# AMap Let an LLM Predict Your Next Stop > Source: > Published: 2026-09-25 16:00:00+00:00 8 min read Open AMap and its homepage may try to predict the place you are heading before you search for it. The navigation app tested whether a small language model could make that guess more accurately than the ranking system already serving the card. The question was practical. AMap already had a production recommender using profiles, recent behaviour, location, time, trending topics and seasonal patterns. Replacing it meant beating a mature baseline without making the app noticeably slower. In a [company-authored paper submitted in July 2026](https://arxiv.org/abs/2607.26073), researchers from [AMap](https://www.amap.com/), part of Alibaba Group, describe the resulting AMap A/B test. Their system, called Gwhere, generated the identifier of a likely next destination instead of recalling and ranking a list of candidates through the usual multi-stage pipeline. Gwhere’s prediction depends on behavioural and location data most companies don’t have at this scale. We covered that constraint in more depth in [our analysis of the data gap keeping AI shopping agents from truly knowing users](https://industrycontents.com/ai-shopping-agents-shopper-memory/). ## The homepage was already personalised AMap is a mapping and navigation platform. The experiment focused on a homepage card called “Guess Where You Go,” which recommends a point of interest, or POI, based on where the user may travel next. The existing system was not a weak control. According to the paper, AMap’s production baseline combined several retrieval channels with coarse ranking, fine ranking and a final cognitive re-ranking stage. It received the same request context as Gwhere and also used exposure history, click history, trending topics and seasonal patterns. That made the growth constraint more specific than improving generic recommendation accuracy. The team needed to find more destinations people would click, while keeping the recommendation relevant enough to avoid negative feedback and fast enough for a homepage used at scale. The paper does not document a formal hypothesis statement, but the test design implies one. Generating a destination from a richer representation of places and mobility patterns could outperform a cascade ranker on the same homepage position and context. That is an inference from the comparison, not a quotation from the researchers. ## Gwhere turned places into short codes A language model cannot work efficiently with tens of millions of raw place identifiers. Gwhere first converted each POI into a short semantic code. The code combined four signals. These were the place description, images, spatial relationships and patterns in how people interacted with it. The model then learned from sequences of previous check-ins, along with time, current location, weather and user-profile information. In production, the system turned that context into a prompt, generated the semantic code for one likely destination and mapped the code back to a real place. The final training stage addressed a familiar problem in recommendation data. An exposed place that received no click is not always disliked. The person may not have noticed it, the request may have arrived at the wrong moment or latency may have affected the interaction. Gwhere’s exposure-aware preference method treated clicks as positive signals while limiting how aggressively unclicked exposures counted as negatives. The company’s offline dataset covered 10 million sampled users, 83.1 million places and 380 million interactions. The final model trained on 10 billion tokens using 200 Nvidia H20 GPUs. Those figures show why the experiment is hard to reproduce outside a platform with a large mobility graph and substantial computing capacity. ## The smallest model won the deployment decision Offline results improved as the researchers moved from a 0.5 billion parameter model to a 7 billion parameter version. The largest version achieved the best prediction accuracy, but it also created a serving problem. Under a load test of 50 queries per second on two Nvidia H20 GPUs, the existing cascade system ran at roughly 20 milliseconds of p99 latency. The 0.5 billion parameter Gwhere model came in at about 30 milliseconds. Larger versions were more accurate offline but too costly for the production constraint described in the paper. AMap therefore chose the smallest model for the live test. This is the decision at the centre of the experiment: the team accepted less offline accuracy in exchange for a system that could run on the homepage at manageable latency. ## One-week and one-month tests used the live system The AMap A/B test compared Gwhere-0.5B with the deployed cascade-ranking system. The paper says both systems used the same traffic allocation, page position, request context and metric definitions. It also says the production deployment served millions of active users. The researchers ran a one-week test for short-term response and a one-month test for longer-term engagement. In the one-week comparison, the company reported a 5.83% relative increase in page click-through rate and a 6.20% increase in user click-through rate over the production baseline. Over the longer test, the paper reports a 9.67% increase in active-user rate, a 2.60% increase in retention rate and an 11.11% reduction in negative-feedback rate. Scroll rate increased by 2.18%. These are relative changes reported by AMap’s own researchers. The paper describes the changes as statistically significant but does not provide absolute baseline rates, confidence intervals or p-values. It also omits the exact traffic split, market coverage and number of users assigned to each condition. Those gaps prevent an independent check of the effect sizes and leave the commercial value of an additional click unknown. ## The preference layer had its own comparison The researchers also reported a version of Gwhere without the exposure-aware preference stage. That version raised page click-through rate by 4.23% and user click-through rate by 4.59% over the production baseline. Its negative-feedback rate fell by 5.85%. With the preference stage included, the reported click lifts reached 5.83% and 6.20%, while negative feedback fell by 11.11%. The comparison supports the team’s claim that learning carefully from exposed-but-unclicked places improved more than clicks alone. The paper does not disclose whether the two Gwhere versions received equal traffic or provide uncertainty around the difference between them, so that mechanism claim remains less secure than the main production comparison. ## A click was only one part of the test The useful feature of this experiment is the metric sequence. AMap did not stop after showing that more people clicked the card. The month-long comparison tracked whether people stayed active, returned and sent fewer negative signals. That matters when a recommendation system can manufacture clicks by becoming more surprising or aggressive. A short-term CTR gain may look like growth while relevance deteriorates. The AMap experiment used negative feedback and activity as guardrails against that failure, although the paper does not define the exact actions counted as negative feedback. The experiment also separated model quality from deployment feasibility. The best offline model did not ship. The team tested the version that could operate within its latency budget, then evaluated it against the full production system rather than an academic baseline. ## What a comparable business can test A company does not need AMap’s scale to borrow the test logic. The transferable move is to define the strongest current decision system as the counterfactual, hold placement and context constant, and follow the immediate action with a longer behavioural measure. For a marketplace, that might mean testing a generative recommendation against the live ranked feed while measuring clicks, saves, returns and hides. For a SaaS product, it could mean predicting the next workflow against the current rules engine while tracking activation and repeated use. The hard part is not simply adding an LLM recommender system. Gwhere depended on years of behavioural history, a large catalogue, spatial data, images, production exposure logs and a serving stack designed for high concurrency. A smaller business may learn more from the experimental structure than from copying the architecture. The remaining question is whether the predicted destination led to a trip, transaction or other valuable outcome. The AMap A/B test establishes a stronger response to the homepage recommendation and reports better engagement guardrails. It does not connect those changes to navigation starts, completed visits or revenue. That missing step defines the next useful experiment. Two related experiments show different sides of the same ranking problem. [Meta had to fix its own recommender after discovering it was crediting clicks for purchases they did not cause](https://industrycontents.com/meta-post-click-purchase-attribution/), and [Wantedly tested whether boosting underexposed candidates in its rankings could reduce churn](https://industrycontents.com/wantedly-retention-experiment-candidate-ranking/) without a clear win. Both cases turn on the same question Gwhere leaves open: whether a ranking change that looks good on engagement actually produces the outcome the business needs. ## Sources - [Guess Where You Go: Generative Next Point-of-Interest Recommendation in Amap](https://arxiv.org/abs/2607.26073) , Penglong Zhai and co-authors, July 2026 - [Alibaba’s public Gwhere repository](https://github.com/alibaba/SimCIT) - [AMap](https://www.amap.com/) - [Penglong Zhai on LinkedIn](https://cn.linkedin.com/in/penglong-zhai-25037a186)