Paper deep dive

Replan, Repair, or Edit? A Unified Empirical Evaluation of Travel Agents for Itinerary Revision under Resource Disruptions

Xiaofei Yuan, Yan Zhang, Shaobo Qiao, Huangleshuai He, Leyan Ni, Mingchen Ju, Lujia Yang, Sijia Xu, Yifu Tang, Zhengyi Yang

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 46

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%

Last extracted: 9/18/2026, 4:32:55 AM

Summary

This paper presents a unified empirical evaluation of three travel itinerary revision strategies under resource disruptions: LLM-Z3 (full replanning), IPyHOPPER (classical hierarchical repair), and iTIMO (LLM-based local revision). Using two TREK-derived benchmark sets (500 single-disruption and 200 compound-disruption cases), the study compares effectiveness, plan stability, and computational cost. Results show LLM-Z3 with Gemini achieved the highest compound-disruption success, while IPyHOPPER and iTIMO preserved significantly more of the original itinerary with fewer edits, though iTIMO incurred higher token costs. IPyHOPPER required no LLM inference.

Entities (9)

IPyHOPPER · method · 98%iTIMO · method · 98%LLM-Z3 · method · 98%DeepSeek-v4-Flash · model · 95%Gemini 3.8 Flash · model · 95%Qwen 3.8 Flash · model · 95%TREK · dataset · 92%Z3 · tool · 90%HTN · concept · 88%

Relation Signals (10)

iTIMO → istype → LLM-Based Revision

confidence 98% · iTIMO local-revision adapter

LLM-Z3 → istype → Full Replanning

confidence 98% · We compare LLM–Z3 full replanning

IPyHOPPER → istype → Classical Plan Repair

confidence 98% · IPyHOPPER hierarchical repair

IPyHOPPER → achievedmetric → High Plan Stability

confidence 95% · IPyHOPPER... preserving substantially more of the accepted itinerary on successful repairs.

LLM-Z3 → achievedmetric → Highest Compound-Disruption Success

confidence 95% · LLM–Z3 with Gemini achieved the highest observed compound-disruption success.

IPyHOPPER → hascharacteristic → No LLM Inference

confidence 95% · IPyHOPPER used no LLM inference

LLM-Z3 → usesmodel → Gemini 3.8 Flash

confidence 95% · LLM–Z3 with Gemini achieved the highest observed compound-disruption success.

IPyHOPPER → usesconcept → HTN

confidence 92% · IPyHOPPER reuses plan structure... Hierarchical Task Network (HTN) planning

iTIMO → hascharacteristic →

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Travel-planning agents generate itineraries that may become infeasible after acceptance because of flight cancellations, hotel unavailability, or attraction closures. Revising these itineraries involves full replanning, classical plan repair, and LLM-based travel-agent revision, whose differing task formulations and evaluation protocols hinder comparison. We conduct a systematic empirical study using two TREK-derived benchmark sets: 500 single-disruption cases, including feasible and infeasible instances, and 200 feasible simultaneous compound-disruption cases. We compare LLM-Z3 full replanning, IPyHOPPER hierarchical repair, and an iTIMO local-revision adapter across effectiveness, plan stability, and computational cost. LLM-Z3 with Gemini achieved the highest observed compound-disruption success. IPyHOPPER nearly matched that configuration's single-disruption overall success, while preserving substantially more of the accepted itinerary on successful repairs. Successful hierarchical and local repairs made fewer edits and retained more accepted commitments than full replanning. Computational profiles differed: IPyHOPPER used no LLM inference, the evaluated LLM-Z3 adapter used compact one-call inference, and the iTIMO adapter consumed substantially more tokens. The study provides practical guidelines for balancing feasibility recovery, commitment preservation, and computational cost within evaluated settings.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

42,221 characters extracted from source content.

