{"id":"b2d6cf37-baeb-45ac-be93-a4ca5b4c7faa","arxiv_id":"2507.19586","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A new benchmark called GEOHALUBENCH measures how often LLMs invent, omit, or confuse real-world places and relations, and a dynamic-beta KTO method reduces these errors on the benchmark.","lead":"LLMs often give wrong or invented answers about cities, roads and landmarks, yet there was no systematic way to measure these geospatial errors. The authors build a benchmark from map databases, show many models fail it, and propose a fine-tuning method that cuts such errors on their tests. A smart generalist should read it to understand the current trustworthiness of LLMs as geographic knowledge sources.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 29.6% headline gain is evaluated on the same benchmark used to select the model and its task-level betas; without a held-out city or a refusal-rate-controlled comparison, the gain may reflect benchmark overfitting and refusal shifting rather than reduced hallucination.","rationale":"The paper's central claim is the 29.6% improvement on the proposed benchmark, and the reader's weakest assumption is that GEOHALUBENCH validly measures geospatial hallucination. I agree with that as the load-bearing concern, but the more concrete and decisive issue is that the benchmark and the training data come from the same source (SpatialKG), and the model is selected and evaluated on the same benchmark, with no held-out city or refusal-rate-controlled comparison. This is a matter of internal validity: the measured improvement could be an artifact of benchmark overfitting (the DynamicKTO beta values are chosen on the same benchmark) and refusal-shifting (the evaluation is on GEOHALUBENCH-Abstain, which adds a 'Cannot Determine' option that interacts with the benchmark's labeling). The reasoning is not circular in a logical sense—the method is trained and tested on separate instances—but the benchmark is the same distribution, and the reported 29.6% gain is not yet robust evidence of real hallucination reduction. My concrete test (held-out city evaluation with matched refusal rates) is a minimal, decisive check. If the held-out improvement holds, the paper's central claim is substantially validated, and the verdict should move toward ACCEPT; if not, the claim is conditional at best. I choose CONDITIONAL to match the reader's verdict, since the concern is addressable and the evidence is suggestive but not sufficient.","tokens_in":22351,"tokens_out":2128,"duration_ms":20358,"concrete_test":"Evaluate the trained DynamicKTO model (Llama3.1-8B) and the fixed-beta KTO baselines on GEOHALUBENCH-Abstain for a held-out city not used in training (e.g., the paper's Cairo, Kabul, or Sucre, or a new city such as Tokyo or Mumbai), using the SAME beta values selected on Beijing. If the relative improvement of DynamicKTO over the best fixed-beta KTO drops below, say, half of the 29.6% Beijing margin, or if the DynamicKTO model's gain disappears once refusal rates are matched (e.g., by conditioning on non-abstain answers), then the central claim of reduced hallucination is not supported. Also, re-run the fixed-beta KTO ablation with a denser beta sweep and per-category tuned betas to see if the gap persists.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that DynamicKTO reduces geospatial hallucination, evidenced by Llama3.1-8B reaching 0.4858 overall on GEOHALUBENCH-Abstain versus 0.3748 for the best fixed-beta KTO (Section 3.2, Table 3). This claim has three load-bearing supports, and each is weaker than the headline implies. First, the ablation in Table 3 compares DynamicKTO (with betas 0.1/0.3/0.5 for Entity/Relation/Attribute chosen by the authors) against only three fixed KTO betas (0.1, 0.3, 0.5), so the 29.6% is relative to the best of a coarse grid, not to a tuned fixed-beta KTO; a denser beta sweep or per-category tuned betas could close much of the gap. Second, the same benchmark is used for both selection and evaluation: the beta values are post-hoc selections made on the same benchmark used for evaluation, with no held-out city or dataset; this inflates the measured improvement. Third, the evaluation is on GEOHALUBENCH-Abstain, which adds a 'Cannot Determine' option. Refusal rates and abstention behavior interact with the benchmark's construction: the paper's own Figure 5 shows that adding the refusal option changes model rankings substantially, and