Breast MRI provides high-resolution volumetric imaging critical for tumor assessment and treatment planning, yet manual interpretation of 3D scans remains labor-intensive and subjective. While AI-powered tools hold promise for accelerating medical image analysis, adoption of commercial medical AI products remains limited in low- and middle-income countries due to high license costs, proprietary software, and infrastructure demands. In this work, we investigate whether the Segment Anything Model 2 (SAM2) can be adapted for low-cost, minimal-input 3D tumor segmentation in breast MRI. Using a single bounding box annotation on one slice, we propagate segmentation predictions across the 3D volume using three different slice-wise tracking strategies: top-to-bottom, bottom-to-top, and center-outward. We evaluate these strategies across a large cohort of patients and find that center-outward propagation yields the most consistent and accurate segmentations. Despite being a zero-shot model not trained for volumetric medical data, SAM2 achieves strong segmentation performance under minimal supervision. We further analyze how segmentation performance relates to tumor size, location, and shape, identifying key failure modes. Our results suggest that general-purpose foundation models such as SAM2 can support 3D medical image analysis with minimal supervision, offering an accessible and affordable alternative for resource-constrained settings.

Contextual hallucinations -- statements unsupported by given context -- remain a significant challenge in AI. We demonstrate a practical interpretability insight: a generator-agnostic observer model detects hallucinations via a single forward pass and a linear probe on its residual stream. This probe isolates a single, transferable linear direction separating hallucinated from faithful text, outperforming baselines by 5-27 points and showing robust mid-layer performance across Gemma-2 models (2B to 27B). Gradient-times-activation localises this signal to sparse, late-layer MLP activity. Critically, manipulating this direction causally steers generator hallucination rates, proving its actionability. Our results offer novel evidence of internal, low-dimensional hallucination tracking linked to specific MLP sub-circuits, exploitable for detection and mitigation. We release the 2000-example ContraTales benchmark for realistic assessment of such solutions.

Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic-based steering vectors. To properly evaluate MTM topics against word-list-based approaches, we propose \textit{topic judge}, an LLM-based pairwise comparison evaluation framework. Across five datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective steering of LLM outputs.

Understanding complex multimodal documents remains challenging due to their structural inconsistencies and limited training data availability. We introduce \textit{DocsRay}, a training-free document understanding system that integrates pseudo Table of Contents (TOC) generation with hierarchical Retrieval-Augmented Generation (RAG). Our approach leverages multimodal Large Language Models' (LLMs) native capabilities to seamlessly process documents containing diverse elements such as text, images, charts, and tables without requiring specialized models or additional training. DocsRay's framework synergistically combines three key techniques: (1) a semantic structuring module using prompt-based LLM interactions to generate a hierarchical pseudo-TOC, (2) zero-shot multimodal analysis that converts diverse document elements into unified, text-centric representations using the inherent capabilities of multimodal LLMs, and (3) an efficient two-stage hierarchical retrieval system that reduces retrieval complexity from $O(N)$ to $O(S + k_1 \cdot N_s)$. Evaluated on documents averaging 49.4 pages and 20,971 textual tokens, DocsRay reduced query latency from 3.89 to 2.12 seconds, achieving a 45% efficiency improvement. On the MMLongBench-Doc benchmark, DocsRay-Pro attains an accuracy of 64.7%, substantially surpassing previous state-of-the-art results.

Large Language Models exhibit powerful few-shot in-context learning (ICL) capabilities, but the performance is highly sensitive to provided examples. Recent research has focused on retrieving corresponding examples for each input query, not only enhancing the efficiency and scalability of the learning process but also mitigating inherent biases in manual example selection. However, these studies have primarily emphasized leveraging Positive samples while overlooking the additional information within Negative samples for contextual learning. We propose a novel method that utilizes Negative samples to better select Positive sample examples, thereby enhancing the performance of few-shot ICL. Initially, we construct Positive and Negative sample corpora based on Zero-Shot-Cot. Then, during inference, we employ a semantic similarity-based approach to select the most similar examples from both the Positive and Negative corpora for a given query. Subsequently, we further retrieve Positive examples from the Positive sample corpus based on semantic similarity to the Negative examples, then concatenating them with the previously selected Positive examples to serve as ICL demonstrations. Experimental results demonstrate that our approach surpasses methods solely relying on the most similar positive examples for context, validating that the additional information in negative samples aids in enhancing ICL performance through improved Positive sample selection.