Expand or collapse full text

Replan, Repair, or Edit? A Unified Empirical Evaluation of Travel Agents for Itinerary Revision under Resource Disruptions Xiaofei Yuan 1 , Yan Zhang 1 , Shaobo Qiao 1 , Huangleshuai He 2 , Leyan Ni 2 , Mingchen Ju 1 , Lujia Yang 3 , Sijia Xu 2 , Yifu Tang 4 , and Zhengyi Yang 3(B) 1 Euler AI; 2 University of New South Wales; 3 University of Sydney; 4 Vecton AI xiaofei.yuan,yan.zhang,shaobo.qiao,mingchen.ju@eulerai.au; huangleshuai.he,leyan.ni,sijia.xu@unsw.edu.au; lyan8923,zhengyi.yang@sydney.edu.au; yves.tang@vectonai.com Abstract. Travel-planning agents generate itineraries that may be- come infeasible after acceptance because of flight cancellations, hotel unavailability, or attraction closures. Revising these itineraries involves full replanning, classical plan repair, and LLM-based travel-agent revi- sion, whose differing task formulations and evaluation protocols hinder comparison. We conduct a systematic empirical study using two TREK- derived benchmark sets: 500 single-disruption cases, including feasi- ble and infeasible instances, and 200 feasible simultaneous compound- disruption cases. We compare LLM–Z3 full replanning, IPyHOPPER hierarchical repair, and an iTIMO local-revision adapter across effec- tiveness, plan stability, and computational cost. LLM–Z3 with Gemini achieved the highest observed compound-disruption success. IPyHOP- PER nearly matched that configuration’s single-disruption overall suc- cess, while preserving substantially more of the accepted itinerary on successful repairs. Successful hierarchical and local repairs made fewer edits and retained more accepted commitments than full replanning. Computational profiles differed: IPyHOPPER used no LLM inference, the evaluated LLM–Z3 adapter used compact one-call inference, and the iTIMO adapter consumed substantially more tokens. The study pro- vides practical guidelines for balancing feasibility recovery, commitment preservation, and computational cost within evaluated settings. Keywords: Travel planning· Itinerary revision· Plan repair· Language agents· Empirical evaluation 1 Introduction Travel itinerary planning coordinates transportation, accommodation, and activ- ities under user preferences, budgets, schedules, and resource availability. Travel- planning agents can generate feasible itineraries, including by combining large language models (LLMs) with formal verification tools [33, 29, 9, 35, 13]. Flight arXiv:2609.19654v1 [cs.AI] 17 Sep 2026 2X. Yuan et al. Itinerary Revision under Resource Disruptions Travel request q Accepted itinerary P Flight Hotel Attraction Resource disruptions ∆ Flight cancelled / hotel unavailable / attraction closed Updated environment K ′ P may become infeasible q unchanged Revision If P is infeasible Feasible revised itinerary P ′ Infeasibility diagnosis ⊥ or Fig. 1. Itinerary revision under resource disruptions, resulting in either a feasible re- vision or an infeasibility diagnosis. cancellations, hotel unavailability, or attraction closures can subsequently inval- idate an accepted itinerary [15, 21], shifting the task from initial generation to revision. Revision starts from accepted commitments and preferences. Restoring fea- sibility may require changes beyond the unavailable resource, but generating a new itinerary can also replace arrangements that remain valid. Evaluation must therefore consider preservation of the accepted itinerary [6] alongside fea- sibility recovery and computational cost, as illustrated in Figure 1. Studies of LLM evaluation and agentic systems across classification, question answering, retrieval, financial analysis, data exploration, and serving provide complemen- tary examples of task-specific quality, reliability, and efficiency assessment [28, 17, 30, 26, 19, 25, 31, 18, 11]. Three paradigms differ in their use of the accepted itinerary: full replanning solves the updated problem anew, classical plan repair reuses plan structure, and LLM-based travel-agent revision edits the itinerary directly [9, 32, 12]. We compare these scopes by effectiveness, preservation, and computational cost. Motivation. Four research gaps motivate the comparison. (1) Limited bench- mark coverage for itinerary revision. Travel-planning benchmarks empha- size itinerary generation and broader planning capabilities [29, 22, 1–3]. Despite work on adaptive planning and itinerary modification [15, 12], benchmark sup- port for revising accepted itineraries under resource disruptions remains limited, particularly for evaluating both single and compound disruptions. (2) Lack of a unified evaluation setting. Full replanning, classical plan repair, and LLM-based travel-agent revision originate from different task formulations and evaluation settings, so previously reported results are not directly comparable. A common revision interface, feasibility criterion, and controlled disruption set- ting are needed to compare them. (3) Incomplete evaluation dimensions. Success rates do not reveal how much of an accepted itinerary is preserved or at Travel Agents for Itinerary Revision3 what computational cost. High repair coverage may come with extensive changes, while local revision may incur high inference or runtime costs despite making few edits. (4) Limited strategy-selection guidance. Existing evaluations provide limited guidance for choosing a revision paradigm when deployments prioritize feasibility recovery, plan preservation, and computational cost. Contributions. We compare LLM–Z3 full replanning, IPyHOPPER hierarchi- cal repair, and an iTIMO-based local revision adapter, with four contributions: – A benchmark for itinerary revision under resource disruptions. Two fixed TREK-derived benchmark sets [22] provide 500 single-disruption cases, including feasible and infeasible instances, and 200 feasible compound- disruption cases. All methods receive the same instances. – A unified evaluation setting for heterogeneous revision strategies. A common interface takes the original request, accepted itinerary, disrup- tions, and updated environment, and assesses complete repairs or infeasibil- ity outcomes against a common feasibility criterion. Unsuccessful attempts count as failures; method-specific restrictions and budgets are explicit. – A multi-dimensional evaluation framework. Effectiveness covers feasi- ble repair, overall success, and correct refusal; stability covers logical edits and preservation; computational cost covers latency, LLM calls, and tokens. – Practical guidelines for revision-strategy selection. Observed trade- offs characterize the strengths, limitations, and suitable deployment settings of the three paradigms for the evaluated implementations and settings. 