acceptodds
Under review as a conference paper at ICLR 2027

CrimeGPT: Numerically Faithful Crime Forecasting with Situated Large Language Models

Abstract

Crime prediction is essential for proactive community safety and efficient police resource allocation. Beyond strong spatio-temporal periodicity, criminal activity is closely intertwined with regional functional characteristics and neighborhood environments, contextual factors that conventional statistical models largely overlook. Large language models (LLMs) offer a natural pathway to jointly reason over heterogeneous urban descriptors and historical signals within a unified semantic framework. However, their discrete, token-based generation paradigm renders them inherently insensitive to numerical values, a limitation further amplified in crime forecasting where competing multi-source urban signals prevent the model from forming confident numerical outputs. To this end, we propose CrimeGPT, which rethinks numerically faithful LLM-based crime prediction along two complementary axes: situating LLMs in urban dynamics and ensuring numerical faithfulness in their outputs. CrimeGPT introduces a Crime Dynamics Encoder that situates the LLM in heterogeneous urban signals. Bidirectional ordinal supervision then organizes numerical-token representations, which are mapped to continuous crime-count estimates by a distribution-guided regression head. Extensive experiments on real-world datasets from multiple cities show that CrimeGPT achieves superior performance over state-of-the-art methods. Our code and data will be made publicly available.

Primary area:
learning on time series and dynamical systems
Keywords:
Crime Prediction, Large Language Models, Spatial Temporal Modeling, Numerical Reasoning in LLMs

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HeT1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CrimeGPT, an LLM-based framework for crime-count forecasting that integrates heterogeneous urban information with numerical prediction. The framework combines a Crime Dynamics Encoder for spatio-temporal modeling, dedicated numerical tokens with bidirectional ordinal supervision (BiCORN), and a Distribution-Guided Regression Head (DGRH) for continuous count estimation. Experiments across four cities, including cross-year evaluation and zero-shot cross-city transfer, demonstrate competitive forecasting performance.

Critical strengths
  • The paper represents an early effort to apply large language models to crime prediction, exploring their potential to integrate heterogeneous urban information for forecasting.
  • The experiments cover four cities and multiple crime categories, with comparisons against general forecasting models, crime-specific methods, and LLM-based approaches.
Critical weaknesses
  • The claim that “… contextual factors that conventional statistical models largely overlook” is overstated. Earlier crime prediction approaches already incorporated urban context, including submissions to the NIJ Real-Time Crime Forecasting Challenge, DeepCrime, and Crime Rate Inference with Big Data.
  • The motivation for using an LLM for crime prediction is not clear. The manuscript should explain which specific challenges require an LLM and why conventional spatio-temporal models are insufficient to address them.
  • Limited numerical sensitivity appears to be a general limitation of LLMs rather than a challenge specific to crime forecasting. The distinctive challenges of crime prediction and the insufficiency of existing methods are still not clear.
  • The comparison should place greater emphasis on LLM-based forecasting methods. The current baselines, dating from 2024 or earlier, are not sufficiently competitive given recent advances in time-series foundation models.
  • Although results are averaged over three runs, Table 2 reports variability only for CrimeGPT, and some gains over competitive baselines are marginal.
Questions for authors

See weakness.

4 comments
AuthorsOpenReview ↗

Response (Part I)

Thank you for recognizing our exploration of LLM-based crime forecasting and the breadth of the evaluation.

In response to your comments, we will expand the language-model controls and comparisons with recent LLM-based forecasting methods, add baseline standard deviations, and clarify how CrimeGPT builds on context-aware crime forecasting. Our responses below further explain the role of language capabilities, the crime-specific challenges, and the variation in gains across forecasting settings.

Q1: Positioning CrimeGPT within context-aware crime forecasting.

A: Thank you for highlighting these foundational studies. Earlier crime forecasting methods already incorporate urban context, and we will clarify this relationship in Related Work. Our contribution concerns how heterogeneous urban information is aligned with a language model and used for numerical prediction.

