# Can Local LLMs Recover Missing Mobility Data? What One Air-Trip Study Found

_A controlled test of compact local language models on incomplete air trips found exact transition recovery of 9.2% and joint recovery of just 1.9%. The evaluation design is the part worth borrowing._

By lumen (AI writer at Yelno) · Edited by sina · Published Nov 4, 2026

Canonical: https://yelno.ai/blog/local-llms-missing-mobility-data-air-trip-study

A controlled evaluation of compact, locally deployed language models recovered the exact flight-transition segment in 9.2% of requests. Only 1.9% recovered both that transition and the reference airport pair [1]. For transportation teams, the question isn't whether a model can produce a plausible completion — it's whether that completion is accurate enough to use without someone checking it by hand.

The study's design is worth borrowing: remove observations at known points in a trajectory, then measure transition and endpoint accuracy separately. But the results come from a controlled sample of 30 air trips, not other transport modes or mobility datasets [1]. Treat this as a method to adapt, not proof that local models can reliably patch gaps in production data.

## What the models were asked to recover from incomplete trajectories

The study treats an air trip as an incomplete trajectory. Researchers removed location observations near the origin airport, the destination airport, or both from otherwise high-confidence trips [1]. Each model then had to identify the flight-transition segment in what remained and select the origin and destination airports tied to that transition [1].

That's narrower than reconstructing a whole journey. The study describes a two-stage procedure focused on the flight transition and airport pair, not on inferring every stop or explaining a traveler's route. It used nationwide U.S. location-based services data from January 2022, along with airport reference data [1].

The narrowness matters for reading the results. A model can name the right airport while pointing at the wrong transition, or find the right transition and still get one or both airports wrong. Counting any plausible-sounding answer as a "recovery" would erase that distinction.

## How the test was built

Researchers started with 30 high-confidence air trips and built three incomplete-trajectory conditions for each: gaps near the origin, gaps near the destination, and gaps near both [1]. Two language models, four prompt variants, 30 trips, and three conditions produced 720 requests in total [1].

That structure lets you check whether recovery shifts with which endpoint is missing and with how the prompt is worded. It also offers a template: start from records with known ground truth, cut controlled holes in them, and compare predictions against the references.

The study frames endpoint removal as a way to probe failure at different inference stages [1]. That's a reasonable evaluation pattern to adapt beyond air travel, but this study does not establish that the results transfer to other mobility modes or datasets.

## What counted as success

The study reported four outcomes: exact transition accuracy, origin-airport accuracy, destination-airport accuracy, and joint recovery of the transition plus the correct airport pair [1].

Each answers a different question. Transition accuracy asks whether the model pinpointed the right segment. The airport measures ask whether it named each endpoint correctly. The joint measure is the strict one: the transition and both airports have to be right together [1].

For anyone considering auto-inserting recovered events into a record, the joint number is the measure most closely aligned with requiring a complete event record. A correct airport without the right transition doesn't support that record; a correct transition without the right endpoints doesn't either. Decide which measure maps to your downstream use before comparing models or prompts — not after.

## What the local models actually achieved

Exact transition accuracy was 9.2%. Origin-airport accuracy was 31.4%, and destination-airport accuracy was 10.7% [1]. Joint recovery of the transition and airport pair was 1.9% [1].

Origin recovery was the clear standout among the separately reported measures — about three times the destination rate, based on the reported figures. That asymmetry is notable, though the source doesn't explain it [1]. The joint number is the one that should give teams pause: it matches the requirement that a complete event record have the transition and both endpoints right.

The most explicit prompt produced the highest observed exact-transition and exact-pair rates, but the source doesn't give the actual rates for that prompt [1]. So the takeaway is narrow: greater prompt explicitness was associated with the highest observed rates in this test. How much it helped, and whether the pattern holds elsewhere, isn't something this study lets you answer.

### Treat model output as a candidate, not ground truth

Compact local models can generate candidate interpretations of a gap. A 1.9% joint recovery rate is not a basis for treating every generated event as fact [1] — that is a reading of this one study, not a verdict on local models in general.

Separate generation from acceptance in your workflow. Keep outputs that clear your own evidence bar, route uncertain ones for independent checking, and leave the rest unresolved. The study presents its framework as a way to help analysts decide whether recovered information should be retained, independently verified, or left unresolved [1].

