ZeroSlop — May 28, 2026
12 stories worth knowing about today — AI breakthroughs, launches, and innovations making a difference.
MarkTechPost
NVIDIA Releases Polar, a Token-Faithful Rollout Framework for GRPO Training Across Codex, Claude Code, and Qwen Code
NVIDIA’s Polar framework is a game-changer for AI agent training—it lets researchers run reinforcement learning on code-generating models without touching the underlying harness, capturing every token interaction to build better training data. By deploying Polar with GRPO on a modest 3.5B model, the team achieved massive gains on real-world coding benchmarks, with some harnesses seeing 22+ point improvements on SWE-Bench Verified. This open-source breakthrough (now available in NeMo Gym) means faster, more flexible agent development across multiple platforms and models.
SecurityWeek
RevEng.AI Raises $15 Million to Hunt for Flaws and Backdoors in Software Binaries
RevEng.AI Lands $15M to AI-Hunt Software Vulnerabilities Before Attackers Do
RevEng.AI is weaponizing AI to reverse-engineer and expose hidden flaws in software binaries—catching security threats that traditional tools miss. With its BinNet model now backed by $15 million in fresh funding, the company is scaling up a critical defense mechanism that could shift how enterprises protect their supply chains and prevent zero-day exploits from ever reaching users.
The Guardian Tech
Are robots nearing their ChatGPT moment? – podcast
Are robots nearing their ChatGPT moment?
A robot just crushed the world record at a half marathon, and it’s a sign of something bigger: we’re entering a pivotal moment where AI-powered machines could transform from lab curiosities into everyday tools, with China betting over £100 billion to make it happen. The real challenge isn’t speed—it’s dexterity and real-world deployment, and experts are diving deep into what it’ll take to get robots into our homes, workplaces, and gardens for good.
AWS Machine Learning
Building AI agents for business support using Amazon Bedrock AgentCore
AWS and Works Human Intelligence just demonstrated how AI agents can slash business support costs by nearly 97% while boosting efficiency—proving that smart agent architecture, not just raw model power, drives real ROI. Using Amazon Bedrock AgentCore, they built two production agents that tackle concrete operational challenges, offering a blueprint for enterprises ready to move beyond chatbots to genuinely autonomous systems. The breakdown of their approach—from design to deployment—shows that the next wave of AI value isn’t coming from bigger models, but from agents purpose-built to handle your actual workflows.
arXiv CS.AI
A Policy-Driven Runtime Layer for Agentic LLM Serving
Researchers have identified a critical gap in how today’s AI serving systems handle multi-agent LLM workloads—and proposed an elegant fix: a new intermediate runtime layer that bridges the disconnect between high-level agent frameworks and low-level serving engines. This architectural innovation could unlock smarter resource allocation, better safety enforcement, and more efficient execution strategies that currently require messy workarounds scattered across the stack. It’s a foundational move that could reshape how production AI systems scale.
arXiv CS.AI
Intelligence as Managed Autonomy: Failure, Escalation, and Governance for Agentic AI Systems
Intelligence as Managed Autonomy: Failure, Escalation, and Governance for Agentic AI Systems
Researchers have cracked a fundamental problem plaguing autonomous AI systems: instead of blindly pushing forward when uncertain, intelligent agents should know when to pause, recover, and hand control back to humans. This breakthrough framework—managed autonomy—transforms how we think about AI reliability by treating the ability to fail gracefully as a core feature of intelligence, not a bug.
arXiv CS.AI
Got a Secret? LLM Agents Can’t Keep It: Evaluating Privacy in Multi-Agent Systems
Got a Secret? LLM Agents Can’t Keep It
Researchers have uncovered a critical privacy vulnerability in multi-agent AI systems: when LLMs interact socially over time, they leak sensitive information at nearly double the rate of isolated models, and this behavior spreads contagiously between agents. The finding reveals that current safety testing falls dangerously short—evaluating agents in the real-world conditions where they’ll actually operate exposes major flaws that single-model tests completely miss.
TechCrunch AI
AI coding startup Cognition raises $1B at $25B pre-money valuation
Cognition’s $1B Funding Round Signals Explosive Growth in AI-Powered Development
Cognition just pulled in $1 billion at a $25 billion valuation—more than doubling its worth in eight months—as its AI coding assistant rockets toward nearly half a billion dollars in annualized revenue. The funding surge underscores how rapidly enterprises are adopting AI tools that genuinely boost developer productivity and ship better code faster. This isn’t just a valuation milestone; it’s proof that AI’s most practical, revenue-generating applications are emerging right now in the developer tools space.
arXiv CS.AI
LaneRoPE: Positional Encoding for Collaborative Parallel Reasoning and Generation
LaneRoPE: Making AI Models Smarter Through Collaborative Generation
Researchers just cracked a major efficiency challenge in AI reasoning: LaneRoPE lets language models generate multiple solution paths simultaneously while learning from each other in real-time, rather than working in isolation. By weaving together inter-sequence attention and a clever positional encoding trick, this breakthrough dramatically boosts accuracy without burning through extra compute—turning parallel generation into genuinely collaborative problem-solving. This is a game-changer for test-time scaling, finally unlocking the full potential of batched reasoning that researchers have been chasing for years.
arXiv CS.AI
DeepSciVerify: Verifying Scientific Claim–Citation Alignment via LLM-Driven Evidence Escalation
DeepSciVerify Tackles AI’s Most Dangerous Flaw: Making Up Citations
Researchers just cracked a critical problem holding back AI in science—LLMs confidently citing papers that don’t actually support their claims. DeepSciVerify, a smart two-stage verification system, combines quick abstract-level checks with deep-dive passage analysis to catch misalignments before they spread, hitting 86.7 Micro-F1 on key benchmarks and proving AI can police its own evidence. This matters because it transforms LLMs from unreliable storytellers into tools trustworthy enough for real scientific and high-stakes work.
arXiv CS.AI
Prefix-Safe Bayesian Belief Tracking for LLM Reasoning Reliability
Researchers have cracked a crucial problem in AI reasoning: reliably predicting whether a language model will arrive at the correct answer mid-reasoning, before it finishes thinking. By developing Sequential Bayesian Belief Tracking (SBBT), they’ve created a unified framework that works across multiple signal types—from numerical scores to text patterns to hidden model activations—and discovered that calibration and ranking accuracy require fundamentally different evidence. This breakthrough matters because real-time confidence estimates could transform how we deploy AI on high-stakes problems like advanced math, enabling systems to know when they’re likely to fail and call for human help before committing to a wrong answer.
OpenAI News
Warp’s big bet on building open source with GPT-5.5
Warp’s big bet on building open source with GPT-5.5
Warp is betting that GPT-5.5 can be the connective tissue between fragmented developer tools—orchestrating coding agents across local machines, cloud platforms, and open-source projects in one unified workflow. This move could eliminate the friction that keeps developers context-switching between incompatible environments, finally making seamless cross-platform collaboration real. If it works, it’s a genuine productivity unlock for teams juggling complexity.