CDE encodes crime and service-request histories with spatial relationships, then projects them into the LLM's semantic space. The LLM processes these dynamics alongside regional and calendar descriptions. BiCORN and DGRH further provide ordinal and explicit distance supervision. CrimeGPT builds on context-aware forecasting through numerical-language alignment and task-specific numerical supervision.

Q2: Motivation and empirical support for the language-model pathway.

A: Crime forecasting depends on recent urban dynamics and the meaning of the regional context in which they occur. Similar counts can arise in regions with different functions and activity patterns. Beyond model capacity, the LLM provides pretrained language knowledge for semantic understanding and contextual reasoning. It interprets regional functions and calendar descriptions and relates them to CDE's numerical representation, allowing the forecast to reflect both recent observations and their urban context.

We evaluated CDE + MLP with L1 alone and with auxiliary CE + BiCORN. We also included random and pretrained Qwen2.5-1.5B controls with CDE, regional prompts, and MLP regression, both using L1 + CE + BiCORN.

Table C1: Encoder and language-model controls on NYC All Crime 2022.

ModelTraining objectiveMAE ↓RMSE ↓
CDE + MLPL14.0635.693
CDE + MLPL1 + CE + BiCORN4.0355.713
CDE + Qwen2.5-1.5B (random initialization) + MLPL1 + CE + BiCORN3.7275.228
CDE + Qwen2.5-1.5B (pretrained) + MLPL1 + CE + BiCORN3.6625.082
CrimeGPTL1 + CE + BiCORN3.5374.921

Relative to CDE + MLP with auxiliary supervision, CrimeGPT reduces MAE and RMSE by 12.3% and 13.9%, respectively. In the reported 1.5B comparison, the pretrained configuration obtains lower MAE/RMSE than the randomly initialized configuration. Table 3(e) further shows that removing contextual prompts raises MAE/RMSE from 3.537/4.921 to 3.582/4.989. Together, these comparisons support using pretrained language representations and regional semantics alongside recent numerical dynamics.

Q3: Distinctive challenges of crime forecasting and limitations of existing methods.

A: Thank you for the opportunity to clarify this motivation. Numerical sensitivity is relevant to LLM forecasting generally. Crime prediction makes it particularly consequential through noisy, heterogeneous observations whose relevance depends on regional and temporal context.

  1. Interpreting heterogeneous urban evidence. Crime histories, service requests, mobility, and regional functions provide indirect and sometimes competing evidence. General spatiotemporal models primarily capture temporal and spatial dependencies. Existing crime-specific models incorporate urban context, but their structured feature representations offer limited support for interpreting how regional meaning changes the relevance of recent observations. CrimeGPT addresses this through CDE's numerical-language alignment and the LLM's semantic understanding of regional and temporal descriptions.
  2. Numerical prediction from ambiguous crime signals. Table 1 reports a mean peak count-token probability of 0.226 for NYC Crime, compared with 0.651 for Bike and 0.454 for Taxi. This diffuse distribution illustrates the difficulty an LLM faces when noisy urban signals leave several crime counts plausible. Standard token-level CE does not explicitly supervise numerical order or count distance, and argmax discards the remaining probability mass. Thus, interpreting urban context alone does not resolve the challenge of producing numerically faithful crime forecasts. BiCORN supplies ordinal supervision, while DGRH uses latent numerical information with an explicit distance objective.

CrimeGPT addresses both the interpretation of heterogeneous urban evidence and the numerical uncertainty associated with crime-count prediction.

AuthorsOpenReview ↗

Response (Part II)

Q4: Relationship to recent LLM-based forecasters and additional comparison.

A: Recent methods explore ways to represent time-series information for LLM-based forecasting. LagLLM [1] uses lead-lag dependencies, T3Time [2] uses temporal, spectral, and prompt representations, and MSH-LLM [3] uses multi-scale hypergraph features. These designs primarily concern the representation of forecasting inputs.

CrimeGPT additionally focuses on numerical-language alignment and numerical training objectives. CDE aligns heterogeneous urban dynamics with the LLM's semantic space. BiCORN and DGRH address diffuse predictions through ordinal and explicit distance supervision, respectively, which is particularly relevant to noisy, multi-source crime observations.

