NeuralStackAI/TLDR’s latest paper roundup highlights a research landscape moving beyond conventional language-model demonstrations. The 73 tracked releases cover systems for biology, mathematics, robotics, coding agents, video world models and model architecture. Their common thread is an effort to make AI systems more capable of acting, reasoning or operating under real constraints.
Anthropic’s “Claude finds ART” describes hundreds of Claude agents searching DNA databases and identifying an enzyme system with CRISPR-like repeats that biologists had missed. In a separate project, Anthropic reported that Claude designed de novo protein binders for 15 targets, with two outside labs building and confirming binders for 14 of them. The same roundup also describes automated alignment researchers in which Claude agents studied prior work, ran training experiments and found fixes that outperformed experienced human researchers.
Mathematics is another prominent test bed. OpenAI’s agents produced a finite-time blowup proof for the three-dimensional Navier–Stokes equations that was verified in Lean, while Anthropic’s research version of Claude raised a classical Riemann zeta bound to 67.2% and supplied a machine-checkable Lean proof. OpenAI’s Astra also cracked 10 open mathematics and computer-science problems and released Lean-verified proofs.
The papers also focus on making AI systems cheaper and easier to deploy. Carnegie Mellon’s CliffCompaction is a drop-in proxy designed to keep long coding-agent sessions within a token budget without modifying the agent. DeepSeek V4.1 Flash is described as a 552B model that keeps only 890 bytes of cache per token, while Edge0 presents a 35B mixture-of-experts model that runs from an SSD in 2.9 GB by loading only the experts needed for each token.
Other work targets the environments in which agents learn. ScienceIDE turns real scientific code repositories into training environments where agents are graded on numerical correctness, and Google’s EnvHarness adapts a static agent benchmark into a training environment without changing the benchmark’s own code. Repo-To-Skill distills 1,000 machine-learning repositories into 5,000 verified, executable skills that research agents can load when needed.
Robotics and embodied AI appear throughout the list. PhysBrain 1.5 is an open 8B model that reads a scene, plans an arm’s next movement and predicts how the scene will look a second later. Show-Harness gives a vision-language model a semantic action vocabulary for controlling a robot arm without robot-specific pretraining, while Google’s AMIE extends medical consultation work to synchronous video visits and guides patients through a virtual physical exam.
World models are also becoming more interactive. SolarWM opens data, training code and checkpoints for video world models designed to remain interactive for hours, while JD.com’s EchoWM combines images, ambient sound and speech in a world model users can move through. These projects suggest that the next challenge is not merely generating a plausible frame, but maintaining a coherent environment as an agent or person navigates it.
Why this matters
The roundup shows AI research spreading across the full cycle of discovery: finding biological systems, proposing and verifying mathematics, training agents, controlling physical devices and reducing infrastructure costs. Machine-checkable proofs and outside laboratory validation provide stronger forms of evaluation than demonstrations alone, while efficient caching, SSD-based inference and token-budget proxies address the expense of running long-lived agents. If these approaches mature, progress may depend as much on reliable tools, environments and verification loops as on larger base models.
The forward path remains open-ended. The highlighted work includes early reports, previews and research systems, so their broader usefulness will depend on replication, wider testing and integration into dependable workflows. But the direction is clear: AI research is increasingly treating models as collaborators that search, plan, experiment and operate—rather than as systems that only answer prompts.