Brief

Researchers have introduced a novel optimization framework utilizing non-Archimedean polydisc spaces, inspired by Berkovich geometry. These spaces, formed by products of closed balls over non-Archimedean fields, offer a blend of hierarchical structure and geometric properties suitable for optimization. The work includes theoretical developments on geodesic uniqueness and the approximation properties of specific function classes, alongside an open-source Julia library for implementing these optimization procedures. AI

Researchers have developed a novel algorithm for supervisory switching control in partially-observed linear dynamical systems. This data-driven approach adapts multi-armed bandit algorithms to a control setting, aiming to identify and deploy the correct controller from a pool of candidates. The algorithm provides finite-time guarantees and can identify the appropriate controller within $O(N \log^2 N)$ steps while simultaneously achieving finite $L_2$-gain. AI

Anthropic's Claude Sonnet 4.6 achieved 100% comprehension on a newly developed data format called GCF, outperforming its sibling model Opus 4.6 which scored 96.2%. In tests involving 10 different models across three providers, GCF demonstrated superior performance in both comprehension and generation tasks compared to standard formats like JSON. The evaluation also found that Claude models could generate valid GCF output with minimal prompting, indicating strong adaptability. AI

Researchers have introduced a new framework called Interleaved Structural Chain-of-Thought (IS-CoT) to address the issue of long-form content generation collapse in Large Language Models. This framework embeds a dynamic Plan-Write-Reflect cycle within the generation process, allowing for continuous adaptation and alignment without external agents. A model trained with this method, IS-Writer-8B, has demonstrated state-of-the-art performance on long-form benchmarks, showing improved length compliance and coherence compared to existing models. AI

Researchers have developed a new framework for silent speech synthesis that combines surface electromyography (sEMG) and lipreading data. This approach uses modality masking during training to improve robustness against sensor failure or signal degradation. The masked multimodal system significantly reduced word error rates compared to unimodal methods, particularly for vowels and certain consonant groups, demonstrating its effectiveness for assistive technology. AI

Researchers have developed a multi-agent framework using large language models to simulate Chinese civil court proceedings. This system organizes role-based interactions through a five-stage trial process, incorporating memory and statute retrieval for complex adjudications. Experiments demonstrate the framework's ability to produce reliable civil judgments, particularly in liability allocation and multi-item adjudication, with memory quality significantly impacting simulation outcomes. AI

A new research paper introduces a privacy evaluation framework for medical language models, focusing on realistic threat models beyond simple text recovery. The framework assesses verbatim memorization and semantic leakage of sensitive diagnoses under varying levels of adversarial access. When applied to a model trained on clinical notes, it revealed high rates of memorization for encounter metadata and significant recovery of sensitive diagnoses like abortion and HIV, though some memorized tokens were templated. AI

Researchers have introduced UXBench, a new benchmark designed to evaluate the user experience of AI assistants. This benchmark focuses on preference alignment and dialogue generation, utilizing over 70,000 interaction logs from a Chinese AI assistant. UXBench includes three tasks—UX Judge, UX Eval, and UX Recovery—and has been tested on 26 large language models, revealing insights into how well these models understand and improve user experience. AI

Researchers have identified decoder inconsistencies in the Whisper ASR model that lead to higher word error rates for Dravidian and other low-resource languages. They found that these languages have longer words, greater vocabulary diversity, and less repetition, causing sparse token distributions and substitution errors. To address this, the paper proposes two decoder enhancements: Weighted-Attention to balance linguistic and acoustic cues, and Self-Conditioning to improve token consistency by reinjecting intermediate predictions. These methods demonstrated reduced word error rates for agglutinative and low-resource languages. AI

Researchers have developed a novel method called Self-Harness, enabling LLM-based agents to autonomously improve their own operational harnesses. This iterative process involves identifying model-specific failure patterns, generating targeted harness modifications, and validating these changes through regression testing. When applied to three different base models on the Terminal-Bench-2.0 benchmark, Self-Harness significantly boosted performance, demonstrating a path toward self-optimizing AI agents. AI