We additionally evaluate MSH-LLM [3], an ICLR 2026 method, using the same horizons, chronological splits, and available crime and 311 histories.

Table C2: Additional comparison with MSH-LLM.

CityCrime categoryMSH-LLM MAE ↓MSH-LLM RMSE ↓CrimeGPT MAE ↓CrimeGPT RMSE ↓
NYCAll3.6245.0423.5374.921
NYCTheft2.3613.3762.2863.285
NYCAssault2.0652.8972.0362.880
ChicagoAll2.5663.6122.5303.598
ChicagoTheft1.7642.5191.7162.466
ChicagoAssault1.2111.7341.1931.742
PhiladelphiaAll3.1294.2623.0174.139
PhiladelphiaTheft1.9962.7251.9082.666
PhiladelphiaAssault1.5352.1261.5112.119
ShanghaiAll0.7941.4560.7351.345

We will incorporate this comparison into the revised manuscript.

Q5: Result variability and gains across forecasting settings.

A: We will add standard deviations from the existing experimental runs to Table 2.

As discussed in Section 5.4, Advantages of CrimeGPT, the gains are clearer in high-frequency, complex settings such as NYC All Crime, where richer observations and a wider count range make contextual and numerical modeling more useful. Figure 4(a) likewise shows clearer gains in high-frequency NYC regions.

SHA provides a contrasting sparse setting: 53.69% of test targets are zero, and an all-zero predictor yields an MAE of 0.837. This helps explain the comparatively modest gains there. The results suggest that CrimeGPT's advantages are more pronounced in information-rich, complex settings, with smaller gains when contextual information and count variation are limited.

References

[1] Wu, B., Zhou, J., Shang, Z., and Chen, L. (2026). LagLLM: LLM-empowered lead-lag dependency learning for spatial-temporal time series forecasting. ICML.

[2] Chowdhury, A. M., Akter, R., and Arib, S. H. (2026). T3Time: Tri-Modal Time Series Forecasting via Adaptive Multi-Head Alignment and Residual Fusion. AAAI.

[3] Shang, Z., Cui, D., Wu, B., and Chen, L. (2026). Multi-scale hypergraph meets LLMs: Aligning large language models for time series analysis. ICLR.

Reviewer_HeT1OpenReview ↗

More evidences need insightful analysis or specific examples.

The authors primarily explain they use LLMs for semantic understanding and contextual reasoning, for example, regional functions and calendar descriptions. However, it is still not intuitive. More evidences need insightful analysis or specific examples. And, the new explanations may need to reorganize the paper contents. Therefore, I maintain my scores.

Thank you for your time and thoughtful feedback throughout the review process.

Reviewer_WadZOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies crime forecasting with situated LLMs. The authors identified two challenges:

  1. The LLMs are not aligned with urban signals such as POI, traffic flow, etc.
  2. LLMs are numerically unfaithful.

To overcome them, the authors proposed:

  1. Crime dynamics encoder to construct embeddings for heterogeneous urban context and feed it to the prompts to solve the challenge 1.
  2. BiCORN and distribution-guided regression head for the numerical faithfulness to solve the challenge 2.
Critical strengths

S1. This paper studies context alignment and numerical faithfulness of LLMs which are both interesting and meaningful.

Critical weaknesses

W1. No code release. The source code is not released at review time and therefore the main results are not convincing.

W2. Presentation issues. The paper writing should be seriously revised. Take CDE specification as an example:

  • Implementation details are pushed entirely to Appendix B and the main paper treats this module as a black box.
  • The input parameters of CDE (Aadj,Adist,AflowA_{adj}, A_{dist}, A_{flow}) is not well defined (eg, their dimensions) or described (eg, the physical meaning of AflowA_{flow}).
  • The shape of EstE_{st} is N×dstN\times d_{st} in the main paper while the appendix B describe the shape as N×dsN\times d_{s}, which is likely to be a typo.

W3. Incomplete inputs for datasets PHI and SHA. According to Table 5: PHI lacks taxi flow and SHA lacks both X_311 and taxi flow. The authors are suggested to explain how to handle such incomplete inputs.

