Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Pre-Act: Multi-Step Planning and Reasoning Improves Acting in LLM Agents

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that an LLM agent which writes and revises a multi-step plan before each action gets far better tool-use accuracy, and that a 70B model fine-tuned on such plans beats GPT-4 on out-of-domain agent tasks.

desk verdict Useful ReAct variant with honest vanilla-model results, but the fine-tuning headline is not yet supported: no ReAct fine-tuned control, so the gains may come from the proprietary data and QLoRA, not from Pre-Act. read the letter →

arxiv 2505.09970 v2 pith:BKZDHBA2 submitted 2025-05-15 cs.AI

Mrinal Rawat , Ambuje Gupta , Rushil Goomer , Alessandro Di Bari , Neha Gupta , Roberto Pieraccini This is my paper · ORCID
classification cs.AI
keywords LLMagentsmulti-stepplanningReActtoolcallingcurriculumfine-tuninggoalcompletionLLM-as-a-judgemilestonegraph
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's central claim is that an LLM agent acts more reliably when, before each action, it writes out a multi-step plan of the remaining actions and revises that plan after every tool result, a method the authors call Pre-Act. Against ReAct, which reasons only about the immediate next action, Pre-Act reports a 70% average improvement in action recall across five models on the Almita dataset and a 102% improvement on the authors' proprietary dataset. The paper further claims that small models fine-tuned in two stages on Pre-Act traces can match or beat much larger proprietary ones: the fine-tuned Llama 3.1 70B beats GPT-4 by 69.5% in turn-level action accuracy and 28% in end-to-end goal completion on out-of-domain Almita. A reader should care because the result points to plan structure, not model scale, as a lever for agent reliability, and because the paper pairs it with a two-level evaluation that separates correct tool calls from actually finishing the task.

What carries the argument

The load-bearing object is the multi-step plan written inside the model's internal 'Thought'. Pre-Act defines a structured format: at each turn the model must produce Previous Steps, a summary of executed actions and observations, and Next Steps, a numbered plan with detailed reasoning for each upcoming action ending with the final answer. The plan is regenerated after each observation, so the model is always committing to the remaining trajectory rather than only the next tool call. Training uses curriculum fine-tuning: stage one teaches tool calling on a large ReAct-formatted dataset, and stage two uses expert-annotated Pre-Act traces on a small proprietary dataset with parameter-efficient LoRA updates to preserve stage-one skills. Evaluation's key object is the milestone dependency graph, generated by GPT-4, human-verified, with functional milestones tied to tool names and non-functional milestones tied to workflow states; goal completion is measured by how far along this graph the agent's actions are judged to have advanced.

What would settle it

Re-score the end-to-end Almita conversations with human annotators instead of the GPT-4 judge, comparing milestone completion against the same human-verified graphs; if human scores do not reproduce the reported 0.82 versus 0.64 goal-completion gap, the end-to-end claim is an evaluation artifact.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that planning should come before acting and should be revised after every observation. Whereas ReAct generates a thought tied to the immediate action, Pre-Act makes the model commit to the full remaining trajectory at each step: it writes the previous steps already taken and a numbered plan of next steps, executes the next tool call, folds the observation into the plan, and repeats. The paper reports that this change alone raises action recall over ReAct by 102% on its proprietary dataset and 70% on Almita, averaged across five models. It then shows that small models trained in two stages—first to imitate ReAct tool-calling at scale, then to imitate expert-annotated Pre-Act plans on a small dataset—retain the tool-calling skill and gain the planning skill; the fine-tuned 70B model reaches 0.9238 action recall on Almita and 0.82 average goal completion across five end-to-end use-cases, compared to 0.5449 and 0.64 for GPT-4 with Pre-Act.

Load-bearing premise

The end-to-end result rests on the assumption that GPT-4's milestone-scoring, human-checked only at graph-construction time, agrees with what a human would count as task completion; the paper itself concedes 'volatility in LLM-as-a-judge assessments' in its conclusion.

