{"id":"5f148e87-52ed-416f-a385-b2e14e751e1d","arxiv_id":"2510.14819","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CORE embeds trajectories by letting an LLM describe POIs around roads, encoding each road-segment choice with a mixture-of-experts, and outperforming 12 prior TRL methods on six urban tasks.","lead":"GPS trajectories are turned into vectors by treating each trip as a chain of route choices shaped by nearby shops, parks, and road types. The authors show this \"route choice\" view improves six urban prediction/retrieval tasks over 12 prior trajectory-embedding methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CORE's headline 'consistently outperforms' is contradicted by its own Table VII: Chengdu path-ranking MAE is worse than TRACK, and no error bars support the positive claims.","rationale":"The reader's weakest_assumption focused on POI distributions as proxies for environmental semantics—a reasonable architectural limitation but not the most immediate threat to the central claim. My concern is more direct: the paper's own reported results contradict its headline assertion. The claim 'consistently outperforms' is the key takeaway, and it is falsified by a single negative cell in Table VII (Chengdu path-ranking MAE: CORE 0.0934 vs TRACK 0.0927). This is an internal inconsistency, not a matter of scientific consensus or external validity, and it is precisely the kind of load-bearing issue a stress test should surface. The absence of error bars compounds the problem: the paper says results are averaged over five trials (Section V-A3), but no variance or significance information is given, so even the positive claims cannot be properly evaluated. I credit the paper for thorough ablations, multiple datasets, and honest discussion of POI sensitivity in Fig. 12, and I am not suggesting misconduct. But the central claim, as written, is too strong. The appropriate disposition is CONDITIONAL: accept the novel architecture and most of the empirical work, but require the authors to (a) reconcile the abstract/main-text discrepancies, (b) provide error bars and significance tests, and (c) either weaken 'consistently outperforms' to a more accurate statement or demonstrate that the Chengdu PR-MAE deficit is within noise. This aligns with the reader's CONDITIONAL verdict but for a more central reason than the POI-proxy assumption.","tokens_in":34268,"tokens_out":3580,"duration_ms":32767,"concrete_test":"Reproduce the path-ranking evaluation on Chengdu with at least five seeds, computing paired differences between CORE and TRACK on MAE; run a paired t-test or bootstrap 95% CI. Also scan every cell in Tables III–VIII and count all dataset/metric combinations where CORE fails to improve over the best baseline. If the negative Chengdu result persists or is not statistically negligible, the headline should be revised from 'consistently outperforms' to 'outperforms on most tasks,' with confidence intervals reported throughout. Optionally, rerun TTE with all segment-time embeddings in Eq. 21 zeroed except the first timestamp to test the label-leakage ambiguity raised by the reader.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim—repeated in the abstract, introduction, and conclusion—is that CORE 'consistently outperforms 12 state-of-the-art TRL methods' with an average improvement of 9.79% over the best baseline. This claim requires CORE to beat every baseline on every dataset and metric. The paper's own Table VII (Path Ranking) provides a direct counterexample: on Chengdu, CORE's MAE is 0.0934, while TRACK achieves 0.0927; the improvement column reports -0.76%, explicitly acknowledging a loss. Thus 'consistently' is false in the reported data. This is not a trivial wording issue; the entire contribution is framed around this consistency. Moreover, Section V-A3 states that all results are averaged over five independent trials, yet no error bars, standard deviations, or significance tests are reported anywhere. Without these, we cannot determine whether the many positive differences—several of which are small, e.g., Chengdu destination prediction Acc@1 0.4296 vs GREEN 0.4060, or Porto TTE MAE 1.3519 vs TRACK 1.4001—are statistically reliable or within noise. The abstract's inconsistent numbers (15 baselines/9.20% vs 12 baselines/9.79% in the main text) further weaken confidence in the precision of the headline. For the central claim to hold, CORE must be meaningfully better than all baselines on all tasks; the Chengdu path-ranking result directly contradicts that condition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CORE, a trajectory representation learning framework that treats a GPS/map-matched trajectory as a sequence of context-aware route choices rather than a passive spatiotemporal sequence. CORE first constructs a context-enriched road network by using an LLM to summarize POI distributions at fine (road-segment) and coarse (grid/functional-hotspot) granularities, then encodes route-choice behavior