Researchers have developed a new self-supervised learning method called DecSelfMask to improve the performance of decoder-only models on classification tasks, particularly in domains with limited annotated data like healthcare. This approach uses relevance attribution to identify key text portions, masks them, and trains the model to reconstruct them, thereby transferring knowledge from unlabeled data. Experiments on clinical notes demonstrated significant gains over standard supervised fine-tuning and other self-learning techniques. AI

Researchers have developed AbstRAG, a new method to address abstraction gaps in retrieval-augmented generation systems. AbstRAG explicitly models abstraction as a retrieval object, decomposing the gap into components like expression and intent. The system uses reflective refinement, where a critic identifies retrieval failures, suggests patches, and accepts them under control mechanisms to improve relevance and generation accuracy. AI

Researchers have developed MUDIDI, a two-stage framework designed to digitize multilingual dictionaries, particularly those for low-resource languages. The framework addresses challenges like varied scripts, complex layouts, and the preservation of lexicographic structure. MUDIDI's first stage assesses character recognition and markup preservation, while the second stage segments dictionary entries into a machine-readable format. Experiments show that large language models (LLMs) outperform traditional OCR and vision-language models in this task, with performance further enhanced by providing additional contextual information like dictionary introductions. AI

Researchers are exploring the adaptation of Voice Activity Projection (VAP) models to predict turn-taking in sign language interactions. An initial study using the Public DGS Corpus adapted a VAP architecture to sign language, utilizing pose data from hands and facial regions. While the model showed promise in predicting SHIFT/HOLD actions, particularly with hand cues, predicting the precise SHIFT remains challenging, indicating a need for sign-language-specific event definitions. AI

Researchers have developed a new evaluation metric for grounded generation models that addresses the limitations of existing precision-focused methods. The new metric, which incorporates recall alongside precision, was tested using Formula 1 telemetry and NOAA weather forecasts, domains with complete ground truth data. Results showed that current frontier models, while precise, cover less than half of the relevant facts, highlighting the need for coverage-aware evaluation. AI

Researchers have developed LexRubric, a new benchmark designed to evaluate the performance of large language models on open-ended legal tasks in Chinese. The benchmark includes 649 instances covering legal consultation and judicial examination, with over 12,000 expert-written scoring criteria across six dimensions. Initial tests on 18 LLMs revealed varying capability profiles, indicating that current models still struggle with complex legal reasoning. AI

Researchers have developed a novel speech-to-LLM interface called Convex Gate (C-Gate) that constrains speech representations to the LLM's input embedding manifold. This approach ensures compatibility with pretrained LLMs while preserving continuous expressivity, unlike previous methods that either lost paralinguistic information or allowed representations to drift. C-Gate demonstrated strong joint performance in automatic speech recognition and emotion recognition, improving word error rate by up to 48.7% and matching single-task emotion accuracy. The study suggests that the geometry of time-resolved trajectories in the embedding space, rather than discrete token identities, is crucial for multimodal integration in frozen LLMs. AI

Researchers explored the effectiveness of prompting API-accessed Large Language Models for Ukrainian grammatical error correction. Their study found that while fine-tuned models still lead, certain commercial LLMs, particularly Claude and Gemini, showed significant improvement with Ukrainian-specific prompts and minimal-edit strategies. The best configuration achieved over 90% of the gap to the state-of-the-art, though some models exhibited overcorrection patterns related to Ukrainian linguistics. AI

Researchers have developed a new hybrid metaheuristic approach to solve the Traveling Salesman Problem (TSP), a complex optimization challenge. This method integrates the Dragonfly Algorithm, known for its global search capabilities, with Tabu Search, which uses memory to refine solutions locally. The combined strategy aims to improve tour quality by exploring broadly and then fine-tuning promising results, showing better performance than individual algorithms on benchmark instances. AI

Researchers have developed a new method for stylometric analysis inspired by genome-wide association studies (GWAS). This approach uses logistic regression to identify statistically significant lexical markers associated with individual authors across English, German, and Russian text corpora. The technique aims to provide a more interpretable way to understand authorship attribution. AI

