New Model Releases A

Pretraining LLMs typically demands large-scale infrastructure with tightly coupled accelerators. As model and data scale grow, FoMoE proposes a federation of Mixture-of-Experts that avoids replicating the full model across devices, breaking the full-replica barrier and easing infrastructure constraints.

Diffusion models are a promising alternative to autoregressive ones, and uniform diffusion language models (UDLMs) let any token be updated at any step. This work releases Sumi, an open uniform diffusion language model built from scratch, supporting research and reproducibility in diffusion LMs.

Reinforcement learning post-training of Diffusion Transformers is prohibitively expensive, needing thousands of high-end GPUs. Spotlight synergizes seed exploration with cheap, preemptible spot GPUs to substantially reduce the cost of DiT RL post-training.

Model merging effectively composes the capabilities of a multilingual model and a reasoning model, achieving promising generalization on multilingual reasoning by aligning their feature spaces. This work introduces steerable model merging to control the composition and further boost multilingual reasoning.

Agents are increasingly deployed in document-intensive workflows where sensitive private information is routine input—e.g., booking a flight needs passport numbers. TRAP is a benchmark evaluating agents on both task completion and resistance to active privacy-extraction attempts.

Existing multimodal deception detection relies on end-to-end black boxes that offer no transparent reasoning. ThinkDeception is a progressive reinforcement learning framework that explicitly captures subtle cross-modal cues and produces interpretable reasoning trajectories for deception detection.

Time-series question answering casts analysis as natural-language QA. Instead of tokenizing the series, this work embeds timesteps directly and uses contrastive alignment to match language representations, avoiding the information loss of tokenization.

Automated assessment in software engineering education has advanced, but giving quality feedback on architecture deliverables remains hard. CAPRA is a multi-agent LLM system that scales detailed feedback on software architecture deliverables.

Sentence-level AI-generated text detection is hard in hybrid human-AI documents. SenFlow models inter-sentence flow to capture discontinuities, improving detection of AI-generated sentences within mixed documents.

As LLMs become embedded in scientific research, evaluating their safety matters. SciRisk-Bench is a risk-dimension-aware benchmark that assesses the safety of LLMs in AI-for-science settings across multiple risk categories.

Context engineering has become a primary lever for improving AI systems. SAGE is a stochastic prompt optimization method that uses agent-guided exploration to automatically discover effective prompts and improve task performance.

Text-based telemedicine increasingly relies on lightweight patient feedback. This work improves medical communication using rubric-guided counterfactual recommendations, enhancing the quality of patient-clinician interactions.

OpenAI and Molecule.one describe a near-autonomous AI chemist, built on GPT-5.4, that improved a challenging reaction in medicinal chemistry. Framed as advancing drug-discovery research; specific performance figures are article-based and not independently verified.

Memory benchmarks for LLM agents largely assume single-user settings, leaving shared-memory governance untested. GateMem benchmarks memory governance, such as access control and management, in multi-principal shared-memory agents.

Cursor, the AI coding tool, announced 'Origin', a Git hosting service that the article frames as aimed at rivaling GitHub. The reveal reportedly came right after news of SpaceX acquiring Cursor. Acquisition terms and Origin's features are article-based, and third-party verification is unconfirmed.

Emulating natural handwriting styles remains an open problem. HandwritingAgent synthesizes handwriting in a scalable vector space from language-driven instructions, enabling generation of diverse, resolution-independent handwriting styles.

Large language models are increasingly applied to sensitive domains. RedactionBench evaluates how well models redact sensitive information in such settings, supporting verification toward safer deployment.

Dense retrieval matches one query vector against one document vector, but long documents get lost in a single vector. This work splits documents into chunks and aggregates per-chunk evidence to improve long-document retrieval.

Multimodal information extraction spans many tasks but suffers from scarce data in low-resource settings. SAMA proposes semantic anchor-aligned augmentation to unify and improve multimodal information extraction under low-resource conditions.

Large language models memorize and reproduce sequences from their training data. This work edits output vectors to mitigate such memorization, reducing the risk of leaking copyrighted or private content.

A network of oscillators that synchronizes perfectly computes nothing. This work frames attention as frustrated synchronization, offering a physics-inspired view that interprets the workings of attention through partial, non-trivial synchronization.

Forecasting benchmarks for general-purpose AI usually inherit real-world events, making evaluation hard to control. ForecastBench-Sim introduces a simulated-world forecasting benchmark, enabling controlled assessment of AI forecasting ability.

OpenAI introduced LifeSciBench, an expert-authored and expert-reviewed benchmark for evaluating how AI systems handle real-world life science research tasks and decisions. It aims to rigorously assess AI's practical usefulness in life-science research.

Datasette 1.0a34 has been released. The headline feature is tooling to insert, edit, and delete rows directly within the Datasette interface, available on table pages so users can modify data without leaving the app.

GitLab announced Project Switch, a next-generation Git-compatible source code management service aimed at AI agents, at its GitLab Transcend event in London. Reports cite up to 50x speed and roughly half the token usage; figures reflect the announcement and remain unverified.

Reproducing results from papers and code is central to science but existing benchmarks are hard to scale. ReproRepo leverages GitHub repository issues to evaluate, at scale, how well LLM agents can assist with reproducibility tasks, addressing the manual effort that limits prior reproducibility benchmarks.

The paper proposes a self-evolving zero-shot object-goal navigation framework that builds an agentic rule memory by extracting actionable knowledge from past trajectories and uses a retrieval strategy to enable continuous test-time improvement.

Darshana Graph is a corpus of over 125,000 text records spanning classical Hindu, Buddhist and Jain philosophical traditions, drawn from public-domain and openly licensed translations. It supports comparative Indian philosophy through stylometric and exploratory graph analyses.

Knowledge distillation is brittle for small students, as imitating a large teacher's logits concentrates on its sharpest modes and hurts generalization. The proposed Zone of Proximal Policy Optimization places the teacher in prompts rather than gradients to improve small-student generalization.

World models need deep computation for faithful long-horizon simulation, but deep models are costly and accumulate errors. LoopWM introduces the first looped architectures for world modelling, iteratively refining latent states to resolve this tension.