The ReAct (Reasoning + Action) capability in large language models (LLMs) has become the foundation of modern agentic systems. Recent LLMs, such as DeepSeek-R1 and OpenAI o1/o3, exemplify this by emphasizing reasoning through the generation of ample intermediate tokens, which help build a strong premise before producing the final output tokens. In this paper, we introduce Pre-Act, a novel approach that enhances the agent's performance by creating a multi-step execution plan along with the detailed reasoning for the given user input. This plan incrementally incorporates previous steps and tool outputs, refining itself after each step execution until the final response is obtained. Our approach is applicable to both conversational and non-conversational agents. To measure the performance of task-oriented agents comprehensively, we propose a two-level evaluation framework: (1) turn level and (2) end-to-end. Our turn-level evaluation, averaged across five models, shows that our approach, Pre-Act, outperforms ReAct by 70% in Action Recall on the Almita dataset. While this approach is effective for larger models, smaller models crucial for practical applications, where latency and cost are key constraints, often struggle with complex reasoning tasks required for agentic systems. To address this limitation, we fine-tune relatively small models such as Llama 3.1 (8B & 70B) using the proposed Pre-Act approach. Our experiments show that the fine-tuned 70B model outperforms GPT-4, achieving a 69.5% improvement in action accuracy (turn-level) and a 28% improvement in goal completion rate (end-to-end) on the Almita (out-of-domain) dataset.

We present a hierarchical framework to solve robot planning as an input control problem. At the lowest level are temporary closed control loops, ("tasks"), each representing a behaviour, contingent on a specific sensory input and therefore temporary. At the highest level, a supervising "Configurator" directs task creation and termination. Here resides "core" knowledge as a physics engine, where sequences of tasks can be simulated. The Configurator encodes and interprets simulation results, based on which it can choose a sequence of tasks as a plan. We implement this framework on a real robot and test it in an overtaking scenario as proof-of-concept.

Real-world multi-modal problems are rarely solved by a single machine learning model, and often require multi-step computational plans that involve stitching several models. Tool-augmented LLMs hold tremendous promise for automating the generation of such computational plans. However, the lack of standardized benchmarks for evaluating LLMs as planners for multi-step multi-modal tasks has prevented a systematic study of planner design decisions. Should LLMs generate a full plan in a single shot or step-by-step? Should they invoke tools directly with Python code or through structured data formats like JSON? Does feedback improve planning? To answer these questions and more, we introduce m&m's: a benchmark containing 4K+ multi-step multi-modal tasks involving 33 tools that include multi-modal models, (free) public APIs, and image processing modules. For each of these task queries, we provide automatically generated plans using this realistic toolset. We further provide a high-quality subset of 1,565 task plans that are human-verified and correctly executable. With m&m's, we evaluate 6 popular LLMs with 2 planning strategies (multi-step vs. step-by-step planning), 2 plan formats (JSON vs. code), and 3 types of feedback (parsing/verification/execution). Finally, we summarize takeaways from our extensive experiments. Our dataset and code are available on HuggingFace (https://huggingface.co/datasets/zixianma/mnms) and Github (https://github.com/RAIVNLab/mnms).

Unlike myopic HPO methods, planning based approaches fundamentally require building models of the future to assess the impact of a current decision on later timesteps. Though these methods also rely on a GP as a surrogate model, each point in multi-step planning involves fantasizing/imagining an updated GP posterior ( ft 1 xt),…,( ft h xt, xt 1,…, xt h 1) based on simulated choices from lookaheads {( xt, yt),…,( xt h 1, yt h 1)} (Lam et al., 2016; Jiang et al., 2020). Note that we use xt to represent a fantasized decision, while xt is the actual choice made at timestep t. Whilst multi-step planning is promising, constructing the posterior of a GP model requires matrix inversion which is a compute-intensive operation (Cormen et al., 2022). Even outside of this limitation, traditional planning based approaches are compute intensive due to (i) poor scaling behavior of the search tree--O(qh) where q is the number of choices at each decision point for each lookahead step (Lam et al., 2016; Lam and Willcox, 2017)--which forces most methods to explore short horizons, typically h {1,2}, and (ii) nested expectation and maximization: marginalizing future observation yt j,j h and global search on the acquisition function to obtain query xt j at every lookahead step.