ReasonAct: Interleaving Reasoning and Action
- ReasonAct is a family of computational frameworks that interleave reasoning traces with external actions to refine decision-making and gather evidence.
- The framework enables models to leverage explicit reasoning for planning, prediction, and environment interaction, reducing errors like hallucination.
- Variants such as Focused ReAct, PreAct, and Co-ReAct demonstrate improvements in accuracy, runtime, and decision validation across applications.
ReasonAct denotes a family of reasoning–acting frameworks in which a model interleaves explicit reasoning traces with actions that query, manipulate, or inspect an external environment, then conditions later reasoning on the resulting observations. Across the literature, the term appears both as a generic label for this paradigm and as part of specific method names. In its canonical large-language-model form, ReasonAct is associated with ReAct, while later work extends the same basic idea to graph inference, video reasoning, rubric-guided search, multi-agent adjudication, proactive reward modeling, and planner–executor architectures (Yao et al., 2022, Yu et al., 8 May 2026, Liu et al., 3 Aug 2025, Kang et al., 22 May 2026).
1. Origins and conceptual scope
The foundational formulation is ReAct, introduced as “Reason + Act,” where a frozen LLM is prompted to generate both reasoning traces and task-specific actions in an interleaved manner. The central claim is that reasoning traces help the model induce, track, and update action plans as well as handle exceptions, while actions let it interface with external sources such as knowledge bases or environments to gather additional information. ReAct was evaluated on question answering, fact verification, and interactive decision-making benchmarks, including HotpotQA, FEVER, ALFWorld, and WebShop, and it was positioned against both pure chain-of-thought prompting and pure acting or planning approaches (Yao et al., 2022).
Within this literature, ReasonAct is not a single algorithmic object. Some works use it as an umbrella for any framework that alternates reasoning with action; others instantiate it as a task-specific training method or controller. The terminology therefore spans prompt-only systems, supervised and reinforcement-trained agents, judge-based extensions, planning-first variants, and domain-specialized action spaces. A common denominator is that intermediate reasoning is operational rather than merely expository: it selects, constrains, or revises action.
2. Canonical ReAct formulation
The canonical loop is the interleaving of reasoning, acting, and observation. In the graph-oriented reformulation of the paradigm, this is written as
and the original ReAct paper formalizes the agent context at time as
with ReAct extending the ordinary action space to
where denotes environment actions and denotes language actions, or “thoughts” (Yao et al., 2022).
This formulation distinguishes ReasonAct from two nearby traditions. Relative to chain-of-thought prompting, ReasonAct does not treat reasoning as a closed internal process; it permits search, lookup, navigation, clicking, calculator use, and other environment-facing actions. Relative to pure acting systems, it does not suppress explicit internal state updates; natural-language thoughts decompose goals, track subgoals, summarize observations, and select the next action. In ReAct’s own analysis, this coupling reduced hallucination in retrieval-heavy tasks and produced human-like, interpretable trajectories, while later work retained the same skeleton but altered the controller, action space, or training regime.
3. Architectural variants and control strategies
Subsequent ReasonAct systems differ mainly in where they intervene: in prompt control, planning, verification, action scheduling, or reward assignment. The recurrent design problem is that a vanilla ReasonAct loop can lose focus, repeat actions, stop too early, or reason over poorly structured evidence.
