Retrieval pipelines, vector search, chunking and reranking: how models are grounded in a corpus instead of their weights.
arXiv:2609.00293v1 Announce Type: new Abstract: We investigate how vision-language models (VLMs) handle context-memory conflicts; that is, situations in which the model is given information in contex...
The paper introduces a method for generating multilingual reasoning traces for medical question answering using large language models. It creates 500,000 reasoning traces in English, Italian, and Spanish by retrieving medical information from Wikipedia and applies them to MedQA and MedMCQA datasets extended into Italian and Spanish. The approach improves performance in both few‑shot in‑context learning and supervised fine‑tuning, achieving state‑of‑the‑art results for 8B‑parameter LLMs and releasing all resources for further research.
The paper "Evaluating Style-Personalized Text Generation: Challenges and Directions" examines the difficulties of assessing text that is tailored to individual users’ styles. It critiques common metrics such as BLEU, embeddings, and LLM-as-judges, and introduces a style discrimination benchmark covering domain discrimination, authorship attribution, and LLM-generated personalized versus non-personalized discrimination across eight writing tasks. The study finds that ensembles of diverse evaluation metrics outperform single-evaluator approaches and offers guidance for reliable assessment of style-personalized generation.
arXiv:2609.00454v1 Announce Type: new Abstract: Pretrained transformer-based language models achieve strong performance across a wide range of NLP tasks but remain limited in encoding geo-locational...
The paper introduces X‑SG$^2$S, a feed‑forward framework that embeds 1D to 3D watermarks into 3D Gaussian Splatting (3DGS) scenes without altering the original rendering pipeline. It splits watermark messages into patches, uses a self‑adaptive gate to choose injection points, and an XD injection head to embed multi‑modal messages into sorted 3DGS points. A learnable gate and XD‑extraction heads recover the hidden messages, achieving robust watermarking with minimal interference to scene fidelity.
arXiv:2603.02974v2 Announce Type: replace Abstract: DINO models provide rich patch-level representations that have recently enabled strong performance in unsupervised anomaly detection (UAD). Most ex...
The paper introduces a hybrid insider threat detection framework that combines multi-agent simulation, layered SIEM correlation, trust‑adaptive thresholds, behavioral and communication forensics, and Theory‑of‑Mind reasoning. It evaluates four variants—Layered SIEM‑Core, Cognitive‑Enriched SIEM, Evidence‑Gated SIEM, and an Enron‑calibrated version—showing progressively higher actor‑level F1 scores and reduced false positives, especially with evidence gating. Domain‑shift tests reveal that an Enron‑trained email classifier does not transfer to other domains, but fine‑tuning achieves high F1, and scalability tests confirm stable performance up to 1,000 agents.
MemoryWalker addresses the conditioning problem that arises when training agents with compressed context during rollout. It introduces two exact, gradient‑equivalent corrections—LogitTree, a segmented K‑forward traversal, and a packed 4D attention mask—alongside SDCC, a self‑distillation method that reduces the train‑deployment gap by minimizing KL divergence at each eviction. Experiments on seven web‑search benchmarks show that SDCC significantly lowers logit drift and boosts rollout rewards compared to naive training.
arXiv:2609.00396v1 Announce Type: new Abstract: Histopathological whole slide images (WSIs) are central to cancer diagnosis, but their gigapixel scale, tissue heterogeneity, weak slide-level supervis...
arXiv:2609.01059v1 Announce Type: new Abstract: As Vision-Language Models (VLMs) tackle dynamic 3D spatial reasoning, ego-motion perception becomes essential to resolve monocular scale ambiguity. How...
The paper introduces CRAFT, a data‑centric fine‑tuning approach that aligns small language models (SLMs) for pre‑hoc reasoning in AI‑native 6G radio access networks (RAN). By automatically generating verified (input, trace, label) triplets and fine‑tuning with low‑rank adaptation, CRAFT achieves high accuracy and F1 scores on TRACTOR and IC xApp datasets while avoiding parse failures that plague RL methods like GRPO. It also reduces energy consumption by 59% compared to GRPO baselines, offering a more sustainable path to auditable AI in 6G RAN.
The paper introduces Drift‑Aware Sparse Routing (DRS), a method for routing requests in a multi‑model language service while respecting compute, latency, memory, or cost budgets. DRS estimates reward and resource use from a rolling audit window, routes using pessimistic reward and optimistic cost estimates, updates resource shadow prices online, and applies a hard meter before commitment. The authors provide theoretical regret bounds that separate control from statistics, showing how the method adapts to non‑stationary prompt distributions and model changes.
The paper introduces a benchmark for topic matching in real-world ASR transcripts from contact centers, where noisy, punctuation‑free speech data must be classified into predefined topics. It presents a human‑annotated dataset of topic‑utterance judgments and evaluates three matcher types—regex, zero‑shot sentence embeddings, and Gemini‑based LLMs—using two topic representations: keyphrases and natural language descriptions. Experiments show that lightweight LLM matchers outperform the other methods, especially when natural language descriptions are used.
arXiv:2609.00591v1 Announce Type: new Abstract: An image may be worth a thousand words, but most captioning models describe it in only a few. Modern vision-language models produce fluent high-level c...
EvoFlint is an evolutionary atlas that maps multi‑turn vulnerabilities in large language models by treating attack discovery as a search problem rather than a generation task. It uses evolutionary quality‑diversity search to evolve phased conversation plans, employing Pareto fitness for success rate and severity, novelty search for diversity, and a generation‑level memory to incorporate model insights. The resulting risk‑indexed archive, tested on HarmBench, shows high attack success rates across models such as Claude Sonnet, GPT‑5, and Qwen3, revealing which harm categories each model’s safety training covers or misses.
The paper introduces embedded Conditional Independence Tests (eCITs) to assess whether large language model (LLM) outputs contain additional information about an attribute beyond the source text. By embedding the source and output texts and applying existing CITs to these representations, the authors show that the null hypothesis can transfer from the original variables to their embeddings under sufficient conditions. Using a semi‑synthetic simulation design, they evaluate type I error control and power, and apply eCITs to German Parliament speeches, finding that LLM summaries reveal speaker faction and gender information beyond the original speech content.
The paper introduces TrustPropRAG, a method that builds a document relation graph to propagate reliability signals across documents in retrieval‑augmented generation systems. By anchoring this propagation with a limited set of human feedback on document reliability, the approach estimates trust scores through an optimization problem that incorporates pairwise relations and user input. These trust scores are then used to select more reliable documents and generate answers that are aware of document trustworthiness, leading to improved retrieval quality and exact match performance compared to baselines, even with sparse or noisy feedback.
arXiv:2609.01564v1 Announce Type: cross Abstract: Large language models (LLMs) struggle to classify text into taxonomies with many semantically similar labels, as the distinctions are domain-specific...
APEX-EM is a non‑parametric experience memory that stores full procedural‑episodic traces in a typed Procedural Knowledge Graph and retrieves them via semantic search, structural‑signature matching, and graph traversal. It uses a Plan‑Retrieve‑Generate‑Iterate‑Ingest workflow to produce, quality‑gate, and commit experiences, indexing both successes and failures so the agent learns what to reuse and what to avoid. Evaluations on five benchmarks with a shared GPT‑4o backbone show significant performance gains, such as +7.6 pp on BigCodeBench transfer and +1.4 pp on Lifelong Agent Bench, demonstrating that the memory adds to model capability rather than replacing it.
arXiv:2609.00023v1 Announce Type: cross Abstract: In this paper, we introduce ES-AHD, a novel framework that fundamentally integrates Evolution Strategy (ES) into Large Language Model (LLM)-driven Au...