WALT introduces a method to align latent trajectories with pretrained driving world models, creating a compact generative trajectory space that preserves action-relevant semantics without altering the original model. The approach uses a dual-branch autoencoder to map raw waypoints into this latent space and transfers visual world knowledge into trajectory representations. Experiments on NAVSIM benchmarks show modest performance gains and a 30.5% reduction in planner FLOPs, indicating that maintaining world representations while extracting action-relevant information can improve trajectory planning efficiency.
By Mingkai Jia, Jiaxin Guo, Zhijian Shu, Jiawei Xu, Mingxiao Li, Jintao Cheng, Ping Tan, Wei Yin
The paper introduces TokenAdapt, a model‑agnostic tokenizer transplantation method that uses a hybrid heuristic to initialize new token embeddings, and a novel pre‑tokenization learning approach for multi‑word Supertokens to improve compression. TokenAdapt combines local subword decomposition and global semantic similarity to preserve semantics while reducing retraining needs. Empirical results show that TokenAdapt outperforms existing baselines such as Transtokenizer and ReTok, achieving lower perplexity ratios and significant compression gains.
By Shaurya Sharthak, Vinayak Pahalwan, Adithya Kamath, Adarsh Shirawalmath
The paper introduces PromptCCZSL, a framework that enables vision‑language models to continually learn new attributes, objects, and their unique compositions while avoiding forgetting. It uses a frozen VLM backbone with prompt‑based techniques, recency‑weighted multi‑teacher distillation, and several loss functions (CAL, OPL, IDL) to maintain prior knowledge and promote diverse, distinct embeddings. Experiments on UT‑Zappos and C‑GQA show significant performance gains over existing VLM‑based and non‑VLM baselines, establishing a new benchmark for continual compositional zero‑shot learning.
By Sauda Maryam, Sara Nadeem, Faisal Qureshi, Mohsen Ali
Forensic Twins introduces a Self‑Supervised Residual Learning (SSRL) framework that trains on real images only, using a frozen forensic residual extractor to generate two disjoint crops per image. The pretext task suppresses semantic content, focusing the model on the stationary fingerprint of the image acquisition pipeline, and achieves 56.61% accuracy in attributing AI‑generator sources, outperforming prior zero‑shot methods. When combined with an offline Gaussian Mixture Model, the approach reaches 97.99% AUC across 27 unseen AI generators, including GANs, diffusion models, and commercial systems.
By Javier Mu\~noz-Haro, Ruben Tolosana, Ruben Vera-Rodriguez, Aythami Morales, Julian Fierrez
SkillFlow is an open, multi-stage retrieval system that helps AI agents selectively load relevant skills from a large library of community-contributed SKILL.md definitions. The pipeline uses dense retrieval, two rounds of cross-encoder reranking, and LLM-based selection to balance recall and precision. Evaluations on SkillsBench and Terminal-Bench show that SkillFlow improves performance when high-quality skills are available, but retrieval alone does not help if the corpus lacks executable skills for the target domain.
By Fangzhou Li, Pagkratios Tagkopoulos, Ilias Tagkopoulos
The paper introduces Federated Martingale Posterior (FMP) sampling, a one‑shot, embarrassingly parallel method for federated Bayesian neural networks. FMP replaces the traditional prior–likelihood framework with a predictive distribution, allowing each client to upload only a small set of trainable data embeddings while the server performs predictive sampling centrally. Experiments demonstrate that FMP closely matches the performance of a centralized approach and achieves the lowest mean expected calibration error among evaluated one‑shot federated methods.
By Boning Zhang, Matteo Zecchin, Mingzhao Guo, Dongzhu Liu, Osvaldo Simeone
The paper investigates why retrieval‑based open‑ended evaluation fails in medical fact verification. By creating two detailed taxonomies—one for retrieval‑stage errors across five quality dimensions and another for verifier‑reasoning errors across six steps—the authors automatically label evidence quality and reasoning errors using an LLM‑as‑Judge pipeline. Their large‑scale stress tests across multiple retrieval methods and verifier models show that increasing model size, reasoning effort, source breadth, or medical fine‑tuning does not eliminate these failure modes, indicating fundamental limits of the retrieve‑then‑verify paradigm in open‑ended medical contexts.