Multimodal IoT systems coordinate diverse IoT devices to deliver human-centered services. The ability to incorporate new IoT devices under the management of a centralized platform is an essential requirement. However, it requires significant human expertise and effort to program the complex IoT integration code that enables the platform to understand and control the device functions. Therefore, we propose AutoBridge to automate IoT integration code generation. Specifically, AutoBridge adopts a divide-and-conquer strategy: it first generates device control logic by progressively retrieving device-specific knowledge, then synthesizes platformcompliant integration code using platform-specific knowledge. To ensure correctness, AutoBridge features a multi-stage debugging pipeline, including an automated debugger for virtual IoT device testing and an interactive hardware-in-the-loop debugger that requires only binary user feedback (yes and no) for real-device verification. We evaluate AutoBridge on a benchmark of 34 IoT devices across two open-source IoT platforms. The results demonstrate that AutoBridge can achieves an average success rate of 93.87% and an average function coverage of 94.87%, without any human involvement. With minimal binary yes and no feedback from users, the code is then revised to reach 100% function coverage. A user study with 15 participants further shows that AutoBridge outperforms expert programmers by 50% to 80% in code accuracy, even when the programmers are allowed to use commercial code LLMs.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, with different models excelling in distinct domains and specific abilities. Effectively combining the predictions of multiple LLMs is crucial for enhancing system robustness and performance. However, existing ensemble methods often rely on simple techniques like voting or logits ensembling, which overlook the varying confidence and reliability of models in different contexts. In this work, we propose LENS (Learning ENsemble confidence from Neural States), a novel approach that learns to estimate model confidence by analyzing internal representations. For each LLM, we train a lightweight linear confidence predictor that leverages layer-wise hidden states and normalized probabilities as inputs. This allows for more nuanced weighting of model predictions based on their context-dependent reliability. Our method does not require modifying the model parameters and requires negligible additional computation. Experimental results on multiple-choice and boolean question-answering tasks demonstrate that LENS outperforms traditional ensemble methods by a substantial margin. Our findings suggest that internal representations provide valuable signals for determining model confidence and can be effectively leveraged for ensemble learning.

Despite advances in large language model (LLM)-based natural language interfaces for databases, scaling to enterprise-level data catalogs remains an under-explored challenge. Prior works addressing this challenge rely on domain-specific fine-tuning - complicating deployment - and fail to leverage important semantic context contained within database metadata. To address these limitations, we introduce a component-based retrieval architecture that decomposes database schemas and metadata into discrete semantic units, each separately indexed for targeted retrieval. Our approach prioritizes effective table identification while leveraging column-level information, ensuring the total number of retrieved tables remains within a manageable context budget. Experiments demonstrate that our method maintains high recall and accuracy, with our system outperforming baselines over massive databases with varying structure and available metadata. Our solution enables practical text-to-SQL systems deployable across diverse enterprise settings without specialized fine-tuning, addressing a critical scalability gap in natural language database interfaces.

This research work demonstrate s that current AI systems fail catastrophically on auditory tasks that humans perform effortlessly. Drawing inspiration from Moravec's paradox ( i.e., tasks simple for humans often prove difficult for machines, and vice vers a), we introduce a n auditory Turing test comprising 917 challenges across seven categories: overlapping speech, speech in noise, temporal distortion, spatial audio, coffee - shop noise, phone distortion, and perceptual illusions. Our evaluation of state - of - the - art audio models including GPT - 4's audio capabilities and OpenAI's Whisper reveals a striking failure rate exceeding 93%, with even the best - performing model achieving only 6.9% accuracy on tasks that humans solve d at 7.5 times higher success (52%). These results expose focusing failures in how AI systems process complex auditory scenes, particularly in selective attention, noise robustness, and contextual adaptation. Our benchmark not only quantifies the human - machine auditory gap but also provides insights into why these failures occur, su ggesting that current architectures lack fundamental mechanisms for human - like auditory scene analysis. The traditional design of audio CAPTCHAs highlight s common filters that humans evolved but machines fail to select in multimodal language models. This work establishes a diagnostic framework for measuring progress toward human - level machine listening and highlights the need for novel approaches integrating selective attention, physics - based audio understanding, and context - aware perception into mult imodal AI systems. Artificial intelligence has made great strides in language understanding and multimodal perception, yet machines still struggle with basic auditory tasks that humans perform successfully [1 - 20] . A striking example is the cocktail party effect [21 - 22 ] - the human ability to focus on a single conversation in a noisy room - which remains a formidable challenge for AI.

Large language models (LLMs) have changed the reality of how software is produced. Within the wider software engineering community, among many other purposes, they are explored for code generation use cases from different types of input. In this work, we present an exploratory study to investigate the use of LLMs for generating smart contract code from business process descriptions, an idea that has emerged in recent literature to overcome the limitations of traditional rule-based code generation approaches. However, current LLM-based work evaluates generated code on small samples, relying on manual inspection, or testing whether code compiles but ignoring correct execution. With this work, we introduce an automated evaluation framework and provide empirical data from larger data sets of process models. We test LLMs of different types and sizes in their capabilities of achieving important properties of process execution, including enforcing process flow, resource allocation, and data-based conditions. Our results show that LLM performance falls short of the perfect reliability required for smart contract development. We suggest future work to explore responsible LLM integrations in existing tools for code generation to ensure more reliable output. Our benchmarking framework can serve as a foundation for developing and evaluating such integrations.