistics. PMLR. 430–438.arxiv.org/abs/1402.5886
  16. [16]Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. 2023. Llm-blender: Ensembling large language models with pairwise ranking and generative fusion. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 14165–14178.arxiv.org/abs/2306.02561
  17. [17]Katarzyna Kobylczak, Nicolas Astorga, Tennison Liu, and Mihaela van der Schaar. 2025. Active task disambiguation with llms. In International Conference on Learning Representations, Vol. 2025. 37823–37847.arxiv.org/abs/2502.04485
  18. [18]Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2022. Claim: Selective clarification for ambiguous questions with generative language models. arXiv preprint arXiv:2212.07769 (2022).arxiv.org/abs/2212.07769
  19. [19]Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv preprint arXiv:2302.09664 (2023).arxiv.org/abs/2302.09664
  20. [20]Dongryeol Lee, Segwang Kim, Minwoo Lee, Hwanhee Lee, Joonsuk Park, Sang-Woo Lee, and Kyoming Jung. 2023. Asking clarification questions to handle ambiguity in open-domain QA. In Findings of the Association for Computational Linguistics: EMNLP 2023. 11526–11544.
  21. [21]Yiyang Li, Weixiang Sun, Tianyi Ma, Kaiwen Shi, Zheyuan Zhang, and Yanfang Ye. 2025. Food4All: An Agentic Framework and Benchmark for Food Resource Navigation with Adaptive User Understanding. arXiv preprint arXiv:2510.18289 (2025).arxiv.org/abs/2510.18289
  22. [22]Yiyang Li, Weixiang Sun, Tianyi Ma, Kaiwen Shi, Zheyuan Zhang, and Yanfang Ye. 2026. HoosierHelp: Benchmarking LLM Agents for Social Service Navigation. arXiv preprint arXiv:2608.09946 (2026).arxiv.org/abs/2608.09946
  23. [23]Yaokun Liu, Yifan Liu, Daniel Yue Zhang, Ruichen Yao, Zelin Li, and Dong Wang. 2026. PropUQ-MAS: Propagation-Aware Uncertainty Quantification for LLM Multi-Agent Systems. arXiv preprint arXiv:2608.22130 (2026).arxiv.org/abs/2608.22130
  24. [24]Bríelen Madureira and David Schlangen. 2023. Instruction clarification requests in multimodal collaborative dialogue games: Tasks, and an analysis of the CoDraw dataset. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. 2303–2319.
  25. [25]Seow Min, Julian Michael, Hanne­nah Hajishirzi, and Luke Zettlemoyer. 2020. AmbigQA: Answering ambiguous open-domain questions. In Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP). 5783–5797.arxiv.org/abs/2004.10645
  26. [26]Paul Morello. 2021. How food banks and food pantries get their food. Feeding America. https://www.feedingamerica.org/hunger-blog/how-food-banks-and-food-pantries-get-their-food Published December 29, 2021; accessed October 23, 2026.
  27. [27]OpenAI. 2025. gpt-oss-120b & gpt-oss-20b Model Card. arXiv:2508.10925 [cs.CL] https://arxiv.org/abs/2508.10925
  28. [28]OpenAI. 2025. Introducing GPT-4.1 in the API. https://openai.com/index/gpt-4-1/
  29. [29]Sudha Rao and Hal Daumé III. 2018. Learning to ask good questions: Ranking clarification questions using neural expected value of perfect information. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2737–2746.arxiv.org/abs/1805.04655
  30. [30]Allen Z Ren, Anushri Dixit, Alexandra Bodrova, Sumeet Singh, Stephen Tu, Noah Brown, Peng Xu, Leila Takayama, Fei Xia, Jake Varley, et al. 2023. Robots that ask for help: Uncertainty alignment for large language model planners. arXiv preprint arXiv:2307.01928 (2023).arxiv.org/abs/2307.01928
  31. [31]Zachary B Rubenstein. 2020. Dynamic Management of Food Redistribution for 412 Food Rescue. (2020).
  32. [32]Zachary B Rubenstein. 2020. FoodGo Test Dataset [supporting dataset]. (2020).
