PA3: Policy-Aware Agent Alignment through Chain-of-Thought

TL;DR

This paper introduces a multi-stage alignment approach for large language models to better adhere to business policies during reasoning, reducing prompt length and improving accuracy without including full policies in context.

Contribution

The paper presents a novel method for models to recall and apply policies during reasoning, along with a new reward and penalty to enhance policy adherence without lengthy prompts.

Findings

Model outperforms baseline by 16 points

Surpasses in-context baselines by 3 points

Uses 40% fewer words in prompts

Abstract

Conversational assistants powered by large language models (LLMs) excel at tool-use tasks but struggle with adhering to complex, business-specific rules. While models can reason over business rules provided in context, including all policies for every query introduces high latency and wastes compute. Furthermore, these lengthy prompts lead to long contexts, harming overall performance due to the "needle-in-the-haystack" problem. To address these challenges, we propose a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time, without including the full business policy in-context. Furthermore, we introduce a novel PolicyRecall reward based on the Jaccard score and a Hallucination Penalty for GRPO training. Altogether, our best model outperforms the baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 points, while using 40% fewer words.