2 Background and Problem Formulation This section formalizes the travel-itinerary revision problem considered in this study. We define the planning environment, resource disruptions, revision out- comes, plan stability, and the feasibility-first repair objective. 2.1 Travel Itinerary Planning A travel request q specifies the user’s requirements and preferences for a trip, such as destinations, dates, and budget. The planning environment K contains knowledge-base records, resource availability, and execution conditions needed to assess executability. An itinerary X is a complete plan comprising flights or other transportation between destinations, accommodation, local transportation where applicable, and attractions or activities with temporal and spatial assign- ments. Let C(q,K) be the set of applicable constraints. Request constraints require the itinerary to satisfy the trip requirements in q. Grounding requires its re- sources and associated information to be supported by the planning environ- ment. Persona or preference constraints capture the traveller’s specified needs, while budget constraints bound the cost of the trip. Temporal consistency con- cerns the timing and compatibility of scheduled components; spatial consistency 4X. Yuan et al. concerns their locations and the connections between them. These constraints determine the feasible set F(q,K) =X | c(X) = true for every c∈ C(q,K).(1) Initial itinerary generation seeks any complete X ∈F(q,K). Once accepted, the itinerary is denoted by P, with P ∈F(q,K) before disruption. 2.2 Resource Disruptions and Revision Outcomes A resource is a travel entity used in an itinerary; this study considers disruptions to flights, hotels, and attractions. A disruption δ makes a previously available resource unavailable. Let ∆ =δ 1 ,...,δ k denote the set of disruptions:|∆| = 1 is a single disruption, and |∆| > 1 is a compound disruption. Compound dis- ruptions in this study are presented simultaneously, rather than arriving as a sequence of online changes. They produce the updated environment K ′ = Update(K,∆). The request q remains unchanged, but disruptions may render P infeasible in K ′ through temporal and spatial dependencies. For example, a replacement flight may shift the arrival time and prevent a scheduled attraction visit. The common revision interface takes the original request, accepted itinerary, disruptions, and updated environment: f(q,P,∆,K ′ )−→ P ′ , ⊥, or fail.(2) Here P ′ denotes a complete revised itinerary; it is a successful repair only if P ′ ∈F(q,K ′ ). The outcome⊥ denotes a correct infeasibility diagnosis, requiring F(q,K ′ ) =∅. The outcome fail records an attempt that produces neither a valid repair nor a justified infeasibility outcome within the method’s execution procedure or budget. Examples include timeouts, execution errors, exhausted search or correction budgets, and malformed or invalid candidates. Failure to find a feasible repair does not establish infeasibility. The original user requirements are not silently relaxed to make the updated problem feasible. 2.3 Plan Stability Feasible revisions can differ in how much of P they retain. Plan stability is assessed through logical itinerary components, or logical slots, each representing a semantically meaningful commitment rather than a raw JSON field or textual string. Examples include a flight assigned to a route, a hotel booking for a city, and an attraction assignment for a day. The logical edit measure d(P,P ′ ) counts semantic itinerary changes, not tex- tual edits. Let S(P) denote the logical slots of P. Preservation is the fraction of slots in S(P) whose assignments remain unchanged in P ′ . Fewer edits and higher preservation generally indicate greater stability. The measures are comple- mentary rather than mathematically equivalent. Section 4.1 specifies the bench- mark’s matching and counting rules and reports stability on successful feasible repairs. Travel Agents for Itinerary Revision5 2.4 Feasibility-First Repair Objective Feasibility is the primary requirement: retaining accepted commitments does not compensate for an invalid itinerary. When F(q,K ′ ) is nonempty, an ideal minimum-edit revision satisfies P ∗ ∈ arg min X∈F(q,K ′ ) d(P,X).(3) This conceptual preference favors fewer changes among feasible alternatives; it is neither an objective implemented by all methods nor a guarantee of globally minimum-edit repair. 3 Related Work and Representative Strategy Selection Existing work relevant to itinerary revision can be organized into three broad lines according to how an accepted plan is used after the environment changes. First, travel-planning and solver-assisted approaches typically formulate the up- dated request as a new planning problem and generate a complete feasible itinerary, with limited emphasis on preserving the previously accepted plan [29, 9]. Second, classical plan-repair methods explicitly reuse an existing plan, retain- ing prior assignments or structural decompositions while revising the parts af- fected by changed goals or execution conditions [6, 32]. Third, recent LLM-based itinerary-modification methods operate directly on the current itinerary through semantic editing operations rather than reconstructing a formal plan [12]. These lines differ mainly in their use of the accepted itinerary and the scope of revision. We therefore group the literature into full replanning, classical plan repair, and LLM-based travel-agent revision, and evaluate one representative per category. 