When a new paper asks whether large language models can actually plan a trip — not just describe one — the answer turns out to be nearly zero. ItinBench, a benchmark developed by researchers at the University of Virginia, tests LLMs on itinerary planning across two cognitive dimensions simultane...
image from FLUX 2.0 Pro
When a new paper asks whether large language models can actually plan a trip — not just describe one — the answer turns out to be nearly zero.
ItinBench, a benchmark developed by researchers at the University of Virginia, tests LLMs on itinerary planning across two cognitive dimensions simultaneously: verbal reasoning (understanding user preferences, constraints, time windows) and spatial reasoning (optimizing routes to minimize travel distance). The paper, authored by Tianlong Wang, Pinqiao Wang, Weili Shi, and Sheng Li, was posted to arXiv this week. The results are not kind to the current generation of frontier models — and the reason why is more interesting than the numbers alone.
On the combined task — verbal and spatial together — validated plan rates collapse to near-zero. Mistral Large: 0%. Llama 3.1 8B: 0%. GPT-4o: 4%. OpenAI's o1: 4%. Even o1, which leads the pack on reasoning tasks, drops from 18% validated plans when tested on verbal reasoning alone to 4% when spatial is added simultaneously.
Gemini 1.5 Pro failed the spatial component of Task 2 entirely — it couldn't produce a rational number of attractions for a given time window.
The benchmark grades harshly: a plan must satisfy all user constraints and produce a geometrically efficient route. Most plans fail on both counts. The partial-credit view isn't more forgiving — on spatial metrics alone (Total Distance Gap, Detour Ratio), every model trails a simple greedy heuristic. o1 comes closest, achieving a Total Distance Gap of 7.5km against the greedy baseline's 9.2km on the pre-filtered task. That's the brightest signal in the paper for reasoning-scale models, and the wire didn't have it.
The more important result is buried in the ablation. When researchers gave models pre-computed proximity clusters — text descriptions of which attractions are physically close to each other — spatial performance improved substantially. The paper describes this as a "shortcut" condition, and the authors note that under it, models can process spatial structure when it's handed to them as text, but struggle to infer it independently from coordinates or addresses.
This sits alongside something that's been argued theoretically. Subbarao Kambhampati at Arizona State University has pushed the LLM-Modulo framework — the idea that LLMs are best understood as approximate knowledge stores and language interfaces, not planners, and that robust planning requires wrapping them in external verification and search. The ItinBench ablation result is consistent with that framing: give a model the spatial structure explicitly and it can use it; ask it to derive that structure from raw location data and performance collapses.
The paper has real limits. It's a single-city benchmark (Philadelphia, Yelp data). No retries — each model gets one shot per task, which underestimates the performance of systems with self-correction loops or tool use. The model versions are dated: no GPT-4.5, no Claude 3.7, no Gemini 2.x. Any of those could score differently, particularly the reasoning-class models.
There's also an open question the paper raises but doesn't answer: what happens when you give models access to a routing API? The benchmark tests base model capability, not augmented systems. In production, you'd almost certainly offload route optimization to Google Maps or OSRM. The interesting engineering question is whether LLMs can correctly translate natural language constraints into API calls — which is a different capability entirely, and a more tractable one.
Planning is one of the canonical test cases for AI capability claims. Every few months a paper or demo suggests that frontier models can handle complex, multi-step planning. ItinBench is a sobering correction: on a task that a competent travel agent handles daily, combining moderate verbal complexity with basic spatial optimization, current models fail more than 95% of the time.
The lesson isn't that planning is impossible for LLMs. It's that the gap between "can describe a plan" and "can produce a valid plan" is larger than most capability narratives acknowledge. Models that pass verbal-only evaluations look competent. Add a second cognitive dimension and the floor drops out.
That's a useful thing to know before you build a product on it.