| Variant | Main addition | Representative result |
|---|---|---|
| Focused ReAct | Reiteration of the original question at each step; early stop on repeated actions | Accuracy gains of 18% to 530% and runtime reduction of up to 34% on HotPotQA (Li et al., 2024) |
| PreAct | Adds prediction of possible feedback types and contingent plans | Improves success rate over ReAct across AgentBench tasks; gains increase when more historical predictions are retained (Fu et al., 2024) |
| Reasoning Court | Two ReAct-style agents plus a judge over full trajectories | HotpotQA EM 36.2 → 42.2; FEVER EM 64.8 → 74.0; MuSiQue EM 30.4 → 36.0 (Wu et al., 14 Apr 2025) |
| Co-ReAct | RL-trained rubric generator guides each step and verifies it | DRB Avg 34.23 → 36.92 and SQA-CS-V2 Avg 72.73 → 76.05 with Qwen3-14B search agent (Kang et al., 22 May 2026) |
| RP-ReAct | Reasoner-planner supervises one or more ReAct executors; context-saving for large tool outputs | Higher Combined Performance Score on several hard ToolQA domains, especially SciREX, Airbnb, and Coffee (Molinari et al., 3 Dec 2025) |
| ProRe | Proactive reward via reasoner–actor collaboration and targeted probing | Reward accuracy 93.7 and F1 83.0 on average; policy success rate improvement of up to 22.4% (Dai et al., 26 Sep 2025) |
Focused ReAct is the smallest intervention: it keeps the ReAct trace format but restates the original question at every reasoning step and terminates when the wrapper detects an identical repeated action. PreAct instead augments the loop with prediction, forcing the model to enumerate possible next observations and corresponding handling strategies. Reasoning Court introduces a distinct judgment phase in which a judge selects between or synthesizes answers from multiple reasoning–retrieval trajectories. Co-ReAct inserts a trained rubric generator into the action-selection loop, so that each step is guided by explicit criteria and then checked against them. RP-ReAct separates strategic planning from low-level ReAct execution, primarily to control trajectory instability and context blowup in enterprise-style tool use. ProRe repurposes the same Reason–Act split for reward modeling: a reasoner schedules state-probing tasks and evaluator agents act in the GUI environment to collect missing evidence before a final reward is assigned.
4. Training-based ReasonAct and the compilation of reasoning routines
A distinct line of work treats ReasonAct not as a prompting pattern but as a training target. In video understanding, “ReasonAct: Progressive Training for Fine-Grained Video Reasoning in Small Models” uses a three-stage curriculum: Text-only Foundational Reasoning Enhancement, Video-Specific Chain-of-Thought Fine-Tuning, and Temporal-Aware Reinforcement Learning with sub-action rewards and temporal consistency. Using Qwen2.5-VL-3B-Instruct as the base model, the full system reaches 67.2% on HMDB51, 94.1% on UCF-101, and 78.9% on Kinetics-400, improving over the baseline by 17.9, 15.8, and 12.3 points respectively (Liu et al., 3 Aug 2025).
A closely related agentic line is Pre-Act, which extends ReAct by generating a multi-step execution plan with detailed reasoning for each step and then refining the remaining plan after each tool output. On the Almita dataset, turn-level evaluation averaged across five models shows that Pre-Act outperforms ReAct by 70% in Action Recall, and a fine-tuned Llama 3.1 70B model is reported to outperform GPT-4 with a 69.5% improvement in action accuracy and a 28% improvement in goal completion rate on Almita (Rawat et al., 15 May 2025).
Another training-based direction is ReST-style self-improvement for ReAct agents. “ReST meets ReAct” defines a search agent with decision, search, summarization, answer, and self-critique stages, then iteratively improves it with growing-batch reinforcement learning using AI feedback over previously generated trajectories. After two iterations, a fine-tuned small model is reported to achieve comparable performance on compositional question-answering benchmarks with two orders of magnitude fewer parameters than the prompted large model (Aksitov et al., 2023).
A more structural approach is Reasoning Primitive Induction, which mines successful ReAct traces, clusters recurrent reasoning moves, and converts frequent moves into typed pseudo-tools that a standard ReAct loop can call at test time. The induced libraries outperform the source ReAct agent by +44 percentage points on RuleArena NBA, +30 on MuSR team allocation, and +22 on NatPlan meeting planning, indicating that recurrent scratchpad routines can be externalized as reusable reasoning operators rather than rediscovered instance by instance (Lei et al., 2 Jun 2026).
5. Domain-specific instantiations and empirical behavior
GraphReAct is the clearest domain-specific reconstruction of the ReasonAct idea. It defines a graph action space with topological retrieval, semantic retrieval, and context refinement, thereby replacing unstructured search actions with graph-aware operations. In a cross-dataset, zero-shot node-classification setting on text-attributed graphs, it is best on most target datasets, including Cora 0.273 versus 0.202 for TEA-GLM, History 0.645 versus 0.528, Photo 0.523 versus 0.497, and Sports 0.483 versus 0.404, while TEA-GLM is slightly better on PubMed, 0.848 versus 0.819. The same study reports that reduces to TEA-GLM, that is used in experiments, and that graph-internal evidence outperforms a standard Wikipedia Search action (Yu et al., 8 May 2026).