3.1 Full Replanning: LLM–Z3 Full replanning seeks a complete feasible itinerary for an updated problem with- out necessarily preserving P. TravelPlanner [29] and AgentTravel [34] address itinerary generation and knowledge-augmented planning. External reasoning supports constraint satisfaction through verification in LLM-Modulo [8], mixed- integer optimization in To the Globe [14], formalized programming in LLMFP [10], and LLM–solver integration in Personal Travel Solver [24]. RETAIL [5] and ATLAS [4] address complex or changing requirements without centering accepted-itinerary preservation. Hao et al. [9] translate requests into executable steps and code invoking a Satisfiability Modulo Theories (SMT) solver. Formal solving separates constraint satisfaction from language generation and supports full replanning without a preservation objective. Our LLM–Z3 Full Replan adapter uses a fixed Python/Z3 model [20] rather than generated solver code. One LLM call extracts structured requirements; the model solves updated resource constraints and decodes P ′ . It does not optimize similarity to P, which is retained for post-hoc stability evaluation. 6X. Yuan et al. 3.2 Classical Plan Repair: IPyHOPPER Classical repair reuses plans under changed goals, constraints, or execution con- ditions. Related work spans assignment reuse in dynamic constraint satisfaction [27], localized plan adaptation [7], planning-based repair [16], and plan stability and minimum-change repair [6, 23]. Hierarchical Task Network (HTN) planning decomposes tasks into subtasks and primitive actions, providing structure for repair methods including SHOP-FIXER, IPyHOPPER, and REWRITE [32]. IPyHOPPER [32] reuses plan structure while adapting repair scope. Given decomposition tree T and primitive sequence π = plan(T), simulation identifies a failed action and replaces an ancestor task’s decomposition. If the result is inapplicable, backtracking expands repair to a higher-level task. This reuse does not guarantee minimum edits. Our adapter reconstructs T from P before activating ∆. Updated conditions identify the affected action, and the repaired primitive sequence is decoded into P ′ . Candidates target disrupted slots; the worker uses domain constraints with- out the benchmark scorer. Simulation is internal and does not imply that the traveller has executed a trip prefix. 3.3 LLM-Based Travel-Agent Revision: iTIMO LLM-based revision edits an itinerary semantically rather than reconstructing a formal plan or repairing an HTN hierarchy. iTIMO [12] provides an operation- based formulation over points of interest (POIs): O =o add ,o replace ,o delete .(4) These operations add, replace, or delete a POI. The source benchmark evalu- ates POI-level perturbations based on popularity, spatial distance, and category diversity. iTIMO’s local modification task motivates our editing baseline; its formulation is not a general resource-disruption repair algorithm. Our iTIMO adapter receives q, P, ∆, and K ′ and selects operations and tar- gets, leaving untouched components unchanged. Unlike the source’s single POI- level modification, it permits corrections for simultaneous disruptions within the budget in Section 4.1. Corrections modify the preceding result and do not rep- resent new arrivals of ∆; the adapter remains a bounded local-repair baseline. 3.4 Comparison of Representative Strategies Table 1 highlights the main differences among the three representative strategies. Full replanning treats the updated problem globally and does not use the ac- cepted itinerary P as a repair structure, whereas IPyHOPPER explicitly reuses the decomposition derived from P and adapts the repair scope hierarchically. The iTIMO adapter instead edits P directly through bounded local operations. These differences yield three distinct revision scopes—global, adaptive hierar- chical, and bounded local—and motivate their comparative evaluation under a common revision setting. Travel Agents for Itinerary Revision7 Table 1. Conceptual comparison of the evaluated revision strategies. Paradigm Representative Role of acceptedP Repair scope Core mechanism Full replanning LLM–Z3 adapter Used only for stability evaluation GlobalFormal constraint solving Classical plan repair IPyHOPPER Decomposition from P reused Adaptive hierarchical HTN repair and backtracking LLM-based revision iTIMO adapterP directly editedBounded local ADD, REPLACE, DELETE 4 Experimental Evaluation Figure 2 summarizes the unified evaluation setting. We evaluate the three rep- resentative strategies using identical fixed instances under both single and si- multaneous compound resource disruptions. The evaluation covers effectiveness, plan stability, and computational cost, capturing repair coverage, preservation of accepted commitments, and execution or inference overhead. 