digest: 2026-05-21

[JustinPerea]

Summary

9 items, 1 conflict flagged (AI-as-reviewer vs AI-as-author asymmetry).

Focus areas covered

• Agent memory (1): Mem-π — generative memory with decision-content decoupled RL; >30% web nav improvement over retrieval baselines. 13th orthogonal memory-mechanism axis on the ledger.
• Context engineering / Local & open models (3): OScaR (HKU, INT2 near-lossless KV, 5.3× memory reduction, code released), OCTOPUS (Stability AI, octahedral rotation codec, multimodal), Mix-Quant (NUS, prefill-only FP4, 3× prefill speedup).
• Agent orchestration / Tools & connectors (2): IBM Temporal Semantic Cache for MCP plan-execute (1.67× workflow speedup, 30.6× median cache-hit speedup, fixes time-sensitive cache failure mode); DeepMind Co-Scientist (Nature publication; specialized hypothesis-generation / critique / ranking agents).
• Coding agents / Evaluation (2): SpecBench (UCL; reward hacking in long-horizon coding agents; 28-pp gap per 10× code size; 2,900-line memorizing hash-table "compiler" failure mode), AI Reviewers Nature paper (GPT-5.2 60% > 48.2% top human, 16 recurring AI-only weaknesses).
• Reasoning / inference-time compute (1): RELEX — RLVR weight trajectories are rank-1; 15% steps for full performance, 10–20× extrapolation.

Conflicts flagged

• AI-as-reviewer (item digest: 2026-05-12 #9) vs. AI-as-author (last week's [arxiv:2605.19156] ResearchArena) — asymmetry, not contradiction. GPT-5.2 reviewer beats top human reviewer (60.0% vs 48.2%, p = 0.009) on Nature-family papers; same period, 0/117 generated papers under artifact-aware review meet top-tier venue standards. Both can be true: AI is competitive as reviewer, still failing as author. Operational signal: review/critique products are closer to production-ready than full-artifact-generation products. Open empirical question: does the AI-reviewer advantage hold when AI is reviewing AI-authored papers?

Cross-week sharpenings (not conflicts)

• KV-cache codec convergence — OScaR (digest: 2026-05-06 #2, per-channel canalized rotation) and OCTOPUS (digest: 2026-05-07 #3, octahedral triplet quantization) land same week with different geometries. Track for third entrant that benchmarks both head-to-head.
• Three independent low-rank linearity findings in the agentic LLM stack — [arxiv:2605.07990] (tool calling linearly readable in logits), [arxiv:2605.14038] (tool-necessity orthogonal probes), RELEX (digest: 2026-05-10 #7, rank-1 weight updates). Three different surfaces converging on dominant linear structure.
• Agent-eval-credibility cluster now at 10+ papers in 4 weeks with SpecBench (digest: 2026-05-09 #6) adding the 28-pp/10×-code-size visible↔held-out gap.
• Multi-agent role separation pattern now production-blessed at Nature-publication tier — DeepMind Co-Scientist (digest: 2026-05-11 #8) joins PREPING / SkillsVote in a three-way convergence on the role-separation + explicit-gating multi-agent pattern.

Local-model corner

• OScaR (try-now, code released) — Qwen3-30B-A3B at Q4_K_M + INT2 KV could fit 1M-context within 64 GB unified memory. Constraint flips from "how much context fits" to "decode speed at 1M context".
• Mix-Quant — argues prefill-asymmetric quant (FP4 prefill / BF16 decode) is the production-realistic agentic-inference shape. NVIDIA-specific today, principle portable to MLX.
• Compositional argument — OScaR (KV) + ZEDA (active experts) + Mix-Quant (prefill) are three orthogonal optimizations on the same memory budget. Open question whether they compose multiplicatively.

Sources noted as quiet

• anthropic.com/research: no new posts since May 14
• microsoft.com/research/blog: no new posts since May 15
• meta.ai/research: no new posts since April 8
• mistral.ai/news: no new posts since April 29
• openai.com/news: HTTP 403, persistent for 4+ weeks
• allenai.org/blog: only Earth-science cross-post (OlmoEarth v1.1)

Sources flagged for next-run follow-up

• DeepMind blog dates ambiguous for several May 2026 posts (Gemini 3.5 with action, Gemini Omni, Antigravity 2.0) — appear to be Google I/O 2026 cluster, dates need verification
• NVIDIA Developer Blog 4-post agent series (May 19–20) — NVIDIA-Verified Agent Skills with MCP signing / OWASP LLM / MITRE ATLAS pattern is the first major-vendor productization of MCP skill governance; devblog-tier but worth tracking if Anthropic / Google adopt parallel framework
• Google Research: Empirical Research Assistance (ERA, May 19) — product-tier framing, needs benchmark verification

Test plan

• inbox/2026-05-21.md exists with 9 verified items, all primary sources opened
• _meta/processed.json updated with 9 new IDs (validated as valid JSON)
• _meta/claims-ledger.md appended with 9 new ledger entries
• AI-Reviewer vs ResearchArena conflict cross-linked in both ledger entries
• All 7 picked arXiv abstracts + DeepMind Co-Scientist blog post opened and verified directly
• OScaR code-released claim verified at github.com/ZunhaiSu/OScaR-KV-Quant

https://claude.ai/code/session_01D6RZUnnUE2dMv6N2sZySDg

---

Generated by Claude Code