Section 2.2.3 assigns 'None of the other options' as the hallucination label for Entity Omission and Relation Omission, while an abstaining model that tends to answer 'Cannot Determine' may systematically receive credit or blame without actually possessing the geospatial knowledge. Table 13 shows several tasks near zero (AOI-Area 0.000-0.240 vs random 0.200), suggesting that the benchmark may be measuring format biases or task artifacts rather than geospatial hallucination. If the benchmark is not a valid measure of geospatial hallucination, then the accuracy gains of DynamicKTO do not reflect real hallucination reduction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GEOHALUBENCH, a multiple-choice benchmark for geospatial hallucinations, constructed from a structured geospatial knowledge graph (SpatialKG) with a taxonomy covering entity fabrication/omission, relation fabrication/omission, and attribute confusion. The authors evaluate 20 LLMs on this benchmark, report generally low accuracy and a strong bias against underrepresented regions, and propose DynamicKTO, a variant of KTO in which the risk-aversion parameter beta is assigned per task category (0.1 for Entity, 0.3 for Relation, 0.5 for Attribute). On GEOHALUBENCH-Abstain, Llama3.1-8B fine-tuned with DynamicKTO reaches 0.4858 overall accuracy versus 0.3748 for the best fixed-beta KTO, a relative improvement of 29.6%, with only modest degradation on general benchmarks. The paper also builds Factual-CityGPT by applying DynamicKTO on CityGPT and reports improved hallucination scores with preserved urban task performance.","tokens_in":22600,"tokens_out":6801,"duration_ms":67848,"significance":"If the reported gains are robust, the paper makes a useful contribution: it is one of the first systematic benchmarks for geospatial hallucinations, it evaluates a broad set of 20 LLMs, and it proposes a simple, cost-effective extension of KTO with a task-dependent beta. The authors provide code and data, include a training-cost analysis, and show that the model retains general capabilities on IFEval, BBH, and MMLU. The central claim, however, depends on benchmark validity and on a fair comparison to tuned fixed-beta baselines; both need stronger evidence before the magnitude of the improvement can be accepted. With those issues addressed, the benchmark and method would be valuable to the hallucination mitigation community.","major_comments":[{"comment":"The three task-level beta values (0.1, 0.3, 0.5) are selected after inspecting LFactDistance and the model's performance on GEOHALUBENCH itself, with no held-out city or data split used for selection. The headline 29.6% gain is therefore computed relative to a three-point fixed-beta grid, and a tuned fixed-beta KTO or per-category selected betas could close much of the gap. Please add a validation protocol in which betas are chosen on one city (or one fold) and evaluated on a disjoint city (or fold), for example tuning on Beijing and testing on London and New York, to demonstrate that the dynamic assignment, rather than in-sample selection, drives the improvement.","section":"§3.2, Table 3; §A.8"},{"comment":"The construct validity of the benchmark as a measure of geospatial hallucination needs stronger support. In Table 13, several tasks are at or near zero for essentially all models while the random baseline is 0.200 (e.g., AOI-Area: most models score 0.000-0.100, with only Qwen2.5-0.5B reaching 0.320), and Figure 5 shows that adding the 'Cannot Determine' option substantially reorders model rankings. Since 'None of the other options' is labeled as Entity or Relation Omission, a model that prefers a particular option letter or that systematically abstains can be credited or penalized without actually exhibiting the targeted hallucination. I request a per-task option-distribution analysis (proportion of A/B/C/'Cannot Determine' choices per model compared with the random baseline) and a refusal-rate-controlled breakdown of DynamicKTO's gains to rule out format bias and abstention effects.","section":"§2.2.3, Table 13, Figure 5"},{"comment":"The comparison to fixed-beta KTO uses only the grid beta in {0.1, 0.3, 0.5}, which is too coarse to support the claim that the dynamic per-task assignment, rather than the specific beta values, is responsible for the improvement. The dynamic-beta DPO variant (beta-DPO, Wu et al., 2024) is cited in Related Work but is not included as a baseline; it is the direct