4.1 Experimental Setup The experimental setup specifies the benchmark instances, evaluation metrics, and implementation settings used for all compared strategies. Benchmark and tasks. All methods receive identical fixed TREK-derived in- stances [22]: a request, accepted itinerary, and disruptions in an updated envi- ronment. Reference repairs and feasibility labels are hidden. The 500 single-disruption cases comprise 375 feasible flight, optional-hotel, and optional-attraction disruptions and 125 infeasible required-hotel conflicts, testing repair and infeasibility detection under unchanged requirements. The 200 feasible compound cases comprise 50 each of AB, AC, BC, and ABC (A/B/C: flight/hotel/attraction disruptions), with multiple components invalidated si- multaneously. The benchmarks share 191 source queries but are reported sepa- rately; compound groups use different itineraries without difficulty matching. Evaluation metrics. We evaluate each configuration along three complemen- tary dimensions: effectiveness, plan stability, and efficiency. Together, these met- rics capture whether a revision succeeds, how much of the accepted itinerary it preserves, and the computational cost required to obtain it. Effectiveness. Valid repairs are complete itineraries passing applicable TREK checks for requirements, grounding, persona, budget, and spatio-temporal con- sistency, and avoiding disrupted resources. Single-disruption feasible repair rate, correct refusal rate, and overall success rate are percentages: valid repairs di- vided by 375, correct refusals divided by 125, and their sum divided by 500, respectively. For the 125 required-hotel conflicts, correct refusal checks recog- nition of the constructed conflict, not a general-purpose proof of infeasibility. 8X. Yuan et al. Unified Evaluation of Representative Revision Strategies Common revision instance(q,P,∆, K ′ ) Same fixed instance for all methods; method-specific representations 500 single-disruption cases 375 feasible + 125 infeasible 200 compound-disruption cases All 200 feasible LLM–Z3 Full Replanning IPyHOPPER Hierarchical Repair iTIMO Local Revision GLOBALADAPTIVE HIERARCHICALBOUNDED LOCAL Solve updated problem Complete plan P ′ Formal constraint solving No similarity objective on P Reconstruct hierarchy from P HTN repair + backtracking Repair may expand upward Direct semantic edits to P HA F ′ ADD / REPLACE / DELETE Preserve untouched components Common Evaluation Effectiveness repair / correct refusal / overall success Plan Stability logical edits / preservation Computational Cost calls / tokens / latency Fig. 2. Unified evaluation of the three representative revision strategies across effec- tiveness, plan stability, and computational cost. Errors, invalid outputs, and timeouts in retained final records count as failures in the denominators, not correct refusals. All compound cases are feasible: overall success equals repair rate over 200 cases; refusal is inapplicable. Plan stability. Over successful feasible repairs, mean logical edits averages d(P,P ′ ); preservation rate averages the percentage of original logical slots retaining the same resource or value at the corresponding key. Flights are keyed by route, hotels and cars by city, and city assignments and attraction sets by day. Daily attraction edits count as max(|removed|,|added|). These implement Section 2.3’s semantic measures. Table rows use each configuration’s successful feasible subset; paired comparisons use joint successes. Efficiency. Final iTIMO releases replaced 283 DeepSeek single-disruption provider-balance errors via targeted reruns and 43 Gemini compound-disruption execution-error records through official-endpoint reruns. Metrics describe re- tained final outputs; discarded attempts are excluded from consumption aggre- gation. Where records are available, mean calls and tokens cover all retained cases, including failures. Qwen iTIMO figures use reported summaries. End- to-end latency is summarized in seconds by mean, median, and 95th percentile (P95), linearly interpolated where raw records are available. IPyHOPPER’s zero calls/tokens mean no LLM inference, not zero computation. Implementation settings. LLM–Z3 and the iTIMO adapter use Gemini 3.8 Flash, DeepSeek V4 Flash, and Qwen 3.8 Flash service aliases. LLM–Z3 uses one call per case, temperature 0, a 2,048-token output limit, no retry, and disabled thinking for DeepSeek/Qwen. It checks up to 25 resource combinations without a Travel Agents for Itinerary Revision9 similarity objective. IPyHOPPER uses no LLM, scorer-free disrupted-slot candi- dates, a 300-second limit, and no retry. The iTIMO adapter permits two planning rounds for one disruption and k + 1 for k simultaneous disruptions; format cor- rection may add calls. Documented Gemini/DeepSeek settings use temperature 0 and a 2,048-token output limit. Qwen iTIMO adapter raw records, sampling settings, and P95 interpolation details are unavailable. 4.2 Single-Disruption Evaluation Design. The single-disruption experiment evaluates isolated resource failures using 375 feasible repair cases and 125 infeasible cases, with effectiveness, sta- bility, and efficiency assessed jointly. Table 2. Single-disruption effectiveness and plan stability. Stability metrics use suc- cessful feasible repairs. MethodLLM Overall success (%) Feasible repair (%) Correct refusal (%) Mean edits Preserved slots (%) LLM–Z3 Gemini 3.8 Flash97.096.00100.0 6.62865.35 LLM–Z3 DeepSeek V4 Flash84.690.1368.0 6.61865.25 LLM–Z3 Qwen 3.8 Flash91.088.5398.4 6.57265.36 IPyHOPPER –96.895.73100.0 1.00092.67 iTIMO adapter Gemini 3.8 Flash87.883.73100.0 1.00392.59 iTIMO adapter DeepSeek V4 Flash88.684.80100.0 1.00092.33 iTIMO adapter Qwen 3.8 Flash67.073.3348.0 1.00091.95 Effectiveness and plan stability. Table 2 gives effectiveness in its first three metric columns and stability, conditional on successful feasible repairs, in its final two. LLM–Z3 with Gemini achieved 485/500 overall successes (97.0%) ver- sus IPyHOPPER’s 484/500 (96.8%); this one-case difference alone supports no meaningful ranking. With DeepSeek, LLM–Z3 repaired 338/375 cases and cor- rectly refused 85/125 conflicts, whereas the iTIMO adapter repaired 318/375 and refused all 125. Overall success therefore reversed their feasible-repair ordering. Full replanning averaged about 6.6 edits and 65% preservation, versus one edit and 92% for successful IPyHOPPER and iTIMO adapter repairs. On the 346 jointly successful Gemini LLM–Z3/IPyHOPPER cases, respective means were 6.809 versus 1.000 edits and 64.70% versus 92.68% preservation. Restricting to joint successes removes differences from the methods succeeding on different case subsets; the comparison remains conditional on both returning valid repairs. 