# Free Dig Sampler — 2026-09-16 (Asia/Shanghai)

**Theme:** Skill self-evolution · repair loops · multi-turn IF · native routing · phase-aware KV · continual self-improvement infra

> **Free sampler day.** This is one densest day from the Agentic Engineering Dig Pack. The full paid pack covers a ~7-day archive (briefings + mermaid maps + keepers). Soft upgrade: https://app.anyway.sh/pay/PL26XI1CXFWLJADE?utm_source=sampler&utm_medium=landing&utm_campaign=dig_pack&utm_content=free-day

---

## BRIEFING

- **SkillLift（2609.15396）**：skill 自演化若每候选靠完整 rollout 当 oracle，会锁死在失败补丁；用 **bilevel + dense rubric**（内环零 oracle 修订、外环少量 rollout 对齐秩相关）摊销成本，报相对前沿 evolving **少 40–70% token**。
- **RESKILL（2609.15684）**：失败修复不是 one-shot reflection——维护跨轮 **hypothesis↔patch↔retest** 显式 repair state；六组设置均最强（+3.7pp vs direct / +3.3pp vs hypothesis-conditioned）——归因必须与选择、再测耦合。
- **MTAC-IFBench（2609.14992）**：多轮 agentic coding 过程指令遵循欠评；平均 7.04 轮 / 91.33 约束，checklist+脚本核验——现有 code agent **随会话变长 IF 快速退化**。
- **Gavel / Router Within（2609.15982）**：冻结 LLM mid-layer 已有路由信号，两张线性 map（Glance+Verdict）无需 skill 正文进上下文；相对 progressive disclosure / retrieve-rerank 最高 +13.4 / +21.9——**路由是独立能力**。
- **AgentKV（2609.14872）**：agentic 生成违反「未来注意力≈最近注意力」；按 think/act/tool 相位 buffer 打分，相对 R-KV +5.5、吞吐最高 ~1.80×——**agent KV 要相位感知**。
- **Reef（HF 09-15）**：开源 continual self-improvement 基建——stateful inference + harness/model 联合演化 + 评测门控发布；**自改进首先是推理基建**（对齐 RSI / Continual Harness）。

```mermaid
flowchart LR
  skillLift[SkillLift<br/>rubric bilevel]
  reskill[RESKILL<br/>attribution+retest]
  mtac[MTAC-IFBench<br/>multi-turn IF]
  gavel[Gavel<br/>native skill routing]
  agentkv[AgentKV<br/>phase-aware KV]
  reef[Reef<br/>stateful inference+evolve]
  skillLift --> reskill
  gavel --> skillLift
  mtac --> gavel
  agentkv --> reef
  reskill --> reef
```

## Keepers (public links only)

1. SkillLift — https://arxiv.org/abs/2609.15396
2. RESKILL — https://arxiv.org/abs/2609.15684
3. MTAC-IFBench — https://arxiv.org/abs/2609.14992
4. The Router Within (Gavel) — https://arxiv.org/abs/2609.15982
5. AgentKV — https://arxiv.org/abs/2609.14872
6. Reef (HF blog) — https://huggingface.co/blog/quao627/your-inference-server-is-secretly-a-learner-reef

## One-liners you can reuse

- Repair that doesn’t retest is just journaling.
- Mid-layer routing can beat stuffing skill text into context.
- Agentic KV ≠ chat KV — phase matters.
- Continual self-improvement starts at the inference harness, not the model card.

## What this sampler is / isn’t

| Is | Isn’t |
| --- | --- |
| One free day of outcome-first BRIEFING + mermaid + public keepers | The full week archive |
| Honest taste of Dig Pack density | A newsletter subscription or private tip feed |
| Safe to share / forward | Ops notes, tip SHAs, or workspace internals |

**Upgrade:** full Dig Pack (~7 days) → https://app.anyway.sh/pay/PL26XI1CXFWLJADE?utm_source=sampler&utm_medium=landing&utm_campaign=dig_pack&utm_content=free-day