  33. [33]Kaiwen Shi, Weixiang Sun, Zheyuan Zhang, Lichao Sun, Nitesh V Chawla, and Yanfang Ye. 2026. CiteAudit: You cited it, but did you read it? A benchmark for verifying scientific references in the LLM era. arXiv preprint arXiv:2602.23452 (2026).
  34. [34]Kaiwen Shi, Zheyuan Zhang, Han Bao, Colby Nelson, and Yanfang Ye. 2026. Confidence Laundering in Agent Systems: Why Uncertainty Needs a Latent Carrier. arXiv preprint arXiv:2606.20662 (2026).arxiv.org/abs/2606.20662
  35. [35]Kaiwen Shi, Zheyuan Zhang, and Yanfang Ye. 2026. SAGE: Answer-Conditioned Uncertainty Targets for Verbal Uncertainty Alignment. arXiv preprint arXiv:2606.11512 (2026).arxiv.org/abs/2606.11512
  36. [36]Kaiwen Shi, Zheyuan Zhang, Zhengqing Yuan, Keerthiram Muregusan, Vincent Galassi, Chuxu Zhang, and Yanfang Ye. 2026. NG-router: Graph-supervised multi-agent collaboration for nutrition question answering. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 7508–7527.
  37. [37]Kaiwen Shi, Zheyuan Zhang, Caiqi Caiqi Zhang, Zheyuan Liu, Zehong Wang, Shanglin Wu, Haoran Wang, Xufei Wang, Yijun Ma, Han Bao, et al. [n.d.]. UNCERTAINTY IN LLM AGENTS: COMPREHENSIVE SURVEY. [n.d.].
  38. [38]Sola S Shirai, Oshani Seneviratne, Minor E Gordon, Ching-Hua Chen, and Deborah L McGuinness. 2021. Identifying ingredient substitutions using a knowledge graph of food. Frontiers in Artificial Intelligence (2021).DOI
  39. [39]Jon M Stauffer, Manoj Vanajakumari, Subodha Kumar, and Theresa Mangapora. 2022. Achieving equitable food security: How can food bank mobile pantries fill this humanitarian need. Production and Operations Management 31, 4 (2022), 1802–1821.DOI
  40. [40]Manan Suri, Puneet Mathur, Nedim Lipka, Franck Dernoncourt, Ryan A Rossi, and Dinesh Manocha. 2026. Structured uncertainty guided clarification for LLM agents. In Findings of the Association for Computational Linguistics: ACL 2026. 40811–40838.
  41. [41]Qwen Team. 2025. Qwen3 Technical Report. arXiv:2505.09388 [cs.CL] https://arxiv.org/abs/2505.09388
  42. [42]Albert Testoni and Raquel Fernández. 2024. Asking the right question at the right time: Human and model uncertainty guidance to ask clarification questions. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 258–275.arxiv.org/abs/2402.06509
  43. [43]U.S. Department of Agriculture, Economic Research Service. 2026. Food Security in the U.S. – Measurement. https://www.ers.usda.gov/topics/food-nutrition-assistance/food-security-in-the-us/measurement Updated March 30, 2026.
  44. [44]USDA Economic Research Service. 2026. Food Security in the U.S.: Key Statistics and Graphics. https://www.ers.usda.gov/topics/food-nutrition-assistance/food-security-in-the-us/key-statistics-graphics Updated March 30, 2026. Statistics refer to calendar year 2024.
  45. [45]Wenjie Wang, Ling-Yu Duan, Hao Jiang, Peiguang Jing, Xuemeng Song, and Liqiang Nie. 2021. Market2Dish: health-aware food recommendation. TOMM (2021).DOI
  46. [46]Siheng Xiong, Ali Payani, and Faramarz Fekri. 2026. Enhancing language model reasoning with structured multi-level modeling. In International Conference on Learning Representations, Vol. 2026. 36557–36610.arxiv.org/abs/2611.00616
  47. [47]Siheng Xiong, Ali Payani, Yuan Yang, and Faramarz Fekri. 2025. Deliberate reasoning in language models as structure-aware planning with an accurate world model. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 31900–31931.arxiv.org/abs/2410.03136
  48. [48]Tiany­ang Xu, Dan Zhang, Kushan Mitra, and Estevam Hruschka. 2026. Function-aware planning for multi-agent systems. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 7528–7546.
