breaking papers · 86 analyzed
AI-powered analysis of breakthrough research from arXiv and beyond. We surface the work that matters before it hits the news cycle.
A Tsinghua research group splits legal reasoning from final answers, training preference-reward signals (the same approach used to align chat models) on style, elements, and chain quality, in work that has not been independently replicated.
A probe that reads a model's own internal signals caught more reward hacking — shortcuts agents learn when a test is easy to game — than conventional reasoning monitors on one leading AI model, and fewer on another.
A new preprint says each design candidate can be cut down to one upfront computation per problem plus a cheap per-design run, taking the best average rank among generative methods on 47 trade-off tasks, though the result is benchmark-only until
AREX-2, an agent trained to revise its own work across many rounds, with 27 billion parameters, released as an open-weight model (its trained weights are downloadable) by the Beijing Academy of Artificial Intelligence (BAAI) and trained on long
A new arXiv preprint on multi-objective workflow generation describes a method that returns different AI agent workflows for accuracy, cost, latency, robustness, or consistency trade-offs at inference time, without retraining — a distinct mechanism
A controlled study finds that language models copying each other under a shared budget explored less than solo learners at the same cost.
A preprint finds that three diagnostic signals — operator entanglement (how a quantum state's information spreads across the circuit), stabiliser entropy (a structural measure of the quantum state), and out-of-time-order correlators, or OTOCs (a
A spiral-shaped soft robot hit 100% on 50 real grasps and 30 directed throws by using its whole body as the gripper, on simple 2D (planar) tasks in one lab.
Fiatlux, a new simulation benchmark, asks a humanoid to climb a ladder, swap a lightbulb, and dispose of the old one without breaking it. Policies with no task-specific training clear none of its twelve subtasks.
Northeastern researchers propose PANDA, a multi-agent design that drops the central registry and lets agents self-form teams. The 8x benchmark claim is single-dataset.
Adversaries targeted leading AI models' step-by-step reasoning through normal-looking API traffic, scaling to 16,000 requests from more than 4,000 accounts in two days — with a related cluster spanning more than 15,000 more accounts before
Vision-Language-Action models like π0.5 clear a demo in seconds, but production robots must hit a real-time control cycle, share a processor, pass a safety case, and clear a bill of materials.
Data Agent Benchmark(DAB)测的是"模型+智能体+数据系统"端到端协同,覆盖 PostgreSQL、MongoDB、SQLite、DuckDB 与金融、生物医学、政务等领域;首个破 90% 的提交用 OceanBase 加智谱 GLM-5.2 大模型组合,Scout 是 OceanBase 这次提交 DAB 的内部代号,仍非上线产品。
An arXiv impossibility theorem shows bounded-degree quantum annealing solvers (a class of quantum optimization techniques restricted to low-order penalty interactions) can tie their math to thirteen decimal places and still overflow the plate (the