PreAct: Prediction Enhances Agent's Planning Ability

TL;DR

PreAct addresses the gap between forecasted observations and actual results in LLM agent planning by introducing a prediction-augmented loop that synchronizes prediction, reasoning, and action. The method prompts the LLM to generate future observations $p$ at each step and incorporate $p$ into history, improving planning diversity and directional strategy. Experimental results across AgentBench datasets HH, OS, DB, LTP and HotpotQA TOT show that PreAct outperforms ReAct, with larger gains in permanent and reflexion modes, and gains persist as historical predictions accumulate; improvements hold across TOT and memory variants, though LTP can suffer from refusals under safety constraints. Ablations reveal a sustained positive role for $p$-history, increasing success rates on multiple tasks and suggesting two planning-evaluation metrics for future reinforcement learning reward shaping. The work points to future directions combining PreAct with longer-term memory and trajectory-based fine-tuning to further enhance agent planning.

Abstract

Addressing the disparity between forecasts and actual results can enable individuals to expand their thought processes and stimulate self-reflection, thus promoting accurate planning. In this research, we present **PreAct**, an agent framework that integrates **pre**diction, **rea**soning, and **act**ion. By utilizing the information derived from predictions, the large language model (LLM) agent can provide a wider range and more strategically focused reasoning. This leads to more efficient actions that aid the agent in accomplishing intricate tasks. Our experimental results show that PreAct surpasses the ReAct method in completing complex tasks and that PreAct's performance can be further improved when paired with other memory or selection strategy techniques. We presented the model with varying quantities of historical predictions and discovered that these predictions consistently enhance LLM planning.The variances in single-step reasoning between PreAct and ReAct indicate that PreAct indeed has benefits in terms of diversity and strategic orientation over ReAct.

Figures (6)

• Figure 1: Comparison between ReAct and PreAct. The scene start from $Env_{k-1}$ and $Pred_{k-1}$, The $Obs_{k-1}$ comes from the $Env_{k-1}$ and $Action_{k-1}=go\; right$. Env = environment, Obs=observation, Pred=prediction.
• Figure 2: Historical Prediction's Influence. 0 refers to ReAct's history, 1 refers to immediate mode history and all refers to permanent mode history.
• Figure 3: Two representative examples in DB and HH set between ReAct and PreAct. We omit unimportant information in the example. Act=Action, Obs=observation, Pred=prediction.
• Figure 4: Overall Diversity Comparison between ReAct and PreAct
• Figure 5: Correlation Analysis in HH Dataset