Brody AI News
DAILY INTEL · 20260806 · EDITORIAL

今天只留能變現的訊號

Self-evolve rank 20260806

84sources ranked
—capability drill
99skills in motion
Top signal

主訊號 · Rank

  • TurnSight | 工具推理的細粒度信用分配：從軌跡層級提升到「轉向」層級 | 落地：工具鏈的細粒度信用分配
  • Video-DeepResearch | 從靜態圖延伸到連續影片：需要密集的時空定位與開源網探索 | 落地：影片層級的時空定位與開源網探索
  • Type or Talk | 人機輸入的特徵：鍵盤的正字形雜訊 vs. 語音的不流利與重構 | 落地：人機輸入的特氣與重構
  • Game Theory for FM | 基礎模型的合作原理：透過相似度推論確保安全與合作 | 落地：基礎模型的合作原理
  • ParVL | 多模模型的平行擴張：打破參數與序列推論的固定分配 | 落地：多模模型的平行擴張
  • SocietyBench | 社會世界的反事實預測：模型對真實社會演化的理解 | 落地：模型對社會世界的反事實預測
Models

雙模狀態

8080 maker 8081 checker local only

  • Web Elite v2 skill 已掛載
  • Editorial news shell 啟用
  • channel strip + hard_audit
Labs

Verify / Labs

  • 將工具鏈切分為單個 Turn，標記每個 Turn 的 Context 與 Tool Result。
  • 計算每個 Turn 對最終結果的邊際貢獻 (Marginal Contribution)。
  • 生成高信用與低信用的 Preference Pairs。
  • 支援本地 Llama 進行 SFT 或 DPO 微調。
  • 失敗時回退至軌跡層級 (Trajectory-level) 的簡單 Reward。
Ship

Ship notes

  • items: 84
  • raw: D:\ai-daily-data\evolution\daily\raw_20260806.json
  • rank: D:\ai-daily-data\evolution\daily\rank_20260806.md
  • draft: D:\ai-daily-data\evolution\skill-drafts\self_evolve_20260806_050001.md
  • verify: D:\ai-daily-data\evolution\daily\verify_20260806.md

Papers & sources

今日排序 · 可掃讀

2026-08-04
TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning
arxiv
src
2026-08-04
Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent
arxiv
src
2026-08-04
Should We Type or Talk to LLM Agents? A Comprehensive Study of Voice and Keyboard Input Perturbations
arxiv
src
2026-08-04
A game theory for foundation models shows new paths to rational cooperation through similarity inference
arxiv
src
2026-08-04
ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs
arxiv
src
2026-08-04
SocietyBench: Forecasting Counterfactual Social-World Evolution
arxiv
src
2026-08-04
Logic Before Language: Pre-pretraining on Formal Derivations Fosters Skill Acquisition and Compressibility
arxiv
src
2026-08-04
Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility
arxiv
src
2026-08-04
FOUND-AF: Benchmarking ECG Foundation Models for Atrial Fibrillation Detection
arxiv
src
2026-08-04
Test-Time Augmentation for Tabular-to-Image Classifiers under Distribution Shifts
arxiv
src
2026-08-04
Approximate Speculative Decoding
arxiv
src
2026-08-04
SciRet: A Compute-Aware Empirical Study of Retrieval and Reranking for Scientific RAG
arxiv
src
2026-08-04
DiagChain: A Diagnostic Benchmark for Evaluating LLM Agents on Evidence-Grounded Attack Chain Reconstruction
arxiv
src
2026-08-04
Training Documents Reranker with Search Rubrics for Deep Research Agent
arxiv
src

Skills distilled

可落地能力

UI

web-elite-2026 v2

GSAP · Lenis · Three · OKLCH · 反 AI-slop · zero-build/Vite 雙軌

News

commercial-news-ui-2026

編輯氣質 bento 情報站殼 · 8081 ≥9 才 deploy

draft

_report_20260806_0507

# Skill Evolve 20260806_0507 - ops/youtube-quick-summary -> D:\ai-daily-data\evolution\skill-drafts\ops_youtube-quick-summary__20260806_0507.md - ops/ai9-fast-deploy -> D:\ai-daily-data\evolution\ski

draft

ops_ai9-fast-deploy__20260806_0507

version 1.1 changelog: 注入失敗回退、驗證、一鍵指令與歸因路徑。 # AI9 快速部站 ## 伺服器指紋 - IP: 119.14.175.24 / host ai9 / Win Server 2025 / i5-9500 / 16GB - SSH: rovk@119.14.175.24 - 網域: brodyaitw.com - 靜太根: `D:\sit

draft

ops_youtube-quick-summary__20260806_0507

## SKILL.md (Upgraded) version: 1.1 changelog: - 引入「失敗回退」機制：讀取 sub_err.txt 並自動換成 en 重試 - 新增「驗證」步驟：在模型精煉前先驗證 summaries/{id}.md 是否存在 - 歸因路徑：所有歸因資料統一存於 `D:/ai-daily-data/evolution/attribu

draft

self_evolve_20260806_050001

---SKILL--- # TurnSight: 工具推理的細粒度信用分配 ## name TurnSight ## description 將傳統軌跡層級 (Trajectory-level) 的強化學習信用分配，提升至「轉向層級」(Turn-level)。透過識別工具鏈中具體的「決策點」，精確評估每個工具調用對最終結果的貢獻，解決複雜 Agent 任務中稀疏獎勵的信用分配問題。 ## v

Living rules

組織記憶

Living rules (auto-evolved)

20260725

Decouple harness (prompts/tools/loop) from base model training.

Capture harness-native signals (tool calls, reasoning, rewards).

Train harness layer independently of base model.

Fall back to joint training if decoupling fails.

Avoid retraining the base model when only harness changes.

Agentic Context Management