By Heyuan Huang, Jirui Dai, Alexandra DeLucia, Sonal Joshi, Mahsa Yarmohammadi, Jie Gao, Bernal Jim\'enez Guti\'errez, Mark Dredze
The paper introduces an LLM-as-a-judge method to quantify perturbation strength for assessing self-consistency in LLM-generated explanations. It evaluates explanations from various LLMs under controlled perturbation strengths, comparing input and chain-of-thought (CoT) perturbations. Results show the proposed measure outperforms embedding- and probability-based approaches, and that input perturbations impact LLMs more than CoT perturbations, implying fair self-consistency judgments require consistent perturbation types.
By Phuong Q. Le, Kemal Kurniawan, Jey Han Lau
REALMS is a conversational system that provides real‑time, exact audience sizing for digital marketers. It uses embedding‑based vector search to retrieve relevant categorical attributes, an LLM‑powered NL2SQL pipeline for accurate query generation over complex nested schemas, and schema standardization for industry‑agnostic deployment. Evaluations on real enterprise data show high recall, accurate SQL execution, and low latency, enabling interactive audience insights that previously required hours.
By Haixu Ma, Aditya Bansal, Shubham Lohiya, Sumit Ranjan
The study investigates how different transcript compression methods affect large language model (LLM) detection of medical misinformation in Japanese YouTube videos. Four input designs were compared: full transcripts, LLM-generated summaries, RAPTOR-based retrieval‑augmented generation (RAG), and a Screening approach that extracts candidate medical sentences. Results show that full transcripts yield the best classification accuracy, while all compressed inputs increase false negatives, with summaries causing the largest performance drop and Screening performing best among compressed methods yet still missing many relevant sentences. Linguistic analysis indicates that compression reduces affective, social, temporal, cognitive, and conversational cues, and increases the prominence of institutional and technical terms, thereby making fake videos appear more coherent and authoritative.
By Yuya Wake, Sho Tsugawa, Toshiyuki Amagasa
CG-Probes are a method for extracting guardrail directions—Medical Urgency, Psychological Urgency, and Topic Sensitivity—from patient query embeddings. By treating each risk axis as a linear direction in the normalized embedding space, the approach uses difference‑in‑means probing and contrastive pairs generated from clustered Czech oncology search queries. Evaluated on 200 queries graded by oncologists, the probes perform comparably to open‑weight and frontier LLMs while offering lower latency and requiring only search logs, axis definitions, and black‑box embedding access.
By Marko \v{R}eh\'a\v{c}ek, V\'it\v{e}zslav Du\v{s}ek, Martin Rusinko, V\'it Nov\'a\v{c}ek
The paper introduces "stale‑document poisoning," a temporal alignment failure where outdated retrieval evidence causes retrieval‑augmented generation models to produce incorrect answers even when the model could answer correctly without retrieval. A benchmark of 317 verified knowledge reversals across medicine, law, software, and platform policy shows that outdated evidence flips 30–37% of Llama and Qwen answers, rising to 66–75% when models are explicitly instructed to trust the document. The study demonstrates that providing explicit validity information and a recency‑aware re‑ranker can substantially reduce poisoning, highlighting the need for models to assess whether retrieved evidence remains applicable.
By Md Shamim Ahmed, Lukas Galke Poech, Richard R\"{o}ttger
The paper evaluates whether music‑text models truly capture fine‑grained musical meaning by introducing attribute‑swap perturbations that exchange properties such as timbre or order between instruments in a caption. Four contrastive models and one large audio‑language model were tested to see if they would score higher on the original caption than on the perturbed one. The results show that none of the contrastive models reliably distinguish the captions, and the audio‑language model’s advantage stems mainly from language priors, indicating that CLAP scores behave like a bag‑of‑words and fail to reflect attribute bindings.
