thinkingbox grades ai agents on database state, not words
microsoft's thinkingbox benchmark runs 507 business workflows 20 times each to measure whether ai agents leave correct backend records, not just plausible replies.
machine learningtag machine learning
summaries filed under machine learning
autosynthdata turns agent failures into training data
servicenow's autosynthdata generates synthetic tasks from model failures to improve enterprise agents.
machine learningopen-sourcing astabrief, a fast report-generation model
ai2 releases astabrief 8b, an open-weights model that generates cited scientific reports from research questions and literature excerpts, cutting generation time by 3.5x.
machine learninggoogle's tee-based federated learning adds verifiable privacy
google research announces a new federated learning system using trusted execution environments to provide externally verifiable privacy guarantees and faster training.
researcholmo-core 3 opens scalable moe training
allenai releases olmo-core 3, an open framework for training trillion-parameter mixture-of-experts models with improved throughput.
machine learningdiffusion controller steers image models without retraining
google research introduces a lightweight steering network that improves prompt alignment in text-to-image models, even for closed-source systems.
researchshopify's continual learning loop cuts serving costs 96%
shopify built a daily flywheel that turns production failures into model weights, beating frontier models and slashing costs.
machine learningvllm adds hardware-agnostic layers for portability
vllm introduces new hardware-agnostic layers to keep torch.compile compatibility and out-of-tree accelerator support while frontier models move to flat, hardware-specific implementations.
machine learningliquid ai speeds up vision-language models with dspark drafter
liquid ai releases an experimental dspark draft model for lfm2.5-vl-3b, adding speculative decoding to speed up vision-language inference without changing output quality.
machine learninggoogle's multi-agent framework for coherent long-form video
google research introduces a multi-agent system that plans visual continuity to generate minutes-long videos without identity drift.
researchgemini 3.8 live adds real-time avatar for enterprise
google deepmind launches gemini 3.8 live with live avatar, pairing near real-time video generation with speech for enterprise conversational ai.
researchtransformers adds gguf support for local inference
hugging face transformers can now load and run gguf quantized models efficiently on apple silicon, reusing llama.cpp kernels.
machine learningnvidia nemotron 3 diarization tops voicearena leaderboard
nvidia releases an open-weight 100m-parameter model for real-time multi-speaker diarization, ranking first on voicearena's diarization-bench with a 14.72% der.
machine learningclaude opus 5.5 and gpt-6 sol and luna launch with price cuts
anthropic and openai released new models with significant price reductions, intensifying competition in the ai market.
toolsagents pass once but fail on repeat runs
a new consistency analyzer finds decision points where an ai agent flips between runs, then generates guidelines that cut the repeat-failure gap in half.
machine learninglow precision flash attention 4 for blackwell
pytorch extends flash attention 4 with mxfp8 support, reaching 2.85 petaflops forward on blackwell gpus.
machine learninggoogle launches gemini 3.8 live voice models
google deepmind released gemini 3.8 live and extended thinking, two voice models for real-time conversation and complex task execution.
researchgoogle's retrieve-for-train speeds up ai search
google research introduces a framework that trains a lightweight diffusion model to generate diverse search results without slow autoregressive reasoning.
researchzgcm-1: a fully open 7b model for math and agentic search
a new open 7b model uses efficient training and tool use to match much larger models on math and search tasks.
researchasync grpo with lora across hf jobs
a bucket, a proxy, and no nccl: training lora adapters asynchronously across separate hugging face jobs.
machine learning