Butter-Bench: Evaluating LLM-Controlled Robots for Practical Intelligence

Quick Overview

The "Butter-Bench" paper reveals that Large Language Models (LLMs) struggle significantly with practical intelligence tasks, scoring only 40% completion on tasks requiring physical manipulation and showing a critical gap compared to human common sense, as demonstrated by the failure of advanced models like GPT-4 and Gemini 2.5 Pro on simple tasks like identifying butter in a picture or performing safety checks.

Key Points: The best-performing LLM tested, Claude Opus, only managed 40% completion on tasks involving physical manipulation. When performing real-world tasks like identifying butter in a picture or checking status, LLMs failed to exhibit common sense reasoning. The study found a significant gap between LLM analytical/coding skills (scoring 95%) and practical reasoning skills (scoring 40% or lower). Specific models tested, including GPT-4 and Gemini 2.5 Pro, exhibited similar shortcomings in practical reasoning, with Gemini 2.5 Pro scoring only 27% on robotics data tasks. Failures often stemmed from an inability to handle ambiguity, social cues, or the complexity of real-world physics, rather than just lacking basic skills. The paper suggests that current fine-tuning methods, which often rely on internal monologue or self-correction, do not effectively bridge the gap to real-world physical intelligence.

Context: This video summarizes findings from the research paper "Butter-Bench," which evaluates the practical intelligence of Large Language Models (LLMs) when controlling robots. The evaluation specifically contrasts the LLMs' high performance in analytical and coding tasks against their performance in tasks requiring common sense reasoning about the physical world, such as navigating environments or interacting with objects like butter.

Detailed Analysis

The discussion centers on the "Butter-Bench" paper, which evaluates LLM-controlled robots across five main categories: tool use, movement driving/turning, kinematics control, environmental perception, and social interaction. The paper finds a major disconnect between LLMs' high analytical abilities (like coding) and their capacity for practical intelligence. For instance, while LLMs excel at abstract reasoning, they struggle profoundly with real-world tasks. The best model, Claude Opus, scored only 40% on physical manipulation tasks. When tested on specific failures, like identifying butter in a picture or assessing if a robot could carry an item on a tray, advanced models like GPT-4 and Gemini 2.5 Pro performed poorly, sometimes scoring as low as 27% on certain robotics data sets. A key finding is that models often fail when encountering ambiguity, social cues, or unexpected physical situations, suggesting that current fine-tuning methods that emphasize internal monologue do not adequately teach the practical common sense needed for real-world deployment.