Researchers have developed NüshuVoice, the first text-to-speech system designed to revive the endangered Nüshu language. This system addresses the challenge of extremely limited audio data by creating a new sentence-level dataset. It utilizes a pitch-aware VITS framework, Nüshu-PitchVITS, which incorporates Nüshu's five-level pitch notation to improve speech synthesis accuracy and intelligibility. AI

Researchers have identified paradoxes within the rejection rate metric used to evaluate social learning performance in decentralized decision-making systems. Their analysis reveals this metric is unsuitable for accurately measuring performance. The study then focuses on error probability for a binary Gaussian problem, deriving a formula that highlights an irreducible, agent-dependent gap between decentralized and centralized error probabilities. AI

Researchers have developed a new framework for multimodal sentiment analysis that improves performance by aligning representations from different modalities, such as text and images. The proposed method uses vision-language models to convert visual content into textual descriptions, creating a shared linguistic space for analysis. This approach, combined with a hybrid learning strategy, has achieved state-of-the-art results on several benchmarks, demonstrating the importance of representation alignment for effective multimodal learning. AI

This paper analyzes the effectiveness of 11 different stopping criteria within the CMA-ES black-box optimization algorithm. Researchers quantitatively evaluated these criteria on the BBOB function set, focusing on their ability to accurately determine when to halt the search process without wasting computational resources. The study found that `tolflatfitness` and `tolfun` were frequently the first criteria to be triggered, while `tolfunhist` and the combined portfolio of criteria achieved the highest stopping accuracy. AI

Researchers have developed EviProp, a novel method for retrieving relevant pages from long, visually rich documents. Unlike existing approaches that score pages independently, EviProp models documents as multimodal Chunk-Page graphs. It uses seeded relevance diffusion, combining query-page similarity with chunk-level signals to improve retrieval accuracy. Experiments on benchmark datasets show EviProp outperforms traditional methods and leads to better downstream question-answering performance. AI

Google DeepMind has released findings from a randomized controlled trial in Sierra Leone, evaluating an AI tool called Guided Learning within the Gemini platform. The study, involving 1,763 students over eight weeks, indicated that the AI augmented, rather than replaced, teachers, leading to significant improvements in math scores. Students using the AI tool showed gains equivalent to 1.2 to 2.5 years of learning, with conversations prioritizing conceptual understanding over simple answer-seeking. AI

The AI news landscape saw significant developments in coding benchmarks and agent development. Cognition introduced FrontierCode, a new benchmark that evaluates code mergeability and maintainability, revealing that even top models like Opus 4.8 struggle with complex tasks. The concept of 'loops' is gaining traction as a dominant metaphor for controlling coding agents, emphasizing clear goals and iterative structures, though practitioners caution against naive implementation and highlight the continued need for human oversight. Agent ergonomics are also improving with new tools for observability and orchestration, alongside practical advice for operators on measurable outcomes and bounded autonomy. AI

Researchers have developed a new opticalmechanics framework for estimating the dynamics of multibody systems without direct contact force sensors. This approach uses image-measured kinematic data as non-contact inputs to a constrained multibody model. A genetic algorithm identifies unknown joint torques by minimizing discrepancies between predicted and measured kinematics, demonstrating potential for dynamic estimation in challenging environments. AI

Researchers have introduced Gryphon, a new recommendation system architecture designed to improve the accuracy of generative retrieval. Unlike previous methods that optimize for token sequence likelihood, Gryphon incorporates an item-level scoring component trained directly on user relevance. This dual approach allows Gryphon to re-score items generated from the same semantic ID, leading to better recall and a more streamlined system. In an A/B test on a music service, Gryphon maintained listening time while simplifying the candidate generation pipeline. AI

Researchers have developed a multilingual fact-checking system for Factiverse, utilizing fine-tuned compact models for efficiency and scalability. The system employs a three-stage pipeline involving claim detection, evidence retrieval, and veracity prediction. Comparative experiments showed that fine-tuned models like XLM-RoBERTa-Large and mmBERT-base offer strong performance across numerous languages, remaining competitive with larger LLMs such as GPT-5.2 and Claude Opus 4.6 in terms of accuracy and significantly outperforming them in latency and cost-efficiency for production deployments. AI

