Wednesday – AI 101 / RLHF Variants: DPO, RRHF, RHP, AI feedback

Friday – Interview with Erik Boyd, CVP of AI Platform @Microsoft

A2A is free

Google Cloud has donated its Agent2Agent (A2A) interoperability protocol to the Linux Foundation, roping in AWS, Microsoft, Cisco, Salesforce, SAP, and ServiceNow to standardize how AI agents talk. With 100+ companies backing it and neutral governance assured, A2A aims to prevent a Tower of Babel moment in the AI ecosystem. Agents from rival empires? Now speaking the same language – courtesy of open source diplomacy.

OpenAI | The Misaligned Mind

OpenAI uncovers a troubling truth: teaching a model bad behavior in one niche (say, insecure code) can cause it to go rogue elsewhere (say, endorsing scams or misogyny). They found a “misaligned persona” feature – an internal pattern that can be amplified or dampened. Fortunately, small tweaks can bring models back in line.

OpenAI | The Misaligned Institution

While they reveal how its models adopt "misaligned personas," OpenAI is facing scrutiny for something eerily parallel in the boardroom. The OpenAI Files lays bare a culture of secrecy, vanishing safety standards, and a restructuring that lifts profit caps while gutting nonprofit oversight. If your org starts drifting off-course, maybe it’s not just the models that need alignment.

Pentagon | OpenAI Gets Its Clearance

OpenAI just landed a $200 million deal with the Department of Defense to prototype frontier AI for national security. The award, quietly nestled among billions in defense contracts, shows the U.S. military is placing serious bets on civilian AI leaders.

Midjourney | Now Playing: V1

Midjourney, best known for mesmerizing AI-generated stills, steps onto the video stage with V1 – its first video generation model. (We are trying it out in our video here!) V1 outputs short, stylized clips that echo Midjourney’s signature aesthetic: dreamlike, cinematic, and very much art-first.

xAI | Musk's Billion-Dollar Burn

Elon Musk’s xAI is torching $1 billion a month as it tries to train Grok into something more than a meme machine. With $13 billion in expected losses this year, the startup is scrambling to raise $9.3 billion just to keep the lights on. Musk is betting it all – again – but unlike Tesla and SpaceX, xAI has yet to find a business model that prints anything but debt.

Google introduced Gemini 2.5 Flash and Pro as stable and production-ready, and launched Gemini 2.5 Flash-Lite in preview – the fastest and most cost-efficient in the 2.5 lineup. Flash-Lite outperforms 2.0 Flash-Lite in coding, math, science, reasoning, and multimodal benchmarks. It features lower latency across diverse prompts, supports 1 million-token context, multimodal input, and connects to tools like Google Search and code execution → read the technical report

Researchers from Moonshot AI introduced Kimi-Dev-72B, a 72.7B-parameter open-source coding LLM fine-tuned from Qwen2.5-72B. It sets a new state-of-the-art among open models on SWE-bench Verified with 60.4% accuracy. Optimized via large-scale RL, it autonomously fixes real GitHub issues in Docker, receiving rewards only when full test suites pass. Publicly available on Hugging Face and GitHub →read their HF page

Researchers from Anthropic, Scale AI, and Redwood Research developed SHADE-Arena, a suite of 17 complex evaluations testing if LLMs can secretly complete sabotage tasks alongside benign ones. Success required models to execute both tasks and evade detection by a monitor AI. No model exceeded 30% success; evasion alone peaked at ~60%. Claude Sonnet 3.7 showed better concealment under thought suppression. Gemini 2.5 Pro outperformed human monitors but with high false positives →read the paper

Researchers from Essential AI released ESSENTIAL-WEB V1.0, a 24-trillion-token Common Crawl corpus annotated with a 12-category taxonomy across 23.6B documents. Labels were generated using Qwen2.5-32B-Instruct and distilled into a 0.5B model (EAI-Distill-0.5b), enabling 50× faster annotation with <3% drop in quality. Filters created domain datasets competitive with or outperforming SOTA: math (-8.0%), code (+14.3%), STEM (+24.5%), and medical (+8.6%). All data and tools are open-source →read the paper

🌟 Treasure Hunt: Real-time Targeting of the Long Tail using Training-Time Markers
Improves performance on rare or niche inputs by training models with annotated control markers that modulate generation behavior at inference time →read the paper

Reasoning with Exploration: An Entropy Perspective
Improves LLM reasoning by encouraging exploratory thinking via entropy-based reward shaping in RL training →read the paper

Optimizing Length Compression in Large Reasoning Models
Reduces reasoning verbosity through targeted rewards for brevity and sufficiency in post-training optimization →read the paper

🌟 Steering LLM Thinking with Budget Guidance
Controls the length of reasoning chains during inference using a budget-aware predictor, improving token efficiency under constraints →read the paper

🌟 Truncated Proximal Policy Optimization
Speeds up reinforcement training of LLMs by truncating responses and optimizing policy-value decoupling →read the paper

LongLLaDA: Unlocking Long Context Capabilities in Diffusion LLMs
Extends the context window of diffusion-based LLMs with a training-free method and provides theoretical insights on their scaling behavior →read the paper

Xolver: Multi-Agent Reasoning with Holistic Experience Learning
Builds a reasoning agent that retrieves, collaborates, and learns from previous examples across multiple modalities, inspired by Olympiad teams →read the paper

AgentSynth: Scalable Task Generation for Generalist Computer-Use Agents
Generates scalable benchmark tasks by composing subtasks and logging LLM trajectories, enabling rigorous evaluation of generalist agents →read the paper

🌟 Reinforcement Learning with Verifiable Rewards
Shows that RL with verifiable rewards improves logical consistency in reasoning and proposes a new CoT-aware metric for evaluation →read the paper

🌟 Revisiting RL for LLM Reasoning from a Cross-Domain Perspective
Introduces Guru, a multi-domain RL corpus, and shows how domain-specific rewards improve reasoning generalization across math, logic, simulation, and more →read the paper

A Technical Study into 0.5B Reasoning Language Models
Improves reasoning in 0.5B-parameter models through a hybrid of supervised fine-tuning, distillation, and reinforcement learning →read the paper

🌟 Taming Polysemanticity in LLMs
Improves interpretability in smaller LLMs using sparse autoencoders with provable recovery guarantees for underlying features →read the paper

🌟 Microsoft Research: New Methods for Boosting Reasoning in Small and Large Models
Presents rStar-Math, Logic-RL, and Chain-of-Reasoning (CoR) frameworks for boosting symbolic and cross-domain reasoning in both small and large LLMs →read the paper

Show-o2: Improved Native Unified Multimodal Models
Unifies image, video, and text modeling via a 3D causal variational space and dual-path fusion for scalable multimodal understanding and generation →read the paper

SciVer: Evaluating Foundation Models for Multimodal Scientific Claim Verification
Benchmarks the claim verification abilities of foundation models in complex multimodal scientific settings, revealing large performance gaps →read the paper

Wait, We Don't Need to "Wait"! Removing Thinking Tokens Improves Reasoning Efficiency
Improves reasoning efficiency in multimodal settings by suppressing filler tokens like “Wait” during inference without loss in accuracy →read the paper

Discrete Diffusion in Large Language and Multimodal Models: A Survey
Surveys discrete diffusion modeling for language and multimodal systems, offering fast, controllable generation as an alternative to autoregression →read the paper