Daily digest

4 items · ~4 min · Week 2026-W31

Worth knowing (3)

Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning

NVIDIA
Research official + media 3 src. ~1 min

NVIDIA open-sources Molt, a PyTorch-native asynchronous RL training system built to stay small and readable enough for a single researcher to hold in their head, while matching the throughput of Megatron-based stacks. It supports multimodal and mixture-of-experts policies and enforces on-policy training by never training on a token the current policy did not generate. Uses Ray for placement/async queues, vLLM for rollout, and NVIDIA AutoModel + FSDP2 for training.

Why it matters
Top-voted paper on HuggingFace Daily Papers for 2026-07-27 with 605 upvotes, reflecting strong community interest in simplifying agentic RL infrastructure at scale; ships as an open-source repo (NVIDIA-NeMo/labs-molt) rather than just a paper.

AREX: Towards a Recursively Self-Improving Agent for Deep Research

Research official + media 2 src. ~1 min

AREX is a family of deep-research agents (4B dense and 122B-A10B MoE) trained with reinforcement learning on verified synthetic tasks. It runs a dual-loop process: an inner loop gathers evidence and drafts provisional answers, an outer loop audits the answer constraint-by-constraint and dispatches targeted follow-up searches for unresolved claims. A learned context-update tool autonomously compresses growing interaction history into a compact state that preserves verified evidence.

Why it matters
141 upvotes on HuggingFace Daily Papers; the self-auditing outer loop and autonomous context compression address two persistent weaknesses of long-horizon research agents, with competitive results against larger models on BrowseComp, WideSearch, DeepSearchQA and Humanity's Last Exam.
For reference (1)

K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs

Research official + media 2 src. ~1 min

The authors build K12-KGraph, a knowledge graph extracted from Chinese K-12 textbooks (math, physics, chemistry, biology), and derive two resources from it: K12-Bench, a 23,640-question benchmark spanning five task categories testing curriculum and prerequisite understanding, and K12-Train, a fine-tuning dataset with text and visual question-answer pairs. Evaluation shows current LLMs struggle with prerequisite and concept-relationship reasoning, while domain-specific fine-tuning on K12-Train substantially closes the gap.

Why it matters
Provides one of the first curriculum-structured, prerequisite-aware benchmarks for educational LLMs rather than flat QA-style tests, useful for gauging whether models actually understand pedagogical dependency structure.