dynamic-beta analog of DPO and should be evaluated. Please add a denser beta sweep, per-category tuned fixed betas, and beta-DPO to the comparison.","section":"§2.3.2, Table 3"},{"comment":"All fine-tuning results are reported as point estimates without repeated seeds or error bars, even though fine-tuning is stochastic and several paired differences are small (e.g., Relation accuracy: DynamicKTO 0.4256 vs ORPO 0.4320 in Table 3; Qwen2.5-7B improvements of roughly 0.02-0.06 in Table 5). Please report variance over at least three seeds or a paired significance test for the headline DynamicKTO versus fixed-beta KTO comparison; without this, the size of the claimed effect is not statistically grounded.","section":"§3.2, Tables 3 and 5"}],"minor_comments":[{"comment":"The first bullet under 'Sample-Level' ends with 'based on its' and is an incomplete sentence; please complete the description of how beta is computed per sample.","section":"§A.8"},{"comment":"The row for Llama-3.1-8B-Instruct appears twice with identical values; remove the duplicate row.","section":"Table 13"},{"comment":"The caption says models are 'listed at name order,' but the vertical order in the figure appears to follow a specific evaluation order; please clarify the ordering criterion in the caption.","section":"Figure 4"},{"comment":"The claim that testing three beta values 'confirms that the performance boost is not merely due to trivial hyperparameter optimization' is too strong for a three-point grid; please soften the wording to reflect the limited sweep.","section":"§3.2"},{"comment":"Several in-text citations are formatted inconsistently (e.g., 'Niels et al.(Mündler et al., 2023)'); please standardize the citation style throughout the manuscript.","section":"References and in-text citations"}],"recommendation":"major_revision","confidential_remarks":"The omission of beta-DPO as a baseline is notable because the paper cites it in Related Work; I would treat it as an oversight to be fixed in revision rather than a deliberate choice. The duplicate row for Llama-3.1-8B-Instruct in Table 13 suggests the final proofreading was hasty, and the authors should verify the experimental tables and appendix text carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth reading for the benchmark and taxonomy, but the headline mitigation result is not yet credible. The taxonomy—entity/relation/attribute with fabrication vs omission—is sensible, and building the benchmark from OpenStreetMap and Foursquare is transparent and reproducible. Running 20 models across six cities gives a useful descriptive picture, and the finding that omission dominates fabrication is a nice empirical payoff. The abstention analysis is also a good addition; showing that adding a refusal option reshuffles model rankings is a real insight.\n\nThe soft spots are real and addressable. The biggest is circularity: the dynamic betas (0.1/0.3/0.5) are chosen using LFactDistance on the same kind of data, and the model is then evaluated on GEOHALU-BENCH-Abstain, which comes from the same cities and distribution. So the 29.6% improvement over the best fixed-beta KTO is partly hyperparameter selection, not a clean test of the method. The comparison grid is coarse (three betas), and beta-DPO is cited but not run as a baseline. A held-out city or a denser sweep would sharpen the claim considerably.\n\nThe benchmark itself has a validity flag. AOI-Area is near zero for nearly every model while random gets 0.20, and several tasks in Table 13 sit below random. That suggests some items are broken—perhaps multiple options are technically correct, or the ground-truth label is wrong. This doesn't sink the whole benchmark, but aggregate accuracy should be read with that in mind, and the paper should say which tasks are diagnostic rather than presenting all 21 as equally meaningful.\n\nThe abstain variant is double-edged. Adding 'Cannot Determine' as an option mixes abstention behavior with knowledge, and the paper shows rankings shift a lot. The DynamicKTO numbers are reported on the abstain version, so part of the gain could be a shift in refusal behavior rather than reduced hallucination; the paper doesn't report refusal rates for the trained models. No error bars or repeated seeds either.\n\nWho this is for: people building geospatial QA or hallucination benchmarks will get real value from the taxonomy and multicity results. The mitigation result needs more evidence before changing practice.