with an MoE-based Route Choice Encoder that uses selected and unselected adjacent segments, historical transition likelihoods, and directional deviation. A Transformer encoder and contrastive pretraining produce the final trajectory embedding. The authors evaluate on four real-world datasets across six downstream tasks (road label prediction, travel time estimation, similar trajectory retrieval, destination prediction, path ranking, and trajectory generation), reporting large gains on several tasks and claiming that CORE \"consistently outperforms\" 12 state-of-the-art baselines with an average improvement of 9.79%.","tokens_in":34623,"tokens_out":5357,"duration_ms":49957,"significance":"If the reported gains are reliable, the core idea is a useful and timely contribution: injecting behavioral route-choice semantics and LLM-distilled POI context into trajectory embeddings is a natural inductive bias that appears to help most downstream tasks, especially road-label prediction and similar-trajectory retrieval. The paper is also commendable for releasing code, for including detailed ablations and sensitivity analyses, and for transparently reporting the API cost and wall-clock time of the LLM preprocessing stage. However, the central \"consistently outperforms\" claim is not supported by the paper's own numbers, and the absence of error bars or significance tests makes it difficult to judge the robustness of the many positive differences, several of which are small. The contribution is substantive enough to warrant a major revision rather than rejection, but the overclaim and missing statistical grounding must be fixed.","major_comments":[{"comment":"CORE's MAE is 0.0934 while TRACK achieves 0.0927, and the reported Improvement is -0.76%. This directly contradicts the abstract, introduction, and conclusion claim that CORE \"consistently outperforms\" all 12 state-of-the-art baselines. Even if every other table entry favors CORE, one counterexample falsifies the strict claim. The claim should be weakened to \"outperforms the baselines on most datasets and tasks\" or otherwise qualified, with a concrete explanation for the Chengdu path-ranking degradation. This is not a wording issue; the consistency claim is the paper's headline conclusion.","section":"Table VII (Path Ranking, Chengdu row)"},{"comment":"The paper states that all results are averaged over five independent trials, but no standard deviations, confidence intervals, or significance tests are reported anywhere. Several of the favorable differences are small (e.g., Table VI Chengdu Acc@1: 0.4296 vs. 0.4150 for JGRM; Table IV Porto TTE MAE: 1.3519 vs. 1.4001 for TRACK; Table XI Chengdu MAE: 1.3681 vs. 1.3675 for DeepSeek-V3.2-Exp). Without variability information, the reader cannot determine whether these differences are reliable or within trial-to-trial noise. Please add error bars or standard deviations for at least the headline tables and report paired significance tests for the main claim on each dataset/task.","section":"Section V-A3 / Section V-B"},{"comment":"The abstract reports \"15 state-of-the-art TRL methods\" and a \"9.20%\" average improvement, while the title page, introduction, and conclusion report \"12 state-of-the-art TRL methods\" and a \"9.79%\" average improvement. This numerical inconsistency in the central claim must be corrected, and the aggregation rule (how the average improvement is computed across tasks, metrics, and datasets) should be stated explicitly.","section":"Abstract vs. Introduction"},{"comment":"The historical transition likelihood P(r_c|r_i) is computed from counts N(r_i->r_c), but the paper does not state whether these counts are restricted to the training split or are computed over the entire trajectory corpus before the chronological 7:1:2 split. If the latter, transition probabilities would include test-set route frequencies, leaking future information into CORE's inputs and giving it an unfair advantage over baselines that do not use this signal. Please clarify that the counts are computed from the training split only (or from an independent historical period), and confirm that the ablation \"w/o historical transition\" remains valid under that setting.","section":"Section IV-B, Eq. (13)"}],"minor_comments":[{"comment":"The task description is vague: path labels are said to be the IoU of each candidate with the \"optimal\" path, but the ground-truth path is the observed trajectory. It would help to define how the candidate set is generated and how the ground-truth path is chosen, so that the reader can interpret what the ranking score is measuring.","section":"Section V.B.5 (Path Ranking)"},{"comment":"The efficiency and data-efficiency experiments compare CORE with only a subset of baselines (JGRM, TRACK, GREEN, HOSER, STEGA). Please justify the selection of these baselines for these analyses, since the full comparison set is larger.","section":"Table IX / Figure 6"},{"comment":"The labels in Figure 3 (visit count distributions) contain garbled glyph sequences in the version I reviewed, making the figure difficult to interpret. Please ensure the final PDF renders the axis labels and legend correctly.