digest: 2026-05-24

[JustinPerea]

Summary

• 10 items; UTC date 2026-05-24 (Sunday — slow announcement day, but harvested a missed frontier release + 9 substantive arXiv papers from the May 22 listing not picked up yesterday)
• Focus areas covered: Frontier models, Long-running / autonomous loops, Tools & connectors, Context engineering, Agent orchestration, Evaluation, Reasoning & inference-time compute
• Conflicts flagged: 1 (compare) — Gemini 3.5 Flash 76.2% Terminal-Bench 2.1 vs. [arxiv:2605.22535] TerminalWorld 62.5% in-the-wild ceiling with Pearson r = 0.20 between the two benchmarks

Item highlights

1. Gemini 3.5 Flash (Google, May 19) — missed by prior digests; frontier-fast model claiming 76.2% Terminal-Bench 2.1, 84.2% multimodal, ~4× faster than other frontier models
2. MOSS (arxiv:2605.22794) — agent self-evolution via source-level code rewriting; 0.25 → 0.61 on OpenClaw in a single cycle; 6th layer above SkillOS / SkillsVote / EvolveMem / Ratchet / Continual Harness
3. HarnessAPI (arxiv:2605.22733) — single typed Python source-of-truth auto-derives MCP tools + streaming HTTP endpoints; 74% boilerplate reduction; pip install harnessapi — try-now
4. Gated DeltaNet-2 (NVIDIA, arxiv:2605.22791) — channel-wise decoupled erase + write gates; beats Mamba-2/Gated DeltaNet/KDA/Mamba-3 at 1.3B on long-context RULER NIAH
5. LCGuard (arxiv:2605.22786) — first paper defining KV-cache leakage as a security boundary in multi-agent systems; fourth angle on the agent-safety surface
6. Boiling the Frog (arxiv:2605.22643) — multi-turn agentic safety bench; 44.4% aggregate ASR; 93.3% on loss-of-control scenarios; Claude Haiku 4.5 strongest (20.5%), Gemini 3.1 Flash Lite weakest (92.9%)
7. Capability as Liability (arxiv:2605.22672) — inverse scaling on tail-risk forecasting; threshold scoring reverses sign vs. tail-inclusive scoring on identical outputs
8. Search-E1 (arxiv:2605.22511) — minimal GRPO + offline self-distillation recipe; Qwen2.5-3B reaches 0.440 avg EM on 7 QA benchmarks
9. Shor (arxiv:2605.22505) — priority-ranking step-level eval of harness optimizers; seventh paper in the agent-eval credibility thread, attacks evaluator cost
10. VPO (arxiv:2605.22817, MIT — Khattab/Agrawal) — vector-valued rewards for diversity-preserving RL; solves problems GRPO models cannot solve at all under evolutionary search

Conflicts flagged

• compare — Gemini 3.5 Flash 76.2% Terminal-Bench 2.1 vs. [arxiv:2605.22535] TerminalWorld 62.5% in-the-wild frontier ceiling. The r = 0.20 Pearson correlation TerminalWorld reported between the curated benchmark and in-the-wild distribution means the 76.2% headline is unverified as a real-world capability signal until a TerminalWorld run is published for Gemini 3.5 Flash. Extends the existing agent-benchmark credibility thread (AuditRepairBench, Rollout Cards, Evidence-Supported Bounds, BenchJack, HarnessAudit-Bench, Counterfactual Trace Auditing, Shor) to frontier-model release announcements.

Operational notes

• OpenAI news / research: HTTP 403 on both openai.com/news/ and openai.com/research/ — could not access — retry next run
• Local-model corner: no purely local model release today; HarnessAPI (Mac-installable framework) and Gated DeltaNet-2 (1.3B architecture pattern) are the indirect items
• Sunday is the natural slow point in the arXiv announcement cycle (no Sat/Sun announcements in most cs.* categories) — today's batch is the residual of Friday's listing not covered in yesterday's digest

https://claude.ai/code/session_01TQtJp3suuBwXXx2W2SMAPf

---

Generated by Claude Code