The 10th AI City Challenge, held alongside ECCV 2026, celebrates a decade of benchmarking for intelligent transportation, smart cities, and physical AI. Since its 2017 inception focused on vehicle detection, classification, and tracking, the challenge has expanded into a comprehensive benchmark suite covering multi‑camera perception, multimodal reasoning, synthetic‑to‑real learning, generative forecasting, and privacy‑preserving evaluation. The 2026 edition saw 325 registered teams from 26 countries, with six main tracks—spanning multi‑camera 3D perception, transportation safety captioning and VQA, traffic anomaly reasoning, text‑based person anomaly search, generative traffic video forecasting, and cross‑city object detection—plus two out‑of‑domain leaderboards for fisheye traffic‑violation understanding and pedestrian situated‑intent VQA.
By Zheng Tang, Shuo Wang, David C. Anastasiu, Ming-Ching Chang, Anuj Sharma, Quan Kong, Munkhjargal Gochoo, Jun-Wei Hsieh, Tomasz Kornuta, Zhedong Zheng, Renran Tian, Judah Goldfeder, Fulgencio Navarro, Yuxing Wang, Yizhou Wang, Sameer Satish Pusegaonkar, Anqi Li, Nalin Dadhich, Ridham Kachhadiya, Dhanishtha Patil, Haoquan Liang, Jiajun Li, Han Zhang, Yilin Zhao, Zaid Pervaiz Bhat, Shuyu Yang, Ashutosh Kumar, Rong Wang, Rafael Martin Nieto, Peter Christiansen, Ahmed Abduljawad, Mohanrasu Shanmugam, Nadeem Shaik, Sujit Biswas, Xunlei Wu, Vidya Murali, Rama Chellappa
The study evaluates four pretrained molecular language models on six virtual libraries covering drug discovery, organic materials, and catalysis. It finds that native embeddings vary widely in performance, while molecular fingerprints remain consistently strong. Fine‑tuning the models on library‑specific data markedly improves sample efficiency, with several adapted encoders outperforming others across all tasks.
By Henrik Wille, Luis-Finley Sch\"utz, Felix Strieth-Kalthoff
Sparse Coverage is an unsupervised semantic retrieval framework designed for patent prior‑art search. It maps local span embeddings to a sparse vocabulary of embedding‑space centers chosen via a coverage‑oriented k‑center objective, allowing spans to activate nearby centers and produce sparse representations that work with inverted‑index retrieval. Experiments on CLEF‑IP 2013 demonstrate that Sparse Coverage matches or surpasses dense patent encoders in document‑level recall while remaining competitive at the passage level, making it an effective first‑stage retrieval approach for patent search.
By You Zuo (ALMAnaCH), Kim Gerdes (LISN, Qatent, STL), \'Eric de la Clergerie (ALMAnaCH), Beno\^it Sagot (ALMAnaCH)
The paper investigates the finite‑iteration behavior of exact asynchronous recursions used in categorical distributional temporal‑difference (TD) learning. It analyzes both scalar categorical TD in the Cramér geometry and multivariate signed‑categorical TD in the maximum mean discrepancy geometry, showing that these methods can be viewed as single‑state stochastic‑approximation recursions that contract in a block‑supremum norm. The authors develop a restricted‑domain theory, derive discounted bounds under i.i.d. and Markovian sampling, and extend the analysis to undiscounted fixed‑horizon policy evaluation with horizon‑stacked categorical methods under episodic sampling, thereby providing a unified non‑asymptotic analysis across various settings.
By Ege C. Kaya, Abolfazl Hashemi
The paper introduces MASS, a hierarchical data selection method that first groups data using low‑dimensional manifold coordinates learned by a dense autoencoder, then applies a TopK sparse autoencoder for quality‑aware feature coverage within each group. This approach addresses the shortcomings of traditional diversity metrics that mix semantic, supervisory, and noise signals. Experiments on Vision Flan and LLaVA‑CoT demonstrate that MASS outperforms existing baselines across various budgets and can match or exceed full‑data training with only a small subset.
By Peng Sun, Yi Yang, Antong Zhang, Chunxiao Li, Yanbo Wang, Dianbo Liu, xin chen, Kai Yu, Lu Chen, Tianfan Fu
FairNVT is a lightweight debiasing framework that injects calibrated Gaussian noise into sensitive embeddings learned by adapters for pretrained transformer-based encoders. By reducing sensitive-attribute leakage through orthogonality constraints and fairness regularization, it improves fairness metrics such as demographic parity difference and equalized odds while maintaining competitive task performance across vision and language datasets.
