Chain of Mindset (CoM): Adaptive Reasoning

• Chain of Mindset (CoM) is a training-free, agentic reasoning framework that dynamically orchestrates multiple cognitive modes to tackle multi-step problems.
• It employs a sequential decision process with specialized modules—Spatial, Convergent, Divergent, and Algorithmic—to enhance accuracy and efficiency.
• Empirical results show CoM improves performance by up to 4.96% over baselines, reducing token usage while mitigating context pollution through bidirectional gating.

Chain of Mindset (CoM) is a training-free, agentic reasoning framework that orchestrates step-level adaptation among heterogeneous cognitive modes, termed "mindsets," to improve multi-step problem solving with LLMs. Drawing inspiration from human experts who fluidly deploy varying cognitive strategies—such as spatial visualization, focused analysis, creative ideation, and precise algorithm execution—CoM decomposes the reasoning process into a sequential decision problem, dynamically selecting the optimal mindset at each step. The architecture leverages a Meta-Agent and bidirectional Context Gate, supporting state-of-the-art performance across mathematical, code, scientific, and multimodal reasoning tasks while maintaining a favorable accuracy–efficiency trade-off (Jiang et al., 10 Feb 2026).

1. Motivation and Problem Formulation

Human problem-solving is characterized by continual shifts between cognitive modes, dictated by the evolving structure and requirements of each distinct subproblem. In contrast, prevailing LLM reasoning approaches typically employ a static, monolithic mindset—for example, unvarying chain-of-thought (CoT), fixed programmatic pipelines, or coarse episode-level meta-reasoning. This inflexibility constrains multi-step reasoning and limits attainable robustness and accuracy.

CoM structures complex problem solving as a sequential decision-making process. At step $t$, the agent's state is

$s_t = (q,\, \mathcal{H}_{<t}),$

where $q$ is the query and $$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$ records the history of mindset calls, outputs, and distilled insights. The agent selects a mindset $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$ and invokes its subroutine $c_t$ to yield output and insight $(o_t, i_t) = c_t(q, \mathcal{H}_{<t})$. The formulation emphasizes three challenges: (1) deciding when to switch mindsets, (2) choosing the best mindset for a subproblem, and (3) avoiding detrimental cross-mode interference.

2. Heterogeneous Mindset Modules

CoM implements four distinct modules, each delineated by specific prompts, reasoning styles, and I/O modalities:

Mindset	Core Strategy	Example Task
Spatial	Visualization, diagram edits	Fermi estimation anatomy diagram
Convergent	Deep, rigorous deduction	Symbolic simplification
Divergent	Parallel branch exploration	Multiple proof strategies
Algorithmic	Code execution and repair	Programmatic series summation

• Spatial ($m_{\text{spat}}$): Converts text or code into images (e.g., via Nano-Banana-Pro or matplotlib), grounding problem relations visually. Redraws and reference updates facilitate spatial reasoning tasks.
• Convergent ($m_{\text{conv}}$): Produces focused, logically complete analysis for subproblems, explicitly stating assumptions and missing information before reaching conclusions.
• Divergent ($m_{\text{div}}$): Generates $s_t = (q,\, \mathcal{H}_{<t}),$0 solution branches, each explored via independent LLM calls to break impasses or tackle open-ended tasks.
• Algorithmic ($s_t = (q,\, \mathcal{H}_{<t}),$1): Implements a code-centric generate–execute–repair loop with bounded repair attempts ($s_t = (q,\, \mathcal{H}_{<t}),$2), enabling computational verification and complex calculations.

3. Meta-Agent Orchestration

The Meta-Agent $s_t = (q,\, \mathcal{H}_{<t}),$3 governs adaptive mindset selection through an iterative process:

1. Observe $s_t = (q,\, \mathcal{H}_{<t}),$4.
2. Score $s_t = (q,\, \mathcal{H}_{<t}),$5 for each $s_t = (q,\, \mathcal{H}_{<t}),$6.
3. Formulate selection probabilities:

$s_t = (q,\, \mathcal{H}_{<t}),$7

1. Select $s_t = (q,\, \mathcal{H}_{<t}),$8 (or sample stochastically).
2. Invoke $s_t = (q,\, \mathcal{H}_{<t}),$9 and summarize $q$0 to obtain $q$1.
3. Augment $q$2 with $q$3.
4. Halt upon selection of $q$4 or terminal condition.