Researchers are exploring new approaches to developing cancer vaccines, moving beyond traditional preventive methods. The focus is on therapeutic vaccines administered to individuals already diagnosed with cancer. Despite decades of attempts and a history of limited success, a renewed sense of optimism is emerging in the field, driven by recent advancements and a deeper understanding of the immunological mechanisms involved. AI

Researchers have developed a new jam-absorption driving strategy inspired by police car swerving maneuvers to mitigate stop-and-go traffic waves. This strategy, termed Single-Vehicle Double-Detector Jam-Absorption Driving (SD-JAD), aims to suppress traffic congestion by having a dedicated vehicle perform "slow-in" and "fast-out" actions. The proposed method is designed for practical implementation, utilizing only two roadside detectors to measure key parameters and has demonstrated success in simulations without causing secondary waves. AI

A paper titled "The Moral Inefficacy of Carbon Offsetting" is discussed, questioning the moral consistency of its author's logic. The discussion centers on whether the author's arguments align with their own reasoning, particularly concerning the efficacy of carbon offsetting. AI

Researchers have introduced SFILES 2.0, an enhanced text-based notation for representing chemical process flowsheets. This new version addresses limitations of the original SFILES, enabling unambiguous descriptions of essential configurations and control structures crucial for process operation. The development includes open-source software for converting between graph-based flowsheets and SFILES 2.0 strings, aiming to establish a standard for a FAIR (Findable, Accessible, Interoperable, Reusable) database of chemical process flowsheets. AI

The James Webb Space Telescope has captured new images revealing intricate details of the cosmic web, the large-scale structure of the universe. These observations provide unprecedented insights into the distribution and evolution of matter across vast cosmic distances. The findings are expected to advance our understanding of galaxy formation and the underlying scaffolding of the cosmos. AI

Researchers have identified limitations in the mesoscopic traffic flow model used by the Simulation of Urban MObility (SUMO) software. The existing model, based on Eissfeldt's 2004 work, does not fully adhere to the Lighthill-Whitham-Richards (LWR) model principles, leading to inaccurate congestion dynamics and underestimation of congestion magnitude. To address these issues, a new discrete-time implementation of the link transmission model is proposed, which more accurately captures queue spillback phenomena and aligns with kinematic wave theory and microscopic SUMO simulations. AI

Researchers have developed a new method for automatically detecting quotation errors in academic papers using fine-tuned large language models. This approach aims to improve the accuracy and efficiency of identifying inconsistencies between cited information and its original source. The study found that incorporating the full text of cited literature, particularly the abstract, significantly enhanced detection performance. Additionally, the researchers utilized the TokenSHAP tool to analyze the interpretability of the model's predictions. AI

Researchers have detailed the process by which transformer language models, which operate on subword fragments, aggregate these into word-level representations. They identified a two-stage detokenization process primarily occurring in early to middle layers, involving attention transmitting token-specific signals and MLPs composing them with local embeddings. This mechanism was found to be consistent across twelve models from eight different families, with the depth of the process varying based on positional encoding types. AI

Researchers have developed OpenOpt, an open-source framework for optimizing SRAM architecture and transistor sizing. This framework utilizes equivalent circuit models to achieve significant simulation speedups while maintaining high accuracy for read/write delays and power consumption. The system integrates various optimization algorithms and has demonstrated substantial improvements in static noise margin, area, and peak power. AI

Researchers have developed Ishigaki-IDS, an open-weight large language model specifically designed to assist in drafting Information Delivery Specification (IDS) files for Building Information Modeling (BIM) projects. This model integrates continued pretraining on BIM/IDS data, supervised fine-tuning, and reinforcement learning with validator feedback to generate machine-checkable IDS drafts. Ishigaki-IDS significantly outperforms existing LLMs like Claude Opus 4.5 on key metrics and has been shown to reduce authoring time by over 50% in user studies, easing the practical burden of creating these specifications. AI

A new research paper reveals that large language models fine-tuned for specific linguistic communities do not necessarily align with the expected political orientation. The study found that a Ukrainian-oriented model was less resistant to Russian disinformation when queried in Russian, while a Russian-oriented model showed stronger rejection of such narratives. The research suggests that factors like corpus composition and prompt format are more influential than nominal cultural alignment in determining an LLM's susceptibility to disinformation. AI

