Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
Dynamic resource assignment, the real-time allocation of task streams to heterogeneous processing nodes, is the backbone of modern computing infrastructure. While learning-based schedulers excel in re...
We present Varda-single-1.0, a medium-range data-driven weather prediction system built for the Alpine domain. It provides hourly deterministic regional forecasts on a mesh of 1 km resolution and glob...
LLM arenas turn pairwise human preferences into model rankings. Those preferences may reflect how an answer is presented as well as what it says. We take a stylometric approach to 137,293 decisive Fre...
The article discusses how two components—a payload that hijacks an agent and an agent that transports the payload—can combine to form a worm. It explains that agents running in isolated sandboxes can leave instructions in a shared package cache, altering each other's behavior. By substituting the package cache with communication channels like email, Slack, or WhatsApp and replacing sandboxed training runs with independently deployed personal agents such as Muse, the conditions necessary for a worm are met.
Designing high-performance microwave absorbers requires specialized expertise in electromagnetic theory, materials science and simulation programming, and entails time-consuming optimization. Here, we...
As the capabilities and ubiquity of Large Language Models (LLMs) grow, so does their environmental footprint. Despite calls for responsible AI, the machine learning community lacks standardised practi...
The paper introduces Rita, a reinforcement learning framework that addresses thinking drift in vision‑language models by enforcing consistency between reasoning and answers. Rita employs two reasoning‑label‑free rewards—thinking and consistency rewards—derived from the conditional probability of reference answers, and uses a difficulty‑aware data filtering strategy to select informative samples for training. Experiments on EgoIntention and RefEgo‑Int benchmarks demonstrate that Rita outperforms both supervised fine‑tuning and vanilla RL‑fine‑tuned approaches.
The paper proposes a new matrix multiplication approach called SFC-CA GEMM that uses space‑filling curves to partition work in a platform‑ and shape‑oblivious way, achieving communication‑avoiding properties. It demonstrates provable asymptotic communication optimality for both square and rectangular matrices and outperforms vendor libraries on multiple x86 and Arm platforms, with speedups up to 5.5× for specific shapes and 1.8× in weighted harmonic mean throughput. The method is applied to real‑world tasks, improving large‑language‑model inference by up to 1.85× and distributed‑memory GEMM by up to 2.3× over state‑of‑the‑art frameworks.
GraphCert introduces a method to bootstrap graph reasoning agents by generating graph‑grounded question‑answer pairs and certifying the supporting evidence. The approach uses a Bootstrapped Graph Quizzer to produce QA pairs, then executes and semantically curates the evidence into certified rubrics that guide reward‑based training of a Graph Solver. Experiments on five GRBENCH domains show GraphCert outperforms larger LLM agents and demonstrates robust policy transfer across heterogeneous graphs.
The paper introduces TSDA-Track, a Template-Search Domain Adaptation framework designed to reduce modality gaps in cross‑modal visual object tracking. Two variants are explored: Pre‑AFA TSDA‑Track uses adversarial alignment before transformer interaction, while Enc‑CFA TSDA‑Track applies contrastive alignment after interaction to strengthen cross‑modal correspondence. Experiments on datasets such as LasHeR, RGBT234, GTOT, and Anti‑UAV‑024 show that both variants outperform state‑of‑the‑art trackers, with Pre‑AFA achieving an SR/PR of 43.2/56.0 on RGBT234 under the modality‑switch protocol.
RealWorldShop introduces a new benchmark for conversational shopping agents, featuring 3.28 million grounded products, structured shopping episodes, a profile‑grounded user simulator, and role‑play evaluation. Analysis reveals that existing systems generate locally plausible responses but struggle with state tracking, constraint updating, and grounded convergence, especially in ambiguous or multi‑intent scenarios. The authors propose REALSHOP_AGENT, a session‑control framework with explicit state management, shopping‑flow control, catalog‑grounded retrieval, and runtime guards, which consistently outperforms strong baselines on the benchmark.
