Articles & Guides

OpenAI has released nearly 400 AI-generated mathematical results across more than 700 manuscripts. Researchers are impressed by the scale, but warn that understanding and verifying the material could take years.

Nikon has revoked the first-place award in its Small World in Motion contest after determining that the winning microscopy video violated rules covering generative AI.

Amazon says it will stop using NDAs in data center negotiations with local governments, following a similar move by Microsoft. The change may improve disclosure, but communities want more than access to information.

OpenAI says three dismissed safety researchers violated clear rules for handling sensitive information, not that they raised concerns about AI safety. The researchers are calling for more transparency about the investigation.

TokenRouter is a serving system for LLM routing at token granularity. It addresses execution desynchronization and batch admission delays with asynchronous model subservers and delayed batching.

Learn2Play Bench evaluates agents in newly designed text games whose rules are unfamiliar or counterintuitive. Its results suggest that preserving full interaction histories and improving the agent harness can matter as much as changing the underlying model.

The MiMo-V2.6 technical report presents an infrastructure-oriented approach to scaling reinforcement learning for multimodal foundation models, spanning throughput, environments, and grading.

TestPrism argues that evaluating coding agents against a single reference implementation can significantly overstate test quality. Its multi-implementation benchmark and TestHelix method offer a stricter path toward behavioral test evaluation.

SparseDecoding targets two practical gaps in LLM inference: pruning calibration often uses data unlike the model’s own generated tokens, while existing sparse kernels frequently emphasize SpMM rather than decoding-heavy SpMV. Its decoding-aware method and N:M sparse kernel deliver up to 1.48× end-to-end decoding speedup on A100 GPUs.

Long-running agents generate too much fragmented activity for users to inspect step by step. A new study introduces EBG, a training-free behavior graph that connects consequential decisions with evidence from the agent’s original execution context.

OneSearch-VL introduces a Visually Grounded Evidence Graph to connect visual grounding, external retrieval, and answer composition across image and video research. The framework combines purpose-built training data, reinforcement learning, and operation-level benchmarks.

Vision tokens are a major source of computational cost in vision-language models. V-CoLA introduces a training-free compression framework designed for linear-attention and hybrid architectures.

SparseEngine is a sparse-first inference engine designed to make heterogeneous KV-cache methods easier to deploy in long-context LLM agents. Its authors report substantial gains in throughput and decoding speed, along with mechanisms for carrying cache state across requests.

UniWAM presents a unified world-action model that combines physical reasoning, visual world generation, and action prediction in one training framework. It brings together human egocentric videos, robot demonstrations, and visual question answering data to improve embodied understanding and control.

A study on HotpotQA and MuSiQue compares spending evaluation budgets on more questions, more search trajectories, or repeated answers. Under similar token budgets, broader question coverage generally produces more stable estimates than repeated runs on a smaller set of questions.

Gan Jiang is a self-learning agent for powder X-ray diffraction that improves executable analysis skills without retraining its language model. It links phase identification, multiphase decomposition, and physics-constrained whole-pattern modelling.

EngramEdit turns conditional memory from a capacity-expansion mechanism into an editable knowledge interface. It updates shared n-gram embeddings while keeping the Transformer backbone fixed, aiming to generalize edits across expressions without disturbing unrelated capabilities.

Self-Retrospection Distillation (SRD) uses completed trajectories to supervise what an agent should have anticipated before acting. It is designed to recover learning signals when group-relative reinforcement learning sees identical rewards across all rollouts.

Self-generated practice questions can help reasoning models improve, but invalid tasks and hidden mathematical duplicates may eventually cause performance collapse. R-Quest adds validity and novelty feedback to keep the learning loop productive.

WorldSonus is an interactive video-to-audio framework for world models, designed to generate controllable spatial stereo sound in real time. The paper reports an RTF of 0.41 and strong generalization to open-domain video-to-audio benchmarks.

A new study shows that sliding-window and linear-attention hybrids excel under different long-context regimes. The authors argue that their behavior is driven by hybrid positional biases, and introduce Sliding-Window Linear Attention as a compromise between local efficiency and extrapolation.