Source-linked AI summary
Chain of Mindset: Reasoning with Adaptive Cognitive Modes
Tianyi Jiang, Arctanx An, Hengyi Feng, Naixin Zhai, Haodong Li, Xiaomin Yu, Jiahui Liu, Hanwen Du, Shuo Zhang, Zhi Yang, Jie Huang, Youhua Li, Yongxin Ni, Huacan Wang, Ronghao Chen
TL;DR
Existing LLM reasoning methods generally apply fixed or initially selected strategies, leaving the question of state-dependent mindset choice unresolved. CoM provides training-free step-level orchestration of four specialized mindsets through a Meta-Agent and bidirectional Context Gate, achieving state-of-the-art results across six benchmarks. It outperforms the strongest baseline by 4.96% on Qwen3-VL-32B-Instruct and 4.72% on Gemini-2.0-Flash in overall accuracy.
Problem
Existing LLM reasoning methods do not dynamically select the most suitable mindset as contexts and intermediate reasoning states change.
Method
CoM is a training-free framework in which a Meta-Agent adaptively selects four heterogeneous mindsets and Context Gates filter information exchanged with specialized modules.
Results
CoM achieves state-of-the-art accuracy across six challenging benchmarks, outperforming the strongest baseline by 4.96% on Qwen3-VL-32B-Instruct and 4.72% on Gemini-2.0-Flash.
Takeaways & Limitations
CoM balances reasoning efficiency with consistent generalization across open-source and closed-source models without additional training.
Abstract
from arXiv · showhide
Human problem-solving is never the repetition of a single mindset, by which we mean a distinct mode of cognitive processing. When tackling a specific task, we do not rely on a single mindset; instead, we integrate multiple mindsets within the single solution process. However, existing LLM reasoning methods fall into a common trap: they apply the same fixed mindset across all steps, overlooking that different stages of solving the same problem require fundamentally different mindsets. This single-minded assumption prevents models from reaching the next level of intelligence. To address this limitation, we propose Chain of Mindset (CoM), a training-free agentic framework that enables step-level adaptive mindset orchestration. CoM decomposes reasoning into four functionally heterogeneous mindsets: Spatial, Convergent, Divergent, and Algorithmic. A Meta-Agent dynamically selects the optimal mindset based on the evolving reasoning state, while a bidirectional Context Gate filters cross-module information flow to maintain effectiveness and efficiency. Experiments across six challenging benchmarks spanning mathematics, code generation, scientific QA, and spatial reasoning demonstrate that CoM achieves state-of-the-art performance, outperforming the strongest baseline by 4.96\% and 4.72\% in overall accuracy on Qwen3-VL-32B-Instruct and Gemini-2.0-Flash, while balancing reasoning efficiency. Our code is publicly available at \href{https://github.com/QuantaAlpha/chain-of-mindset}{https://github.com/QuantaAlpha/chain-of-mindset}.
1 Introduction
Human problem-solving coordinates multiple cognitive modes as a task evolves, whereas existing LLM methods typically use one fixed strategy or select one strategy only at the start. Chain of Mindset addresses this gap through step-level adaptive orchestration across heterogeneous mindsets.
- Motivation: Human experts transition between spatial, convergent, divergent, and algorithmic reasoning as intermediate problem states evolve.A geometry proof may visualize the configuration, identify key relationships, explore auxiliary constructions, and verify the result computationally.
- Limitations of prior work: Existing LLM methods either apply one cognitive strategy throughout or choose a strategy at task onset, limiting adaptation to heterogeneous sub-tasks and intermediate results.Neither paradigm supports dynamic, state-dependent switching during reasoning.
- Proposed framework: CoM is a training-free agentic framework that dynamically orchestrates Spatial, Convergent, Divergent, and Algorithmic mindsets during one reasoning process.A Meta-Agent selects and invokes mindsets according to the evolving reasoning state.
- Proposed framework: A bidirectional Context Gate filters information exchanged across mindset modules to reduce cross-module interference while supporting switching.The mechanism selectively controls information flow between isolated reasoning contexts and the main chain.