10X. Yuan et al. Table 3. Single-disruption efficiency over 500 cases. Calls and tokens are per-case means; latency is in seconds. MethodLLMCalls Tokens Mean (s) Median (s) P95 (s) LLM–Z3Gemini 3.8 Flash 1.000 1,369 5.103.89 9.08 LLM–Z3DeepSeek V4 Flash 1.000938 3.251.96 4.83 LLM–Z3Qwen 3.8 Flash1.000999 4.483.31 6.53 IPyHOPPER –00 2.051.98 2.57 iTIMO adapter Gemini 3.8 Flash 1.244 23,469 14.21 13.00 29.72 iTIMO adapter DeepSeek V4 Flash 1.132 19,486 8.418.27 16.35 iTIMO adapter Qwen 3.8 Flash1.430 24,205 11.88 17.24 23.55 Efficiency. Table 3 reports per-case inference consumption and latency, sub- ject to the record-availability qualification above. LLM–Z3 uses one call and 938–1,369 tokens per case. The iTIMO adapter averaged 1.132–1.430 calls and 19,486–24,205 tokens despite approximately one logical edit per successful re- pair. IPyHOPPER required no LLM inference and recorded a 1.98-second me- dian. Few edits do not imply low inference consumption; latency describes these implementations and environments, not intrinsic algorithm speed. Summary. LLM–Z3 with Gemini and IPyHOPPER achieved nearly identical overall success. Successful hierarchical and local repairs preserved more accepted commitments than full replanning, while computational costs differed substan- tially across strategies. 4.3 Compound-Disruption Evaluation Design. The 200 feasible compound cases comprise 50 each of AB, AC, BC, and ABC, representing simultaneous flight–hotel, flight–attraction, hotel–attraction, and flight–hotel–attraction disruptions; refusal is therefore inapplicable. Table 4. Compound-disruption effectiveness and plan stability over 200 feasible cases. Stability metrics use successful repairs. MethodLLM Overall success (%) AB AC BC ABC Mean edits Preserved slots (%) LLM–Z3Gemini 3.8 Flash93.0 47 44 49 46 9.06557.25 LLM–Z3DeepSeek V4 Flash85.5 41 42 48 40 9.05356.96 LLM–Z3Qwen 3.8 Flash84.0 41 41 46 40 8.98257.06 IPyHOPPER –89.0 47 41 49 41 2.23086.73 iTIMO adapter Gemini 3.8 Flash67.0 46 36 33 19 2.14286.16 iTIMO adapter DeepSeek V4 Flash75.5 37 46 36 32 2.21286.67 iTIMO adapter Qwen 3.8 Flash67.0 31 37 42 24 2.17986.35 Travel Agents for Itinerary Revision11 Effectiveness and plan stability. Table 4 reports overall success over 200 cases, combination-level successful counts out of 50, and stability on success- ful repairs. LLM–Z3 with Gemini repaired 186/200 cases (93.0%), the highest observed success, followed by IPyHOPPER at 178/200 (89.0%). The iTIMO adapter repaired 134, 151, and 134 cases with Gemini, DeepSeek, and Qwen. IPyHOPPER and LLM–Z3 with Gemini both repaired 47/50 AB and 49/50 BC cases; the iTIMO adapter with Gemini repaired 46/50 AB but 19/50 ABC cases. These differences do not establish a causal effect of disruption count because the groups use different itineraries. Full replanning averaged about 9.0 edits and 57% preservation, versus 2.1– 2.2 edits and 86%–87% for hierarchical/local repair. On 167 jointly successful Gemini LLM–Z3/IPyHOPPER cases, respective means were 9.132 versus 2.222 edits and 57.28% versus 86.77% preservation. The paired comparison fixes the successful case subset and does not describe stability when either method fails. Table 5. Compound-disruption efficiency over 200 cases. Calls and tokens are per-case means; latency is in seconds. MethodLLMCalls Tokens Mean (s) Median (s) P95 (s) LLM–Z3Gemini 3.8 Flash 1.000 1,594 6.494.49 8.33 LLM–Z3DeepSeek V4 Flash 1.000 1,086 4.852.57 7.77 LLM–Z3Qwen 3.8 Flash1.000 1,173 6.334.09 12.08 IPyHOPPER –00 20.232.32 300.00 iTIMO adapter Gemini 3.8 Flash 2.905 67,631 31.62 30.81 52.29 iTIMO adapter DeepSeek V4 Flash 1.660 34,843 13.19 12.32 23.96 iTIMO adapter Qwen 3.8 Flash1.845 39,182 19.29 16.79 38.76 Efficiency. Table 5 reports per-case calls, tokens, and latency under the same provenance qualifications. LLM–Z3 used one call and averaged 1,086–1,594 to- kens; the iTIMO adapter averaged 1.660–2.905 calls and 34,843–67,631 tokens. IPyHOPPER required no LLM inference, but its 2.32-second median accompa- nied a 20.23-second mean, 300-second P95, and 12/200 timeouts. Median latency alone omits this tail; execution differences preclude an intrinsic speed ranking. Summary. LLM–Z3 with Gemini achieved the highest observed compound suc- cess. Hierarchical and local repairs preserved more accepted commitments, while IPyHOPPER showed a long latency tail and the iTIMO adapter higher token consumption. 5 Practical Guidelines for Revision Strategy Selection No strategy dominates all evaluated dimensions. Table 6 summarizes trade-offs within the evaluated setting. 12X. Yuan et al. Table 6. Strategy-selection guidance within the evaluated setting. PriorityMore attractiveMain trade-off / caution Feasibility recovery LLM–Z3 / IPyHOPPER Preservation / latency cost PreservationIPyHOPPER / iTIMO adapter Hierarchy / lower coverage Limited LLM inference IPyHOPPER / LLM–Z3 None / compact one-call inference Local semantic editing iTIMO adapterBounded scope; locality not controlled LLM–Z3: Feasibility Recovery, Compact Inference. LLM–Z3’s observed coverage favors feasibility recovery when broad revisions are acceptable. Its larger changes are consistent with the absence of a preservation objective, and coverage remains configuration-dependent. IPyHOPPER: Strong Preservation, No LLM Inference. IPyHOPPER is attractive when preservation is prioritized and a usable hierarchy is available, requiring no LLM inference. Its compound-disruption latency tail matters under strict runtime requirements. iTIMO: Local Editing, Higher Inference Cost. The iTIMO adapter sup- ports preservation-oriented local editing, but showed lower compound coverage and substantially higher token consumption in the evaluated configurations; lo- cality was not independently varied. These guidelines are limited to the evaluated implementations and synthetic resource-disruption setting. Repair scope is confounded with feedback, candi- date restrictions, model routing, and budgets, so differences cannot be attributed solely to paradigms; no method is shown to return globally minimum-edit P ∗ . Refusal covers required-hotel conflicts only, while stability is conditioned on suc- cessful repairs and joint successes for paired comparisons. Compound groups use different itineraries and are not difficulty-matched. Runtime and token results de- pend on implementations, environments, and service endpoints. Exact historical equivalence for iTIMO cannot be established because its external knowledge- base/scorer checkout was not fingerprinted; Qwen iTIMO also lacks raw records, sampling settings, and P95 interpolation details. 