170,000 Papers Later: What the First Half of 2026 Reveals About AI Research

When a small research-explainer outfit fed 170,927 arXiv papers into a keyword classifier this summer, it was not trying to referee the field. It was trying to take its temperature. The resulting snapshot — an independent analysis by AI Papers Academy of every paper posted to arXiv's four main machine-learning categories from the start of 2025 through June 26, 2026 — has since ricocheted across research feeds because it puts numbers to something practitioners had only felt: the center of gravity in AI research has moved decisively from building bigger models to making models cheaper, more reliable, and able to act.

The analysis is not an official dataset, and its authors are candid about the method. It matches keywords against titles and abstracts using a curated list of topics, model families, and institutions, then splits the window into three half-year slices — H1 2025, H2 2025, H1 2026. Because the field grew roughly 25 percent overall, the report tracks each topic's share of papers rather than raw counts, so a topic can grow in absolute terms and still lose ground. With those caveats, the picture is unusually coherent.

From "can we build agents?" to "how do we make them plan?"

The single loudest signal is agents — not the concept, which is old news, but the infrastructure beneath it. Mentions of agentic workflows climbed from 4,585 papers to 10,496 over the window. More telling are the specialized building blocks growing two to five times faster. Long-horizon planning, the problem of holding a goal across many steps without losing the thread, rocketed from 264 to 1,611 mentions, a 510 percent jump that made it the fastest-growing topic in the entire dataset.

"The field has moved from 'can we build agents?' to 'how do we make agents plan, reason, use tools, and judge their own outputs?'" the AI Papers Academy analysis concludes. That reframing shows up everywhere in the underlying literature. arXiv is now thick with papers on context compression for long-horizon agents, agent memory benchmarks, tool-selection accuracy, and self-governing context management — the unglamorous plumbing that separates a demo from a deployable system.

The academic conferences confirm the drift. A roundup of ICLR 2026 work by the infrastructure firm Lambda notes that the accepted papers cluster around "making AI systems reliable, efficient, and secure," with recurring themes of "long-horizon agentic planning under sparse rewards, alignment with safety constraints, structured world modeling, and inference-time efficiency." Those four phrases could double as a table of contents for the year.

Efficiency is the new frontier

If agents are the destination, efficiency is the vehicle. The second clear trend across the corpus is a pivot away from scale-for-its-own-sake toward compression, distillation, and selective computation. Researchers are chasing architectures that activate only the parameters a given problem needs — selective activation sparsity — with some reported models matching the reasoning performance of peers three times their size. The economic stakes are concrete: separate mid-year reporting notes that the cost of running large models has dropped by up to 100x in specific settings, and a 30-billion-parameter open model reached 64 percent on a real coding benchmark, territory that belonged to frontier systems a year ago.

The efficiency turn also reshuffled the open-model pecking order. Alibaba's Qwen nearly doubled its footprint in the literature, from 752 to 1,489 mentions, overtaking Meta's Llama (1,085 to 1,232) to become the most-referenced open model family in Western research for the first time. Google's small open model Gemma posted the fastest percentage growth of any family at 147 percent. Researchers, in other words, are voting with their citations for accessible, capable bases they can fine-tune and distill cheaply.

Kevin Indig, writing in his Growth Memo "AI Halftime Report" for H1 2026, framed the broader stakes bluntly: the defining story of mid-2026 is "the agentic boom, Qwen overtaking Llama, and what's fading." What's fading is instructive. Synthetic data lost 24 percent of its share and shrank in absolute terms, a possible signal of quality fatigue and worries about models trained on model output drifting from reality. State-space models like Mamba lost share as their ideas were absorbed into hybrid architectures rather than standing alone.

What the shift actually means

Read together, these trends describe a field growing up. The scaling era rewarded whoever could afford the largest run; the emerging era rewards whoever can make a modest model behave reliably over long, multi-step tasks and prove it on a benchmark that mirrors real deployment. That is why credit-assignment research — figuring out which action in a long chain deserves reward when feedback is sparse — has become a hot subfield spanning reasoning and agentic reinforcement learning, and why evaluation and domain-specific benchmarks are proliferating alongside the models themselves. The prestige is migrating from "biggest" to "most dependable."

What to watch

The obvious caution is that this is one team's keyword analysis, not a census; a classifier that leans on titles and abstracts will overweight whatever language researchers happen to trend toward. Watch whether AI Papers Academy's promised recurring "AI Research Pulse" holds its numbers up to scrutiny across future editions. Watch, too, whether the agent-reliability and efficiency gains translate into shipped products or stay trapped in benchmark tables. And keep an eye on the fading categories: if synthetic data keeps shrinking, the industry's favorite answer to the data-scarcity problem may be quietly running out of road.