JIT-Agent重塑底座
Papers arxiv:2608.
25593 Copy markdown JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution Published on Aug 26 · Submitted by gbz on Aug 27 · National University of Singapore Upvote 47 +39 Authors: Guibin Zhang ,Leo Lu ,Fangzhou Xie ,Kang Zhu ,Junhao Wang ,Zhifei Xie ,Zhaochen Yu ,Zihang Liu ,Zhongxiang Sun ,Qiankun Li ,Yue Liao , Heng Chang ,Xiaobin Hu ,Qibing Ren ,Wangchunshu Zhou ,Shuicheng Yan Abstract JIT-Agent is a trainable model that synthesizes adaptive agent harnesses for off-the-shelf LLMs, improving performance across diverse models and tasks.
Generated by thinkingmachines/Inkling-Small Agent capability is not determined by the model alone.The agent harness, encompassing memory management, planning strategy, action protocol, and tool/skill orchestration, can dominate the contribution of the underlying foundation model.
Yet harness design remains manual, task-specific, and fundamentally unscalable.We present JIT-Agent, a harness intelligence model trained to synthesize task-adaptive agent harnesses on the fly for arbitrary off-the-shelf agentic LLMs.
We formalize the agent harness as a composable, machine-generatable artifact governed by a fixed four-module protocol, and train JIT-Agent to customize harnesses for a given task at hand, repair harnesses for stable and reliable execution, and self-evolve by distilling performance signals from an expanding archive of prior harness configurations.
Equipped with JIT-Agent as a harness helper, DeepSeek-V4-Flash surpasses GPT-5.6 on DeepSearchQA (+9.1) and OdysseyBench (+4.3), while the already strong GLM-5.2 gains up to +20.2 points.
Across controlled evaluations, JIT-Agent-generated harnesses are performance-competitive with mature agent runtimes such as OpenCode and Claude Code and consistently improve multi-scale model families of DeepSeek V4, Mimo-V2.5, and Qwen3.6.
To our knowledge, JIT-Agent is the first model purpose-built for just-in-time harness generation, establishing harness intelligence as a trainable, transferable, and compounding dimension of agent capability orthogonal to model scaling.
View arXiv page View PDF Project page GitHub 24 Add to collection Community greeky Paper submitter about 24 hours ago Model-as-a-Harness; Harness Intelligence Model Website: https://bingreeky.github.io/JIT-site/GitHub: https://github.com/bingreeky/JITHugging Face: https://huggingface.
co/datasets/JIT-Agent/jit-meta-harness Reply EditPreview Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.Tap or paste here to upload images Comment · Sign up or log in to comment Upvote 47 +35 Get this paper in your agent: hf papers read 2608.
25593 Don't have the latest CLI?curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 1 Datasets citing this paper 0 No dataset linking this paper Cite arxiv.org/abs/2608.25593 in a dataset README.md to link it from this page.
Spaces citing this paper 0 No Space linking this paper Cite arxiv.org/abs/2608.25593 in a Space README.md to link it from this page.Collections including this paper 3
Related
相關文章
Google 開源 Harness²：遞迴代理程式框架，無需新模型也能提升效能
Google Cloud AI Research 開源了 Harness²，這是一個遞迴代理程式框架，透過改良凍結模型周圍的程式碼與檔案來增強效能。對開發者而言，重點很簡單：你的代理程式無需更強大的模型也能進步。重點摘要：規模為框架而非模型，訓練參數為0，包含8個可編輯的框架元件，版本0.1.0。執行環境為Linux、Python 3.12或3.13搭配uv，模型呼叫則透過Vertex AI。GPU需求未公開。效能方面：一次改良步驟即可在3個基準測試中，提升所有4個模型與框架組合的表現。最佳成果：GLM-5搭配OpenCode在WorkBuddy-Bench上從64.2提升至75.7（+11.5點）；最差表現：GLM-5搭配Codex在JobBench上僅提升1.2點（23.6至24.8）。什麼是Harness²？它是Google研究推出的即插即用框架。
如何利用NVIDIA結構化資料模型進行零訓練表格預測
In this tutorial, we implement Structured Data Models, NVIDIA’s new library that puts several pretrained foundation models for tables behind one PyTorch interface, and use it to make predictions without training.