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing before you read it. First, the core idea is good: they treat trajectories as sequences of route choices, inject LLM-distilled POI semantics at road-segment and grid level, and encode each step as the selected next segment versus unselected alternatives with an MoE. That is a real departure from the START/GREEN/TRACK line, and the ablations show each piece earns its keep. Second, the paper's own tables undercut the abstract's 'consistently outperforms' claim. On Chengdu path-ranking, their MAE is 0.0934 versus TRACK's 0.0927, an explicit -0.76% in their own improvement column. Tau and rho favor them, so it is not a wholesale loss, but 'consistently' is not true as written. The abstract also says 15 baselines and 9.20% improvement while the main text says 12 and 9.79%; that inconsistency should have been caught before submission.\n\nThe strongest parts are the Environment Perception Module and the Route Choice Encoder. The LLM descriptions are concrete and the prompt/example in the appendix is convincing: the LLM produces genuinely abstract functional semantics ('core commercial district,' 'transit-oriented clustering') that raw text embedding would not give. The coarse-grained hotspot modeling with CNN spillover is a sensible complement. The MoE visualization shows meaningful expert specialization by region and road type. The efficiency analysis is honest: the LLM preprocessing costs about $2 for Beijing, which makes the approach practically feasible.\n\nThe soft spots are real but addressable. No error bars anywhere despite 'averaged over five independent trials'; several headline margins on Chengdu and Porto are small (e.g., 0.4296 vs 0.4060 Acc@1, 1.3519 vs 1.4001 MAE), so without variance or significance tests I cannot tell if those are noise. The TTE setup needs clarification: the architecture adds per-segment time-of-day embeddings in Eq. 21, yet the TTE section says 'We input only the departure time to avoid potential leakage.' Either the pretraining features are dropped during fine-tuning, which would break representation consistency, or the text is inaccurate. The latter is likely, but the ambiguity should not survive peer review. The POI dependency is acknowledged in Fig. 12 and is a fair limitation, not a defect.\n\nWho should read this: anyone working on trajectory representation learning or urban semantics for mobility. It deserves a serious referee. I would send it, but with a request to fix the consistency claim, reconcile the abstract numbers, report error bars, and clarify the TTE input. If those are addressed, this is a solid paper.","headline":"Nice framework with genuine novelty; headline claim overreaches given one negative metric and missing error bars.","tokens_in":35172,"tokens_out":2205,"would_cite":true,"duration_ms":18944,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Trajectory embeddings improve when routes are read as decisions shaped by the surrounding city, not as passive GPS traces.","keywords":["trajectory representation learning","route choice modeling","point of interest","large language models","mixture-of-experts","contrastive learning","road network embedding","urban mobility"],"falsifier":"Compare a CORE variant trained with real POI data against one trained with the same POI labels randomly shuffled across road segments. If shuffling does not degrade downstream metrics, the claimed environmental-semantics mechanism is not the source of the gains. A more localized test: hold out all POIs for a random subset of road segments and check whether road-label prediction on those segments drops relative to covered segments.","tokens_in":34118,"feed_emoji":"🧭","tokens_out":5786,"duration_ms":45314,"temperature":0.7,"pith_summary":"This paper argues that a GPS trace is not just a sequence of positions but the visible outcome of a series of route choices, each conditioned on the surrounding urban environment. Most existing trajectory representation learning treats trajectories as passive spatiotemporal sequences, missing that behavioral layer. CORE instead builds a context-enriched road network by using a large language model to distill semantics from nearby point-of-interest distributions, then encodes each step as a choice between the taken next road segment and adjacent segments not taken, mediated by a mixture-of-experts router and navigational cues. On four real-world datasets and six downstream tasks, the paper reports consistent wins over twelve existing methods, with an average gain over the best baseline of roughly 10%. If this holds, trajectory learning should model decision context, not just geometry.","feed_headline":"Trajectory embeddings get a 10% lift by modeling route choices","feed_subtitle":"Street-level context from nearby places