By Qiaoyue Tang, Sepidehsadat Hosseini, Mengyao Zhai, Thibaut Durand, Greg Mori
The paper introduces a differentiable evaluation engine for linear temporal logic (LTL) that is algebra‑generic and suitable for training soft‑valued systems such as neural policies and adaptive controllers. It presents an executable specification of the algebras it can accept, implements several algebras, and audits their forward and backward behavior, all within the PyTorch library telos.
By Konstantinos Kogkalidis
The paper introduces LSem2Vec, a two‑stage method that first uses a large language model to extract source code semantics and then applies a sentence embedding model to produce vector representations. This approach removes the need for task‑specific training or fine‑tuning, addressing errors in LLM outputs. Experiments on three datasets across multiple programming languages show that LSem2Vec outperforms five state‑of‑the‑art unsupervised methods.
By Zixiang Xian, Chenhui Cui, Rubing Huang, Chunrong Fang, Zhenyu Chen
The paper introduces GADR, a multi‑agent, self‑correcting workflow that extracts architectural decisions from raw meeting transcriptions and produces Nygard‑formatted ADR drafts. A feasibility study using five real project transcripts, expert reviews by four senior architects, and evaluations by fifteen students shows that GADR captures most expert‑identified decisions and yields drafts that participants find clear and useful, outperforming zero‑shot and few‑shot baselines in stability and structural adherence. The study also highlights a trade‑off: RAG‑based enrichment can deepen ADR content but may introduce transcript‑unfaithful material, raising open questions about traceability in automated architectural documentation.
By Lucas Daniel Costa da Silva, Kiev Gama
CoAL‑RAG is a complexity‑aware legal retrieval‑augmented generation method that adapts its retrieval strategy based on a multi‑dimensional evaluation of question essence and retrieval consistency. It quantifies reasoning demand from the logical structure of a question and uses the discrepancy between semantic and keyword retrieval to gauge problem complexity, thereby selecting the most suitable retrieval approach and filtering context dynamically. Experiments show that CoAL‑RAG outperforms baseline models on Chinese legal benchmarks (SocialLawQA, LawBench) with a 42.5% BLEU improvement and 3.6× ROUGE‑L, while also achieving strong cross‑jurisdictional performance on English datasets (LexGLUE, CaseHold).
By Jin Su, Zhuofeng Zhao, Huanhuan Wang, Hao Chen
The paper investigates using a small language model (SBERT) for invoice categorisation, showing that fine‑tuning on a single GPU yields 0.96 accuracy and 0.9 F1 with about 100 client‑specific invoices. It analyses the embedding geometry, finding that the sentence‑embedding space is globally anisotropic but locally isotropic, with clusters strongly linked to vendor identity. The study demonstrates that an in‑house SLM can outperform zero‑shot LLMs and vendor baselines while improving cost, security, and interpretability.
By Emma Ceccherini, Daniel Lawson, Anjulika Salhan
The paper proposes a Learnware-based framework for deploying scene‑specific CSI feedback models in 6G systems. A centralized AI data center maintains a catalog of pre‑trained models, each tagged with semantic and statistical specifications. Base stations retrieve the most relevant model using only statistical fingerprints, which reduces data privacy risks, lowers retrieval latency, and cuts fine‑tuning effort, achieving up to 57.7% performance gains over a general model.
By Xiangyi Li, Jiajia Guo, Chao-Kai Wen, Xin Geng, Shi Jin, Zhi-Hua Zhou
The paper introduces a spatio‑temporal traffic forecasting framework that fuses Graph Neural Networks with semantic knowledge from general-purpose knowledge graphs such as Wikidata. By generating embeddings that capture relationships like nearby points of interest, administrative hierarchies, and functional roles of locations, the framework creates additional adjacency matrices that enrich the sensor graph beyond physical connectivity. Experiments with established forecasting methods demonstrate that this external knowledge improves prediction accuracy and offers a path toward better interpretability.
