Retrieval pipelines, vector search, chunking and reranking: how models are grounded in a corpus instead of their weights.
Skin diseases represent a major global public health burden, yet machine learning tools developed to assist in their diagnosis suffer from two critical limitations: reliance on only one modality for d...
The paper introduces Temporally Decomposable Image Representations (TDIR), a learning algorithm that splits historical photographs into orthogonal date and content components. It establishes theoretical conditions for this decomposition, quantifies errors when conditions are partially met, and demonstrates that temporal and categorical subspaces naturally become orthogonal during joint optimization. TDIR supports transitive operations, allowing temporal information from one image to be transferred to another without label supervision, and is validated on composed image retrieval tasks that combine object content and time period queries.
The paper introduces WONDER, a radio world‑model‑based negotiation framework designed to optimize multi‑UAV coverage for rapid post‑disaster wireless restoration. WONDER employs a Joint‑Embedding Predictive Architecture to forecast the incremental radio impact of candidate UAV trajectories and uses multi‑round negotiation to sequentially commit trajectories while updating the context. Experiments in the RadioDynamics simulation environment demonstrate that WONDER outperforms six other methods, achieving a balanced score of 0.870 and a 0.162 coverage advantage over STACCA while preserving full UAV connectivity.
The paper introduces ADAPT, an adversarial disentangled prompt tuning framework designed to improve the robustness of vision‑language models. ADAPT employs a dual‑prompt strategy: a target prompt learns robust features while a set of decoy prompts capture pseudo‑robust, non‑generalizable shortcuts. By enforcing orthogonality between target and decoy prompts, the method mitigates robust generalization overfitting and provides a theoretical error bound for unseen classes, leading to significant empirical robustness gains.
The paper investigates Retrieval-Augmented Generation fine‑tuning (RAG‑SFT) for generating requirements documents in electronics engineering, comparing two 7B models trained with different data strategies. It introduces a claim‑based evaluation pipeline, C‑FEX, and a new metric, Parametric Knowledge Precision (PKP), to assess factuality of model‑generated claims. Results show that fine‑tuned 7B models can match or surpass a 72B baseline, but standard metrics may mislead, and fine‑tuning reduces hallucination by encouraging more reliable use of parametric knowledge.
The paper introduces a multi‑view contrastive learning framework that creates low‑dimensional spatial embeddings by combining satellite imagery and OpenStreetMap data across Europe. These embeddings align with coordinate‑based encodings, allowing any dataset with latitude‑longitude pairs to be enriched with meaningful spatial features without needing the original spatial inputs. In case studies on French real‑estate prices and Belgian flood claim counts, models using the embeddings outperform those using raw coordinates, improving predictive accuracy and territorial risk classification while offering explainable spatial effects.
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.
The paper introduces Fast MCTS, a Monte Carlo Tree Search approach for knowledge base question answering that replaces costly terminal rollouts with an information gain reward for intermediate states. This reward is computed using a question‑conditioned PPL‑ratio proxy over sanitized interaction histories, leveraging an open‑source instruction LLM without extra training. Experiments on four KBQA benchmarks demonstrate that Fast MCTS consistently outperforms linear baselines and improves the accuracy‑cost trade‑off compared to classic rollout‑based MCTS.
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.
The paper introduces ORPA, a framework that adds a lightweight, feedback-conditioned module to a pretrained robotic manipulation policy, enabling real‑time residual adjustments in joint space without retraining the base policy. ORPA allows immediate correction of execution errors and distribution shifts, improving success rates and recovery on precision‑sensitive tasks compared to baseline policies and rule‑based inverse kinematics. The method is evaluated on the ALOHA platform, showing its effectiveness in real‑time deployment scenarios.
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.
MoRAX is a lightweight framework that augments geospatial foundation model embeddings with functional structure derived from human mobility data. By incorporating mobility flows, MoRAX preserves the coverage and consistency of existing geospatial models while adding information about functional connectivity among urban regions, enabling zero‑shot deployment in unseen cities. Experiments across four cities in two countries show that the MoRAX teacher model outperforms baseline geospatial models on eight socioeconomic and environmental prediction tasks, and the student model—without direct mobility input—approaches the teacher’s performance.
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.
The paper introduces Action-Contrastive Masked Transition Modeling (AC‑MTM), a method that replaces the Gaussian regularizer used in Joint‑Embedding Predictive Architectures (JEPAs) with a contrastive inverse‑dynamics head. AC‑MTM trains a forward latent‑prediction model while an auxiliary inverse‑dynamics task forces the encoder to distinguish actions from latent transitions, preventing collapse without requiring a target network or reconstruction loss. Experiments on pixel‑control and multi‑object visual tasks show that AC‑MTM matches or surpasses the performance of the Gaussian‑based SIGReg regularizer, achieving up to 20–24 point improvements on the OGBench Visual Scene benchmark.
WIP: LLM Odyssey is an open‑source, browser‑based serious gaming platform that offers 13 interactive games to teach Large Language Model engineering concepts such as tokenization, transformer architecture, prompt engineering, RAG, and production deployment. The platform is organized into three learning tiers—Cognitive Core, Systems Forge, and Foundry Arena—aligned with Bloom’s revised taxonomy, and each game employs five pedagogical strategies including immediate feedback, scaffolded hints, progressive difficulty, worked examples, and authentic production scenarios. An initial deployment at a Canadian college in Winter 2026 confirmed functional requirements, highlighted the need for adaptive difficulty, and led to the design of a formal mixed‑methods evaluation protocol for future studies.
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.
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.
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.
GxP-Agent is a multi‑agent system that transforms clinical trial protocols into CDISC‑compliant datasets by encoding the regulatory workflow as a directed acyclic graph (DAG). Each node in the DAG represents a domain‑specific task executed by a worker agent with specialized skill context, validation gates, and conditional retry logic. On the CDISC‑Bench benchmark, GxP-Agent with Claude Sonnet 4.6 achieved a perfect 100 % structural match for 49 variables across 254 records, outperforming single‑agent and flat multi‑agent baselines and enabling weaker models like GPT‑4.1 to reach 59.2 % under the same DAG.
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.