
HUST's MoDA Replaces a Decade-Old AI Building Block With 2.11% Accuracy Gains
HUST's hustvl group open-sources MoDA, replacing residual connections with attention across transformer layers for a 2.11% accuracy gain at only 3.7% extra compute.
20 stories

HUST's hustvl group open-sources MoDA, replacing residual connections with attention across transformer layers for a 2.11% accuracy gain at only 3.7% extra compute.

Node Banana v1.5.0 lands with loops, fallback models, and new video providers, offering a free MIT-licensed alternative to Weavy and Flora.

A radically stripped-down coding agent from the SWE-bench team hits over 74% on SWE-bench Verified using just bash and 100 lines of Python.

GuideLLM crossed 1,000 stars as the vLLM project's benchmarking tool for measuring real-world LLM inference latency, throughput, and SLO compliance.

A new bash-based workflow system ships 113 prebuilt automations for OpenClaw agents, from inbox triage to overnight builds, installable with one curl command.

A new study pinpoints why reinforcement learning with weak reward signals works for some LLMs and fails for others, and offers a concrete recipe to fix it.

EvoMap's Evolver lets AI agents encode what they learn as compact, reusable Genes — not verbose skill docs — and share them across a decentralized evolution network.

A new non-conformity score based on the distance to the median-ranked neighbor makes conformal prediction regions immune to nearly half-sample contamination.

Hyperspace open-sources a P2P autoresearch system where autonomous agents run experiments, gossip findings, and collaboratively train models from any device

MindsDB unifies 200+ databases and SaaS apps behind a single SQL and MCP interface, letting AI agents query live data without ETL pipelines.

A new empirical study shows that for RLVR fine-tuning, mixing question difficulties beats piling on easy examples, delivering up to 5x sample efficiency.

Nous Research's open-source Hermes Agent pairs a persistent learning loop with 20+ messaging surfaces, turning a chatbot into a self-hosted teammate that grows.

Pydantic 2.13 restores polymorphic subclass serialization, adds exclude_if to computed fields, and folds pydantic-core into the main repo.

oh-my-claudecode turns Claude Code into a full multi-agent team with 19 specialists, parallel execution, and zero configuration — hitting 38k GitHub stars

CowAgent (formerly chatgpt-on-wechat) is a self-evolving open-source AI agent with long-term memory, a skills marketplace, and support for 10+ LLMs across 12 messaging platforms

CalibAdv fixes two chronic failures in GRPO training for search agents: unfair penalties on correct intermediate steps and catastrophic language collapse.

A vertical chunking trick on Mamba2 keeps embedding memory flat regardless of document length, matching transformer quality on standard benchmarks.

Google AI Edge ships MediaPipe v0.10.35 with Android NPU acceleration via JIT compilation, API3 migrations, and Vite worker support for web builds.

A new HPCA paper redesigns processing-in-memory around product quantization, cutting KV cache footprint 6.5x and delivering 3.4x speedup over the previous SOTA PIM accelerator.

A TypeScript runtime plus CLI that turns any Model Context Protocol server into typed function calls, generated CLIs, or agent-ready clients.