- Evaluation: Across six challenging benchmarks, CoM outperforms all baselines while maintaining computational efficiency and generalizing across open-source and closed-source models without additional training.The evaluated domains include mathematics, coding, scientific question answering, and spatial reasoning.
2 Method
CoM formulates mindset selection as sequential decision-making over reasoning history and separates meta-cognitive control from specialized execution modules. Its architecture iteratively plans, calls a selected mindset, internalizes insights, and filters context for subsequent steps.
- Problem formulation: CoM formalizes a mindset as a specialized reasoning paradigm with a distinct strategy, isolated context, and structured output.Mindsets are functionally heterogeneous rather than interchangeable strategies.
- Problem formulation: The framework represents reasoning as a trajectory of mindset calls, outputs, and insights, with the next mindset selected from the current problem and accumulated history.At each step, the state is defined by the input problem and preceding reasoning history.
- Mindset definitions: The four mindset modules correspond to spatial imagination, convergent analysis, divergent exploration, and algorithmic computation.Each mindset is instantiated through a corresponding specialized call.
- Architecture: The Meta-Agent orchestrates a Plan-Call-Internalize loop by selecting mindsets, generating call instructions, and revising the remaining plan from newly internalized insights.The loop separates decision-making from concrete task execution.
- Mindset Dispatch: An Input Gate supplies each mindset with a call instruction, relevant reasoning history, and injected images when visual information is required.The filtered tuple enables specialized modules to receive selectively relevant context.
- Mindset Dispatch: The Spatial Mindset uses visual externalization through text-to-image, image-plus-text-to-image, and code-to-image generation modes.These modes are implemented with Nano-Banana-Pro, including image editing or augmentation and execution of returned matplotlib code.
Meta Agent
The Meta-Agent orchestrates specialized mindsets through iterative decisions, filtered context, summarized results, and internalized insights. This process supports adaptive reasoning, including visual grounding, ambiguity resolution, and precise computation.
- Meta-Agent orchestration: The Meta-Agent selects and dispatches subtasks to specialized mindsets, receives summarized results, internalizes insights, and can revise its plan.Its cycle proceeds through cognitive decisions, mindset calls, mindset results, and insights before producing the final answer.
- Specialized mindsets: The Convergent Mindset performs a focused reasoning pass that grounds steps in established facts, identifies missing information, and reaches a clear conclusion.It receives an instruction and filtered relevant context, producing a complete logical derivation.
- Specialized mindsets: The Divergent Mindset explores k ∈[2, 5] distinct solution branches in parallel, with each branch representing a candidate reasoning path and explicit assumptions.Parallel exploration uses separate LLM calls to analyze the generated branches.
- Specialized mindsets: The Algorithmic Mindset uses a code-based generate-execute-repair loop to address precise calculation, with execution errors triggering bounded repair attempts.The procedure initializes Python code and permits Nmax = 2 repair attempts before returning an error state.
- Illustrative execution: In the Fermi example, Spatial visualization extracts a head-to-arm ratio of ≈3.5×, Convergent reasoning maps head size to the Sun’s radius, and Algorithmic reasoning computes the answer.The reported answer is 2,437,190 km, illustrating visual grounding, ambiguity resolution, and computational precision.
- Context Gate: The Context Gate filters relevant inputs and distills verbose outputs so specialized reasoning remains isolated while concise insights re-enter the main chain.The Input Gate extracts relevant history and images, while the Output Gate produces a concise summary for Meta-Agent internalization.
3 Experiments
Experiments evaluate CoM across six benchmarks, two base models, multiple baseline families, ablations, efficiency, and mindset-invocation patterns. CoM achieves the highest overall accuracy while balancing computational cost, with benefits varying by task and mindset.
- Main Results: CoM achieves 63.28% overall accuracy on Qwen3-VL-32B-Instruct and 52.41% on Gemini-2.0-Flash, exceeding MRP by 4.96% and 4.72%, respectively.These are the highest overall accuracies across the two evaluated base models.
- Main Results: CoM exceeds the second-best method on AIME25 by 10.00% with Qwen3-VL-32B-Instruct and outperforms MRP on MAZE by 6.00% and 7.50% across the two models.The largest gains occur on mathematical and spatial tasks requiring flexible mindset adaptation.
