[논문리뷰] INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

링크: 논문 PDF로 바로 열기

The user wants a summary of the paper. I have browsed the content. Now I need to extract the information as per the instructions.

Metadata:

• Authors: Junhan Sun, Hao Zhao, Guofeng Zhang
• Keywords: Need to identify 5-8 keywords. I'll look for recurring technical terms.
  • World Models
  • Intent-to-Action Learning
  • Search-Free Control
  • JEPA (Joint Embedding Predictive Architecture)
  • Latent Dynamics
  • Multi-task Learning
  • Reinforcement Learning (or Offline Reinforcement Learning / Imitation Learning since it uses offline trajectories)
  • Isomorphic Learning

Part 1: Markdown Summary

## 1. Key Terms & Definitions

• INTACT (INtent-To-ACTion): An end-to-end JEPA (Joint Embedding Predictive Architecture) that converts action-labeled, reward-free trajectories into a deployable intent-to-action interface, enabling search-free control.
• JEPA (Joint Embedding Predictive Architecture): A type of world model that predicts in representation space, learning a forward conditional (given state and action, predict next state) without explicit generative objectives.
• Action Quotient: At a fixed current state z, two endpoint conditions are action-equivalent if they induce the same expert action law, forming equivalence classes that are in one-to-one correspondence with the realizable action-law image.
• Motion Intent (m_t): A latent representation of a desired motion, either a local physical transition (m_t^local = z_{t+1} - z_t) or a future goal (m_t^goal = sg(z_g) - z_t), used by the INTACT Predictor.
• Search-Free Control (Direct Control): An inference mode where INTACT directly outputs an action sequence (plan) without requiring expensive test-time sampling and model rollouts (e.g., CEM).

## 2. Motivation & Problem Statement Latent world models, while capable of predicting future states given actions (forward conditional), struggle with inverse control—recovering actions for a desired change. This representation–control asymmetry necessitates expensive test-time search algorithms like CEM or MPPI to invert the model, making deployment inefficient and often starting from generic random actions. The core problem is that existing world models learn what an action will do, but not which action should realize a requested latent change, leading to a semantic uncalibration between action and latent space in the inverse direction needed for control. This gap means that the forward-only bottleneck of traditional world models gathers information for forecasting observations but leaves its action geometry incidental, preventing direct query by a downstream controller.

## 3. Method & Key Results 본 논문은 INTACT (INtent-To-ACTion)를 제안하여 search-free control을 가능하게 하는 isomorphic intent-to-action interface를 학습한다 [Figure 1]. INTACT는 기존 JEPA에 하나의 conditional action operator를 추가하며, 이 operator는 realized physical transition과 deployable goal intent라는 두 가지 조건에 대해 parameter sharing과 feature grammar sharing을 수행한다. Local intent (m_t^local)는 physical reachability를 제공하고 representation shaping을 유도하며, goal intent (m_t^goal)는 stop-gradient deployment anchor 역할을 한다. 이 shared operator는 두 intent families를 conditional action law를 통해 정렬하며, pointwise latent matching 없이도 action-law semantics를 구축한다. INTACT는 Direct control 모드에서 zero candidate search로 동작하며, optional local CEM을 통한 verification도 지원한다 [Figure 2].

Figure 1 — INTACT 학습 및 제어 아키텍처

Figure 2 — INTACT 제어 흐름

핵심 결과로, 단일 태스크(single-task) 설정에서 INTACT는 LeWM의 10배 적은 epoch과 zero candidate sequences로 PushT, Cube, Reacher, TwoRoom 네 가지 공식 벤치마크에서 각각 85.78%, 100.00%, 97.67%, 97.89%의 Direct SR을 달성하여 95.33% macro SR을 기록했다 [Table 2]. 이는 CEM 300x30의 1.48s inference latency를 2.9–5.5ms로 대폭 감소시키는 효과를 가져왔다. multi-task 설정에서 shared encoder를 사용한 Goal-displacement INTACT는 E5에서 89.39% macro Direct SR을 달성했으며, matched shared-encoder LeWM 대비 모든 네 태스크에서 SR을 각각 5.66, 32.23, 12.56, 42.44 포인트 향상시켰다 [Figure 4, Table 3]. predicted–expert action-family kNN과 CKA는 official Direct SR과 각각 r = 0.954, r = 0.897의 높은 상관관계를 보였으며, 이는 pointwise action R^2 (r = 0.815)보다 family-level intent-to-action preservation이 더 중요함을 시사한다 [Table 4, Figure 5].

## 4. Conclusion & Impact 본 논문은 INTACT를 통해 physical successor와 deployable goal intent 사이의 inverse control mismatch를 해결했다. 이 연구는 latent world models의 search-free control을 위한 end-to-end learning framework를 제시하며, forward JEPA와 action-conditioned deployment interface를 통합하는 새로운 방법을 제안한다. INTACT의 기여는 zero-search Direct control을 통해 inference latency를 획기적으로 줄이고 multi-task learning 환경에서도 뛰어난 generalization 성능을 보장한다는 점에 있다. 이는 고주파수 제어(high-frequency control)를 위한 model-based search bottleneck을 제거하며, learned representation이 downstream controller에 의해 직접 쿼리될 수 있도록 semantic calibration을 제공하여 academic 및 industrial 분야에서 efficient하고 robust한 robot learning 및 decision-making systems 개발에 중요한 시사점을 제공한다.

Part 2: Important Figure Information

I will look for Figure 1, Figure 2, Figure 3, Figure 4, Figure 5 and extract their image URLs and a brief Korean caption.

• Figure 1: "Shared-encoder INTACT training and recurrent control."
  • Image URL: https://arxiv.org/html/2607.26056v1/x1.png
• Figure 2: "INTACT converts mandatory search into optional verification."
  • Image URL: https://arxiv.org/html/2607.26056v1/x2.png
• Figure 3: "Single-task action likelihoods shape the representation."
  • Image URL: https://arxiv.org/html/2607.26056v1/x3.png
• Figure 4: "Controlled E5 success across four tasks."
  • Image URL: https://arxiv.org/html/2607.26056v1/x4.png
• Figure 5: "Intent–action relation, rather than task clustering, tracks control."
  • Image URL: https://arxiv.org/html/2607.26056v1/x5.png

I will choose Figure 1 (architecture), Figure 2 (control flow), and Figure 4 (key multi-task results).

Let's double-check the constraints:

• Technical Terminology in English: Checked, used original terms.
• Natural Korean Phrasing with mixed English nouns/verbs: Checked.
• Mandatory Metadata: Authors and Keywords at the top. Checked.
• Output Format: Part 1 (Markdown), Part 2 (JSON) separated by `

Figure 4 — 다중 태스크 E5 SR 결과

⚠️ 알림: 이 리뷰는 AI로 작성되었습니다.