6 Conclusion We evaluated LLM–Z3 full replanning, IPyHOPPER hierarchical repair, and iTIMO-based local revision under single and compound resource disruptions. LLM–Z3 with Gemini achieved the highest observed compound success, while IPyHOPPER reached comparable single-disruption overall success with greater preservation. Hierarchical and local repairs made fewer edits than full replanning, while the strategies exhibited distinct computational costs. No strategy domi- nates all dimensions; selection should balance feasibility recovery, preservation, and computational cost. Travel Agents for Itinerary Revision13 References 1. Chaudhuri, S., Purkar, P., Raghav, R., Mallick, S., Gupta, M., Jana, A., Ghosh, S.: TripCraft: A benchmark for spatio-temporally fine grained travel planning. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). p. 17035–17064. Association for Computa- tional Linguistics, Vienna, Austria (2025). https://doi.org/10.18653/v1/2025.acl- long.834 2. Chen, W., Wang, S., Gao, Z., Hu, K., Ni, W., Di, S., Zhang, C.J., Chen, L.: Trav- elEval: A comprehensive benchmarking framework for evaluating LLM-powered travel planning agents. In: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. p. 8719–8730. Association for Computing Machinery (2026). https://doi.org/10.1145/3770855.3817533 3. Cheng, X., Hu, Y., Zhang, X., Xu, L., Tan, L., Pan, Z., Li, X., Liu, Y.: Beyond itinerary planning—a real-world benchmark for multi-turn and tool-using travel tasks. In: Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). p. 29200–29251. Association for Computational Linguistics, San Diego, Cal- ifornia, United States (2026). https://doi.org/10.18653/v1/2026.acl-long.1347, https://aclanthology.org/2026.acl-long.1347/ 4. Choi, J., Yoon, J., Chen, J., Jha, S., Pfister, T.: ATLAS: Constraints- aware multi-agent collaboration for real-world travel planning. In: International Conference on Learning Representations (2026), https://openreview.net/forum?id=mIYGiBf9Pm 5. Deng, B., Feng, Y., Liu, Z., Wei, Q., Zhu, X., Chen, S., Guo, Y., Wang, Y.: RETAIL: Towards real-world travel planning for large language models. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. p. 14870–14902. Association for Compu- tational Linguistics (2025). https://doi.org/10.18653/v1/2025.emnlp-main.752, https://aclanthology.org/2025.emnlp-main.752/ 6. Fox, M., Gerevini, A., Long, D., Serina, I.: Plan stability: Replanning versus plan repair. In: Proceedings of the 16th International Conference on Automated Plan- ning and Scheduling (ICAPS). p. 212–221 (2006) 7. Gerevini, A., Serina, I.: Fast plan adaptation through planning graphs: Local and systematic search techniques. In: Proceedings of the Fifth International Conference on Artificial Intelligence Planning Systems (AIPS 2000). p. 112–121. AAAI Press (2000) 8. Gundawar, A., Valmeekam, K., Verma, M., Kambhampati, S.: Robust planning with compound LLM architectures: An LLM-Modulo approach. arXiv preprint arXiv:2411.14484 (2024), https://arxiv.org/abs/2411.14484 9. Hao, Y., Chen, Y., Zhang, Y., Fan, C.: Large language models can solve real-world planning rigorously with formal verification tools. In: Proceed- ings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Tech- nologies (Volume 1: Long Papers). p. 3434–3483. Association for Com- putational Linguistics (2025). https://doi.org/10.18653/v1/2025.naacl-long.176, https://aclanthology.org/2025.naacl-long.176/ 10. Hao, Y., Zhang, Y., Fan, C.: Planning anything with rigor: General- purpose zero-shot planning with LLM-based formalized program- ming. In: International Conference on Learning Representations (2025), https://openreview.net/forum?id=0K1OaL6XuK 14X. Yuan et al. 11. Huan, C., Meng, Z., Liu, Y., Yang, Z., Zhu, Y., Yun, Y., Li, S., Gu, R., Wu, X., Zhang, H., Hong, C., Ma, S., Chen, G., Tian, C.: Scaling graph chain-of- thought reasoning: A multi-agent framework with efficient LLM serving. arXiv preprint arXiv:2511.01633 (2025). https://doi.org/10.48550/arXiv.2511.01633, https://arxiv.org/abs/2511.01633 12. Huang, Z., Ma, Y., Zhang, H., Ma, H., Sun, Z.: iTIMO: An LLM-empowered synthesis dataset for travel itinerary modification. arXiv preprint arXiv:2601.10609 (2026). https://doi.org/10.48550/arXiv.2601.10609 13. Jiang, R., Wang, J., Zhao, G., Luo, C., Wang, K., Zhang, W.: Advanc- ing multimodal agent reasoning with long-term neuro-symbolic memory. arXiv preprint arXiv:2603.15280 (2026). https://doi.org/10.48550/arXiv.2603.15280, https://arxiv.org/abs/2603.15280 14. Ju, D., Jiang, S., Cohen, A., Foss, A., Mitts, S., Zharmagambetov, A., Amos, B., Li, X., Kao, J.T., Fazel-Zarandi, M., Tian, Y.: To the Globe (TTG): Towards language-driven guaranteed travel planning. In: Proceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. p. 240–249. Association for Compu- tational Linguistics (2024). https://doi.org/10.18653/v1/2024.emnlp-demo.25, https://aclanthology.org/2024.emnlp-demo.25/ 15. Karmakar, P., Chaudhuri, S., Mallick, S., Gupta, M., Jana, A., Ghosh, S.: TripTide: A benchmark for adaptive travel planning under disruptions. In: Findings of the As- sociation for Computational Linguistics: ACL 2026. p. 40269–40292. Association for Computational Linguistics (2026). https://doi.org/10.18653/v1/2026.findings- acl.2002 16. van der