Machine Learning News

MimicScribe's live-preview transcription, built on a batch ASR model with overlapping-window merging, transcribed 18.8 hours of meetings and earnings calls with 15.5% of words wrong and 96.9% of spoken numbers correct, w…

Researchers behind the paper "Evolutionary Architecture Search through Grammar-Based Sequence Alignment" (arXiv:2512.04992) use sequence alignment to identify the edits connecting two neural architectures and then draw o…

Iluvatar Labs released Kev, an open-source tool on GitHub that converts any LLM into a decision model returning choices and probabilities instead of generated text, starting with Qwen3's 1.7B, 4B, and 8B models. On a 276…

Valen-Team released Valen-Preview-0923, a multimodal decision model built on a Qwen3.5-0.8B or 2B backbone that scores supplied candidates and returns decision probabilities instead of generating answer tokens. On a Soko…

TypeSafe released Jev in September 2026, a model that returns probabilities for typed questions instead of generating text, scoring 67.8% versus GPT-5.6 Terra's 67.9% on TypeSafe's own four-workflow evaluation at roughly…

A developer's technical deep dive explains that Jev's per-call confidence scores are only useful if calibrated, since a system can be highly accurate overall yet badly overconfident on the decisions it gets wrong. The po…

A Stanford University paper introduces Self-Organizing Agent Teams (SAT), a framework in which groups of AI agents learn collaboration strategies from as few as 15 AIME 2024 or 25 GPQA Diamond problems and transfer them …

The DataMinds team built an end-to-end urban flow analytics platform for the SLIIT CodeFest Datathon 2026, processing 48.6 million trip records across 260+ urban zones and finishing as Top 8 finalists. The pipeline used …

Black Forest Labs released FLUX 3 Action, a 7B open-weights model that predicts robot actions and future video frames, on 22 September, and it now tops NVIDIA's RoboLab leaderboard at 42.9% with its slower variant while …

Rerun released version 0.32 of its open-source SDK, its biggest release since the company open sourced the project three years ago, adding a stabilized file format, lower-level read-and-write APIs, a chunk-manipulation A…

A developer benchmarked a typed-decision API (Jev), a 35B mixture-of-experts LLM (Qwen3.5-35B-A3B), and a 421M open-weight model (Laya) on 12,000 real U.S. federal IT solicitations, finding Jev led primary-class accuracy…

A developer built Sentinel, an agentic fraud-investigation system on TigerGraph that reads six months of the Vesta card-transaction graph and produces per-alert verdicts, fraud-pattern labels, approval-routed actions, an…

Microsoft's harrier-oss-v1-270m sentence-transformers model can be fine-tuned with SentenceTransformerTrainer and MultipleNegativesRankingLoss for Portuguese QA retrieval, though no public example combines that exact sta…

Jev is a decision-oriented AI model that returns structured outputs — such as a route, probability and confidence level — rather than generated text, aimed at application workflows that need a discrete decision. The proj…

TypeSafe AI has introduced Jev, a "System One" model that abandons general-purpose text generation in favor of producing typed, structured decisions directly consumable by backend software. Jev exposes three primitives —…

Black Forest Labs launched FLUX 3 Action, an open robotics AI model that uses camera feeds to predict a robot's next action. With 7 billion parameters, FLUX 3 Action sets a record on the RoboLab-120 benchmark and runs up…

Figure reported that pretraining on its Index dataset raised full-task success in 30 previously unseen homes from 9% to 56%, a 47-percentage-point gain, in a September 17, 2026 evaluation that held architecture and downs…

A developer published a reproducible pipeline that reaches 94.25% accuracy on the 77-class Banking77 intent-detection benchmark using frozen sentence-transformer embeddings and a scikit-learn logistic regression classifi…

AWS released a WhisperX Deep Learning Container that packages OpenAI's Whisper with wav2vec2 forced alignment for per-word timestamps and speaker diarization, deployable to Amazon SageMaker AI real-time or asynchronous e…

A multi-dimensional evaluation framework called PotARCin exposed a 25-52 percentage-point performance gap in five state-of-the-art models — Claude-2, GPT-4-Turbo, LLaMA-2-70B, PaLM-2-Chat, and a specialized ARC transform…