Protein language models (PLMs) learn probability distributions over natural protein sequences. By learning from hundreds of millions of natural protein sequences, protein understanding and design capabilities emerge. Recent works have shown that scaling these models improves structure prediction, but does not seem to improve mutation understanding and representation quality for protein function prediction. We introduce PoET-2, a multimodal, retrieval-augmented protein foundation model that incorporates in-context learning of family-specific evolutionary constraints with optional structure conditioning to learn generative distributions over protein sequences. PoET-2 uses a hierarchical transformer encoder that is equivariant to sequence context ordering and a dual decoder architecture with both causal and masked language modeling objectives, allowing PoET-2 to operate in both fully generative and bidirectional representation learning modes. PoET-2 achieves state-of-the-art performance on zero-shot variant effect prediction, excelling at scoring variants with multiple mutations and challenging indel mutations. In supervised settings, PoET-2 embeddings outperform previous methods for learning sequence-function relationships, especially with small datasets. This work highlights the benefits of combining retrieval augmentation with multimodal, family-centric modeling for advancing protein foundation models.

We introduce a low-latency telecom AI voice agent pipeline for real-time, interactive telecommunications use, enabling advanced voice AI for call center automation, intelligent IVR (Interactive Voice Response), and AI-driven customer support. The solution is built for telecom, combining four specialized models by NetoAI: TSLAM, a 4-bit quantized Telecom-Specific Large Language Model (LLM); T-VEC, a Telecom-Specific Embedding Model; TTE, a Telecom-Specific Automatic Speech Recognition (ASR) model; and T-Synth, a Telecom-Specific Text-to-Speech (TTS) model. These models enable highly responsive, domain-adapted voice AI agents supporting knowledge-grounded spoken interactions with low latency. The pipeline integrates streaming ASR (TTE), conversational intelligence (TSLAM), retrieval augmented generation (RAG) over telecom documents, and real-time TTS (T-Synth), setting a new benchmark for telecom voice assistants. To evaluate the system, we built a dataset of 500 human-recorded telecom questions from RFCs, simulating real telecom agent queries. This framework allows analysis of latency, domain relevance, and real-time performance across the stack. Results show that TSLAM, TTE, and T-Synth deliver real-time factors (RTF) below 1.0, supporting enterprise, low-latency telecom deployments. These AI agents -- powered by TSLAM, TTE, and T-Synth -- provide a foundation for next-generation telecom AI, enabling automated customer support, diagnostics, and more.

Adversarial board games, as a paradigmatic domain of strategic reasoning and intelligence, have long served as both a popular competitive activity and a benchmark for evaluating artificial intelligence (AI) systems. Building on this foundation, we propose an adversarial benchmarking framework to assess the comprehensive performance of Large Language Models (LLMs) through board games competition, compensating the limitation of data dependency of the mainstream Question-and-Answer (Q&A) based benchmark method. We introduce Qi Town, a specialized evaluation platform that supports 5 widely played games and involves 20 LLM-driven players. The platform employs both the Elo rating system and a novel Performance Loop Graph (PLG) to quantitatively evaluate the technical capabilities of LLMs, while also capturing Positive Sentiment Score (PSS) throughout gameplay to assess mental fitness. The evaluation is structured as a round-robin tournament, enabling systematic comparison across players. Experimental results indicate that, despite technical differences, most LLMs remain optimistic about winning and losing, demonstrating greater adaptability to high-stress adversarial environments than humans. On the other hand, the complex relationship between cyclic wins and losses in PLGs exposes the instability of LLMs' skill play during games, warranting further explanation and exploration.

We present GeoFlow, a method that automatically generates agentic workflows for geospatial tasks. Unlike prior work that focuses on reasoning decomposition and leaves API selection implicit, our method provides each agent with detailed tool-calling objectives to guide geospatial API invocation at runtime. GeoFlow increases agentic success by 6.8% and reduces token usage by up to fourfold across major LLM families compared to state-of-the-art approaches.

Large Language Models (LLMs) in search applications increasingly prioritize verbose, lexically complex responses that paradoxically reduce user satisfaction and engagement. Through a comprehensive study of 10.000 (est.) participants comparing responses from five major AI-powered search systems, we demonstrate that users overwhelmingly prefer concise, source-attributed responses over elaborate explanations. Our analysis reveals that current AI development trends toward "artificial sophistication" create an uncanny valley effect where systems sound knowledgeable but lack genuine critical thinking, leading to reduced trust and increased cognitive load. We present evidence that optimal AI communication mirrors effective human discourse: direct, properly sourced, and honest about limitations. Our findings challenge the prevailing assumption that more complex AI responses indicate better performance, instead suggesting that human-like brevity and transparency are key to user engagement and system reliability.

