Open-loop digest
June 2, 2026
18 items · 4.5 KB
Raw LLM outputNo human editsModel: Qwen3.6:35B-A3BPosted automatically by cron
Agentic Frameworks, Tooling, Skills
- SkillAdaptor, https://huggingface.co/papers/skilladaptor — Self-adapting skills for LLM agents from trajectories. Enables dynamic skill routing and state transfer across agentic workflows without full adapter reloads or KV cache thrashing, directly stabilizing long-horizon rover control loops. (HF Papers)
- Harness-1, https://arxiv.org/abs/2606.02373 — Reinforcement learning for search agents with state-externalizing harnesses. Provides a structural pattern for halting, saving state, and resuming agentic search trajectories without catastrophic context degradation during extended research aggregation. (arXiv)
- FineVerify, https://huggingface.co/papers/fineverify — Scaling test-time compute with fine-grained self-verification for agentic search. Cuts hallucination drift in multi-hop research pipelines by verifying intermediate steps before committing to tool-use chains, requiring no heavier base models. (HF Papers)
Notable Research
- LongAttnComp, https://huggingface.co/papers/longattncomp — Cross-family context compression for long-context reasoning. Enables 131k+ token agentic contexts on 32GB VRAM by compressing KV states across attention families without accuracy loss, directly addressing your long-context constraint. (HF Papers)
- SafeSteer, https://arxiv.org/abs/2606.02530 — Localized on-policy distillation for efficient safety alignment. Provides actionable guardrails for local agentic tool-use routing, preventing prompt-injection drift in untrusted telemetry or research document parsing. (arXiv)
- RoboSemanticBench, https://huggingface.co/papers/robosemanticbench — Diagnosing semantic grounding in action prediction for VLA models. Offers a standardized benchmark for validating your Kangaroo rover's spatial grounding and aerospace vision-scraper reliability under physical stress conditions. (HF Papers)
Frontier Lab Updates
- Anthropic IPO Readiness & $30B Run-Rate, https://simonwillison.net/ — Confirms Anthropic targeting first profitable quarter with scaled commercial infrastructure. Signals sustained infra scaling and commercial agent routing baselines ahead of public market metrics and competitive pressure on open-weight releases. (Simon Willison)
- Anthropic Production Sandboxing Architecture, https://simonwillison.net/ — Details runtime boundaries via gVisor, Seatbelt (macOS), Bubblewrap (Linux), and full VMs. Establishes hardened execution contexts for future agent deployments, directly impacting trust assumptions for local vs. cloud agent handoffs in production pipelines. (Simon Willison)
- DeepSeek-V4-Flash Scaling, https://huggingface.co/models?sort=trending — 158B parameter model trending on HF. While cloud-only, its rapid adoption signals shifting baselines for open-weight model capabilities and competitive pressure on local 20-50B class performance. (HF Trending)
Models to Download & Try
- JetBrains/Mellum2-12B-A2.5B-Thinking, https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Thinking — 12B MoE (2.5B active) architecture. ~5.5GB VRAM fit. Claims to beat dense 12B/13B reasoning baselines with lower latency; maximizes your 32GB headroom for 131k context windows, ideal for high-throughput agentic tool-dispatch. (HF Trending)
- nvidia/Qwen3.6-35B-A3B-NVFP4, https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4 — 19B parameters with NVFP4 quantization. ~11GB VRAM fit. Claims to beat BF16/FP8 35B-class counterparts in inference speed and KV cache efficiency; direct variant of your base model for stress-testing rover control pipelines. (HF Trending)
- Kwai-Keye/Keye-VL-2.0-30B-A3B, https://huggingface.co/Kwai-Keye/Keye-VL-2.0-30B-A3B — 30B MoE (3B active) vision-language model. ~14GB VRAM fit. Claims to beat dense 30B VLMs in multimodal grounding and layout parsing; native cross-modal routing expands your vision-scraper's aerospace component inspection and telemetry parsing without drowning your context buffer. (HF Trending)
Skipped as Already Covered
- LiquidAI/LFM2.5-8B-A1B & GGUF (sparsity routing & high-throughput dispatch)
- bytedance-research/Lance & NemoStation/Marlin-2B (any-to-any & video-text routing)
- nvidia/LocateAnything-3B & openbmb/MiniCPM5-1B/PaddleOCR-VL-1.6 (vision OCR/telemetry scrapers)
- Qwen3.6-27B/35B GGUF/MTP variants (unsloth/Jackrong/HauhauCS) (agentic routing optimizations)
- AgentDoG 1.5 & Qwen-VLA (robotic/VLA frameworks & safety alignment)
- Claude Code & MCP-Persona (agentic terminal tooling & benchmark harnesses)