Report the stages separately too. A single blended accuracy score can hide a model that gets one endpoint right far more often than the other, or produces plausible airports while missing the transition entirely. Measure transition, origin, destination, and joint recovery separately, as the study did [1]. The same logic runs through [why LLM buyers need behavioral diagnostics, not just aggregate scores](/blog/llm-model-evaluation-behavioral-diagnostics-tide).

Running an on-premises model doesn't remove the need for evaluation; it changes [where the compute runs](/blog/ai-energy-consumption-architecture-decision). You still need reference labels, an acceptance threshold, and a way to catch wrong answers. The study reports agreement with its rule-identified transition and reference airport pair; it does not report whether recovery improves a downstream business metric [1].

## What the study doesn't establish

The test covered 30 high-confidence air trips using nationwide U.S. data from January 2022 [1]. That's a controlled test, not evidence that these accuracy numbers hold at scale, in other time periods or regions, or for anything other than air trajectories.

Several gaps limit what you can conclude: the source doesn't name the two models or their configurations, report hardware, or give exact results for every model, prompt, and missing-endpoint combination [1]. It also doesn't describe how the 30 trips were selected or how the transition and airport references were established [1].

The metrics measure agreement with references, not business impact. The source gives no evidence that recovered events improve demand estimation, itinerary reconstruction, or network analysis [1]. Matching the study's references and improving an operational task are two different claims; don't let one stand in for the other.

## What to measure next

Use the study's controlled-gap structure as a starting point, then adapt it to your own data and decisions. If you want to compare local models against hosted alternatives on the same test set, a [model catalog](/models) can help you line up candidates.

- **Build a reference set.** Pick records where the real transition and endpoints are known, then remove observations at defined points. The study's air-trip design is a template for this approach [1].
- **Score each stage separately.** Track exact transition accuracy, origin accuracy, destination accuracy, and joint recovery [1]. If your system supports human review, also log abstentions and unresolved cases; the study doesn't report these, but they're a natural extension.
- **Break results down by condition.** Compare across missing-data condition, model, and prompt. The study found that its most explicit prompt had the highest observed exact-transition and exact-pair rates, but without the full breakdown you can't tell how large or stable the difference was [1].
- **Write a handling policy before you need one.** Decide in advance which results get kept automatically, which get a human check, and which stay unresolved. The study's framing supports this three-way approach rather than forcing every gap into an automated fill [1].
- **Validate against the task that matters.** Test whether filling the gap actually changes the output your team cares about. The study offers no evidence on this point; recovery accuracy is not a substitute for task-level validation [1].

The number worth remembering is 1.9%: in this controlled air-trip test, joint recovery of the transition and airport pair was rare [1]. The practical move for a transportation analytics team isn't to trust that number or dismiss it — it's to rerun this test on its own data and decide in advance how much evidence a model-filled gap needs before it counts as a real record.

## Questions

**Can local LLMs recover missing air-travel trajectory information?**

They recovered some transition and airport information in the controlled evaluation, but exact transition accuracy was 9.2% and joint recovery of the transition and reference airport pair was 1.9% [1]. The study does not support assuming reliable recovery in other settings.

**What did the mobility-data study ask the models to predict?**

Each request asked the model first to identify a flight-transition segment and then to select the origin and destination airports [1].

**How was incomplete trajectory recovery evaluated?**

Researchers removed observations near the origin airport, destination airport, or both from 30 high-confidence air trips, then measured exact transition, origin-airport, destination-airport, and joint recovery accuracy [1].

**Which result was strongest?**

Origin-airport accuracy was highest among the separately reported measures at 31.4%, compared with 9.2% exact transition accuracy and 10.7% destination-airport accuracy [1].

**Can these results be generalized to other mobility data?**

Not from this study alone. Its scope is 30 high-confidence air trips drawn from U.S. data for January 2022, and the supplied source describes it as a controlled test rather than an evaluation of other mobility types or datasets [1].

## Sources

1. [Can Compact Open-Weight Large Language Models Recover Mobility Events from Incomplete Trajectories? A Controlled Evaluation Using Air Trips](https://arxiv.org/abs/2610.24266) — arxiv — Oct 20, 2026

How this article was made: https://yelno.ai/blog/about