This paper introduces Elmes*, an end-to-end framework for automatically constructing, refining, and applying fine-grained evaluation rubrics for LLMs in educational contexts. It combines a multi-agent engine with a self-evolving module SceneGen to co-optimize evaluation criteria and test data. The resulting Edu-330 benchmark covers 330 scenarios across 11 subjects, 3 grade levels, and 10 task types with over 1,000 indicators. Experiments reveal that educational capability is multidimensional: top LLMs differ mainly in creativity and values integration, knowledge-strong models may fail at Socratic scaffolding, and the education-specialized InnoSpark achieves the best human-evaluated score. LLM judges preserve human-comparable rankings but exhibit biases like self-preference. The framework…

This paper studies parallel Continuous Local Search (CLS) on Boolean satisfiability problems with symmetric pseudo-Boolean constraints. By relaxing the problem to continuous optimization on a hypercube, experiments reveal that redundant constraints inhibit convergence; CLS can quickly complete partial assignments in hybrid settings; and local search rapidly converges to a stable solution quality distribution due to saddle-dense objectives. These findings inform practical uses of CLS for SAT on modern accelerator hardware.

We present Accelerated Fourier SAT (AFSAT), a GPU-accelerated solver for pseudo-Boolean satisfiability based on continuous local search (CLS). AFSAT realises the proof-of-concept approach, FastFourierSAT, into a fully-engineered solver supporting any heterogeneous mixture of symmetric constraint types and lengths within a single problem instance. Using the JAX compiler, AFSAT leverages pure function composition, automatic vectorisation, automatic differentiation, and just-in-time (JIT) compilation to perform massively parallel CLS across batches of candidate assignments. We demonstrate substantially improved numerical stability, runtime performance, and memory efficiency over the proof-of-concept. We achieve this by way of identifying and addressing various limitations that arise from mem…

This position paper argues that AI research should focus on studying the training dynamics that produce model behavior, rather than only analyzing final models post-hoc. It calls for a science of AI that can predict, intervene, and design training procedures to reliably achieve desired properties.

CARVE-Q is a quantum-AI hybrid architecture for certified repair of vetoed driving maneuvers. It uses quantum minimum finding to accelerate repair enumeration while keeping safety certification classical, achieving provable speedup and 100% compliance.

A new study shows that if attackers strategically choose when to attack, the measured safety of AI control evaluations drops significantly. The researchers decompose attack decisions into start and stop policies, and demonstrate on BashArena and LinuxArena that both policies reduce safety by 20 to 28 percentage points at a 1% audit budget, without changing the underlying attack capability. Existing evaluations may overestimate safety, and the study recommends incorporating attack selection for more realistic estimates.

Large language models have made substantial progress on mathematical reasoning, but existing benchmarks typically evaluate well-specified problems with final answers or complete proofs, missing collaborative open-problem solving. CrowdMath is a dataset of 164 expert-annotated progress chains from the MIT PRIMES-AoPS CrowdMath program (2016-2025). Each chain tracks multi-participant forum discussions from problem statement to completed proof, with posts labeled by functional roles. Six frontier models achieve 83-88% accuracy on next-post prediction but only 0.42 macro-F1 on post-role classification, highlighting a gap in understanding collaborative mathematical progress.

This paper proposes Lean4Agent, the first framework that uses Lean4, a dependent-type formal language, to model and verify LLM agent behavior. It includes FormalAgentLib for verification and LeanEvolve for workflow revision. Experiments show verification-passing workflows outperform failing ones by 11.94%, and LeanEvolve further improves SWE performance by 7.47%.

Existing Sudoku solvers either lack correctness guarantees (learning-based) or suffer from long-tail search (symbolic). DiBS combines a complete symbolic solver with a diffusion model for branch ordering, reducing search cost significantly. Theoretical proofs and experiments on the Royle 17-clue benchmark show effectiveness in nodes, backtracks, and long-tail percentiles. Code is open-source.

This paper formalizes bias as a symmetry breaking operation and uses loss-based regularization to restore symmetry, achieving over 90% violation reduction with only about 5% accuracy cost on synthetic datasets. The framework requires no causal graph knowledge, is computationally light, and generalizes to any bit-flip definable sensitive attribute.

NVIDIA and LG Group are building an AI factory to accelerate LG's AI-driven businesses in robotics, autonomous driving, data center technologies, and GPU cloud services. The collaboration integrates NVIDIA's full-stack AI factory platform with LG's leadership in consumer electronics and robotics, aiming to create a unified workflow for physical AI systems.

Apple announces its third generation of Foundation Models, a family of five models built with Google, including on-device and server-based models, with a focus on privacy and new architectures like sparsely activated models and instruction-following pruning. The models power new Siri and intelligent tools, and show significant quality improvements in evaluations.

OpenEnv is a tool for creating an agentic execution environment like terminals, browsers, or anything an agent can interact with. Today, we’re excited to announce that OpenEnv is becoming even more open, to make the future of training agents open source. Starting today, OpenEnv will be coordinated by a committee that so far includes Meta-PyTorch, Reflection, Unsloth, Modal, Prime Intellect, Nvidia, Mercor, Fleet AI, and Hugging Face. OpenEnv now lives at huggingface/OpenEnv. The project focuses on being an interoperability layer for RL environments, not a reward framework or trainer.

Simon Willison releases datasette-agent-edit 0.1a0, a base plugin for Datasette Agent that provides storage-agnostic file-editing tools (view, str_replace, insert) inspired by Claude's text editor.

NVIDIA and Doosan Group are expanding their collaboration to advance new opportunities across physical AI, robotics and AI factory infrastructure, spanning Doosan Robotics, Doosan Bobcat, Doosan Enerbility and Doosan Corporation Electro-Materials BG.