A large airport departures board listing flights, times and gates at Geneva Airport
Departure board at Geneva Airport. Photo: Tiia Monto / Wikimedia Commons, CC BY-SA 3.0

vizmaya desk · 3 Nov 2026

AI trip planning: 10% right alone, 100% valid with rules

10%

That is how often a standard AI model wrote a valid trip plan on its own.

A new paper from the University of Connecticut shows a fix. It works by changing what the model is allowed to write next.

Why does a smart model fail at something this basic?

Source: Su, Xiang and Zhang, FactorSpec, arXiv 2611.02297v1, pages 1 and 2. Model: Qwen3-8B. 999 trip problems of 3 to 7 cities.

An AI writes one small piece at a time

The pieces are called tokens. The model picks one, then the next, then the next.

A format checker watches each pick. It makes sure brackets close and quotes match.

A neat shape sounds like a good plan. Is it?

Source: Fig 1 (simplified), page 1. Format checkers: page 2.

The shape is fine. The plan is wrong.

A good plan has rules. Every hop needs a direct flight. No city twice. Each day number is 1 plus the stays before it.

A format checker cannot see those links. Alone, the model got 10 plans in 100 right. Retries that named the broken rule reached 14 (4 points more, computed).

Would more thinking time fix it?

Source: pages 1 and 2; rules in Fig 4, page 5. Qwen3-8B, 999 problems.

More thinking helped a little

The authors let the model think for up to 8,192 tokens before it answered.

Trip plans: 19% valid. Meeting schedules: 31%. Calendar slots: 95%. Trip plans stayed the hardest.

What if the rules did the sums for the model?

Source: page 9. 64 problems per task, 8 requests at once.

Sometimes the rules leave only one choice

A toy trip from the paper: flights only go Rome to Oslo, then Oslo to Lima. You start in Rome. After Rome, only Oslo is possible, so nothing is left to guess.

Rome, Italy stay 2 days
The Colosseum in Rome seen from outside in daylight
Colosseum. Photo: Jean-Pol Grandmont / Wikimedia Commons, CC BY-SA 3.0
Oslo, Norway stay 3 days
The white angled roof of the Oslo Opera House beside the water
Oslo Opera House. Photo: Helge Høifødt / Wikimedia Commons, CC BY-SA 3.0
Lima, Peru stay 1 day
Plaza Mayor, the main square of Lima, with a fountain and old buildings around it
Plaza Mayor. Photo: Jordan Klein / Wikimedia Commons, CC BY 2.0

Can the rules also work out the day numbers?

Source: example tour, Fig 4, page 5, and the worked step on page 4.

Then the rules do the sums

Rome starts on day 1 and lasts 2 days. So Oslo starts on day 1 + 2 = 3.

Oslo lasts 3 days. So Lima starts on day 3 + 3 = 6. Every piece of the answer is now fixed.

If it is all fixed, why write it one piece at a time?

Source: stays and day rule in Fig 4, page 5; the Oslo day 3 and Lima day 6 step on page 4.

Fixed pieces can be written all at once

Servers already use a trick. A helper guesses several pieces ahead and the big model checks them in one go. It is called speculative decoding.

A guess can be wrong. A piece the rules fix cannot be. In the paper's example, 24 fixed pieces took 2 checking steps.

How much of a real form can rules fix?

Source: Fig 1, page 1; page 4. The 24 steps at one piece per step is computed (24 pieces, 1 per step).

For trip plans, nearly all of it

The authors counted, for six tasks, the pieces where the rules allow exactly one choice.

Trip plans: 99.98%. Calendar slots: 96.1%. Forms for tool calls: 70.2% to 85.8%.

Does a fixed piece turn into speed?

Source: Table 1, page 3 (200 answer-key records per task). Percent = share x 100, computed.

The result

Up to 10.5 pieces per check

On trip plans, the big model now accepts 10.5 pieces per check. The best rival guesser gets 5.9, and a learned guesser 3.9. That is 1.78 times the rival (computed).

Fast is easy. Are the plans right?

Source: Table 2, page 8. 8 requests at once. 10.5 / 5.9 = 1.78, computed.

Every plan was valid

The empty bar from earlier fills to 100%: every record the system wrote followed the rules.

Valid is not the same as the best answer. On trip plans, 98 to 99% matched the answer key. The best rival reached 13.6%.

What happens when many people ask at once?

Source: page 9; Fig 12, page 17. Full sets, 64 requests at once. Rival bars are the best of the baselines.

Rows of server racks with cables and blinking lights in a data center
Server racks. Photo: Mutante / Wikimedia Commons, CC BY-SA 3.0

Under load, the gap grows

Trip requests arrive at 32 per second. Valid plans finished in 10 seconds:

32
valid plans per second, FactorSpec
1.2
valid plans per second, best rival

That is 26.7 times as many (computed). The slowest 1 in 100 requests took 3.8 s, against 26 s.

So what should you do with this?

Source: pages 9 and 10. 32 / 1.2 = 26.7, computed. Rival: SGLang n-gram drafter.

Our reading, not the paper's claim

Ask what your AI tool actually checks

Format or logic? "Structured output" often checks only the shape. The links between fields are left to the model.

Do not lean on retries. Naming the broken rule lifted valid plans from 10% to 14%.

If you know the rules, write them down. Let a checker enforce them while the model writes.

Where does this break?

Source: interpretation by vizmaya desk. Numbers: pages 1 and 2.

The lead fades when the text is free

Add a free-text note to every stay and the rules fix less. With 128-token notes, the gain over a plain format checker shrinks from 1.77x to 1.13x.

The rival guessers fall to 0.72 and 0.73. That is slower than no guessing at all.

What else does it cost?

Source: Fig 7 and text, page 8. Trip plans with notes, 64 requests at once.

These are the authors' own numbers

On meeting schedules, a solver matched the answer key more often: 46 to 56% vs 14 to 16%.

The speed needs about 8 spare CPU cores per GPU (7.4 to 8.3 of 32).

One GPU type, up to 128 requests at once. A preprint, 2 days old. I found no outside check.

Want to check the work yourself?

Source: pages 9, 8 and 10; research file gaps.

A graphics processor chip package with memory stacks mounted beside it on a silicon interposer
An example GPU package (AMD Fiji), not the card the authors tested. Photo: C. Spille / pcgameshardware.de / Wikimedia Commons, CC BY-SA 4.0

Credit and sources

  • Paper. Jianchang Su, Dawei Xiang, Wei Zhang, University of Connecticut. "FactorSpec: Serving Exact Relational Constraints as Speculative Drafts." arXiv 2611.02297v1, 1 Nov 2026. arxiv.org/abs/2611.02297. A preprint, not peer reviewed. No DOI printed.
  • Method. Every number matches a page of the PDF (pages 1, 3, 8, 9, 10, 17). Computed here: 14 - 10 = 4 points; 10.5 / 5.9 = 1.78; 32 / 1.2 = 26.7; 24 pieces at 1 per step = 24 steps; shares x 100 for percents.
  • Not found. Independent replication. Citation counts. A code link in the PDF text.
  • Note. The authors say they used AI tools to edit the paper's writing and reviewed every edit (page 14).
  • More data stories: vizmaya.fyi/stories. Byline: vizmaya desk, 3 Nov 2026.

Photo credits

Flags: flag-icons, MIT licence. Icons: Phosphor, MIT licence. The cities and the trip are a toy example from the paper, not real travel advice.