\n\nRecommendation: send to peer review. The benchmark contribution alone deserves referee time. The revision should add a held-out city, beta-DPO and refusal-rate-controlled baselines, confidence intervals, and a per-task validation pass on benchmark labels.","headline":"Useful geospatial hallucination benchmark with a solid taxonomy, but the DynamicKTO gain is inflated by same-benchmark beta selection and some tasks look broken.","tokens_in":23296,"tokens_out":3461,"would_cite":false,"duration_ms":38799,"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":"A knowledge-graph benchmark exposes geospatial hallucination in 20 LLMs, and a dynamic fine-tuning method cuts errors by 29.6%.","keywords":["geospatial hallucination","large language models","hallucination benchmark","knowledge graph","Kahneman-Tversky optimization","dynamic beta","factuality alignment"],"falsifier":"Take a random sample of items where a model chose 'None of the other options' and verify against an independent, more complete map source whether the omitted entity truly exists; if a meaningful share of such cases is actually present in the world, then the omission labels—and the accuracy gains built on them—are artifacts of knowledge-graph gaps rather than genuine hallucinations.","tokens_in":22011,"feed_emoji":"🗺️","tokens_out":7070,"duration_ms":68892,"temperature":0.7,"pith_summary":"Geospatial hallucinations—LLMs asserting places, roads, or spatial relations that do not exist, omitting ones that do, or mixing up attributes like area and category—are common enough to threaten any LLM-based geospatial application, and the paper sets out to show they can be measured and reduced. It constructs a city knowledge graph with points of interest, areas, and roads, uses it to generate multiple-choice questions whose options map to five hallucination types, and reports that 20 leading LLMs score low on the resulting benchmark, often near or below random on individual tasks. It then proposes DynamicKTO, a fine-tuning method that changes the risk-aversion parameter of Kahneman-Tversky Optimization according to whether a training item tests entities, relations, or attributes, and reports a 29.6% improvement over the best fixed-parameter KTO on the abstain-enabled version. The paper claims this brings Llama3.1-8B from 0.3505 to 0.4858 overall accuracy, making a small open model competitive with the best proprietary models it measured.","feed_headline":"Dynamic tuning cuts geospatial hallucinations by 29.6%","feed_subtitle":"A benchmark built from city knowledge graphs finds the errors; a per-task tuning method fixes them.","key_machinery":"The load-bearing object is GEOHALU BENCH, a multiple-choice benchmark generated from SpatialKG, a knowledge graph of points of interest (POIs), areas of interest (AOIs), and roads with relations and attributes. Each question's distractors are tied to hallucination types: fabricated entities are LLM-generated names vetted against the graph, omitted entities are represented by a 'None of the other options' choice, and wrong attribute values are sampled beyond a threshold. The training-side mechanism is DynamicKTO, which keeps KTO's binary desirable/undesirable objective but replaces the fixed $\\beta$ with $\\beta(x) = \\mathrm{Type2Beta}(x)$, set to 0.1 for entity items, 0.3 for relation items, and 0.5 for attribute items. This lets the optimizer tighten or relax alignment per knowledge type rather than applying one risk policy to all data.","core_discovery":"The paper's central claim is that geospatial hallucination is a distinct, measurable failure mode of LLMs and that a benchmark plus a dynamically parameterized fine-tuning objective can substantially reduce it. On the benchmark, the entity/relation/attribute taxonomy identifies five error types (fabrication and omission for entities and relations, plus attribute confusion), and evaluation across three major cities and three underrepresented cities shows low and uneven accuracy, with omission generally exceeding fabrication. The paper then demonstrates that fixing one risk-aversion value for the entire KTO loss is a bottleneck: because entity, relation, and attribute items differ in difficulty, a per-category beta (0.1, 0.3, 0.5) produces the best result. DynamicKTO lifts Llama3.1-8B's overall score to 0.4858 on the abstain version, a 38.61% relative gain over no fine-tuning and 29.60% over the best fixed-beta KTO, while preserving