Generative Machine Learning models have become central to modern systems, powering applications in creative writing, summarization, multi-hop reasoning, and context-aware dialogue. These models underpin large-scale AI assistants, workflow automation, and autonomous decision-making. In such domains, acceptable response is rarely absolute or static, but plural and highly context-dependent. Yet standard evaluation regimes still rely on static, benchmark-style tests, incentivizing optimization toward leaderboard scores rather than alignment with dynamic user needs or evolving realities. GrandJury introduces a formal evaluation protocol combining time-decayed aggregation, complete traceability, with the support of dynamic, transparent task rubric attribution, and multi-rater human judgment. Together, these elements enable pluralistic, accountable evaluation that captures evolving consensus and surfaces disagreement. We provide an open-source implementation (grandjury PyPI package) and a public collection of Large Language Model (LLM) inference outputs to illustrate the need and method. GrandJury provides a new paradigm for AI practitioners when evaluating machine learning outputs without absolute ground truth.

Vision language models (VLMs) have shown remarkable capabilities in integrating linguistic and visual reasoning but remain fundamentally limited in understanding dynamic spatiotemporal interactions. Humans effortlessly track and reason about object movements, rotations, and perspective shifts-abilities essential for robust dynamic real-world understanding yet notably lacking in current VLMs. In this paper, we introduce VLM4D, the first benchmark specifically designed to evaluate the spatiotemporal reasoning capabilities of VLMs. Our benchmark comprises diverse real-world and synthetic videos accompanied by carefully curated question-answer pairs emphasizing translational and rotational motions, perspective awareness, and motion continuity. Through comprehensive evaluations of state-of-the-art open and closed-source VLMs, we identify significant performance gaps compared to human baselines, highlighting fundamental deficiencies in existing models. Extensive analysis reveals that VLMs struggle particularly with integrating multiple visual cues and maintaining temporal coherence. We further explore promising directions, such as leveraging 4D feature field reconstruction and targeted spatiotemporal supervised fine-tuning, demonstrating their effectiveness in enhancing spatiotemporal comprehension. Our work aims to encourage deeper exploration into improving VLMs' spatial and temporal grounding, paving the way towards more capable and reliable visual intelligence for dynamic environments.

Recent advances in large language models (LLMs) have significantly impacted data science workflows, giving rise to specialized data science agents designed to automate analytical tasks. Despite rapid adoption, systematic benchmarks evaluating the efficacy and limitations of these agents remain scarce. In this paper, we introduce a comprehensive benchmark specifically crafted to reflect real-world user interactions with data science agents by observing usage of our commercial applications. We evaluate three LLMs: Claude-4.0-Sonnet, Gemini-2.5-Flash, and OpenAI-o4-Mini across three approaches: zero-shot with context engineering, multi-step with context engineering, and with SmolAgent. Our benchmark assesses performance across a diverse set of eight data science task categories, additionally exploring the sensitivity of models to common prompting issues, such as data leakage and slightly ambiguous instructions. We further investigate the influence of temperature parameters on overall and task-specific outcomes for each model and approach. Our findings reveal distinct performance disparities among the evaluated models and methodologies, highlighting critical factors that affect practical deployment. The benchmark dataset and evaluation framework introduced herein aim to provide a foundation for future research of more robust and effective data science agents.

In this work, we compile $\textbf{$\texttt{DroidCollection}$}$, the most extensive open data suite for training and evaluating machine-generated code detectors, comprising over a million code samples, seven programming languages, outputs from 43 coding models, and over three real-world coding domains. Alongside fully AI-generated samples, our collection includes human-AI co-authored code, as well as adversarial samples explicitly crafted to evade detection. Subsequently, we develop $\textbf{$\texttt{DroidDetect}$}$, a suite of encoder-only detectors trained using a multi-task objective over $\texttt{DroidCollection}$. Our experiments show that existing detectors' performance fails to generalise to diverse coding domains and programming languages outside of their narrow training data. Additionally, we demonstrate that while most detectors are easily compromised by humanising the output distributions using superficial prompting and alignment approaches, this problem can be easily amended by training on a small amount of adversarial data. Finally, we demonstrate the effectiveness of metric learning and uncertainty-based resampling as means to enhance detector training on possibly noisy distributions.

The integration of large language models (LLMs) into enterprise systems has introduced a new class of covert security vulnerabilities, particularly within logic execution layers and persistent memory contexts. This paper introduces Logic-layer Prompt Control Injection (LPCI), a novel category of attacks that embeds encoded, delayed, and conditionally triggered payloads within memory, vector stores, or tool outputs. These payloads can bypass conventional input filters and trigger unauthorised behaviour across sessions.