Shuffling reasoning steps yields better AI accuracy and faithfulness

A new causality reward scrambles the order of a model's reasoning steps without altering their content, improving out-of-domain accuracy, reasoning quality, and truthful reflection of the model's process.

Industry

Xiao Yu · Justin Chih-Yao Chen · Zaid Khan · Hyunji Lee · Mohit Bansal

UNC Chapel Hill

Research Digest·2 November 2026·2 min read

Illustration: Shuffling reasoning steps yields better AI accuracy and faithfulness

In one line

Scrambling reasoning steps and rewarding drops in answer probability improves model accuracy, reasoning quality, and faithfulness.

Before you read this

Three ideas this paper builds on

reinforcement learning with verifiable rewardsReinforcement learning that uses rewards from objectively checkable signals like correct answers or verification feedback.141 papers →credit assignmentDetermining which actions or steps in a sequence contributed to a final outcome, often for reinforcement learning.174 papers →group relative policy optimizationA reinforcement learning method that optimizes policy by computing advantages based on group-level reward statistics.160 papers →

Read it as

When a large language model (LLM, a text-predicting program) gives an answer, it usually writes reasoning steps first. Standard training rewards only whether that answer is right, so the reasoning can become irrelevant decoration. Then the model's logic is hard to trust.

The authors add a reward based on scrambling the order of those steps without changing their content. If the final answer becomes less likely after scrambling, the reasoning actually mattered, and the model is rewarded.

This raised accuracy on unseen logic and game problems to 46.8%, beating the strongest comparison method. It also made reasoning more faithful: when a hint was present, models acknowledged it 34.0% of the time, far more often than the comparison. Trustworthy reasoning matters for real decisions.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

chain-of-thought causality reinforcement learning with verifiable rewards chain of thought chain-of-thought faithfulness credit assignment group relative policy optimization