Editorial extensions

If this is right

  • In setups similar to the five tested, replacing ReAct's single-step thought with a regenerated multi-step plan should raise the chance that the right tool is called; the averaged action-recall gain reported is 70% on Almita.
  • Fine-tuning a 70B model on Pre-Act traces is enough to beat a much larger proprietary model in this task family: 69.5% higher turn-level action accuracy and 28% higher end-to-end goal completion on Almita, with lower latency and cost.
  • The two-level evaluation separates tool-call correctness from task completion, so a system that picks the right tools but never finishes the workflow will no longer look fully successful.
  • A curriculum that first teaches tool calling and then refines on expert-annotated plan traces avoids the usual forgetting: Glaive action recall drops only 0.80 points for the 8B model and 0.36 for the 70B after stage two.
  • The same Pre-Act prompting also helps GPT-4 in end-to-end settings, raising its average goal completion from 0.32 with ReAct to 0.64 with Pre-Act.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the end-to-end gains should be read with the judge in the loop, because the milestone graphs and the judge are both GPT-4; part of the 28% gap may reflect the fine-tuned model producing traces that match the judge's preferred format, and an independent human re-scoring would settle this.
  • Editorial inference: because Pre-Act's gains appear on out-of-domain tools, the mechanism may be transferable planning behavior rather than memorized tool schemas; a testable extension is to freeze the planner and vary unseen tools to see whether plan quality predicts action accuracy.
  • Editorial inference: the planning format could be combined with inference-time strategies such as sampling multiple plans or backtracking on failed observations, which the paper does not explore; if plan regeneration is what drives recovery, those combinations should compound.
  • Editorial inference: if planning before action is the active ingredient, then forcing a plan after choosing an action should not help; a simple ablation that swaps the order of plan and action at inference would test whether lookahead causes the gain or merely correlates with it.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes Pre-Act, a modification of ReAct in which the agent generates a multi-step execution plan with per-step reasoning before each action, refining the plan as tool observations arrive. The authors evaluate Pre-Act as a prompting strategy on five pretrained models (Llama 3.1 8B/70B, Nemotron 70B, DeepSeek-distill 70B, GPT-4) on three datasets (Glaive, proprietary, Almita) using turn-level metrics (action recall, F1, parameter match, similarity), and as a fine-tuning target for Llama 3.1 8B/70B using a two-stage curriculum (Glaive in ReAct format, then proprietary data in Pre-Act format). They also propose an end-to-end evaluation that uses GPT-4 to build milestone dependency graphs and GPT-4 as a judge to measure goal completion and progress rate on five Almita use-cases. The headline claims are that Pre-Act improves turn-level action recall over ReAct by 70% on average (Almita) and that a fine-tuned 70B model outperforms GPT-4 by 69.5% in turn-level action accuracy and by 28% in end-to-end goal completion rate on Almita.

Significance. If the results hold, Pre-Act is a simple, broadly applicable prompting and fine-tuning scheme that could improve agent performance across model sizes. The turn-level comparisons in Table 2 are internally consistent, the authors provide code and dataset contributions for Almita, and the curriculum-learning retention analysis in Table 4 is a useful check. However, the central claims for the fine-tuned models rely on a missing control: no model is fine-tuned on the same proprietary data in ReAct format, so the reported gains over GPT-4 and vanilla models cannot be attributed to multi-step planning alone. The end-to-end evaluation is a second-order validity risk because GPT-4 generates the milestones and also serves as the judge, and the paper itself concedes judge volatility. The significance of the paper is therefore conditional on whether the authors can disentangle Pre-Act from the effect of additional domain-specific fine-tuning.