Researchers have introduced TRADE, a novel architecture for speech Large Language Models designed to enable efficient streaming inference. By integrating a transducer branch with an LLM, TRADE achieves frame-synchronous acoustic alignment while retaining the LLM's linguistic reasoning capabilities. This approach allows for accurate, streamable, and long-form speech processing, demonstrated by competitive Word Error Rates on various benchmarks and improved end-of-utterance detection. AI

Researchers have developed ToolRec, a new framework designed to improve query recommendation in on-device intelligent assistants. This system addresses the limitations of existing methods by focusing on the rapid invocation of system tools, which is common in assistant usage. ToolRec utilizes a comprehensive repository of system tools and a dual-level calibration mechanism to refine raw user click data, reducing noise from varying activity levels and emphasizing tool-invoking queries. Extensive A/B testing on a platform with over 150 million monthly active users showed significant improvements in click-through rates and total clicks compared to existing baselines. AI

A new research paper explores how personal preferences are revealed through highlighting in shared documents. The study found that while the general selection of highlighted sentences is heavily influenced by social trends and what others mark, an individual's unique choices become apparent when selecting from already highlighted passages. This suggests that personal identity in digital interactions is more about refined selection than initial broad marking. AI

Researchers have identified a critical flaw in multi-agent AI systems, particularly in medical question answering, where consensus on answers can mask underlying reasoning misalignment. They developed CARA, a metric to assess reasoning alignment, and found that debate protocols can create an "consistency illusion," making agents appear more aligned while their reasoning diverges. A new protocol, GDP, was introduced to improve this by requiring agents to commit to specific facts and stances, significantly enhancing reasoning alignment without increasing computational cost. AI

A new study on arXiv explores the effectiveness of query decomposition in multi-condition information retrieval systems. Researchers found that decomposing queries early in the retrieval process can harm performance by diluting semantic meaning. However, decomposing queries during the reranking stage significantly improves accuracy by allowing for more precise constraint verification. To address this, the study proposes a framework that keeps queries monolithic during initial retrieval and uses sub-queries only for reranking, demonstrating improved performance on established benchmarks. AI

Researchers have developed TinyGiantALM, a new 1.5 billion parameter audio-language model designed for resource-constrained environments. This model utilizes an Instruction-Aware Feature Refinement framework, incorporating a Query-guided Projector and Semantic Gating, to better process acoustic signals based on user intent. On the MMAR benchmark, TinyGiantALM achieved 46.4% zero-shot accuracy, outperforming larger models up to 13 billion parameters and demonstrating a viable path for efficient edge-based perception. AI

Researchers have developed a new framework called S$^3$E to evaluate multimodal language models by probing their internal decision states under semantic stress. This method contrasts image-supported captions with semantically similar but incorrect options, analyzing hidden states to detect instability even when the model's external behavior remains correct. Studies on models like Qwen3VL, Gemma3, and InternVL3 revealed that semantic stress can cause significant internal state displacement, suggesting that external correctness alone is insufficient to guarantee stable internal decision geometry. AI

Researchers have developed a novel pipeline using GPT-4.1 to analyze mental health discourse in Arabic-speaking X (formerly Twitter) communities. The study examined 8,147 tweets related to borderline personality disorder, bipolar disorder, and ADHD, identifying distinct linguistic patterns associated with each condition. Findings suggest that bipolar disorder discussions frequently include religious and medical terms, BPD tweets focus on relationships and emotional distress, and ADHD conversations often revolve around practical symptoms and medication. AI

Researchers have developed a new framework called Stigmatized Self-Reflection (SSR) to better simulate patient self-stigma in large language models. This approach incorporates internal monologues into mental health dialogues, allowing AI agents to exhibit more realistic context-sensitive resistance behaviors like avoidance or self-blame. By fine-tuning LLMs with a specialized dataset and using a chain-of-thought method, the SSR framework enables patient agents to dynamically adjust their expression of stigma, leading to more authentic responses for clinical training and empathetic dialogue systems. AI