digest: 2026-06-12

[JustinPerea]

Items (8)

1. DiffusionGemma (DeepMind blog, June 10) — 26B-total / 3.8B-active MoE text-diffusion LM on Gemma 4 backbone; 256 tokens/forward pass via bi-directional attention; 1000+ tok/s H100, 700+ tok/s RTX 5090; MLX support at launch (~18 GB at NVFP4 4-bit); Apache 2.0. try-now + local-fit.
2. EvoArena + EvoMem ([arxiv:2606.13681], HF digest: 2026-05-06 #2 54 upvotes) — first dynamic-environment memory benchmark; agents avg 39.6%; EvoMem +1.5/+6.1/+4.8% on EvoArena/GAIA/LoCoMo. track.
3. HyperTool ([arxiv:2606.13663]) — MCP code-block tool primitive; Qwen3-32B 15.69→35.29% (+19.6pp) on MCP-Universe; schema-compatible with existing MCP servers. try-now.
4. The Illusion of Multi-Agent Advantage ([arxiv:2606.13003]) — auto-MAS underperforms CoT-SC at 10× cost; expert-architected MAS still wins. compare.
5. EurekAgent ([arxiv:2606.13662]) — environment engineering > workflow prescription; SOTA 26-circle packing at ~$11 API. track.
6. SENTINEL ([arxiv:2606.12908]) — failure-as-curriculum RL; Tau2-Bench Retail 66.4→74.9%. track.
7. G-Long ([arxiv:2606.13115]) — associative-graph dialogue memory; +9.8% MSC, 40.8% LME recall. track.
8. MemRefine ([arxiv:2606.13177]) — LLM-judge memory compaction (similarity proposes, content judges). track.

Focus areas covered

Frontier models / Local Apple Silicon (1), Agent memory (3, 7, 8 + 2 partial), Tools & connectors / MCP (3, 6), Agent orchestration (4, 5), Evaluation (2, 4), Reasoning & inference-time compute (6).

Conflicts flagged (2)

• Auto-MAS critique vs expert-architected MAS positive thread — resolved to "different categories"; Arbor/MUSE-Autoskill survive the critique.
• DiffusionGemma 4× speedup is GPU-only — open empirical question whether it transfers to M5 Max under MLX; direct head-to-head vs Qwen3-30B-A3B is the right experiment.

State files updated: _meta/processed.json (8 new IDs newest-first), _meta/claims-ledger.md (8 new entries).

https://claude.ai/code/session_013XDzaQC6LRhiyU91yxzUoq

---

Generated by Claude Code