This policy is realized via system prompts that elicit mindset discrimination from the underlying LLM. The CoM framework remains training-free, but the utility $q$5 could be learned as a ranking head over embeddings.

Pseudocode: $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$8

4. Bidirectional Context Gate Mechanisms

Frequent mindset switching risks context pollution and unnecessary verbosity. CoM mitigates this through bidirectional gates:

• Input Gate ($q$6): Selects a minimal relevant history subset $q$7 and reference images for injection, conditioned on the current call.
• Output Gate ($q$8): Distills the possibly verbose output $q$9 into a concise insight $$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$0.

Mathematically, these are implemented as gated attention masks: $$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$1

$$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$2

where $$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$3 is the sigmoid nonlinearity and $$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$4 denotes elementwise masking. Practically, $$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$5 and $$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$6 are implemented by lightweight LLM prompts mimicking this filtering.

5. Evaluation Benchmarks and Empirical Results

CoM was evaluated across six challenging datasets:

• Mathematical: AIME 2025, Real-Fermi
• Code Generation: LiveCodeBench
• Scientific QA: GPQA-Diamond
• Multimodal: MathVision-Mini, MAZE

Two base models were used: Qwen3-VL-32B-Instruct (open-source) and Gemini-2.0-Flash (closed-source). Comparative baselines included Direct I/O, Zero-shot CoT, Tree of Thoughts, Chain of Code, ReAct, MRP, and Meta-Reasoner.

Headline results (pass@1 accuracy):

Model	CoM	Best Baseline	$$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$7
Qwen3-VL-32B	63.28%	58.32% (MRP)	+4.96
Gemini-2.0-Flash	52.41%	47.69% (MRP)	+4.72

Task-level highlights (Qwen3-VL):

• AIME25: 73.33% (CoM) vs 63.33% (2nd best)
• MathVision: 63.16% vs 58.55%
• MAZE: 85.50% vs 79.00%

Ablations quantified the contribution of each component:

Removed Component	$$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$8 Overall Accuracy
–Context Gate	–8.24%
–Divergent	–5.18%
–Spatial	–5.03%
–Convergent	–3.76%
–Algorithmic	–2.52%

Notably, removing Divergent dropped AIME25 by 16.66%, and removing Spatial impacted MathVision by 9.87% and MAZE by 4.50%.

6. Computational Efficiency and Trade-offs

CoM demonstrates a favorable accuracy–efficiency balance. For Qwen3-VL-32B:

• Direct I/O/CoT: $$\mathcal{H}_{<t} = (c_1, o_1, i_1, \dots, c_{t-1}, o_{t-1}, i_{t-1})$$96K tokens; $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$057% accuracy
• Tree of Thoughts: $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$1142K tokens; $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$247% accuracy
• MRP: $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$349.7K tokens; $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$458.3% accuracy
• CoM: $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$528.4K tokens; $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$663.3% accuracy (Pareto-optimal)

Context Gate removal substantially increased token usage (+87%) and reduced accuracy (–8.24%). Removing Divergent reduced tokens by 26% but at a cost of –5.18% accuracy. This evidences the trade-off between adaptivity and computation, with step-level mindset orchestration achieving strong accuracy at moderate additional cost.

7. Extensions, Limitations, and Outlook

CoM’s architecture is compatible with further extensibility, as suggested in its future directions:

• Adding new mindsets (e.g., analogical or probabilistic) in a plug-and-play manner
• Assigning dedicated expert models per mindset
• Incorporating external tools (e.g., symbolic solvers, search engines)
• End-to-end training of the Meta-Agent’s selection policy $m_t = \pi(s_t) \in \mathcal{M}\cup\{\emptyset\}$7

Identified limitations include cumulative latency from repeated LLM and image generation calls, dependence on prompt engineering for prompt quality, and context window constraints as histories grow, though mitigated by the gating mechanisms.

CoM empirically demonstrates that step-level adaptive mindset switching—across Spatial, Convergent, Divergent, and Algorithmic modes—enables significant gains in multi-domain reasoning accuracy and efficiency without model retraining (Jiang et al., 10 Feb 2026).