By Mattis thor Straten, Yannick Wolker, Steffen Strohm, Prathvish Mithare, Ralf Krestel, Matthias Renz
Intent-Driven Dynamic Chunking (IDC) segments documents by predicting user queries with a Large Language Model and then applying dynamic programming to find optimal chunk boundaries. This method outperforms traditional fixed-length or coherence-based segmentation on five out of six question-answering datasets, improving top-1 retrieval accuracy by 5% to 67% and reducing the number of chunks by 40–60% while maintaining 93–100% answer coverage. IDC demonstrates that aligning document structure with anticipated information needs can significantly boost retrieval performance for long and heterogeneous documents.
By Christos Koutsiaris
The paper introduces DEPT, a method that trains a single decoder-only large language model to both expand queries and encode documents for retrieval. By preserving document embeddings close to their initial cached values while allowing gradients to flow through the generator, DEPT stabilizes retrieval targets and enables efficient index reuse and online hard‑negative mining. Experiments on the BEIR benchmark with Qwen3‑4B‑Instruct‑2507 and LLaMA‑3.2‑3B‑Instruct show that DEPT outperforms training‑free, independently trained, and staged unified baselines, with ablations confirming the benefits of preservation, whitening, end‑to‑end expansion training, and online negatives.
By Jingyuan Wang, Richong Zhang, Zhijie Nie, Mingxin Li, Yanzhao Zhang
DMT‑Dens is a parametric manifold‑visualization technique that uses a latent‑token Transformer encoder to produce two‑dimensional embeddings of high‑dimensional biological data. It preserves sampling density by aligning rank‑based manifold structures and optimizing a Pearson‑correlation loss on k‑nearest‑neighbor log‑radius estimates. Benchmark tests show that DMT‑Dens maintains density fidelity while achieving competitive label separability on biological datasets.
By Ruizhe Wang, Yixuan Dong, Bolin Yang, Bingo Wing-Kuen Ling, Fuji Yang, Zelin Zang
The paper presents a perception-action pipeline for robotic pick‑and‑place in convenience stores, using annotation‑guided visual prompting to identify pickable objects and placement locations via bounding boxes. It replaces traditional step‑by‑step planning with Action Chunking with Transformers (ACT), an imitation learning algorithm that predicts chunked action sequences from human demonstrations. The system is evaluated on success rate and visual analysis of grasping behavior, showing improved grasp accuracy and adaptability in retail environments.
By Muhammad A. Muttaqien, Tomohiro Motoda, Ryo Hanai, Yukiyasu Domae
DOW-KE is an anchor‑free knowledge‑editing method that directly optimizes weight updates across multiple layers by backpropagating the final editing objective through the entire model. Unlike traditional locate‑then‑edit pipelines that optimize intermediate activations and then apply local weight updates—leading to a closure gap—DOW-KE ensures that what is optimized is exactly what is deployed, incorporating preservation constraints within the update parameterization. Experiments on two datasets and three models show that DOW-KE achieves the highest overall Score and neighborhood Specificity in five of six evaluated settings.
By Ran Chen, Junbo Zhang, Qianli Zhou, Xinyang Deng, Wen Jiang
The paper shows that language models trained on natural text develop number representations that exhibit periodic features with dominant periods at T = 2, 5, 10. It identifies a two‑tiered hierarchy: all models learn Fourier‑domain spikes at these periods, but only some acquire geometrically separable features that allow linear classification of numbers modulo T. The study demonstrates that data, architecture, optimizer, and tokenizer influence whether these separable features emerge, and that models can learn them either from co‑occurrence signals in language or from multi‑token addition tasks, illustrating convergent evolution across diverse models.
By Deqing Fu, Tianyi Zhou, Mikhail Belkin, Vatsal Sharan, Robin Jia
The paper argues that measuring diversity in AI-generated content using a single scalar score is inherently ambiguous and often misleading. It reviews existing diversity metrics, demonstrates their limitations through axiomatic and empirical analyses, and introduces diversity profiles—curve-valued, condition-aware summaries that evaluate diversity across a range of thresholds, scales, exponents, or orders. These profiles reveal whether comparisons are robust across resolutions or depend on arbitrary parameter choices, offering a more transparent framework for generative AI evaluation.
By Xiuyuan Hu, Xuege Hou, Guoqing Liu, Yang Zhao, Jieran Li, Dongbiao Sun, Jos\'e Miguel Hern\'andez-Lobato, Hao Zhang, Xue Liu