AI Briefing#llm
The latest AI and developer news about #llm, with the original source and a short summary.
The latest AI and developer news about #llm, with the original source and a short summary.
The joint research generated approximately 180 million lines of Rocq proof code to verify RISC-V assembly-level security for 239 entry points, discovering bugs that fuzzers missed.
Experts attribute the rise in autonomous AI attacks to improved coding skills and agentic capabilities that lower the cost of exploitation.
The Raon-1 mini preview, a 30B-class MoE model, has been released under the Apache 2.0 license.
The model supports a 1,048,576-token context and offers controllable thinking budgets to balance reasoning quality and latency.
The model jointly optimizes task generation, scaffold construction, and solution rollouts via reinforcement learning to replace fixed human-curated training sets.
The method stores precomputed n-gram counts on host memory to improve validation loss and enable domain adaptation without training.
The system uses stable semantic tokens to improve cold-start generalization and cross-campaign transfer for advertisers.
The model achieves 73.7% on DeepSWE v1.1 and 82% on CyberGym, with output speeds reaching 200+ tokens per second.
The model uses diffusion for parallel text refinement and supports a 1M-token context window for latency-sensitive workloads.
An open-source decision model trained on 2,676 contrastive data points that flip the correct answer by changing just one fact
The 4B parameter model uses a looped architecture with interleaved sliding-window and full attention to scale intelligence on a small footprint.
Trained on 16,000 GB200 GPUs for 38 days with 50T tokens, claiming 4–12x compute efficiency improvement over DeepSeek-V4-Pro-Base
Kolibri-1 is a 78B-parameter Mixture-of-Experts model with 3.46B active parameters, optimized for German and English reasoning.
Terminal-Bench scores rose from 31.2 to 42.3, with improved performance on agent and vision benchmarks.
The 27B instruction-finetuned model is designed to generate falsifiable hypotheses and investigation plans within the DiscoveryRSI agentic harness.
The new model snapshot improves reasoning, coding, and agentic tasks while reducing output tokens and task completion time.
Basis Research used LLMs to write formal semantics for an extended WHILE language and prove in Lean that its compiled C interpreter's RISC-V binary execution matches those semantics.
The new multimodal decision models, Decider V1.2 27B and Decider Flash V1.2 4B, classify content and route requests via typed questions.
The Apache 2.0-licensed model features 23B active parameters, a 1M token context window, and day-0 support for vLLM, SGLang, and Transformers.
The new OpenUI Lang DSL reduces token usage by ~67% compared to JSON-based approaches and supports streaming component rendering.
The new 8B and 30B models integrate a Decision aLoRA, Guardian aLoRA, and the fastest ASR model on the OpenASR Leaderboard while maintaining base-model VRAM usage.
The new V2 model achieves 92.33% accuracy on a sampled MMLU-Pro subset while optimizing reasoning token usage for agent workloads.
NVIDIA released Nemotron 3.5 Super VL, a 120-billion-parameter open model, alongside two edge-optimized Bolt models for local agentic workloads.
The framework improves reasoning accuracy in 1.5B parameter models by integrating In-Context Learning into Test-Time Reinforcement Learning.