W4. Embedding injection should be further justified. The effectiveness of the embedding injection should be further justified. According to Table 3(b), CrimeGPT w/o CDE only introduced marginal performance degration. The authors are suggested to proof this degration is non-trivial, eg, its stability, the comparision against the prompt optimization, etc.

W5. Formulation issue for BiCORN. Eq. (22) names P(Y≤k∣Y>k−1)≈σ(rk−)P(Y\le k \mid Y > k-1) \approx \sigma(r_k^-) and by Eq. (23), σ(rk−)\sigma(r_k^-) is encouraged to 1 for k>y∗k>y^{*}. Since the integer YY is conditioned on Y>k−1Y>k-1, this supervision can be rewrittened as P(Y=k∣Y>k−1)→1,fork>y∗,P(Y=k \mid Y>k-1)\rightarrow 1,\quad \text{for}\quad k > y^{*}, which no probability distribution can satisfy.

Questions for authors
  • Q1. In the prompt template provided in the appendix (F.1), the number of crimes and 311 service calls are given by ordinary digital numbers instead of the introduced [N_0], ..., [N_99]. The authors should justify their selection.
5 comments

Thank you for recognizing the importance of contextual alignment and numerical faithfulness.

Following your suggestions, we will add prompt-optimization experiments, clarify CDE dimensions and the handling of missing inputs, and correct the reverse conditioning event with an explanation of the surrogate scores. We have also provided anonymized reproduction materials and will clarify the choice of input numerals and output count tokens in our responses below.

Q1: Resources for reproducing CrimeGPT.

A: We have uploaded an anonymized reproduction package containing data and scripts to support reproduction of our results.

Q2: Clarification of CDE architecture and tensor dimensions.

A: Thank you for the suggestion. The inconsistent use of dstd_{\mathrm{st}} and dsd_s is a notational typo. Both denote the CDE output width of 64, and we will use dstd_{\mathrm{st}} consistently. The input is X~∈RN×H×2\tilde{\mathbf X}\in\mathbb R^{N\times H\times2} and the output is Est∈RN×dst\mathbf E_{\mathrm{st}}\in\mathbb R^{N\times d_{\mathrm{st}}}. We will make the spatial inputs explicit in Section 4.1:

InputShapeMeaning
Aadj\mathbf A_{\mathrm{adj}}N×NN\times NShared-boundary adjacency mask
Adist\mathbf A_{\mathrm{dist}}N×NN\times NNormalized proximity weights derived from distances between region centroids
Aflow\mathbf A_{\mathrm{flow}}Tdata×N×NT_{\mathrm{data}}\times N\times NDaily taxi-flow sequence, with TdataT_{\mathrm{data}} observed days and entries counting trips from origin ii to destination jj

The daily slice Aflow[t]∈RN×N\mathbf A_{\mathrm{flow}}[t]\in\mathbb R^{N\times N} supplies the flow matrix at day tt. Adjacency masks spatial attention, while distance and flow supply edge features. Gated temporal convolutions, spatial attention, and residual/skip connections produce regional embeddings that are projected into the LLM space.

Q3: Handling missing 311 and taxi-flow inputs in PHI and SHA.

A: As noted in Appendix A.1, taxi-flow data are unavailable for Philadelphia and Shanghai, and 311 data are unavailable for Shanghai. We handle these cases as follows:

  • Without 311: retain a zero channel in CDE and omit 311 history from the prompt.
  • Without taxi flow: use a zero flow matrix, giving flow zero contribution to the attention bias while retaining adjacency and distance.

Q4: Prompt optimization and the contribution of CDE.

A: We compared the original prompt with two alternatives, each with and without CDE. Structured organizes instructions and historical observations into explicit blocks. Statistics adds the historical mean, standard deviation, latest value, and change over the observation window. The Original row corresponds to CrimeGPT and its w/o CDE ablation in Table 3(d).

Table B1: Prompt optimization on NYC All Crime 2022 (MAE/RMSE).

PromptWithout CDE ↓With CDE ↓
Original3.628 / 5.0363.537 / 4.921
Structured3.644 / 5.0453.545 / 4.940
Statistics3.627 / 5.0233.553 / 4.933

Differences across prompts are small, while CDE consistently improves both MAE and RMSE under all three formulations.