major comments (4)
  1. [§3.2, Tables 2–4] The attribution of the fine-tuned gains to Pre-Act is not supported by the experimental design. The two-stage curriculum fine-tunes on Glaive in ReAct format (Stage 1) and then on proprietary data in Pre-Act format (Stage 2), but no model is ever fine-tuned on the same proprietary data in ReAct format. The only curriculum control, Table 4, measures retention on Glaive across stages, not the effect of Pre-Act versus ReAct. Consequently, the 69.5% turn-level action-accuracy improvement and the 0.82 goal-completion rate reported for the fine-tuned 70B model in Section 5 could be entirely produced by the additional high-quality proprietary annotations and QLoRA fine-tuning, independent of the multi-step plan. The authors should add a Stage-2 control fine-tuned on the proprietary dataset formatted as ReAct, with the same data split, training hyperparameters, and evaluation protocol, and report the comparison in Tables 2 and 3.
  2. [§4.3.1, Appendices F/H] The end-to-end evaluation is a closed loop: GPT-4 generates the milestone dependency graphs (Appendix F), a GPT-4-based synthetic user simulates the conversation, and GPT-4 as judge scores milestone completion (Appendix H). The paper's conclusion concedes 'volatility in LLM-as-a-judge assessments,' yet no agreement study against human judges, no alternative judge, and no per-run variance are reported. Without such evidence, the 28% goal-completion gain of the fine-tuned model over GPT-4 Pre-Act could be an artifact of the judge favoring responses whose style resembles the model's own milestone descriptions. Please report the distribution and confidence intervals over the 50 simulated runs, and provide a human-annotation agreement analysis or an alternative judge (e.g., a different LLM, or rule-based verification against tool outputs) to establish that the milestone-completion metric measures real task success.
  3. [§5, Tables 2–3] Several headline comparisons lack statistical support. All turn-level metrics are single runs without multiple seeds or error bars, and the E2E averages are reported without variances over the 50 runs. For claims of 70%, 128%, and 69.5% relative improvements on datasets with only a few hundred test instances (Table 1: 378 proprietary, 1100 Almita), the authors should provide bootstrap confidence intervals or significance tests. Otherwise the reader cannot judge whether the differences exceed run-to-run variability. This is particularly important because the fine-tuned rows in Table 2 are compared against prompted baselines rather than equally trained alternatives, as noted in the first comment.
  4. [§5] The framing that the fine-tuned 70B model 'outperforms GPT-4' overstates the scientific comparison: it juxtaposes a model fine-tuned on task-specific proprietary data (plus the Almita annotations created by the same team) with a prompted generalist. This comparison is legitimate as an engineering benchmark but does not isolate Pre-Act as the cause of the improvement. The abstract and conclusion should be reworded to distinguish the effect of the Pre-Act prompting strategy (supported by the vanilla-model rows in Table 2) from the combined effect of fine-tuning and the Pre-Act format, unless the ReAct fine-tuned control from the first comment demonstrates that the fine-tuning benefit is specific to Pre-Act.
minor comments (6)
  1. [Figure 1] In the Pre-Act example, the observation text says 'get_my_headlines' while the tool is 'get_news_headlines'; this typo should be fixed.
  2. [Table 2 caption] The entries '-' for vanilla-model Pre-Act on Glaive are not explained in the caption; please state explicitly that Pre-Act annotations are unavailable for Glaive, so those comparisons are not performed.
  3. [Appendix A] Training details are incomplete: please report the LoRA rank, alpha, target modules, quantization settings, number of GPUs, total runtime, and any distributed training setup for the 70B model, as these are needed for reproducibility.
  4. [§3.2] The text says 'LoRA training' in one place and 'QLoRA' in Appendix A; please clarify whether quantized base models were used and specify the exact configuration.
  5. [Appendix F/G] The milestone dependency graph is described as generated in YAML format, but no machine-readable example is provided; Figure 6 is an image. Please include the actual YAML used for the Almita use-cases.
  6. [§5] The statement that Pre-Act improves action recall by 70% on average 'across five models' should specify that this is the mean of per-model relative improvements, since Table 2 shows the relative gains vary from roughly 23% to 168% across models.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Pre-Act is an empirical prompting/fine-tuning comparison; the headline gains rest on missing controls and judge validity, not on construction.

