This paper presents OpsLLM, a domain-specific large language model for software operations that supports knowledge-based QA and root cause analysis. The framework introduces a Human-in-the-Loop mechanism for data curation, supervised fine-tuning, and a domain process reward model for reinforcement learning. Experiments show improvements of 0.2%-5.7% on QA tasks and 2.7%-70.3% on RCA tasks over existing models. Models of 7B, 14B, and 32B parameters will be open-sourced with a 15K fine-tuning dataset.

A new method called eOptShrinkQ achieves near-lossless KV cache compression by decomposing attention cache into shared context and per-token residual, applying optimal spectral denoising and quantization. It outperforms existing methods in experiments on Llama-3.1-8B and Ministral-8B, saving nearly one bit per entry at equivalent quality.

StateSMix is a fully self-contained lossless compressor that couples an online-trained Mamba-style State Space Model (SSM) with sparse n-gram context mixing and arithmetic coding. It trains from scratch on the file being compressed, requiring no pretrained weights, GPU, or external dependencies. On enwik8, StateSMix achieves 2.123 bpb on 1 MB, beating xz -9e by 8.7%. The SSM is the dominant compression engine, while n-gram tables provide complementary gains. Implemented in pure C with AVX2 SIMD and OpenMP parallelism, it processes about 2,000 tokens per second on commodity hardware.

A new multi-agent system prototype for hydrodynamics uses a Layer Execution Graph to coordinate specialized agents, achieving 93.6% factual precision and graceful degradation, addressing context-saturation issues in single-agent LLM workflows.

Virtual Speech Therapist (VST) is an intelligent agent-based platform that streamlines stuttering assessment and delivers customized therapy planning through automated and adaptive AI-driven workflows. It integrates state-of-the-art deep learning-based stuttering classification and multi-agent large language model (LLM) reasoning to support evidence-based clinical decision-making. The system acquires patient speech samples, classifies stuttering types, then initiates an agentic reasoning process where specialized LLM agents autonomously generate, critique, and iteratively refine individualized therapy plans. A dedicated critic agent evaluates all plans for clinical safety, methodological soundness, and alignment with professional guidelines. The output is a comprehensive therapy draft for…

This paper presents a knowledge-driven decision-support system that combines structured defect knowledge with LLM reasoning for explainable defect diagnosis and mitigation in LPBF manufacturing. The system includes a knowledge base of 27 LPBF defect types, supports fuzzy natural language queries, and features a multimodal image-assessment module. Evaluation achieved a macro-average F1 score of 0.808 and substantial agreement with reference labels.

A formal algebraic framework for governed execution, mechanized in Rocq, axiomatizes governance via three axioms (safety, transparency, properness) and proves that all expressible programs are governed, forming a symmetric monoidal category. The system preserves Turing completeness while excluding unmediated I/O, and extracted OCaml passes extensive property-based testing.

This paper presents a machine-checked formalization of structurally governed AI workflow architectures using Interaction Trees in Rocq 8.19. It proves that effect-level governance can be imposed without reducing internal computational expressivity, establishing seven core properties that demonstrate the orthogonality of governance and expressivity.

This study presents an interoperability framework between FINALES and Kadi4Mat ecosystems to optimize the formation process of sodium-ion coin cells, balancing formation time and end-of-life performance. A multi-objective batched Bayesian optimization active-learning agent efficiently explores the parameter space, demonstrating data-driven optimization across automated systems.

ClinicBot is a novel AI system that provides accurate and verifiable answers for clinical diagnosis by structuring guideline extraction, prioritizing evidence, and offering verifiable citations, addressing LLM hallucination in medical contexts. It was validated on diabetes questions and the ADA 2025 Standards of Care.

Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors in LLMs, is explained by feature superposition geometry. The authors provide gradient-level derivation and empirical tests on multiple models (Gemma-2, LLaMA-3.1, GPT-OSS), using sparse autoencoders to show harmful features are geometrically closer. A geometry-aware filtering method reduces misalignment by 34.5%, outperforming random removal and competing with LLM-as-judge filtering.

This study presents a novel, AI-driven framework for assessing Environmental, Social, and Governance (ESG) performance in European small and medium-sized enterprises (SMEs). Using the n8n automation platform and large language models, the framework achieves high consistency with human expert outputs, supporting effective monitoring and intervention strategies aligned with the European Green Deal.