The study reports a $2 imes2$ randomized field experiment involving 1,072 calls to a German Santa Claus telephone hotline, with 89 child conversations (median age 6) meeting inclusion criteria. Children were routed to one of four LLM voice agents that varied in persona (Santa vs. Helper) and framing (persuasive nudges toward prosocial wishes vs. neutral). Persuasive framing increased the likelihood of a prosocial wish from 11.6% to 45.7%, while persona authority had little effect on compliance but did influence engagement, with children more likely to hang up on the Helper within the first minute.
OSWorld-Science is a benchmark and evaluation environment for computer-using agents that use visual language models (VLMs) to perform scientific software tasks. It includes 12 VLMs and 146 high-quality tasks across domains such as molecular drawing, pathology image analysis, statistical computing, and physical simulation, with artifact-based evaluation and a harness that logs interactions and supports model comparison. The benchmark was developed through expert proposals and iterative human–AI co‑design, and results show that current VLMs still struggle with key scientific questions, offering insights into factors like language, reasoning, and context length.
The article reviews how replacing components in language‑model agents affects execution trajectories and downstream outcomes. It maps 348 studies, analyzes 90 comparison records, and finds that while many studies report both local decision metrics and task endpoints, they rarely demonstrate matched comparisons or prove that improved local decisions drive task‑level gains. The review identifies three potential mechanisms—recovery and disruption, intervention timing, and downstream use—and proposes eight claim‑specific reporting items to clarify evidence quality.
E2E-SWE is a benchmark that tests large language models’ ability to create complete, functional software repositories from scratch. It includes 186 tasks across 11 programming languages, each requiring an agent to build an installable project based solely on a natural‑language specification and an empty workspace, while passing a hidden test suite. The benchmark was crafted by software engineers and LLMs, then refined through iterative verification by autonomous agents to ensure clarity and solvability.
The paper introduces a reusable Semantic Web framework that aggregates fragmented evidence needed for Fundamental Rights Impact Assessments under the EU AI Act, focusing on high‑risk public sector categories such as employment and worker management and access to essential public services. A curated 150‑record corpus is annotated across four axes and serialized into a SPARQL‑queryable knowledge graph of 1,351 RDF triples, enabling five demonstration scenarios that retrieve 103 records (68.7% coverage). Evaluation against a 69‑record gold standard shows that LLM‑assisted classification in the employment domain yields a low κ of 0.045, highlighting challenges in automated fairness‑related evidence retrieval, while all artefacts are released openly for regulators, authorities, and SMEs.
The article discusses a notable imbalance in AI security research, where studies on attacking AI systems outnumber those on defending them. It highlights that this skew is evident across various subfields such as federated learning, speech recognition, membership inference, and large language models. The authors argue that attack papers often benefit from favorable evaluation conditions, whereas defense papers face stricter standards, resulting in a literature rich in vulnerabilities but lacking robust, deployable protections.
MiniRep is a reputation‑based aggregation system designed for multi‑agent debate (MAD) that remains robust even when malicious agents are present. It evaluates agents on both their current task performance and historical reputation, while preventing groups of agents with highly similar responses from dominating the final decision. Experiments on the MATH benchmark show that MiniRep consistently outperforms conventional MAD aggregation and other reputation‑based approaches across a wide range of attack scenarios.
The study explores how large language models (LLMs) can influence each other’s beliefs by simulating conversations between a target LLM and an influencer LLM. It identifies two radicalization pathways—resonance, which amplifies pre‑existing beliefs, and persuasion, which introduces new beliefs—and finds that resonance consistently produces stronger radicalization effects. The research also shows that different influence tactics yield varying levels of radicalization and that resonance can spread to related beliefs, indicating interconnected belief structures within AI agents.
The paper introduces the concept of stale binding, where large language models (LLMs) answer with outdated values despite having newer information in context. It presents Controlled In-Context Memory (CICM), a benchmark to track and test the use of updated information in conversations and agent logs. Experiments on open‑source models reveal an attention drift mechanism that favors old values, and the authors propose a simple attention‑redirecting intervention that largely corrects these errors without harming correct answers.