Latest AI model releases, papers, and open-source projects (May 25 to 26, 2026)

On the paper layer, May 25 was dominated by Microsoft Research's SkillOpt (arxiv 2605.23904, 1.32k Hugging Face upvotes), with Microsoft Lens (181), UCL's SciAtlas (106), Rethinking Cross-Layer Information Routing in DiTs (102), and Alibaba Tongyi's See What I Mean (88) rounding out the top five. May 26 shifted to agent-and-world-model benchmarks: TriSplat (186), DVAO (128), WBench (96), Macaron-A2UI (77), Foundation Protocol (75), with computer-use-agent papers Macaron-A2UI, CUA-Gym (26), and Claw-Anything (21) all trending on the same day. On the open-source agent layer, the macOS computer-use agent Fazm shipped v2.9.37 on May 25 (agent guardrails for system-altering shell commands, a third browser mode No browser MCP, Assrt only) and v2.9.41 on May 26 (16 user-facing changes packaging 80+ commits, the May 25 OAuth fix, the verification-codes skill, and the cold-boot menu-bar fix). At 11:33 PDT on May 26, Fazm also deleted 116 of 120 lines from its standalone Claude OAuth window in a structural refactor that inlined Claude auth into the model picker.

Five papers in this window are directly load-bearing for anyone running a computer-use agent. SkillOpt (May 25) shows that a compact natural-language skill document, treated as the trainable state of a frozen agent, buys +19.1 points inside Claude Code on GPT-5.5 over no-skill baselines; this generalizes to any Claude Code wrapper, including Fazm. Macaron-A2UI (May 26) is a generative-UI model for personal agents and is the right reference for any agent that has to produce or restyle UI on the fly. Foundation Protocol (May 26) is a coordination-layer proposal for inter-agent communication. CUA-Gym (May 26) is a scaling-verifiable training environment for computer-use agents (the exact category Fazm sits in). Claw-Anything (May 26) benchmarks always-on personal assistants. If your agent is a real computer-use agent, three of those four May 26 papers are pointing at you.

Two days, two centers of gravity. May 25's top of mind was efficiency for training (SkillOpt's skill-document optimizer, Lens's training-efficiency rework for text-to-image, the diffusion-transformer routing paper). The next day skewed toward agentic and world-model evaluation: TriSplat for simulation-ready 3D reconstruction (any embodied agent benchmark needs simulation-ready geometry), WBench for interactive video world models, plus three explicitly agent-shaped benchmarks (Macaron-A2UI, CUA-Gym, Claw-Anything). The simple read is that the May 25 batch was about lowering the cost of producing a frontier model, and the May 26 batch was about measuring what current frontier models actually do once you put them inside an agentic harness. Both kinds of work landed in the same 48 hours on the same dated index.

The agent project with the densest commit history in this window is Fazm (github.com/m13v/fazm). It tagged two releases: v2.9.37 on May 25 (the guardrails-and-no-browser-MCP release) and v2.9.41 on May 26 (the OAuth-fix-and-onboarding release). Between the two, 80+ commits landed on the main branch. The structural commit of the window is c304a3c3 on May 26 at 11:33:09 PDT, which deletes 116 of 120 lines from Desktop/Sources/Chat/ClaudeAuthSheet.swift. The standalone Claude OAuth window is gone; auth now renders inline in the same PersonalAccountChooser sheet that Codex and ChatGPT already used. A second wave at 19:46 to 19:59 PDT removes the pulsing 'Connect Claude' button entirely, replacing it with an inline entry in the model picker (the model dropdown now shows 'Claude - Connect...' the way it already showed 'Codex - Connect...' before). The whole 48-hour cycle is a worked example of platform-side change driving structural UX change in a downstream client.

The verification-codes skill is a 73-line skill file at Desktop/Sources/BundledSkills/verification-codes.skill.md that teaches the agent to pull a one-time code (SMS, iMessage, 2FA, OTP) from local channels instead of asking the user to read it. The order of operations is Messages.app via raw SQLite against ~/Library/Messages/chat.db (with a 10-minute freshness window to avoid stale codes), then WhatsApp via the bundled whatsapp MCP, then macOS Notification Center via the macos-use MCP. The reason this is interesting in the May 25 to 26 window is that it is a concrete example of work no headless or cloud computer-use agent can do: reading the Messages SQLite database requires Full Disk Access granted to the local app, and reading WhatsApp requires Accessibility granted to a local app. It is the kind of capability that only ships when the agent is a real native macOS process with the right TCC entitlements, and the Claw-Anything paper from the same week is essentially benchmarking the absence of capabilities like this in current always-on assistants.

Every number on this page traces to a primary source. The Hugging Face dated indexes are at huggingface.co/papers/date/2026-05-25 and huggingface.co/papers/date/2026-05-26; upvote counts on dated indexes shift slowly over weeks, but the ranks and arxiv IDs are stable. The Fazm commit log is at github.com/m13v/fazm/commits/main; filter by date 2026-05-25 to 2026-05-26 and you will see the 80+ commits. The release manifests are in CHANGELOG.json at the repo root (search for 2.9.37, 2.9.41, 2.9.42 entries). The standalone OAuth window deletion is commit c304a3c3 and you can run git diff c304a3c3^ c304a3c3 -- Desktop/Sources/Chat/ClaudeAuthSheet.swift to read the diff directly. The verification-codes skill file is at Desktop/Sources/BundledSkills/verification-codes.skill.md in the same repo.

Same break, different stage. The May 25 entry was the moment of discovery: two commits at 17:50:10 and 17:52:10 PDT on May 25 land the fix in Fazm's acp-bridge module, removing the rejected expires_in field from the OAuth token-exchange body. Those commits stayed un-tagged through the rest of May 25 and most of May 26. v2.9.41 on May 26 is the tagged release that finally ships them to end users, alongside 15 other unrelated user-facing changes that accumulated during the gap. The 24-hour delay between commit and tagged ship is itself a signal: an aggregator that only watches release feeds learned about the May 25 OAuth policy change a full day later than an aggregator that watches commit subjects in real time.

Three places, in order of decreasing aggregator coverage. First, dated trending feeds (huggingface.co/papers/date/2026-05-25, /2026-05-26) plus the major lab news pages. This catches SkillOpt, TriSplat, DVAO, the WBench category, and the v2.9.37 and v2.9.41 tags. Second, the CHANGELOG.json (or equivalent release manifest) of one or two open-source agent projects you actually run. This catches the verification-codes skill, the cold-boot menu-bar fix, the 10-minute inactivity cap removal, and other concrete user-facing improvements that arrive in the body of a release rather than the headline. Third, the git commit log of those same projects, read by date range. This is where the structural refactors live (the May 26 11:33 PDT standalone-window deletion, the 19:46 to 19:59 PDT Connect-Claude peel removal) and where you can see platform-side policy enforcement reflected as same-day client fixes across multiple repos (the May 25 OAuth break is the clean example).

Yes, through the deep-research skill that auto-installs to ~/.claude/skills/deep-research/SKILL.md on first launch. The skill runs an 8-phase pipeline (Scope, Plan, Retrieve, Triangulate, Synthesize, Critique, Refine, Package) with parallel web searches, parallel research subagents, citation verification where possible, and a markdown plus HTML plus PDF report into a dated folder under ~/Documents. Because it runs as your local Claude Code, Codex, or Gemini agent on your machine, the answer is grounded against fresh searches from your IP, not a cached newsletter summary. Combined with the verification-codes skill from v2.9.41 and the persistent-session architecture, the dated-window question is something you can answer on a Tuesday morning without leaving the app.