認識 Kyutai REKI：2.4B VLA 模型為機器人操控帶來高效記憶
Kyutai 與 Inria 及 École des Ponts 合作，推出了 REKI——一個用於機器人操控的 2.4B 參數視覺-語言-動作模型（VLA）。該模型僅使用公開數據進行預訓練，並配備了輕量級的近期記憶功能。在三個分佈外基準測試中，僅有規模約為其兩倍的模型能超越它。所有模型權重已於今日在 Hugging Face 上公開。核心資訊：規模：2.4B 參數（2.2B Qwen3.5-2B VLM + 0.2B 流匹配動作專家）。記憶涵蓋最近 3 幀，間隔 1 秒。運行環境：在 1 張 NVIDIA H100 上進行基準測試。從兩個 256×256 攝像頭輸入，編譯後每個動作區塊耗時 21.8 毫秒。權重以 BF16 格式提供，暫無量化版本。性能：在 LIBERO-Plus 和 RoboTwin Clean2Rand 基準測試中，為公開數據預訓練的 VLA 模型中最優；在 LIBERO-Plus 上達到 90.4% 的最高分。
DeepSeek 發布 V4.1-Flash-1017，DeepSWE 得分 77.0，並開放搜尋 API
DeepSeek 剛發布 DeepSeek-V4.1-Flash-1017，這是其 V4.1-Flash 模型的新檢查點，現已在 DeepSeek 網站、應用程式及 API 上線。DeepSeek 宣稱該版本在性能、任務完成速度與輸出簡潔度上均優於 9 月發布的 V4.1-Flash，並稱之為性能與成本的帕累託改進。同時，DeepSeek 也向所有開發者開放了搜尋與擷取 API。對代理建構者而言，這意味著更便宜的編碼模型加上單一 API 金鑰即可使用的第一方網路層。重點摘要：骨架參數 552B，另加 196B 的 Engram 參數；預填充階段每 token 啟用 8B 參數，解碼階段啟用 16B 參數；上下文長度達 1M tokens，最大輸出 384K tokens。運行於 DeepSeek API 的 deepseek-flash 模型。9 月版的 V4.1-Flash 權重（763B 參數，FP8）採 MIT 授權。
2026年20款Jev替代方案：挑戰TypeSafe的決策模型
TypeSafe於2026年9月15日推出Jev，這是一款決策模型而非聊天機器人。使用者可定義允許的答案，Jev將回傳標籤、分數或是否的機率。約4週內，至少20款Jev替代方案相繼問世，來自Cloudflare、OpenAI、Perplexity、Liquid AI、AWS的Strands Labs及獨立開發者。本指南比較各方案在規模、授權、硬體、價格、Jev API相容性及已發布基準測試的表現。重點摘要：規模範圍從287M（GLiNER2.5-multi-Decide）到35B-A3B（Mapika Decider），Jev規模未公開。執行環境：Von和Laya等編碼器可在CPU上運行，4B適配器適用24 GB GPU或32 GB Mac。效能：多數開源模型在困難推理上落後Jev，但在狹窄分類任務上數款與之匹敵或超越。最佳表現：Cloudflare Clef在BANKING7取得94.20分。
長鑫科技市值兩月內跌去1萬億；阿里加碼AI硬件品牌，天貓、速賣通聯動；國行版蘋果AI有望12月上線，百度、阿里參與合作丨雷峰早報
要聞提示1.長鑫科技市值兩月內跌去1萬億2.年內已至少18次故障！DeepSeek發放宕機補償，有人拿到3000元3.阿里加碼AI硬件品牌，天貓、速賣通聯動4.國行版蘋果AI有望12月上線，百度、阿里參與合作5.可靈AI新一代模型Kling 4.