By Yuan-Chiao Cheng, Alexander Lerch
The paper challenges the assumption that improving cross‑lingual alignment automatically enhances cross‑lingual transfer. Using XLM‑R models aligned on token, sentence, and masked‑language‑modeling objectives across four language pairs, the authors evaluate zero‑shot transfer on part‑of‑speech tagging and sentence classification. They find that embedding‑based alignment metrics poorly predict downstream performance and that alignment and task gradients are often nearly orthogonal, especially when operating at different representational levels.
By Yana Veitsman, Yihong Liu, Hinrich Sch\"utze
arXiv:2609.31558v1 Announce Type: new
Abstract: Contrastive Language--Image Pretraining (CLIP) has emerged as a dominant vision backbone due to its strong transferability and zero-shot capabilities....
By Ahmed Abdelnaby, Mohamed Elmahallawy
This paper introduces a new method for predicting transmembrane protein topology by employing the graph neural network SchNet. The model is trained on the same dataset used for DeepTMHMM, using 5‑fold cross‑validation, and incorporates all atom‑level embeddings rather than just sequence or alpha‑carbon features. Results indicate that GNNs hold significant promise for topological predictions without relying on pre‑trained weights.
By Sitong Chen, Xiaopeng Mao
T‑RoPE introduces a time‑aware Rotary Position Embedding for sequential recommendation, replacing index‑only rotations with timestamp‑based angles, learnable temporal coefficients, multiscale frequency banks, shifted query alignment, and non‑stationary key rotation. The authors prove that standard RoPE is time‑translation invariant and cannot capture seasonal contexts, while T‑RoPE breaks this invariance while preserving the RoPE interface. Across five public benchmarks and an industrial‑scale e‑commerce dataset, T‑RoPE outperforms baselines by up to 130 % in HR@10 and delivers significant online lift in conversion and order count.
By Yang Liu, Noel Loo, Ali Khanafer, Shuying Sun, Akshay Soni, Zhong Wu, Linjun Yang
The paper introduces Hyperbolic Clinical Ontology Embeddings (HCOE), a method that transforms frozen BioBERT embeddings into a Poincaré ball to capture medical code hierarchies. HCOE employs ontology-guided contrastive learning and coarse‑to‑fine ontology‑path aggregation, leveraging ICD, CCS, and ATC hierarchies. Experiments on MIMIC‑IV demonstrate superior performance in clinical relation prediction, hierarchy transfer, and various predictive tasks such as mortality, readmission, medication recommendation, and rare drug prediction.
By Yixuan Li, Weihao Li, Ziyang Song
The paper introduces Hard-Origin Adaptively Softened Memory (HasMem), a memory system for large language model agents that combines frozen hard‑prompt embeddings with a controller, writer, reader, and global module to adaptively resize and re‑encode memory entries. On a reconstruction probe of 535 questions, HasMem achieves a lexical F1 of 95.3, outperforming the hard reference by 4.4 percentage points while maintaining 93.6% of the reference’s memory positions. Across six configurations with similar per‑question budgets, the system surpasses rule‑based re‑encoding by 8.0–23.6 exact‑match points, and on LongMemEval‑S it improves local lexical F1 from 3.4 to 8.9 and reduces answer negative log‑likelihood from 12.257 to 5.274.
By Zihong He, Junxiao Shen, Chen Liang, Hai-Ning Liang
LogicTree-RAG is a retrieval‑augmented generation framework that uses a hierarchical logic tree to guide the creation of long‑form patent drafts. Each node in the tree represents a technical element and is built through evidence‑guided recursive generation, while a hybrid traversal maps the tree into patent sections for balanced, controllable output. Experiments show that this logic‑centric approach improves content quality, language conformity, and token efficiency compared to strong LLM baselines.
By Jiaqi Zhu, Naili Xing, Hexiang Pan, Haotian Gao, Jianwei Yin, Xiaokui Xiao, Beng Chin Ooi