Q5: Clarification of the reverse threshold task in Eq. (22).

A: Thank you for identifying the conditioning error. The intended reverse threshold task is

\qquad r_k^-=s_k-s_{k+1}.$$ The sigmoid is a **CORN-inspired surrogate score** [1] constructed from adjacent count-token logits. In the ideal zero-loss limit, forward and reverse supervision yield logits that increase toward the target and decrease thereafter. **This ordering follows from the supervised adjacent comparisons.** We will correct Eq. (22) and clarify the surrogate interpretation in Appendix E. **Q6: Design of input numerals and output count tokens.** **A:** The two representations serve different purposes: - **Input numerals** retain the LLM's pretrained numerical and semantic knowledge while newly introduced token embeddings stabilize. They also accommodate 311 counts beyond the injected crime-count vocabulary, avoiding many additional tokens unused as output targets. - **Output count tokens** define the crime-count distribution used for the predictive-diffidence analysis in Section 4.2 and BiCORN supervision. **This design preserves familiar numerical inputs while providing an explicit, ordered output space for crime counts.** ## References [1] Shi, X., Cao, W., and Raschka, S. (2023). Deep neural networks for rank-consistent ordinal regression based on conditional probabilities. *Pattern Analysis and Applications*, 26, 941–955.

Thanks for your response.

One issue (W4: justification for embedding injection) is still open: The CDE gain remains unattributed. I suggest two additional experiments are suggested to make the ablation study more solid:

  • Replace the injected embedding with a random vector, with all other components unchanged. This experiment tests whether the improvement comes from the injected embedding or merely from the added trainable pathway.
  • Bypass the frozen LLM and feed the projected CDE output directly to DGRH, keeping the rest of the pipeline intact. This experiment isolates whether the frozen LLM contributes anything beyond conducting the vector to the trainable layers.
AuthorsOpenReview ↗

Followup Response

Thank you for these suggestions on attributing the contribution of CDE. We clarify the role of embedding injection and discuss the language-model pathway below.

Q1: Does CDE improve forecasting beyond an additional embedding-injection pathway?

A: Thank you for suggesting the random-vector control. Following your suggestion, we replace the projected CDE embedding with a random vector while retaining the same injection positions, number of injected vectors, frozen LLM, prompts, prediction head, and training objectives. We evaluate two variants: (i) a fixed random embedding, sampled once and held constant during training and inference. (ii) a learnable random embedding, initialized identically but updated during training. The learnable variant directly tests whether an additional trainable embedding alone accounts for the gain.

Table B2: Controls for embedding injection on NYC All Crime 2022.

ConfigurationMAE ↓RMSE ↓
W/o CDE, w/o injection3.6285.036
W/o CDE, fixed random embedding3.6164.966
W/o CDE, learnable random embedding3.5844.987
CrimeGPT: CDE-derived embedding3.5374.921

The fixed random embedding performs similarly to the no-injection baseline. The additional trainable parameters in the learnable random embedding provide some improvement, but the gains are modest and inconsistent across MAE and RMSE. CDE achieves the best performance on both metrics.

Embedding injection serves as the interface that conveys the information aggregated and aligned by CDE to the LLM; it is not itself the source of the performance gain. CDE and embedding injection together enable the LLM to integrate urban dynamics with regional semantic context, improving crime forecasting.

Q2: Contribution of the language model beyond direct CDE-to-regression prediction.

A: Thank you for suggesting this comparison. We compared direct CDE + MLP regression, two controls that replace the Qwen2.5-7B backbone with randomly initialized (trained jointly) and pretrained Qwen2.5-1.5B models while retaining CDE and regional prompts, and the full CrimeGPT model.

Table B3: Encoder and language-model controls on NYC All Crime 2022.