A separate empirical axis concerns decoding. A systematic analysis of ReAct with CAD, DoLa, and DeCoRe shows that faithfulness-oriented decoding can materially improve a ReasonAct agent’s use of retrieved evidence. On HotpotQA with Qwen2-7B-Instruct, ReAct with standard decoding obtains answer F1 19.5, while ReAct with DoLa reaches 32.6; on 2WikiMultihopQA, ReAct with CAD reaches 51.8 versus 43.1 for standard ReAct. The same paper reports that faithful decoding helps ReAct much more consistently than it helps one-step retrieval, which suggests that retrieval quality and decoding policy interact rather than contribute independently (Murphy et al., 30 Mar 2025).
In GUI environments, ProRe extends the same principle from action selection to reward assignment. Over more than 3K trajectories from AndroidWorld, AndroidLab, and MobileAgentBench, ProRe raises average reward accuracy to 93.7 and F1 to 83.0, compared with 88.4 and 63.6 for the best baseline. When used for test-time scaling, it improves success rate by up to 22.4%, which indicates that better reward models can be a first-order component of the ReasonAct stack rather than a peripheral evaluator (Dai et al., 26 Sep 2025).
6. Limitations, diagnostics, and relation to classical reasoning about actions
The ReasonAct literature is united as much by its recurring failure modes as by its successes. Vanilla ReAct can lose focus on the original question and enter repetitive action loops; Focused ReAct targets exactly those two pathologies. Judge-based systems such as Reasoning Court are strong when agents disagree, but the judge almost never overturns cases where both agents produce the same wrong answer. Co-ReAct improves step quality, but untrained rubrics degrade performance, and verification adds inference-time overhead. GraphReAct reports computational overhead, dependence on node text, and prompt-length constraints. RP-ReAct improves robustness on hard ToolQA domains but requires sufficient base-model competence; the smallest open-weight models in that study fail to follow the protocol reliably. Video ReasonAct depends on a manual biomechanical sub-action library and is evaluated on relatively short clips (Li et al., 2024, Wu et al., 14 Apr 2025, Kang et al., 22 May 2026, Yu et al., 8 May 2026, Molinari et al., 3 Dec 2025, Liu et al., 3 Aug 2025).
A broader diagnostic picture comes from classical reasoning about actions and change. ActionReasoningBench evaluates LLMs on Fluent Tracking, State Tracking, Action Executability, Effects of Actions, Numerical RAC, and Composite Questions, and reports average accuracy rates of 73.55%, 65.63%, 58.73%, and 62.38% on the former four dimensions, but only 33.16% and 51.19% on the latter two, reflecting a 17.9% performance decline. It also introduces ramification constraints for indirect effects and reports that, particularly in handling ramifications, GPT-4o fails to solve any question and o1-preview achieves a score of only 18.4% (Handa et al., 2024).
That diagnostic result helps explain why symbolic approaches remain relevant as antecedents and complements. KALMRA, a factual-English authoring system for rules and temporally indexed actions built on DLV and Simplified Event Calculus, reports 100% correctness on rule authoring for UTI guidelines and more than 99.3% correctness on bAbI action reasoning. This contrast does not imply that symbolic knowledge authoring and ReasonAct agents are interchangeable, but it does suggest a durable research direction: pairing LLM-based reasoning–acting controllers with explicit logical representations of actions, rules, timestamps, and ramifications when precise state semantics are indispensable (Wang et al., 2023).
ReasonAct therefore denotes less a closed method than a persistent architectural thesis: reasoning should remain explicit enough to guide action, action should remain structured enough to ground reasoning, and the interface between them can be shaped by planning, prediction, judges, rubrics, pseudo-tools, external memory, or symbolic state models. The diversity of current systems suggests that the central problem is no longer whether reasoning and acting should be interleaved, but how that interleaving should be controlled, trained, verified, and represented in each domain.