turns each GPS trace into a sequence of decisions, and it beats 12 existing methods.","key_machinery":"Context-aware route choice semantics: each step of a trajectory is treated as a choice among adjacent road segments conditioned on the local urban environment. The load-bearing machinery is (1) an Environment Perception Module that uses a large language model to turn nearby POI distributions into textual descriptions, embeds them, propagates them across the road network with a graph attention network, and fuses them with coarse-grained hotspot semantics via gating; (2) a Route Choice Encoder that, at each segment, encodes the chosen next segment and the rejected adjacent segments, with a mixture-of-experts router selecting specialized subnetworks based on the current state plus navigational","core_discovery":"CORE's central claim is that injecting context-aware route choice semantics as an inductive bias improves trajectory representations at every evaluated level. The framework first constructs a context-enriched road network: fine-grained road-segment semantics come from LLM-generated descriptions of nearby POIs, propagated from the top 20% busiest segments across a graph attention network, and coarse-grained urban functional structure comes from LLM descriptions of POI hotspots on a 1 km grid, both fused with road attributes through gating. A Route Choice Encoder then represents each decision step by comparing the selected next road segment against unselected neighbors, using a mixture-of-expe","pith_inferences":["The authors leave implicit that the mechanism can be tested against other context sources such as congestion, tolls, or road quality; where those factors dominate route choice, POI-derived semantics alone would likely underperform and should be enriched.","A direct extension of their logic is cross-city transfer: if route-choice semantics are universal, an embedding pretrained on one city should transfer to another without retraining; the paper currently reports single-city results and flags this as future work.","The reported sensitivity to POI completeness suggests a testable limit: near-random performance under heavily shuffled or sparse POI data would confirm that environmental semantics, not architecture, drive the gains.","The mixture-of-experts gating visualizations imply that experts specialize by functional region and road type; quantifying this specialization could yield interpretable route-choice summaries for planning applications."],"forward_implications":["Trajectory representation learning should adopt behavioral modeling rather than purely geometric sequence encoding if these results hold.","The largest reported gains appear on road segment classification and similar-trajectory retrieval, suggesting the embeddings carry fine-grained route-choice distinctions that geometry alone misses.","LLM-generated semantic maps are described as cheap enough for city scale, with a reported cost of about $2 and two hours for Beijing, making the context-enriched network practical.","The model retains strong performance under small training-data fractions, indicating the behavioral inductive bias reduces data hunger.","Trajectory generation also improves substantially, so the same semantics used for retrieval and prediction can produce realistic movements."],"fun_headline_variants":["Modeling route choices boosts trajectory embeddings by 9%","Trajectory AI learns from route choices, not just paths","Context-aware route semantics lift trajectory embeddings 9%","Route choice modeling gives trajectory embeddings a 9% boost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"POI distributions are an adequate observable proxy for the urban decision-making context; if the places near a road do not reflect what actually drives drivers' choices, the context-enriched network loses its value.","fun_headline_variants_meta":{"raw":{"variants":["Modeling route choices boosts trajectory embeddings by 9%","Trajectory AI learns from route choices, not just paths","Context-aware route semantics lift trajectory embeddings 9%","Route choice modeling gives trajectory embeddings a 9% boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3646,"prompt_tokens":775,"completion_tokens":2871,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":2804}},"tokens_in":519,"tokens_out":2871,"duration_ms":17058,"temperature":1.0,"reasoning_tokens":2804,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:29:31.314352+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare a CORE variant trained with real POI data against one trained with the same POI labels randomly shuffled across road segments. If shuffling does not degrade downstream metrics, the claimed environmental-semantics mechanism is not the source of the gains. A more localized test: hold out all POIs for a random subset of road segments and check whether road-label prediction on those segments drops relative to covered segments.","supporting_citations":[],"review_version":1}