Wednesday, AI 101: hot from the oven – we discuss Gemini’s AlphaEvolve and Codex from OpenAI

Friday, we continue our Agentic Workflow series with one fascinating development

Agent identity, access, and governance are starting to take shape in practical ways.

Interoperability is improving through emerging protocols like MCP and A2A.

GitHub is becoming a more complete environment for building, testing, and shipping agents (not good for platforms Cursor!)

Azure AI Foundry is positioning itself as a control center for managing agent memory, orchestration, and debugging.

Agents are moving beyond the cloud – running locally, inside secure orgs, and across the open web.

Open source remains a key pillar, with tools like Copilot in VS Code, WSL, and NLWeb now available.

Seed1.5-VL by ByteDance advances agent-centric multimodal reasoning with a compact vision encoder and MoE-based LLM that outperforms competitors on 38 of 60 benchmarks and leads in GUI and gameplay tasks.

BLIP3-o by Salesforce unifies image understanding and generation by using a diffusion transformer for CLIP feature generation and sequential training that preserves comprehension while enabling generation.

Skywork-VL Reward by Skywork provides multimodal reward signals using a preference-aligned model that improves mixed optimization techniques and sets new standards on VL-RewardBench.

🌟 Aya Vision by Cohere scales multilingual multimodal generation using synthetic annotation and cross-modal merging, outperforming larger models in multilingual visual reasoning tasks.

Behind Maya supports cultural and linguistic diversity in VLMs with multilingual image-text pretraining across eight languages to improve low-resource language understanding.

MiMo by Xiaomi unlocks structured reasoning by combining large-scale pretraining with reinforcement post-training on math and coding, beating much larger models on reasoning tasks.

AM-Thinking-v1 by Beike elevates open-source reasoning to the 32B scale by fine-tuning Qwen2.5 using RL and SFT, resulting in state-of-the-art scores in math and code.

🌟 INTELLECT-2 by Prime Intellect Team trains a reasoning model via globally decentralized reinforcement learning, innovating with asynchronous infrastructure and outperforming the 32B reasoning baseline.

🌟 MiniMax-Speech by MiniMax achieves zero-shot high-quality speech synthesis with a learnable speaker encoder and Flow-VAE, excelling in cloning, emotion control, and multilingual support.

🌟 SWE-1 by Windsurf
Accelerate software engineering workflows by integrating tool-aware reasoning, incomplete state tracking, and timeline-based context for superior agentic coding experiences.

🌟 PointArena by researchers from University of Washington, Ai2, Anderson Collegiate Vocational Institute is a new evaluation platform that benchmarks multimodal models’ ability to point in images using natural language. It features three components: Point-Bench (982 tasks across five reasoning types), Point-Battle (live pairwise comparisons with 4,500+ votes), and Point-Act (robotic pointing execution). Molmo-72B outperforms others across all stages. Surprisingly, Chain-of-Thought hurts performance. PointArena reveals that precise pointing – not reasoning – best grounds language in real-world tasks. https://pointarena.github.io

🌟 Beyond 'Aha!': Toward Systematic Meta-Abilities Alignment in Large Reasoning Models proposes aligning models with meta-reasoning abilities (deduction, induction, abduction) to improve reasoning reliability and performance → read the paper

🌟 J1: Incentivizing Thinking in LLM-as-a-Judge via Reinforcement Learning
trains evaluators to produce better judgment through structured reward for thoughtful assessment → read the paper

🌟 The CoT Encyclopedia maps the diversity of reasoning strategies in chain-of-thought generations, helping steer and evaluate model thinking → read the paper

Learning from Peers in Reasoning Models enables collaborative error correction during reasoning by having different reasoning paths share intermediate summaries → read the paper

WorldPM: Scaling Human Preference Modeling establishes scaling laws for preference modeling and shows emergent behavior in large models trained on large-scale preference data → read the paper

Reinforced Internal-External Knowledge Synergistic Reasoning for Efficient Adaptive Search Agent designs agents that decide when to rely on internal knowledge or external retrieval, improving both speed and factuality → read the paper

🌟 System Prompt Optimization with Meta-Learning learns system prompts that generalize well across tasks by jointly optimizing with user prompts → read the paper

🌟 Parallel Scaling Law for Language Models introduces a new scaling paradigm by parallelizing computation instead of increasing model size or tokens → read the paper

🌟 Insights into DeepSeek-V3 explores how hardware-model co-design helps scale massive LLMs efficiently, with new attention and precision techniques → read the paper

Learning Dynamics in Continual Pre-Training for Large Language Models analyzes and models loss curves in continual pre-training to predict and control training behavior → read the paper

Memorization-Compression Cycles Improve Generalization proposes alternating phases of memorization and compression to improve generalization during training → read the paper

🌟 QuXAI: Explainers for Hybrid Quantum Machine Learning Models provides an explainability framework for hybrid quantum-classical ML models using feature attributions → read the paper

DanceGRPO: Unleashing GRPO on Visual Generation (Do read our deep dive into GRPO) adapts a reinforcement learning algorithm for diverse image and video generation tasks across paradigms → read the paper

Unified Continuous Generative Model unifies diffusion, flow-matching, and consistency models into a single framework for more efficient generative training and sampling → read the paper.

Depth Anything with Any Prior combines sparse metric priors and dense relative predictions to build highly generalizable monocular depth models → read the paper

MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning uses pure vision models and a novel one-prompt method for universal anomaly segmentation without language supervision → read the paper

🌟AttentionInfluence selects reasoning-rich training data by analyzing the influence of attention heads, no supervision needed → read the paper

🌟 MLE-Dojo: Interactive Environments for Empowering LLM Agents in Machine Learning Engineering introduces a Kaggle-based gym for agent training in real-world ML workflows, supporting reinforcement and supervised learning → read the paper