This roadmap provides a comprehensive perspective on AI and ML foundations, applications, and emerging directions in smart manufacturing, covering industrial big data analytics, autonomous systems, digital twins, robotics, supply chain optimization, and non-traditional ML approaches like physics-informed AI, generative AI, explainable AI, and LLMs, to guide researchers and practitioners in accelerating innovation towards reliable, sustainable, and scalable manufacturing ecosystems.

OpenAI is expanding ChatGPT ads with a beta self-serve Ads Manager for US advertisers, featuring CPC and CPM bidding, conversion measurement, and aggregated reporting, while keeping ads clearly labeled and separate from ChatGPT answers.

MCP Steroid is an IntelliJ plugin that exposes full IDE APIs, visual state, and runtime environment to AI agents, making them 20–54% faster on coding tasks. It provides access to refactoring, debugging, inspections, test execution, and more. Works with any MCP-compatible client.

Anthropic philosopher argues AI bias may correct historical injustice. The 2023 paper by Amanda Askell and co-authors suggests that AI models can be steered to overcorrect in favor of marginalized groups, with experiments showing a 7% positive discrimination toward Black students after human feedback training.

Canadian fiddler Ashley MacIsaac files $1.5M defamation lawsuit against Google after AI Overview falsely identified him as a sex offender, leading to a concert cancellation.

An open-source Android app running Gemma 4 entirely on-device, enabling chat, image understanding, and phone control without internet or cloud APIs.

Skelm is an open-source TypeScript framework for building secure, long-running AI agent workflows with a default-deny permission model. It supports typed pipelines, multiple scheduling options, and native MCP integration.

Google is developing an AI agent codenamed "Remy" that can take actions on behalf of users, not just answer questions or generate content. It is currently being tested internally by employees.

FFmpeg developer richardpl filed a GitHub issue accusing OxideAV of using AI to 'launder' the license of his code, potentially to bypass the original GPL restrictions. The issue, titled 'questionable LICENSE', asks for the location of a reverse-engineering document, raising concerns about the project's compliance with open source terms.

Telus Digital has deployed AI technology that alters the accents of customer service agents in real time to improve clarity. The tool, from Tomato.ai, modifies pronunciation while preserving voice identity. Unions criticize it as deceptive, while Rogers and Bell say they do not use similar technology.

OpenAI announces scalable low-latency voice AI technology, enabling real-time voice interactions at large scale and enhancing user experience.

Fedora 44 is now available with GNOME 50, bringing parental controls, accessibility improvements, graphics enhancements (VRR, fractional scaling, NVIDIA boost, HDR), and a polished experience that rivals mainstream operating systems.

Harrison Chase argues that agent observability should go beyond debugging to power learning loops. Observability alone provides traces of what happened, but feedback—user signals, indirect metrics, LLM-as-judge, and deterministic rules—is essential to turn those traces into actionable insights for improving the model, harness, and context. The post explores learning at multiple levels and outlines what an observability platform needs: store traces, store feedback, and generate feedback.

Inworld AI has released Realtime TTS-2, a new voice model that operates as a closed-loop system, taking full audio context of the conversation rather than just transcripts, to understand user tone, pacing, and emotion. Key features include natural language voice direction, crosslingual support across 100+ languages, advanced voice design without reference audio, and natural disfluencies. The model is available as a research preview via Inworld API and ranks #1 on the Artificial Analysis Speech Arena.

OpenAI has upgraded the default ChatGPT model to GPT-5.5 Instant, available to free users. Key improvements include a 52.5% reduction in hallucinations, a new 'memory source' feature, more concise answers, and a warmer, more natural tone. Sam Altman specifically encourages users who have been relying on deep thinking models to come back and try it.

Synthegy, a new AI system developed at EPFL, allows chemists to guide synthesis and reaction planning using natural language. It combines large language models with traditional algorithms to score and explain the best pathways, outperforming in double-blind studies with 71.2% agreement with chemists.

IBM launched Bob, an AI coding assistant, at Think 2026. It routes tasks to multiple models (Claude, Mistral, Granite), aiming to automate software development while keeping humans in control. Bob is positioned as an alternative to Claude Code and Codex. IBM also highlighted AI as core strategy, with 150 prebuilt agents in Watsonx Orchestrate and Concert AIOps expansion.

SpecMD is a standardized framework developed by Apple researchers to benchmark and evaluate expert caching policies for Mixture-of-Experts (MoE) models. The study reveals that MoE expert access patterns do not follow temporal locality, leading to the proposal of a new eviction policy called Least-Stale, which reduces collision misses by up to 85× compared to LRU and achieves 88% hit rates with 34.7% reduction in time-to-first-token on OLMoE.