loyiv/ITP

Code of Paper: Imagine-then-Plan: Agent Learning from Adaptive Lookahead with World Models

What's novel

Code of Paper: Imagine-then-Plan: Agent Learning from Adaptive Lookahead with World Models

Code Analysis

19 files read · 4 rounds

Implements the Imagine-then-Plan (ITP) method for LLM agents: a training-free 3-step pipeline (adaptive K-selection, world-model imagination, reflect-and-act) and a trained variant with K-head/V-head predictors on top of a CausalLM, evaluated on ALFWorld and other text-based environments.

Strengths

The ITP-R training architecture (K-head + V-head at a <CTRL> token position) and the RAP/MCTS planner with composite rewards are genuinely non-trivial implementations. The core ITP-I pipeline is clean, well-structured, and has robust action validation with multiple fallback strategies.

Weaknesses

Significant code duplication between itp/ and eval/foresight_eval/ (LLM backends, policy model, world model all reimplemented with slightly different interfaces), and zero test coverage across 138 source files. The training script (54KB) is a monolithic file that would benefit from decomposition.

Score Breakdown

Innovation

7 (25%)

Craft

54 (35%)

Traction

8 (15%)

Scope

85 (25%)

Signal breakdown

Innovation

Not Fork+1

Code Novelty+1

Concept Novelty+3

Craft

Ci+0

Tests-1

Polish+1

Releases+0

Has License+5

Code Quality+17

Readme Quality+15

Recent Activity+7

Structure Quality+5

Commit Consistency+0

Has Dependency Mgmt+5

Traction

Forks+0

Stars+6

Hn Points+0

Watchers+0

Early Traction+0

Devto Reactions+0

Community Contribs+2

Scope

Commits+8

Languages+5

Subsystems+13

Bloat Penalty+0

Completeness+7

Contributors+6

Authored Files+15

Readme Code Match+3

Architecture Depth+7

Implementation Depth+8

Evidence

Commits

Contributors

Files

117

Active weeks

TestsCI/CDREADMELicenseContributing

Repository

Language

Python

Stars

Forks

License

MIT