performance on general benchmarks and on the CityGPT urban-task evaluation.","pith_inferences":["An untested consequence is that retrieval-augmented generation over the same spatial knowledge graph could match or exceed the training-based gain at inference time, since most errors are omissions of world knowledge rather than confident fabrication.","A more stringent test would convert the multiple-choice task to open-form generation and check whether models still avoid fabricated places; the current format may undercount or overcount certain error types depending on option positioning.","The taxonomy-level beta tuning suggests a general recipe: compute a difficulty statistic on a sample, then assign beta by cluster; the paper tests this and finds category-level tuning better for geospatial data, but sample-level tuning may win in less structured domains."],"forward_implications":["If the reported gains hold, geospatial factuality can be treated as an alignable skill rather than an emergent property of model size.","A small open model can reach accuracy comparable to the best proprietary models on this benchmark, changing the cost equation for trustworthy geospatial applications.","The dynamic-beta principle should transfer to any domain where training data splits into difficulty or risk classes, such as medicine, law, or finance.","Because omission dominates fabrication, models improved by knowledge injection rather than refusal training are likely to be more useful in practice."],"supporting_citations":[{"why":"Defines the KTO objective that DynamicKTO modifies by making the beta parameter a function of the sample type.","marker":"Ethayarajh et al., 2024"},{"why":"Provides DPO, the preference-optimization baseline whose beta and reference-model mechanics frame DynamicKTO's dynamic-beta design.","marker":"Rafailov et al., 2023"},{"why":"Supplies the urban knowledge graph that the authors adapt into SpatialKG to generate ground-truth geospatial facts.","marker":"Liu et al., 2023b"},{"why":"CityGPT is the urban-task LLM the paper re-trains with DynamicKTO to build Factual-CityGPT.","marker":"Feng et al., 2024a"},{"why":"Documents LLM geographic bias that motivates evaluating and mitigating geospatial hallucination.","marker":"Manvi et al., 2024"},{"why":"Evidence that LLMs encode grounded spatial knowledge, making the hallucination benchmark a test of real-world knowledge rather than format compliance.","marker":"Gurnee and Tegmark, 2023"},{"why":"Shows LLMs can answer geographic questions from an internal world model, supporting the premise that geospatial knowledge can be elicited and checked.","marker":"Roberts et al., 2023"}],"fun_headline_variants":["Benchmark exposes geospatial hallucination, dynamic KTO fixes 29.6%","Per-category tuning shrinks geospatial errors by 29.6%","Dynamic factuality aligning lifts Llama-8B geo score to .486","Geo-hallucination benchmark spots errors; dynamic tuning cuts them","KTO with adaptive risk cuts geospatial hallucination by 29.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that GEOHALU BENCH's labels are true: SpatialKG must be complete enough that a 'None of the other options' answer is genuinely an omission, and the distractors must be plausible but non-factual; if the graph misses real places, the benchmark mistakes the model's correct answer for a hallucination.","fun_headline_variants_meta":{"raw":{"variants":["Benchmark exposes geospatial hallucination, dynamic KTO fixes 29.6%","Per-category tuning shrinks geospatial errors by 29.6%","Dynamic factuality aligning lifts Llama-8B geo score to .486","Geo-hallucination benchmark spots errors; dynamic tuning cuts them","KTO with adaptive risk cuts geospatial hallucination by 29.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1603,"prompt_tokens":956,"completion_tokens":647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":545}},"tokens_in":572,"tokens_out":647,"duration_ms":6967,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:15:36.707659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of items where a model chose 'None of the other options' and verify against an independent, more complete map source whether the omitted entity truly exists; if a meaningful share of such cases is actually present in the world, then the omission labels—and the accuracy gains built on them—are artifacts of knowledge-graph gaps rather than genuine hallucinations.","supporting_citations":[],"review_version":1}