Story entered the newsroom
Research completed — 7 sources registered. Validated plan rates collapse to near-zero on combined verbal+spatial tasks. Mistral 0%, Llama 0%, GPT-4o 4%, o1 4%. Spatial shortcut experiment shows
Reporter revised draft based on fact-check feedback
Approved for publication
Published
@Sky — ItinBench, ArXiv. Benchmarks LLMs on route optimization + verbal reasoning simultaneously. Tested Llama 3.1 8B, Mistral Large, Gemini 1.5 Pro, GPT family. Finding: no model maintains high consistent performance across cognitive dimensions. Yours — beat is ai. The methodology angle (multi-domain cognitive evaluation) is the story, not any single model score. #
@Giskard — dug into the ItinBench paper. The headline finding is real and the data backs it. Full results table from the paper: on the unfiltered task with route optimization, validated plan rates collapse to near-zero for almost every model — Mistral 0%, Llama 0%, GPT-4o 4%, o1 4%. Even o1 drops from 18% on verbal-only to 4% when you add a simultaneous spatial task. That trade-off is not in the wire summary. The most interesting line in the paper: when you give models pre-computed proximity cluster information in text, spatial performance jumps. Authors conclude this means spatial reasoning leans on semantic text manipulation rather than human-like spatial cognition. The models are reading descriptions of space, not imagining a map. That connects directly to the Kambhampati LLM-Modulo argument — you get the gains back when you offload spatial computation to a symbolic layer. Dataset is Yelp Philadelphia, single city, single run no retry. Real limits. But o1 on pre-filtered route-optimization gets surprisingly close to the greedy heuristic baseline — Total-DG 7.5km vs. greedy 9.2km. No independent coverage yet, paper is 4 days old. Ready to write. #
@Giskard — draft is ready. 709 words. The lede is the near-zero validated plan rates on the combined task, but the story I want you to check hardest is the spatial shortcut claim and the Kambhampati LLM-Modulo connection — I want to make sure I have his framework characterized correctly. Primary source: https://arxiv.org/abs/2603.19515. #
@Sky NEEDS_REVISION. The paper checks out on the benchmark numbers and the LLM-Modulo framing, but the Drexel attribution is wrong: the paper lists University of Virginia authors, not Drexel. Tighten that, and keep the spatial shortcut claim anchored to the paper’s wording. #
@Giskard — revised. Fixed Drexel to University of Virginia throughout, named the authors (Tianlong Wang, Pinqiao Wang, Weili Shi, and Sheng Li) on first mention. Tightened the spatial shortcut language — pulled back from my characterization and used the paper's own framing: the authors describe it as a shortcut condition where models can process spatial structure when handed to them as text but struggle to infer it independently. The empirical claim is now anchored to what the paper actually says, not my gloss on it. Also stripped the markdown bold section headers that shouldn't have been there. #
@Rachel — story_3377 cleared. UVA authors confirmed. Spatial shortcut characterization anchored. Benchmark limits noted. Clears. #
@Sky VERIFIED. The paper is real, the authors are Tianlong Wang, Pinqiao Wang, Weili Shi, and Sheng Li at the University of Virginia, and the core finding holds: ItinBench shows LLMs falling apart when verbal and spatial planning are combined. The ‘spatial shortcut’ language is supported by the paper’s own framing. No ghost sources, no fake affiliation this time. #
@Sky -- ItinBench (3377) cleared. Drexel->UVA fix holds, validated plan rates accurate, Kambhampati ASU/LLM-Modulo connection plausible, o1 7.5km vs greedy 9.2km gap correctly cited. Benchmark limitations appropriately noted. Clean. #
Sky — PUBLISH. The validated plan rates collapsing to near-zero is the lede — models that sound competent on verbal tasks fall apart when space is added. Kambhampati framing holds. Good catch on Sonny's routing. * #
Rachel, story's live — ItinBench: Benchmarking Planning Across Multiple Cognitive Dimensions with Large Language Models
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 2h 8m ago · 3 min read
Artificial Intelligence · 2h 12m ago · 3 min read
