On-policy RL has driven the biggest leaps in training coding agents. Extending it to machine learning engineering agents should be a natural next step.

But it almost never works.

The recipe is right there standard trajectory-wise GRPO, the same that worked for SWE.

However, the problem is that one rollout step on an MLE task may take hours because the agent has to actually train a model on a real dataset at every step (preprocessing, fitting, inference, scoring). So even with the N rollouts in a group running in parallel, a single GRPO run may still take days.

Meta shared a new paper, SandMLE, which fixes this with a move that sounds almost too reckless to work.

DeepSeek is rolling out a limited V4 gray release. A new mode switcher now appears in the chat UI with three options: Fast Mode (default), Expert Mode and Vision Mode

Fast Mode:
• File uploads → text-only extraction
• Likely a lightweight, low-latency model optimized for speed

Expert Mode:
• No file uploads supported
• Restriction likely for compute/cost control, since heavy models + file tokens are expensive
• Likely routes to a larger, more powerful reasoning model

Vision Mode:
• Enables multimodal inputs
• Builds on earlier OCR tests
• May signal DeepSeek’s multimodal capability is moving toward end users

Attackers can exfiltrate user files from Cowork by exploiting an unremediated vulnerability in Claude’s coding environment, which now extends to Cowork.

The vulnerability was first identified in Claude.ai chat.

Anthropic just introduced Project Glasswing: an urgent initiative to help secure the world’s most critical software.

It’s powered by newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.

Anthropic partnered with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.

Let's talk about Claude Mythos Preview

This chart shows an evaluation testing models’ ability to successfully exploit vulnerabilities in Firefox 147.

There are three grade levels: 0 for no progress, 0.5 for partial control (controlled crash), and 1.0 for full code execution.

From system card:
1. in ~29% of evaluations, it realized it was being tested, and didn't say so.

2. when an LLM was used to judge its work and kept rejecting it, Mythos identified the evaluator is an LLM, and prompt-injected it.

3. in one test, it saw the answer to a problem it was solving, and intentionally widened the confidence interval to not raise suspicion.

4. when it needed a file permission it didn't have, it found and used a "privilege escalation vulnerability" and then programmed it to delete itself so it doesn't show in the logs.

5. it escaped a sandbox container (escaping sandbox test so not unexpected), then emailed the researchers about it, and without being told to, posted the details to some hard-to-find but public websites, bragging about its success.

6. when Claude Code blocked it from using some permissions, the model acknowledged the block was valid, but then immediately tried to perform the same operation using different commands.

7. when asked to find security bugs, earlier versions planted bugs in the code, and reported them as pre-existing.

The capability slope we’re going to keep seeing from the frontier labs is going to open up all new use cases in finance, healthcare, legal, consulting, supply chains, and more.

Make sure you’re building something that can take advantage of these upcoming improvements, or you’ll be in a tough spot strategically.

It’s a big. Morgan Stanley officially announced the launch of its spot Bitcoin ETF

Morgan Stanley Investment Management is the first U.S. bank-affiliated asset manager to offer a cryptocurrency ETP, and reflects a continued, firmwide focus by Morgan Stanley to develop digital asset solutions designed to meet evolving client demand.

Meta just released Muse spark, the first model from MSL team

Muse spark is a natively multimodal reasoning model w/ support for tool-use, visual chain of thought, & multi-agent orchestration. Through its training process, team saw predictable scaling across pretraining, RL, & test-time reasoning.

Also released contemplating mode, which orchestrates multiple agents that reason in parallel designed to handle complex scientific & reasoning queries. In testing team found it competitive w/ other extreme reasoning models such as Gemini Deep Think & GPT Pro.

Also found muse spark demonstrated strong refusal behavior across high-risk domains such as biological and chemical weapons.

Meta ai now handles quick answers and deep reasoning with instant and thinking modes.

Shopping mode is new too it picks up on the creators, brands, and styling content across our apps and turns that into recommendations.