Krogt, R.P.J., de Weerdt, M.M.: Plan repair as an extension of planning. In: Proceedings of the Fifteenth International Conference on Automated Planning and Scheduling (ICAPS 2005). p. 161–170 (2005). https://doi.org/10.5555/3037062.3037083 17. Lai, L., Luo, C., Lou, Y., Ju, M., Yang, Z.: Graphy’our data: Towards end- to-end modeling, exploring and generating report from raw data. In: Compan- ion of the 2025 International Conference on Management of Data. p. 147–150. SIGMOD/PODS ’25, ACM (Jun 2025). https://doi.org/10.1145/3722212.3725106, https://doi.org/10.1145/3722212.3725106 18. Liu, J., Chen, L., Yang, Z., He, C., Ju, M., Han, B., Liu, R., Zhou, X.: HyperSU: Corpus-driven semantic-unit hypergraph for retrieval-augmented generation. arXiv preprint arXiv:2606.28351 (2026). https://doi.org/10.48550/arXiv.2606.28351, https://arxiv.org/abs/2606.28351 19. Liu, J., Chen, Z., Qiao, S., Ju, M., Zhang, D., Han, B., Yu, S., Shu, X., Wu, J., Wen, D., Cao, X., Liu, G., Yang, Z.: A2RAG: Adaptive agen- tic graph retrieval for cost-aware and reliable reasoning. In: 2026 IEEE 42nd International Conference on Data Engineering Workshops (ICDEW). p. 187–196. IEEE (May 2026). https://doi.org/10.1109/ICDEW71238.2026.00024, https://doi.org/10.1109/ICDEW71238.2026.00024 20. de Moura, L., Bjørner, N.: Z3: An efficient SMT solver. In: Tools and Algorithms for the Construction and Analysis of Systems. Lecture Notes in Computer Science, vol. 4963, p. 337–340. Springer (2008). https://doi.org/10.1007/978-3-540-78800- 3_24 21. Oh, J., Kim, E., Oh, A.: Flex-TravelPlanner: A benchmark for flexi- ble planning with language agents. arXiv preprint arXiv:2506.04649 (2025). https://doi.org/10.48550/arXiv.2506.04649 Travel Agents for Itinerary Revision15 22. Qi, J., Zhang, W., Ng, S.M., Xu, F., Chen, Y., Li, Y., King, I.: TREK: A travel rea- soning and evaluation kit for LLM agents in complex trip planning. arXiv preprint arXiv:2607.26977 (2026). https://doi.org/10.48550/arXiv.2607.26977 23. Saetti, A., Scala, E.: Optimally stable plan repair. The Knowledge Engineering Review 40, e7 (2025). https://doi.org/10.1017/S0269888925100076 24. Shao, Z., Wu, J., Chen, W., Wang, X.: Personal Travel Solver: A preference- driven LLM-solver system for travel planning. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). p. 27622–27642. Association for Computational Linguis- tics, Vienna, Austria (2025). https://doi.org/10.18653/v1/2025.acl-long.1339, https://aclanthology.org/2025.acl-long.1339/ 25. Shu, X., Ju, M., Chen, Z., Ding, Y., Zhang, W., Wen, D., Yang, Z.: ForexAgent: Identifying trading strategies in forex markets with large language models. In: 2026 IEEE International Conference on Big Data and Smart Computing (BigComp). p. 55–62. IEEE (Feb 2026). https://doi.org/10.1109/BigComp68355.2026.00019, https://doi.org/10.1109/BigComp68355.2026.00019 26. Tang, X., Chen, L., Yang, W., Yang, Z., Ju, M., Shu, X., Yang, Z., Tang, Y.: Tabular-textual question answering: From parallel program generation to large lan- guage models. World Wide Web 28(4), 42 (2025). https://doi.org/10.1007/s11280- 025-01351-1, https://doi.org/10.1007/s11280-025-01351-1 27. Verfaillie, G., Schiex, T.: Solution reuse in dynamic constraint satisfaction prob- lems. In: Proceedings of the Twelfth National Conference on Artificial Intelligence (AAAI). p. 307–312 (1994) 28. Wu, J., Tang, X., Yang, Z., Hao, K., Lai, L., Liu, Y.: An experimental evaluation of LLM on image classification. In: Chen, T., Cao, Y., Nguyen, Q.V.H., Nguyen, T.T. (eds.) Databases Theory and Applications. p. 506–518. Springer Nature Singapore, Singapore (2025). https://doi.org/10.1007/978-981-96-1242-0_37 29. Xie, J., Zhang, K., Chen, J., Zhu, T., Lou, R., Tian, Y., Xiao, Y., Su, Y.: Trav- elPlanner: A benchmark for real-world planning with language agents. In: Pro- ceedings of the 41st International Conference on Machine Learning. Proceed- ings of Machine Learning Research, vol. 235, p. 54590–54613. PMLR (2024), https://proceedings.mlr.press/v235/xie24j.html 30. Yang, W., Yang, Z., Chen, L., Yan, R., Yang, Z., Zhang, L., Tang, Y.: Parallel program generation for hybrid tabular-textual question answering. In: Web and Big Data: 8th International Joint Conference, APWeb-WAIM 2024, Proceedings, Part I. Lecture Notes in Computer Science, vol. 14961, p. 121–137. Springer Nature Singapore (2024). https://doi.org/10.1007/978-981-97-7232-2\_9 31. Yang, Z., Yang, W., Liu, G., Qin, L.: RAIDS: Rethinking data systems as responsible intelligent infrastructure. arXiv preprint arXiv:2606.21831 (2026). https://doi.org/10.48550/arXiv.2606.21831, https://arxiv.org/abs/2606.21831 32. Zaidins, P., Goldman, R.P., Kuter, U., Nau, D.S., Roberts, M.: HTN plan repair algorithms compared: Strengths and weaknesses of different methods. In: Proceedings of the International Confer- ence on Automated Planning and Scheduling. vol. 35, p. 297– 305. AAAI Press (2025). https://doi.org/10.1609/icaps.v35i1.36131, https://ojs.aaai.org/index.php/ICAPS/article/view/36131 33. Zhang, C., Goh, X.D., Li, D., Zhang, H., Liu, Y.: Planning with multi-constraints via collaborative language agents. In: Proceedings of the 31st International Con- ference on Computational Linguistics (COLING 2025). p. 10054–10082. Associ- ation for Computational Linguistics (2025), https://aclanthology.org/2025.coling- main.672/ 16X. Yuan et al. 34. Zhao, J., Feng, J., Li, Y.: AgentTravel: Knowledge-augmented LLM agent frame- work for urban travel planning. In: Proceedings of the 1st Workshop on Knowledge Graphs & Agentic Systems Interplay (NORA’25). CEUR Workshop Proceedings, vol. 4162, p. 55–67 (2025), https://ceur-ws.org/Vol-4162/paper5.pdf 35. Zhao, Z., Wang, S., Hou, Y., Xu, Y., Sheng, Y., Xie, X., Zhang, W., Shin, W.Y., Cao, X.: TRACE: Tourism recommendation with ac- countable citation evidence. arXiv preprint arXiv:2605.07677 (2026). https://doi.org/10.48550/arXiv.2605.07677, https://arxiv.org/abs/2605.07677

Instant research discovery