  49. [49]Yanfang Ye, Zheyuan Zhang, Tianyi Ma, Zehong Wang, Yiyang Li, Shifu Hou, Weixiang Sun, Kaiwen Shi, Yijun Ma, Wei Song, et al. 2025. LLMs4all: A review of large language models across academic disciplines. arXiv preprint arXiv:2509.19580 (2025).
  50. [50]Wenbin Yue, Zidong Wang, Jieyu Zhang, and Xiaohui Liu. 2021. An overview of recommendation techniques and their applications in healthcare. IEEE/CAA Journal of Automatica Sinica (2021).
  51. [51]Chunhan Zhang, Zhenglin Wan, Xingrui Yu, Jingxuan Wu, Qi Wen, Pengfei Zhou, Wangbo Zhao, and Ivor Tsang. 2026. Calibration Is Not Control: Why LLM-Agent Oversight Needs Intervention. arXiv preprint arXiv:2606.21399 (2026).arxiv.org/abs/2606.21399
  52. [52]Michael Zhang, W Bradley Knox, and Eunsol Choi. 2025. Modeling future conversation turns to teach llms to ask clarifying questions. In International Conference on Learning Representations, Vol. 2025. 60722–60742.arxiv.org/abs/2410.13788
  53. [53]Michael JQ Zhang and Eunsol Choi. 2025. Clarify when necessary: Resolving ambiguity through interaction with llms. In Findings of the Association for Computational Linguistics: NAACL 2025. 5541–5558.arxiv.org/abs/2311.09469
  54. [54]Zheyuan Zhang, Yiyang Li, Ni Ha Lan Le, Zehong Wang, Tianyi Ma, Vincent Galassi, Keerthiram Muregusan, Nuno Moniz, Werner Geyer, Nitesh V Chawla, et al. 2025. Nqa: a nutritional graph question answering benchmark for personalized health-aware nutritional reasoning. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 5934–5966.
  55. [55]Zheyuan Zhang, Manqing Mao, Hong Wang, Zhuoer Wang, Samson Koelle, Jie Yuan, Yanjun Lin, James Feng, Nikki Lijing Kuang, Yanfang Ye, et al. 2026. Optimizing what policies learn from: Recoverability-aware rollout intervention learning. arXiv preprint arXiv:2608.05080 (2026).
  56. [56]Zheyuan Zhang, Kaiwen Shi, Han Bao, Zehong Wang, Tianyi Ma, and Yanfang Ye. 2026. Why semantic entropy fails: Geometry-aware and calibrated uncertainty for policy optimization. arXiv preprint arXiv:2605.21801 (2026).arxiv.org/abs/2605.21801
  57. [57]Zheyuan Zhang, Kaiwen Shi, Zhengqing Yuan, Zehong Wang, Tianyi Ma, Keerthiram Muregusan, Vincent Galassi, Chuxu Zhang, and Yanfang Ye. 2026. Agent-router: A knowledge-graph-guided llm router for collaborative multi-agent question answering. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 788–809.
  58. [58]Zheyuan Zhang, Zehong Wang, Shifu Hou, Evan Hall, Landon Bachman, Jasmine White, Vincent Galassi, Nitesh V Chawla, Chuxu Zhang, and Yanfang Ye. 2024. Diet-odin: A novel framework for opioid misuse detection with interpretable dietary patterns. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6312–6323.arxiv.org/abs/2403.08820
  59. [59]Zheyuan Zhang, Zhenghong Wang, Tianyi Ma, Varun Sameer Taneja, Sofia Nelson, Nhi Ha Lan Le, Keerthiram Muregusan, Mingxuan Ju, Nitesh V Chawla, Chuxu Zhang, et al. 2025. Mopi-hrfs: A multi-objective personalized health-aware food recommendation system with llm-enhanced interpretation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2860–2871.arxiv.org/abs/2412.08847
  60. [60]Sen Zhao, Ruiqi Kong, Zuyu Zhang, Lifeng Shen, Xinyu He, Xu Zhang, and Qinghua Zhang. 2026. Topological Coherence for Self-evolving Multi-agent Systems. arXiv preprint arXiv:2609.39753 (2026).

Paper details

vlnc:W_dnmsvwwu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents