H
Howardism
Plate IIInteraction & Multimodal機器翻譯 · machine-translated過時翻譯 · stale translationENHOWARDISM

Interaction / Background Model Split

PublishedMay 13, 2026FiledConceptDomainInteraction & MultimodalTagsLLM ArchitectureAgent EngineeringMultimodalReading3 minSourceAI-synthesised

雙模型架構:具時間感知的互動模型維持即時在場;非同步背景模型處理深度推理與工具使用;豐富上下文封包委派;「以非思考延遲達成推理模型級規劃」

Interaction / Background Model Split 概念插圖

資料來源#

摘要#

Interaction Models 的架構由兩個協作模型組成:

  • 一個具時間感知的互動模型,維持即時在場——在持續迴圈中感知並回應(參見 Time-Aligned Micro-Turns);
  • 一個非同步背景模型,負責持續推理、工具使用及更長期的工作。

其效益在於:使用者同時獲得即時回應深度——「以非思考模型的回應延遲,達成推理模型的規劃、工具使用與 agentic 工作流程。」

委派機制#

  • 當任務需要的推理深度超出即時可產出的範圍時,互動模型會委派給背景模型,由其非同步執行。
  • 交接內容是一個豐富的上下文封包——不是獨立的查詢,而是完整的對話
  • 互動模型全程保持在場——回答後續問題、接收新輸入、維持對話脈絡。
  • 背景模型產出結果後會串流回傳;互動模型會在適當時機將更新穿插進對話中——配合使用者當下正在做的事,而非突兀地切換上下文。

兩端皆具智慧#

這並非「笨前端、聰明後端」的設計。互動模型本身「在互動性與智慧基準測試上皆具競爭力」——參見 Interactivity Benchmarks(例如 TML-Interaction-Small 在 Audio MultiChallenge APR 上擊敗所有非思考基線,即使未使用背景 agent;標記 * 的基準測試使用背景 agent 處理推理/工具任務)。

與其他多模型模式的關係#

這是多模型編排中延遲與深度的軸線,有別於:

  • Client-Side Agent Optimization 中的角色導向模型選擇(在 agent 圖中依角色分配便宜/昂貴模型)——該分割以成本為驅動且為靜態;此處則以延遲為驅動且逐輪動態調整;
  • 三 agent / 全新上下文審查者模式(Deep Modules for AgentsAgent Harness Engineering)——該分割目的是上下文隔離;此處則是為了時間性考量(保持回應 vs. 深度思考)。

開放議題 / 已知限制#

TML 稱背景 agents 為「不可或缺的能力」,且他們「才剛開始探索」——既要將背景 agentic 智慧推向前沿,也要探索背景 agents 如何與互動模型協同運作。

相關連結#

資料來源#

§ end
About this piece

Articles in this journal are synthesised by AI agents from a curated wiki and are refreshed automatically as new concepts arrive. Topics, framing, and editorial direction are curated by Howardism.

Cited by 14
  • Full-Duplex Interaction×3

    While listening and speaking, the model can simultaneously call tools, search, browse, or generate…

  • Interaction Models×3

    Time Aligned Micro Turns, Interaction Background Model Split, Encoder Free Early Fusion — the three…

  • TML-Interaction-Small×3

    Inkling-Small — the preview sibling of TML's open-weights release — is a 276B MoE with 12B active:…

  • GPT-Live×2

    When deeper reasoning or tool use is needed, GPT-Live delegates to frontier models such as GPT-5.5…

  • Inkling×2

    Inkling-Small is a 276B MoE with 12B active — exactly the shape of Tml Interaction Small, TML's May…

  • Interactivity Benchmarks×2

    in the table = result reported with the background agent enabled (used for benchmarks that need…

  • Live-Path Minimalism×2

    Interaction Background Model Split — the two-model architecture whose delegation half this page's…

  • Agent Harness Engineering

    Interaction Background Model Split — the same multi-agent split, but for temporal concerns (stay…

  • Client-Side Agent Optimization

    Interaction Background Model Split — another axis of multi-model design: there cost-driven and…

  • Deep Modules for Agents

    Interaction Background Model Split — the async background model is a deep module hiding reasoning…

  • Encoder-Free Early Fusion

    Interaction Background Model Split — the other half of the architecture

  • Interaction & Multimodal

    Interaction Background Model Split — Dual-model architecture: time-aware interaction model stays…

  • Thinking Machines Lab

    Inkling (July 2026) — their first from-scratch model, released with full weights: 975B/41B-active…

  • Time-Aligned Micro-Turns

    Interaction Background Model Split — micro-turns keep the interaction model present; deep reasoning…

Related articles
  • Interaction Models

    Thinking Machines Lab (May 2026): models that handle audio/video/text interaction natively in real time instead of via…

  • Full-Duplex Interaction

    Perceive-and-respond simultaneously across modalities; proactive interjection, visual-cue reactions, simultaneous speec…

  • TML-Interaction-Small

    TML's first interaction model: 276B MoE / 12B active, audio+video+text in / text+audio out, 200ms micro-turns, async ba…

  • Turn-Based Interface Bottleneck

    Why current AI interfaces limit collaboration: single-thread turn-taking is a bandwidth bottleneck; humans pushed out b…

  • Time-Aligned Micro-Turns

    The core interaction-model move: input/output as continuous streams in ~200ms interleaved chunks, no turn boundaries; s…