- Ablation Study: The Context Gate removal causes the largest overall ablation drop of 8.24%, while removing Divergent lowers AIME25 accuracy by 16.66%.Spatial removal most affects MathVision and MAZE, whereas Algorithmic removal lowers LiveCodeBench All by 2.19%.
- Ablation Study: On Fermi estimation, removing Divergent, Convergent, or Context Gate slightly improves results, while Algorithmic and Spatial remain essential.The reported changes are +1.18%, +1.81%, and +1.37%, respectively, suggesting task-dependent mindset subsetting.
- Efficiency Analysis: CoM reaches 63.28% accuracy at 28.4k tokens, while Tree of Thoughts uses 142.5k tokens and Meta-Reasoner uses 49.7k tokens with 37.25% accuracy.The comparison places CoM on the reported accuracy-efficiency Pareto frontier.
- Efficiency Analysis: Removing the Context Gate increases token consumption by 87%, whereas removing Divergent reduces tokens by 26% with moderate accuracy loss.The full CoM is reported to provide the best overall accuracy-efficiency trade-off.
- Mindset Invocation Patterns: Overall, 59.7% of problems invoke at least two distinct mindsets, with task-specific patterns including Algorithmic-Convergent use for Fermi and Spatial use for multimodal tasks.Algorithmic is invoked on 91.2% of Fermi problems, while Spatial is invoked on 80.6% of MathVision and 100% of MAZE problems.
4 Conclusion
Chain of Mindset enables step-level adaptive orchestration of four heterogeneous mindsets through a Meta-Agent and bidirectional Context Gate. Across six benchmarks, it achieves state-of-the-art accuracy while maintaining efficiency and cross-model generalization.
- Framework: CoM dynamically selects Divergent, Convergent, Algorithmic, and Spatial mindsets according to the evolving problem state.The Meta-Agent orchestrates mindset transitions, while the bidirectional Context Gate filters information across modules.
- Results: 4.96% and 4.72% overall-accuracy gains over the strongest baselines are reported for Qwen3-VL-32B-Instruct and Gemini-2.0-Flash, respectively.These results come from experiments across six challenging benchmarks.
- Results: CoM maintains computational efficiency and generalizes across both open-source and closed-source models without additional training.The architecture is training-free and designed to support adaptive reasoning without parameter scaling.
- Implications: The framework positions dynamic cognitive switching as a promising paradigm for more adaptable reasoning systems.Its modular architecture also supports experimentation with new mindsets and policies.
- Implications: CoM makes reasoning trajectories more inspectable and supports targeted safety interventions through explicit traces and structured switching.The paper presents these properties as opportunities for auditability and guidance of cognitive trajectories.
A Related Work
Related work studies cognitive behaviors, prompt-based reasoning structures, and meta-reasoning for strategy selection. Existing meta-reasoning methods differ in whether they select strategies at the task level or adapt actions during reasoning.
- Cognitive behaviors: Research identifies skill recognition and behaviors such as verification, backtracking, subgoal setting, and backward chaining as relevant to LLM self-improvement.A related taxonomy further organizes 28 cognitive elements.
- Prompt-based reasoning: Prompt-based methods either generate explicit intermediate steps or expand reasoning structures into richer topologies.Chain-of-Thought and Decomposed Prompting exemplify intermediate-step generation, while other approaches explore expanded reasoning structures.
- Meta-reasoning: Meta-reasoning methods select strategies either at task onset or dynamically during the reasoning process.Task-level methods maintain an initial strategy, whereas step-level methods adapt control actions as reasoning unfolds.
B Future Directions
The paper identifies several directions for extending CoM beyond its current four-mindset, shared-model, tool-free, and untrained-dispatch configuration.
- Mindset scope: Future work could add cognitive primitives through CoM’s plug-and-play architecture.The current framework instantiates four mindsets representing established cognitive primitives.
- Model allocation: A heterogeneous expert allocation could assign specialized models to different mindsets instead of sharing one base model.The current implementation uses the same base model for all mindsets.
- Tooling: Tailored tools could be added for individual mindsets, such as symbolic solvers for Algorithmic and search tools for Convergent reasoning.These extensions are proposed as ways to enhance mindset-specific capabilities.