full rationale

I walked the claimed derivation chain: Pre-Act's contribution is a prompting/training format (Sec 3.1-3.2), and its evidence is (i) turn-level comparisons of ReAct vs Pre-Act on the same pretrained models (Table 2) and (ii) E2E comparisons of GPT-4 and fine-tuned Llama (Table 3). Neither comparison computes a target from a fitted parameter: action recall is measured against ground-truth actions, and the fine-tuned models are not evaluated on their own training labels. The Almita test annotations were produced by the same team that created Pre-Act training annotations, and the E2E pipeline uses GPT-4 to generate and then judge milestones; these are methodological validity/contamination risks, not cases where the result is equal to an input by definition. The cited self-work (Levin and Pieraccini 1997; Pieraccini and Huerta 2005) is background, not load-bearing. The absence of a ReAct fine-tuned control means the fine-tuning gains cannot cleanly be attributed to Pre-Act, but a missing control is a confound, not circularity. I therefore find no circular step.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters or invented entities; the method has no theoretical derivation. The central empirical claims rely on the validity of the datasets, the annotation process, and the LLM-judge scoring.

assumptions (3)
  • domain assumption Almita transformed by the authors' team into Pre-Act format is a valid out-of-domain test set.
    The test distribution matches their annotation format, so success may partly measure format alignment rather than raw task ability.
  • domain assumption GPT-4 as LLM-judge reliably detects milestone completion and rejects hallucinated tool calls.
    No validation against human judgments or inter-annotator agreement is reported; authors note volatility in future work.
  • domain assumption Action recall on ground-truth tool calls is a meaningful proxy for task success.
    The turn-level metric only checks the predicted action and parameters, not the downstream outcome of that action.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pre-Act: Multi-Step Planning and Reasoning Improves Acting in LLM Agents." pith.science (2026). https://pith.science/paper/BKZDHBA2

@misc{pith2026250509970,
  author       = {Pith},
  title        = {Pith review of: Pre-Act: Multi-Step Planning and Reasoning Improves Acting in LLM Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BKZDHBA2}},
  note         = {Machine review of arXiv:2505.09970}
}
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2505.09970 by the authors.

Figure 1
Figure 1. Side-by-side comparison of ReAct and Pre-Act (our approach) for an example use-case from the Glaive [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Sample Output for Glaive Dataset - ReAct [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 4
Figure 4. Prompt Template used for Pre-Act : Multi-Step Planning [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Prompt Template for Milestones Creation terconnections between various milestones within a workflow. It highlights both Functional Milestones (FC), which correspond to specific tool functions, and Non-Functional Milestones (NFC), represent￾ing conditions, states, or tr…
Figure 6
Figure 6. Figure 6: Sample Milestone Dependency Graph H Prompt Template for GPT-4 as Judge (E2E evaluation) As shown in [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Prompt Template for GPT-4 as Judge (E2E evaluation) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents

    cs.AI 2026-08 conditional novelty 6.0 of 10

    TrajWiki stores long-dialogue facts as evolving, source-linked claim histories organized into wiki pages, improving long-horizon QA on LoCoMo and MedMT-Bench in reported experiments.

  2. NeSyFS: A Neuro-symbolic Fast-Slow Thinking Framework for LLM Agent under Partial Observability

    cs.AI 2026-07 conditional novelty 6.0 of 10

    NeSyFS combines a knowledge-graph belief state, a reflection step, and TSMC-style particle planning to improve LLM agent success rates on ALFWorld, WebShop, and ScienceWorld.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages · cited by 2 Pith papers

  1. [1]

    action":

    I will now proceed with the final answer because ... (explanation) Follow this format (flow): Question: input question to answer Thought: consider previous and subsequent steps and conversation. Summary for what you did previously (ONLY IF function calls were made for the last user request) and create the multi-step plan. Action: ``` $JSON_BLOB ``` Observ...

    work page

  2. [2]

    Otherwise, provide Previous Steps: NA and Next Steps: .. Action: ``` {

    I will do ... (action2) with the detailed reasoning. k. I will do ... (actionk) with the detailed reasoning. k+1. I will now proceed with the final answer because ... (explanation) Example Output: When responding to human, please output a response only in one of two formats (strictly follow it): **Option 1:** If function calls were made for the last human...

    work page

  3. [3]

    This checkpoint ensures that the proper branch is taken based on the order status

    Contextual Notes: - Include any additional information that helps clarify why this milestone is significant in the workflow (e.g., "This checkpoint ensures that the proper branch is taken based on the order status")

    work page

  4. [4]

