Research map / Large language models / Language, safety and interpretability

Agent Instructs Large Language Models to be General Zero-Shot Reasoners

ICML 2024 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Zero-shot AgentInstruct, where a GPT-4 ReAct-style agent with web search generates one set of task-specific instructions per dataset, which are prepended to zero-shot chain-of-thought prompts for smaller LLMs (Vicuna-13b, Llama-2-70b-chat, GPT-3.5 Turbo) to improve zero-shot reasoning across 29 datasets.

Problem

Zero-shot reasoning of LLMs is well studied for arithmetic tasks, but unclear for general language understanding tasks spanning generation, classification, and reasoning.

Approach

An agent (GPT-4 with ReAct and web-retrieval tools) generates a single task-level set of step-by-step instructions once per dataset, and these instructions replace zero-shot CoT's fixed "Let's think step by step" prompt for reasoning LLMs, decoupling agent planning from per-instance reasoning and acting as a cost-effective distillation setup.

Key results

• Zero-shot AgentInstruct wins on 20 of 29 datasets and improves average performance over zero-shot CoT across Vicuna-13b, Llama-2-70b-chat, and GPT-3.5 Turbo.
• Llama-2-70b-chat with zero-shot AgentInstruct outperforms standard zero-shot GPT-3.5 Turbo on average.
• On AddSub with GPT-4, zero-shot AgentInstruct outperforms zero-shot GPT-4 by 8.6% and ties zero-shot CoT GPT-4 at about one-tenth the cost.

Measured effects reported in the paper

Surprising findings

• Merely prepending dataset information to zero-shot CoT does not consistently improve performance, while synthesizing it into agent instructions does.
• Zero-shot AgentInstruct reaches competitiveness with 5-shot prompting, beating it on AddSub and NarrativeQA.

Limitations

• Errors dominated by incorrect reasoning (32%) and answer ambiguity (22.7%) in error analysis.
• Instructions can be taken too literally or partially ignored by the reasoning model.
• Single run per experiment due to cost; results may vary due to temperature-based agent generation.

Experimental setup

Datasets: HELM core scenarios AddSub AQuA BoolQ CivilComments CNN/Daily Mail Coin Flip CommonsenseQA Date Understanding GSM8K HellaSwag IMDB

Benchmarks: AddSub AQuA BoolQ CivilComments CNN/Daily Mail Coin Flip CommonsenseQA Date Understanding GSM8K HellaSwag IMDB Last Letter Concatenation

Metrics: accuracy quasi-exact match quasi-prefix exact match F1 ROUGE-2

Baselines: zero-shot prompting zero-shot chain-of-thought few-shot prompting self-consistency ReAct zero-shot GPT-4

Builds on: ReAct zero-shot chain-of-thought self-consistency

Closest papers by meaning

• Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.89)
• Improving Factuality and Reasoning in Language Models through Multiagent Debate (ICML 2024 · similarity 0.89)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.88)
• From Passive to Active Reasoning: Can Large Language Models Ask the Right Questions under Incomplete Information? (ICML 2025 · similarity 0.88)
• T1: Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling (ICML 2025 · similarity 0.87)
• Absolute Zero: Reinforced Self-play Reasoning with Zero Data (NeurIPS 2025 · similarity 0.87)