Diffusion language models, which generate entire sequences at once and iteratively refine them, have become competitive with autoregressive models on quality since 2024, with releases from Inception L…
Anthropic removed over 80% of Claude Code's system prompt for the Claude 5 generation with no measurable loss on coding evaluations, while an ICLR 2026 paper found that context collapse from rewriting…
A mortgage broker who built AI workflows for underwriting document review found that inconsistent AI output stems from probabilistic models and can originate at any of seven architectural layers. The …
At ICLR 2026, program chairs desk-rejected papers containing hallucinated references, confirming that nonexistent citations can pass through the submission process of a major machine-learning conferen…
A diffusion model's creativity is a mathematical consequence of neural networks learning a 'smoothed' version of the score function, driving the model to interpolate between training data points along…
Four months after Google Research and DeepMind published TurboQuant, a training-free algorithm suite claiming 6x memory reduction for KV cache, the promised official code has not shipped. Instead, the…
Researchers reported on July 8, 2026, that AI peer-review systems can be manipulated by rewriting papers to satisfy model-generated reviewer preferences rather than improving scientific quality. Tests…
KakaoBank's Financial Technology Research Institute had four papers accepted at major AI conferences in 2026, including a prompt injection detection method at ICLR 2026 and studies at LREC 2026 and AC…
Researchers introduced PCB-Bench, the first comprehensive benchmark for evaluating large language models on printed circuit board placement and routing tasks. The benchmark includes over 3,700 questio…
Sakana AI launched Fugu, a multi-agent system delivered as a single model that dynamically orchestrates frontier models to tackle complex multi-step tasks. The system is accessible via the Sakana API …
Google Research developed MoGen, an AI model that generates synthetic neuron geometries to improve brain mapping. The model reduced reconstruction errors by 4.4%, potentially saving 157 person-years o…
CEM888.AI announced that its agent Vetta achieved the highest published scores on the MemoryAgentBench benchmark at ICLR 2026, with 99.9% on AR Retrieval and 77.2% on BEAM Memory, outperforming GPT-4.…
Researchers have developed DiffusionBlocks, a method that trains neural networks by dividing them into blocks and training only one block at a time, reducing memory requirements to roughly 1/B of stan…