Today in AI

The agentic AI wave is accelerating faster than the infrastructure to support it, and today's research reveals both the strain and the ingenuity emerging from that tension. The standout is Tev, a training framework that proves dense progress rewards can let a 27B parameter model punch at the weight class of frontier systems like GPT-6 and Claude Opus 5 on long-horizon tasks. This is not incremental; it's a genuine efficiency breakthrough that suggests the next leap in capable agents may come from better reward shaping rather than brute-force scaling. Equally practical is ChunkAct, which attacks the inference cost problem head-on by chunking robot actions and cutting LLM calls by nearly 3x while actually improving success rates. For anyone building production robotics, this is the kind of result that changes roadmaps. Meanwhile, Hybrid Reasoning Agents shows that the field is finally getting serious about token economics—delivering double-digit performance gains while slashing token use by up to 55% through clever switching between latent and explicit reasoning modes.

What ties these threads together is a maturing focus on *sustainable* agentic intelligence: systems that don't just work in demos but can run economically and adapt in the wild. Self-Evolving Agent Harness pushes this furthest, offering a framework where agents rewrite their own prompts, memory structures, and workflows without touching model weights—complete with formal reliability guarantees. This is the closest we've seen to genuinely autonomous self-improvement that doesn't invoke the usual safety hand-wringing. On the scientific front, the recursive self-improvement approach to Earth observation in "The Last Remote Sensing Model Built by Humans?" is audaciously titled but earnestly so: it demonstrates how AI agents can automate model development for satellite imagery with minimal human labeling, essentially bootstrapping their own supervision. The rest of today's papers—from data-free diffusion for lattice field theories to aesthetic code generation—fill in the methodological edges, but the core story is agents learning to do more with dramatically less.

**Bottom line:** The frontier is shifting from raw capability to agentic efficiency—today's best research proves that smarter training, cheaper inference, and self-modifying architectures are now delivering gains that rival or exceed another generation of model scaling.