Bigger models are already in development with infrastructure scaling to match.

Private api preview open to select partners today, with plans to open-source future versions.

Alibaba published a paper that shows AI is moving beyond bug finding and into actually proving software is exploitable.

This paper asks a simple question with hard consequences: can LLMs confirm software vulnerabilities by actually building working exploits?

The authors’ answer is yes, but only when the model stops acting like a single genius and starts acting like a team.

Meta presented a world model that models the computer

What if AI could invent enzymes that nature hasn’t seen? Meet DISCO: Diffusion for Sequence-structure CO-design

14 rounds of directed evolution and over a year of wet lab work.

That's what it took to engineer an enzyme for selective C(sp³)–H insertion, one of the most challenging transformations in organic chemistry.

DISCO surpasses this with a single plate. No pre-specified catalytic residues, no template, no theozyme, no inverse folding, just joint diffusion over protein sequence and structure.

Paper
Code

Tencent released HY-Embodied-0.5, a family of foundation models for real-world embodied agents. The 2B model is now open source.

The suite includes:

2B for edge deployment
32B for complex reasoning

Key innovations:
1. Mixture-of-Transformers (MoT) architecture for modality-specific computation
2. Latent tokens for improved perceptual representation
3. Self-evolving post-training
4. On-policy distillation from large to small models

Across 22 benchmarks, the 2B model outperforms similarly sized SOTA systems on 16 tasks. The 32B model approaches frontier-level performance.

GitHub
Hugging Face

Polygon Labs is in early talks to raise up to $100 million to launch a new stablecoin payments business, according to sources.

It's rare for a blockchain developer to enter regulated payments business. With this move, Polygon hopes to drive stablecoin volume on its blockchain.

In Jan., Polygon Labs agreed to acquire Coinme and Sequence, positioning to compete with the likes of Stripe

Cool work. R-Zero - self-evolving LLM from zero external data.

One base model, two roles:

1. Challenger generates hard problems

2. Solver solves them.

Challenger is rewarded when Solver fails. Co-evolve with GRPO. Challenger learns to probe for weaknesses, not just generate hard problems.

+6.49 math, +7.54 general reasoning on Qwen3-4B-Base. 3 iterations, no human data.

China defined what an AI Hospital is

It is a new type of smart healthcare model in which AI is embedded into the system itself, linking offline medical expertise with the broader reach of online services to deliver more proactive and continuous care.

Patients become the point-of-care with the help of AI.

Google presented Sparse Selective Caching, an architecture with growing effective memory (similar to attention) but with almost constant inference cost per token (similar to RNNs).

In the paper team mainly discuss:

1) the shared foundation for both softmax attention and fixed-size long-term memory modules (or RNNs) that helped design an architecture with best of both worlds;

2) different variants of memory caching, including a variant whose effective memory is growing while the decoding cost still remains “constant”;

3) a unifying perspective to understand hybrid models, in which attention and recurrent models are combined.

Together AI presents Introspective Diffusion LM

The first DLM to match the quality of AR while outperforming DLMs in both model quality and serving efficiency.

Delivering about 3× higher throughput than prior SotA DLMs.

GitHub
Model

Turns out we can get SOTA on agentic benchmarks with a simple test-time method

Meet LLM-as-a-Verifier

Test-time scaling is effective, but picking the "winner" among many candidates is the bottleneck. This way to extract a cleaner signal from the model:

1. Ask the LLM to rank results on a scale of 1-k
2. Use the log-probs of those rank tokens to calculate an expected score

You can get a verification score in a single sampling pass per candidate pair.

Code

Tether launches self-custodial wallet for end users

The wallet supports USDT, USAT, XAUT and Bitcoin across Ethereum, Polygon, Arbitrum, Plasma, and Bitcoin / Lightning Network, and enables transfers via human-readable usernames such as name@tether.me.

Tether said more than 570 million wallets were already using its technology as of March 2026.