HySparse2 explores KV sharing for long contexts
A post published on September 23, 2026 describes two levels of KV sharing intended to reduce prefill cost and cache size, while improving retrieval in long-context workloads.
Read articleKnowledge you can apply
Radar
Daily bookmarks: relevant news, facts and discoveries.
NVIDIA launches Nemotron 3 Diarization
Announced on September 23, 2026, Nemotron 3 Diarization identifies who spoke and when, including in overlapping speech, and supports up to eight speakers.
Read articleDSec: sandboxes for training agents at scale
On September 23, 2026, an article on DSec described elastic, stateful environments for agent training and evaluation, as well as defenses against agents that exploit training environments.
Read article
Streaming speech: protect personal data before transcription
Real-time ASR and translation models can process speech containing personal data; Rota Nacional’s privacy barrier applies the organization’s policy before a model runs.
Read articleZlaya: Laya CPU inference in Zig
Zlaya is a CPU inference engine for Laya, implemented in Zig and described in its repository as fast and compact.
Read articleiSDFT explores continual learning in LLMs
Published on September 22, 2026, iSDFT describes a continual-learning method for LLMs based on on-policy self-distillation and information proximity. Its repository provides training code, checkpoints, and evaluation results.
Read articlemini-AGI explores continual learning with 8 GB of VRAM
Published on September 22, 2026, the account describes a model trained from scratch on a notebook with 8 GB of VRAM, using batch size 1, disk-paged weights, and learning-rate adjustments to reduce forgetting.
Read articleMixture-of-Depths allocates compute per token
An article published September 22, 2026 presents a method for Transformers to allocate compute to sequence positions while respecting a per-layer budget.
Read articleonPanda: model annotation and inspection
Published on September 22, 2026, this entry describes an open-source tool for annotating LLM data and inspecting models, including token-level edits and inspection of token probabilities and decoding.
Read articleQuiet-STaR explores reasoning before an answer
Published on September 22, 2026, the brief describes a training approach that teaches language models to generate internal reasoning before visible text.
Read articleProduction sandboxes: reported scale
A September 22, 2026 publication cites a paper on sandbox infrastructure and reports production-scale operating figures.
Read articleXeno-interpretability and native representations
A paper investigates whether language models use internal distinctions for which humans lack adequate concepts. It calls these structures xeno-representations and examines the limits of interpreting them through familiar human concepts.
Read articleCodeMidas creates RL tasks from code
Published on September 21, 2026, the report introduces Xiaomi’s CodeMidas: agents extract executable tasks from open-source codebases. The system generated 5,545 tasks across 23 languages.
Read articleCodeMidas proposes RL tasks for coding
Published on September 21, 2026, the announcement describes a pipeline that turns features from open-source projects into executable reinforcement-learning environments.
Read articleMiMo Pro raises its DeepSWE benchmark score
A post reports that MiMo Pro's DeepSWE score rose from 58.41 to 72.57 after reinforcement-learning runs, nearing comparative results cited in the post.
Read articleTR-DPO adds a KL penalty to the DPO loss
A September 21, 2026 publication describes Trust Region DPO (TR-DPO), which incorporates a KL divergence penalty into the DPO loss to limit deviation from the reference model.
Read articleVoice of Reason answers spoken math problems
In a September 21, 2026 post, Kyutai Labs says its model reasons about spoken math problems and answers by voice, without an intermediate transcript or text LLM. The post reports 77.1% on GSM8K, versus 27.3% for GLM-4-Voice.
Read articleBITCOS reduces ternary weights to 1.485 bits per weight
A September 20, 2026 report on Intel’s BITCOS method says it stores exact ternary weights at 1.485 bits per weight and reports decode-throughput gains of up to 18% on CPUs and 27% on GPUs.
Read articleGated Linear Attention combines gates and linear attention
Published on September 20, 2026, the item describes a transformer formulation with linear attention, data-dependent gates, and an equivalent representation as an RNN with matrix-valued hidden states.
Read articleJamba combines Transformer and Mamba layers
Published on September 20, 2026, the item describes Jamba, a base language model with a hybrid Transformer-Mamba architecture and mixture-of-experts.
Read articleRota Nacional
Bring privacy into your workflow.
30 days, no card, with a starting quota. After that, Pix credit from R$ 10,00.