- Dispatch policy: Training the Meta-Agent’s dispatch policy is proposed as another route to improve performance.The current dispatch policy is not optimized through training.
C Baseline Implementation Details
The implementation compares CoM with standardized reasoning baselines and specifies its modular prompts, mindset roles, context filtering, and execution protocols.
- Baseline comparison: All baseline methods use identical inference settings for fair comparison, including temperature and maximum tokens.The evaluated baselines include Direct I/O, Zero-shot CoT, Tree of Thoughts, Chain of Code, ReAct, MRP, and Meta-Reasoner.
- Baseline comparison: Tree of Thoughts generates three candidate thoughts per step, evaluates them, and expands the best branch to a maximum depth of 10 steps.The implementation uses BFS or beam search.
- Baseline comparison: ReAct uses a Thought-Action-Observation loop with Python-Sandbox and ImageGeneration tools, allowing up to 10 interaction turns.The Python-Sandbox timeout is 30 seconds.
- CoM architecture: CoM consists of a Main Agent, four specialized Mindset Experts, and Context Gates for information filtering.The Meta-Agent plans mindset calls, executes them in order, and can revise unexecuted plans.
- Context filtering: The Input Gate extracts task-relevant history and selects relevant images, while the Output Gate returns the generated image path and notes to the Main Agent.The Spatial Mindset routes processed context to an image-generation model; the gates preserve key numbers, data, coordinates, and prior results while omitting the original question.
E Case Studies
The case studies demonstrate CoM across mathematical reasoning with dynamic re-planning and multimodal geometry with visual input.
- E Case Studies: CoM is illustrated on mathematical reasoning with dynamic re-planning and multimodal geometry with visual input.Special tokens such as <cognitive_decision> and <call_convergent> expose the meta-cognitive control flow; a Fermi estimation example additionally demonstrates Spatial Mindset image generation.
E.1 Case Study 1: Mathematical Reasoning with Dynamic Re-planning (AIME) AIME 2025: Base Divisibility
In the base-divisibility case, CoM revises an initially algorithmic plan after recognizing that algebraic simplification can make enumeration efficient, yielding the correct answer 70.
- AIME 2025: Base Divisibility: The convergent step rewrites 9b + 7 as 9(b + 7) − 56, reducing the condition to (b + 7) | 56.This reduction enables computation by filtering divisors of 56 and converting each qualifying divisor d into b = d − 7.
- AIME 2025: Base Divisibility: 70 is the final answer after CoM identifies b = 21 and b = 49 and sums them.The algorithmic module reports 21 + 49 = 70, followed by the final-answer insight.
- AIME 2025: Base Divisibility: CoM dynamically revises its plan from Algorithmic testing to Convergent simplification followed by Algorithmic computation.The system first formulates the divisibility condition, then recognizes that algebraic reduction should precede enumeration.
- AIME 2025: Base Divisibility: The case exemplifies state-dependent cognitive switching rather than commitment to a fixed reasoning strategy.Intermediate results trigger a more efficient path within the same solution process.
E.2 Case Study 2: Multimodal Geometry with Visual Input (MathVision) MathVision #1167: Zig-Zag Angles
In the zig-zag geometry case, an initially incorrect calculation produces 44, prompting CoM to explore alternative principles and obtain the correct option A, 11.
- MathVision #1167: Zig-Zag Angles: 44 is rejected because the initial convergent calculation produces an answer absent from the options.The insight mechanism detects the inconsistency and triggers re-planning.
- MathVision #1167: Zig-Zag Angles: CoM switches from Convergent to Divergent and then Algorithmic reasoning after the initial approach fails.The Divergent step generates several geometric principles, identifying the zig-zag theorem as promising before calculation.
- MathVision #1167: Zig-Zag Angles: 11 matches option A after the Algorithmic mindset applies the zig-zag theorem.The calculation uses the equality between left-pointing and right-pointing angles.
- MathVision #1167: Zig-Zag Angles: The case demonstrates error recovery through mindset switching to escape a reasoning dead-end.The example connects inconsistency detection, alternative-principle generation, and final calculation.