granite 4.2: how ibm built its reasoning llms
ibm released granite 4.2, a family of dense reasoning llms in 3b, 8b, and 30b sizes, trained from scratch on 15t tokens with a multi-stage rl pipeline.
machine learningtag machine learning
summaries filed under machine learning
kvboost reuses kv cache chunks anywhere in prompts
kvboost cuts llm prefill latency by reusing key-value cache chunks at arbitrary prompt positions, not just shared prefixes.
researchdeepmind partners with eve online studio for ai research
google deepmind and fenris creations are collaborating to build ai agents that can learn, remember, and plan in the persistent world of eve online.
researchConstraint-Aware GPU Allocator Boosts Utilization by 33 Points
A new GPU scheduler from Dharma-AI improves cluster utilization and priority-weighted output by planning allocations across the entire scheduling horizon instead of using FIFO order.
machine learningBenchmark optimization found in top speech recognition models
New probes show leading ASR models reproduce benchmark transcripts even when audio contradicts them, inflating scores.
machine learningGoogle adds mobility data to language models for better place understanding
Google Research introduces ME-POIs, a framework that blends text and mobility data to improve predictions about places like opening hours and busyness.
researchGoogle's Biomarker Discovery Framework prioritizes wearable sensor candidates
Google Research introduces a multi-agent AI system that structures biomarker discovery from wearable data through iterative hypothesis generation, statistical analysis, and literature-grounded reasoning.
researchAgent memory dosage depends on model capability
IBM Research shows that the right amount of agentic memory varies by model tier, with curated retrieval boosting weaker models and full guideline sets helping stronger ones.
machine learningLiquid AI Releases DSpark Draft Models for Faster LFM2.5 Inference
Liquid AI published DSpark draft checkpoints for three LFM2.5 models, enabling speculative decoding that speeds up inference up to 3.2x on GPU and 2.87x on-device without changing output quality.
machine learningLiquid AI Releases QAD Q4_0 Checkpoints for LFM2.5 Models
Liquid AI releases quantization-aware distillation checkpoints that recover 97% of BF16 accuracy for LFM2.5 models at Q4_0 memory and speed.
machine learningSentence Transformers v6.0 Adds Multi-Vector Encoder
Sentence Transformers v6.0 introduces MultiVectorEncoder for ColBERT-style late interaction retrieval, supporting PyLate, Stanford-NLP ColBERT, and ColPali checkpoints.
machine learningGoogle's PhotoScan estimates body composition from smartphone photos
Google Research demonstrates a deep learning method that predicts insulin resistance from smartphone imagery with accuracy close to DXA scans.
researchFP8 Training on AMD GPUs Now Upstreamed in PyTorch
AMD and Meta engineers upstreamed FP8 training optimizations for AMD Instinct GPUs into TorchAO and TorchTitan, delivering up to 13.4% throughput gains on dense models and recovering 89% of quantization overhead on MoE models.
machine learningOpen models shift toward China, small models, and agents
Hugging Face data from early 2026 shows Chinese labs dominating frontier open models, small models driving downloads, and coding agents becoming a major user base.
machine learningWhat a 2,200-paper ICML reproduction effort found
A community hackathon used coding agents to reproduce over 2,200 ICML 2026 papers, finding that 23% had at least one falsified or contested claim.
machine learningGoogle DeepMind Launches Gemini 3.7 Flash
Google DeepMind introduces Gemini 3.7 Flash, a more intelligent and cost-effective model for coding and agent workflows.
researchGoogle DeepMind brings sign language AI to Pixel phones
Google DeepMind's SL2T model powers sign-to-text dictation in Gboard and Live Transcribe on Pixel 11, starting with ASL to English.
researchRecall, not encoding, limits LLM factuality
Google Research introduces knowledge profiling to show frontier LLMs encode nearly all facts but struggle to recall them, with thinking recovering many failures.
researchMeta releases Muse Glimmer for on-device agentic AI via ExecuTorch
Meta open-sourced Muse Glimmer, a 30B-parameter model distilled from Muse Spark, optimized for on-device agentic workflows with ExecuTorch support for NVIDIA GPUs and Apple Silicon.
machine learningMeta releases Muse Glimmer, a 30B open-source multimodal model for local agents
Meta’s Muse Glimmer is a 30B-parameter multimodal model optimized for local, privacy-aware agentic tasks like coding and document analysis, released under Apache 2.0 with day-0 support in transformers and llama.cpp.
machine learning