RIFT: Reordered Instruction Following To Evaluate Instruction Following in Singular Multistep Prompt

Quick Overview

The RIFT paper demonstrates that large language models (LLMs) struggle significantly with multi-step instruction following when the steps are reordered, leading to a substantial drop in accuracy (from 85% baseline to 13.9% in one test) due to a failure to maintain a global state, suggesting structural flaws rather than mere capacity issues in current models.

Key Points: LLMs demonstrated a significant drop in performance when instruction order was reordered (RIFT test), with accuracy plummeting from an 85% baseline to 13.9% for the 120B parameter model. The failure is attributed to an architectural flaw, specifically the inability to maintain a global state across steps, rather than a capacity issue. The 120B parameter model performed slightly better (85% baseline accuracy) on the linear instruction set compared to the 30B parameter model's performance on the same test. When tested with jumping prompts (non-sequential instruction following), the 120B model's accuracy dropped significantly, while the smaller 30B model performed nearly at random (0% accuracy). The research suggests that while larger models are better at reasoning and data retention, they remain fragile against structural challenges like context window shifts or non-linear instructions. The failure mode observed in the jumping prompt test was characterized by the model either stopping early or aggressively moving to the end of the sequence, indicating a breakdown in tracking progress. The paper concludes that the current architectural approach of relying heavily on positional bias (what is physically nearby in the text) is a major limitation for complex, multi-step reasoning.

Context: The video discusses findings from the research paper titled "RIFT: Reordered Instruction Following To Evaluate Instruction Following in Singular Multistep Prompt," which investigates the robustness of large language models (LLMs) when the order of instructions within a prompt is intentionally scrambled. The research, conducted by Emory University researchers including Andrew Jaffe, Noah Risen, and Genio DeTroy, highlights a critical weakness in how current state-of-the-art models handle sequential dependencies, contrasting their performance on linear versus non-linear instruction sets.