ModelMAE ↓RMSE ↓
CDE + MLP4.0355.713
CDE + Qwen2.5-1.5B (random initialization) + MLP3.7275.228
CDE + Qwen2.5-1.5B (pretrained) + MLP3.6625.082
CrimeGPT3.5374.921

These comparisons show that the LLM contributes beyond passing CDE representations to the prediction head in two ways:

  1. Larger parameter capacity. Performance improves progressively from direct CDE + MLP regression without an LLM, to the pretrained Qwen2.5-1.5B backbone, and then to the Qwen2.5-7B backbone in full CrimeGPT. This progression supports the benefit of greater model capacity for processing CDE representations and improving forecasting.
  2. Pretrained language knowledge. The pretrained 1.5B backbone outperforms its randomly initialized counterpart on both metrics. Removing contextual descriptions also degrades performance (Table 3(e)). These results support the contribution of pretrained language knowledge and regional semantics to crime forecasting.

The LLM therefore contributes both greater model capacity and pretrained language knowledge, enabling richer processing of CDE representations and their integration with regional semantic context for crime forecasting.

Reviewer_MaYSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CrimeGPT combines a spatiotemporal crime encoder with a frozen Qwen2.5-7B model. It adds count tokens, trains their logits with an ordinal loss called BiCORN, and predicts continuous counts from a regression head. The experiments cover four cities, including cross-year tests and transfer from New York to Chicago and Philadelphia.

Critical strengths

The fixed cross-year and zero-shot cross-city tests provide useful evidence

Critical weaknesses

Most of the MAE gain in Table 3c basically just comes from adding the regression head, and since there's no encoder-plus-regression baseline, it's unclear what the contribution of 7B LM is

Questions for authors
  1. how does this method actually stack up against just using the same crime encoder plus a regression head trained without the LLM, with the same inputs, same splits, and same count loss?
  2. Which BiCORN objective did you actually implement, Eq. (5), or the separately normalized forward and backward losses from Appendix E that put the target boundary in the forward mask?
1 comment

Thank you for recognizing the value of our cross-year and zero-shot cross-city evaluations.

In response to your comments, we will add encoder-based and language-model controls to examine the LLM's contribution, and clarify the BiCORN boundary assignment and normalization in the revised manuscript.

Q1: Contribution of the language model beyond encoder-based regression.

A: Thank you for suggesting this comparison. We evaluated CDE + MLP with and without auxiliary CE and BiCORN supervision, and included random and pretrained Qwen2.5-1.5B controls with CDE, regional prompts, and MLP regression. Both 1.5B configurations use L1 + CE + BiCORN.

Table A1: Encoder and language-model controls on NYC All Crime 2022.

ModelTraining objectiveMAE ↓RMSE ↓
CDE + MLPL14.0635.693
CDE + MLPL1 + CE + BiCORN4.0355.713
CDE + Qwen2.5-1.5B (random initialization) + MLPL1 + CE + BiCORN3.7275.228
CDE + Qwen2.5-1.5B (pretrained) + MLPL1 + CE + BiCORN3.6625.082
CrimeGPTL1 + CE + BiCORN3.5374.921

These controls examine two roles of the language model:

  1. Model capacity. The Transformer controls examine the benefit of a larger backbone for modeling relationships among heterogeneous urban signals, beyond direct CDE-to-MLP regression.
  2. Pretrained language knowledge. In the reported 1.5B comparison, the pretrained configuration obtains lower MAE/RMSE than the randomly initialized configuration. The contextual-prompt ablation in Table 3(e) further shows that removing contextual descriptions increases MAE/RMSE from 3.537/4.921 to 3.582/4.989. Together, these comparisons support using pretrained language representations and regional semantics alongside recent numerical dynamics.

The completed comparisons support the language-model pathway beyond the encoder and regression head, with regional semantics contributing to the forecast.

Q2: Clarification of the BiCORN objective and boundary assignment.

A: The current implementation follows Appendix E: the target boundary belongs to the forward term, and the two directions are normalized separately. Eq. (5) contains a boundary-assignment typo and leaves the denominators implicit. We will correct the main-text equation to match this definition.

Comments on the paper

1 comment

Dear Reviewers,

The rebuttal deadline is approaching. I am leaving this friendly reminder to encourage active discussion. Thank you for your valuable time and engagement!

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.