    (action1) with the detailed reasoning

    I will first do ... (action1) with the detailed reasoning

    work page

  5. [5]

    ================================================================ Example 1 : Workflow with instructions : You are an expert customer service agent specializing in order tracking

    The milestone graph generated should be in the YAML format. ================================================================ Example 1 : Workflow with instructions : You are an expert customer service agent specializing in order tracking. Your primary goal is to assist customers with their order tracking inquiries. Here is the workflow you need to follow:

    work page

  6. [6]

    - For non-functional milestones (NFC), this is usually a descriptive state or phrase

    name (str): - For functional milestones (FC), this must exactly match the tool’s function name. - For non-functional milestones (NFC), this is usually a descriptive state or phrase

    work page

  7. [7]

    FC" - Functional milestones are exactly function name given in the tool description. For example

    type (str): - "FC" - Functional milestones are exactly function name given in the tool description. For example "getBalance" or "Pay using credit card" is a functional milestone. - "NFC" - Non Functional Milestones are milestones that are not direct but are derived from the workflows. For example :-

    work page

  8. [8]

    User Agree to pay is a non functional milestone

    NFC can be natural language conditions derived from workflow - "User Agree to pay is a non functional milestone" to represent a condition derived from the workflow

    work page

Show all 24 references
  1. [9]

    Initialization

    State or Checkpoint Information - Describe the current state or phase in the process (e.g., "Initialization", "Validation complete", "Awaiting user confirmation")

  2. [11]

    Flow Transitions:- Explain how this NFC milestone links different parts of the process, such as transitions between major steps or conditional branches

  3. [12]

    description (str): - A clear description explaining the purpose or action of the milestone

  4. [13]

    **Important Instructions**

    dependencies (list): - A list of milestone names (from other nodes) that must occur before this one. **Important Instructions**

  5. [14]

    Start" milestone (type

    Always start with a "Start" milestone (type "NFC")

  6. [15]

    End" milestone (type

    Always include an "End" milestone (type "NFC")

  7. [16]

    If it is a Functional Milestone (FC) then you should have the exactly same name as the tool

  8. [17]

    From the workflow description you need to find out conditional branches, states, flows and then make FC and NFC accordingly

  9. [19]

    Get the order details from the system and inform the customer with all the details including order number, status, tracking and delivery details

  10. [20]

    Are you satisfied with the information provided, and can we wrap up the call?

    Try to wrap up the call by using the phrase "Are you satisfied with the information provided, and can we wrap up the call?"

  11. [21]

    - If the customer confirms non-receipt, transfer to human to escalate the situation

    If the system shows that the product was delivered: - Ask the customer to double-check the delivery location. - If the customer confirms non-receipt, transfer to human to escalate the situation

  12. [22]

    - If the customer accepts expedited shipping, mark problem resolved

    If the information shows that the order is still at the warehouse: - Provide an updated estimated delivery date and offer a discount or expedited shipping as a gesture of goodwill. - If the customer accepts expedited shipping, mark problem resolved

  13. [23]

    Are you satisfied with the information provided, and can we wrap up the call?

    Always use the phrase "Are you satisfied with the information provided, and can we wrap up the call?" and tool mark_problem_resolved to wrap up calls

  14. [24]

    Use transferToHuman if there is any query beyond the provided tool abilities

    Remember: In transit orders can't be expedited as it is a third party delivery service, and it takes usually 5-7 working days. Use transferToHuman if there is any query beyond the provided tool abilities. ALWAYS END the conversation using mark problem resolved tool: - If the c...

  15. [2000]

    Henry Lieberman

    A stochastic model of human-machine interac- tion for learning dialog strategies.IEEE Transactions on speech and audio processing, 8(1):11–23. Henry Lieberman. 1997. Autonomous interface agents. InProceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems, ...

    1997 arXiv

  16. [2023]

    action":

    Least-to-most prompting enables complex reasoning in large language models.Preprint, arXiv:2205.10625. A Training Details We used the TRL library (von Werra et al., 2020) from Hugging Face to train the Llama 3.1 8B and 70B models. Training was performed with the QLoRA (Dettmer...

    2020 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.