DAILY INTEL · 20260725

Agent 失敗多是上下文管理問題，而非單純的推理能力

今日頭條聚焦 OpenForgeRL 與 Agentic Context Management 兩大方向——讓 Agent 在現成的 inference harness 中端到端訓練，將 context 視為生命週期與架構問題進行管理，而非僅靠單純文字推論。

今日篇數 40
Drill 分 7.0
技能 web-elite-2026
+ commercial-news-ui-2026
Bento 4 格
Top Signal
Self-evolve Rank 20260725 — 可執行 TOP10
OpenForgeRL 讓 Agent 在複雜的 inference harness 中端到端訓練；Agentic Context Management 診斷失敗源於上下文而非推理；Test-Time Scaling via Error Localization 透過錯誤定位擴充推理時運算。其餘包括 Windowed-MTP、Euclid-MCP、SANA-Video 2.0 等。
Models
SANA-Video 2.0
混合線性注意力與殘差，在單一 GPU 上生成高品質影片。
Labs
Euclid-MCP
用 Prolog 提供確定性的邏輯推理，補足 LLM 在安全與合規領域的弱點。
Ship
Windowed-MTP
移除全上下文的 draft-KV 稅金，讓模型在百萬 token 下仍能高效預測。
Ignore
GS-Agent / SalesLoop / Uni-XAS
分別專注 4D 物理模擬、CRM 銷售線索、X 射線吸收光譜，與 agent 核心架構無直接關聯。
Papers 12 篇
01
OpenForgeRL: Train Harness-native Agents in Any Environment arxiv·2026-07-23
P0 02
GS-Agent: Creating 4D Physical Worlds With Generative Simulation arxiv·2026-07-23
P2 03
Same Dangerous Objective, Opposite Advice: Direct Exposure versus Multi-Agent Mediation arxiv·2026-07-23
P2 04
Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems arxiv·2026-07-23
P0 05
Surprisal Theory is Tautological (without Rational Grounding) arxiv·2026-07-23
P1 06
RUMBA: Russian User Memory Benchmark arxiv·2026-07-23
P2 07
Euclid-MCP: A Model Context Protocol Server for Deterministic Logical Reasoning via Prolog arxiv·2026-07-23
P1 08
news-crawler-LM: A Small Long-Context Model For High-Quality News Crawling arxiv·2026-07-23
P2 09
Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context arxiv·2026-07-23
P1 10
Test-Time Scaling via Error Localization arxiv·2026-07-23
P0 11
A Diffusion-Model Subpopulation Digital Twin for Mobile Health Deployment arxiv·2026-07-23
P2 12
Training Large Language Models for Self-Explanation Faithfulness arxiv·2026-07-23
P2
Skills 雙欄
web-elite-2026
全棧前端技能框架，涵蓋零構建單檔 HTML、OKLCH 色票系統、GSAP 動畫、響應式設計與商業級資訊密度。
commercial-news-ui-2026
編輯新聞 UI 模板，對標 Linear changelog × Stripe blog × Bloomberg density，支援 bento 佈局、閱讀進度條、prefers-reduced-motion。
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.
20260725
Agentic Context Management — Treat context as a lifecycle/architecture problem, not just a buffer. Version tool schemas as first-class architectural citizens. Implement compression (summarization/retrieval/selective retention). Monitor context health (size, tool usage, drift).
20260725
讓 Agent 在現成的 inference harness 中端到端訓練，而非僅靠單純文字推論。將任務拆解為可觀察的 reward signal (任務完成、工具使用正確)。失敗時回退到單步推理訓練。
Apple Tree
探索更多 AI 情報與深度分析
→