AI Daily — 2026-10-28

English 中文

Multiscale policy optimization improves long-horizon agent training, Trinity enables multimodal recursive self-improvement, and VibeVideoBench benchma

Covering 24 AI news items

🔥 Top Stories

1. Ψ-0.6 Unifies Vision, Language, and World Modeling With Adaptive Reasoning

Researchers introduced Ψ-0.6, a unified vision-language world model that reasons over words, depth, motion, camera pose, and imagined scenes and uses RL to learn when to use each modality. It reports zero-shot gains on spatial reasoning and object-motion prediction, plus transfer from simulated rooms to real robot tabletops. The release of code and paper makes it a useful testbed for adaptive multimodal reasoning and embodied world models. Source-reddit

2. Papers with Code Relaunches as Open-Source Apache 2.0 Project

Papers with Code has relaunched as an open-source Apache 2.0 project hosted on GitHub under Hugging Face, with a FastAPI backend, React frontend, public API, and MCP server. The relaunch restores a key discovery layer for papers, benchmarks, and SOTA results while inviting community contributions. For practitioners, it could become a more open and collaborative research index if maintainers sustain momentum. Source-reddit

3. Reviewer Warns Half of AI Conference Peer Reviews Are AI-Generated

A longtime reviewer for NeurIPS, ICML, AAAI, and ICLR reports that roughly half the peer reviews they now see appear fully or partly AI-generated, alongside more AI-written submissions. They describe the reviewing process as in “critical condition” and call for far more attention to safeguards and process design. The anecdote is not a formal study, but it highlights a growing governance risk for AI conferences and research integrity. Source-reddit

📰 Featured

Agents & Training

Multiscale Policy Optimization Boosts Long-Horizon LLM Agent Training — MPO aligns credit assignment and policy updates at the turn level to improve long-horizon agent training; the approach targets a known weakness in critic-based RL for multi-turn agents. Source-huggingface
The Trinity Enables Recursive Self-Improvement for Multimodal Agents — The framework exploits shared capability requirements across modalities and task forms to reduce costly environment construction and manual harness iteration. Source-huggingface

Multimodal, Embodied AI & Benchmarks

VibeVideoBench Benchmarks Multimodal Agents on Code-Driven Video Generation — The benchmark evaluates agents that write code to render frames and synthesize audio across 239 English and Chinese tasks, targeting precise outputs that diffusion video models often miss. Source-huggingface
PDRVLA Adds Physical Dynamics Tokens to Vision-Language-Action Models — A lightweight physical dynamics module extracts tokens from vision-language features to condition robot action generation without the latency of online video generation. Source-huggingface
ChronoReplay Tackles Interface Lock-In in Streaming Spatial VLMs — The paper identifies post-training specialization to one readout interface and proposes replaying heterogeneous supervision to preserve multiple spatial readouts. Source-huggingface

Audio & Benchmarking

EmbeddingGemma-2 Hits #1 on MAEB Audio Benchmark via Null-Space PCA — A zero-training PCA projection into the model’s unused null space lifted a 577M embedding model from rank #26 to #1 on a 30-task audio benchmark, outperforming much larger multimodal models. Source-reddit

Tools & Frameworks

Superpowers Offers Agentic Skills Framework for Coding Agents — The open-source framework and methodology guides coding agents to clarify specifications before coding and works across Claude Code, Codex, Cursor, Gemini CLI, and GitHub Copilot CLI. Source-github

⚡ Quick Bites

HKUDS Releases DeepTutor for Lifelong Personalized AI Tutoring — DeepTutor is an open-source system for lifelong, personalized AI tutoring. Source-github
GitHub Collection Publishes 31 Jailbreak Prompts Targeting 25+ AI Models — The repository documents jailbreak prompts aimed at more than 25 AI models. Source-github
Generic feedback matches self-critiques in hindsight self-distillation for financial reasoning — A study finds generic feedback can match self-critiques in hindsight self-distillation for financial reasoning. Source-reddit
Cocktail-Party Audio CAPTCHA Shows Humans Beat Speech Models Narrowly — A reverse Turing test using cocktail-party audio finds humans narrowly outperform speech models. Source-reddit
Pandora M1: Desktop Monitor for Distributed Training Across 64 Nodes — Pandora M1 is a desktop monitoring box for distributed training across 64 nodes. Source-reddit
Harness Records LLM Agent Tool Argument Provenance Before Execution — The harness records where an LLM agent’s tool arguments came from before execution. Source-reddit
AI Design Skills Collection for Claude Code, Cursor and Other Coding Tools — The repository collects AI design skills for Claude Code, Cursor, and other coding tools. Source-github
Harvard’s Stephen Casper Hosts AMA on AI Safeguards and Governance — Stephen Casper is hosting an AMA on AI safeguards and governance. Source-reddit
Reddit User Asks How to Add Reasoning and Tool Use to CPT’d Qwen 9B Legal Model — A practitioner asks how to add reasoning and tool use to a continued-pretrained Qwen 9B legal model. Source-reddit
Manual Autograd for Transformers Built Step-by-Step in PyTorch — A tutorial builds transformer autograd manually step by step in PyTorch. Source-reddit
Could AI Detection Hit a Fundamental Mathematical Limit? — A discussion asks whether AI detection could eventually face a fundamental mathematical limit. Source-reddit
Open-Source Repo Documents 20 NCM and Transformer Interpretability Studies — The repository documents 20 studies on next-concept modeling and transformer interpretability. Source-reddit
PhD Student Seeks Guidance on Starting Agentic AI for Workflow Automation — A PhD student asks how to start with agentic AI for a specific workflow automation use case. Source-reddit
Reddit User Seeks Non-EMA Teacher Methods for Self-Distillation — A practitioner asks about teacher methods other than EMA for self-distillation. Source-reddit

Generated by AI News Agent | 2026-10-28