Enable language models to dynamically switch between four cognitive modes (spatial, convergent, divergent, algorithmic) during problem-solving. Meta-agent observes state and selects optimal mode per step, improving reasoning across math, coding, and spatial tasks without requiring model training.
Enable language models to dynamically switch between four cognitive modes (spatial, convergent, divergent, algorithmic) during problem-solving. Meta-agent observes state and selects optimal mode per step, improving reasoning across math, coding, and spatial tasks without requiring model training.
Chain of Mindset: Adaptive Multi-Modal Reasoning
Single problem-solving strategies fail across diverse task types. Mathematical derivations need convergent logic; creative ideation needs divergent exploration; spatial tasks need visualization; calculations need algorithmic precision. Chain of Mindset (CoM) enables training-free dynamic mode switching, where the model selects appropriate cognitive approaches per reasoning step rather than committing upfront to one strategy.
A Meta-Agent observes reasoning progress and decides which of four specialized mindsets to engage next: Spatial (visualization), Convergent (focused analysis), Divergent (parallel exploration), Algorithmic (precise calculation). Context Gates prevent information pollution between modes while maintaining efficiency.
Core Concept
Standard approach: apply single reasoning strategy throughout problem. Inefficient for multi-faceted problems.
CoM approach: decompose problem into steps, use Meta-Agent to assign optimal cognitive mode per step. Modes collaborate asynchronously: spatial mode generates diagrams, convergent mode analyzes alternatives, divergent mode explores options, algorithmic mode executes calculations.
Key insight: problem-solving naturally decomposes into stages requiring different cognitive approaches. No training needed—just better prompting structure.
Architecture Overview
Four Cognitive Modes:
Spatial: Visualize abstract concepts through diagrams, spatial relationships
Convergent: Focused logical analysis on specific questions
Divergent: Parallel exploration of multiple solution paths
"""You are a convergent reasoning mode. Focus on rigorous logical analysis.
Identify constraints, dependencies, and hierarchical relationships. Reason systematically from axioms.
Output: Detailed logical derivation."""
'divergent'
"""You are a divergent reasoning mode. Explore multiple solution paths in parallel.
Generate alternative approaches, creative interpretations, and unconventional methods.
Output: List of distinct solution strategies."""
'algorithmic'
"""You are an algorithmic reasoning mode. Execute precise calculations,
formal procedures, and symbolic manipulations. Use mathematical notation and step-by-step computation.
Output: Numerical result or formal proof."""
def
select_cognitive_mode
reasoning_state, meta_agent_prompt=None
"""
Meta-agent selects appropriate cognitive mode based on reasoning state.
Args:
reasoning_state: Current reasoning progress (text)
meta_agent_prompt: Optional custom selection prompt
Returns:
selected_mode: One of ['spatial', 'convergent', 'divergent', 'algorithmic']
"""
if
is
None
"""Given the current reasoning state below, which cognitive mode would be most helpful?
- Spatial: For visualizing patterns, geometric relationships, or spatial structures
- Convergent: For focused logical derivation and constraint satisfaction
- Divergent: For exploring alternative approaches and creative solutions
- Algorithmic: For precise calculations and symbolic manipulation
Current state:
{state}
Which mode? Answer with exactly one word: spatial, convergent, divergent, or algorithmic."""
# Query model to select mode
format
# Parse response
'spatial'
'convergent'
'divergent'
'algorithmic'
for
in
if
in
return
return
'convergent'
# Default fallback
class
ContextGate
"""Filters context between cognitive modes."""
def
__init__
self
self
def
filter_for_mode
self, full_context, target_mode
"""Extract relevant context for specific mode."""
# Simple filtering: remove incompatible information
# Initialize reasoner
model = load_language_model("gpt-4")
reasoner = AdaptiveMindsetReasoner(model)
# Complex problem requiring multiple cognitive modes
problem = """A farmer has a rectangular field divided into 4 equal quadrants.
He wants to plant crops such that adjacent quadrants have different crop types.
He has 3 crop types. How many distinct planting patterns exist if rotations and
reflections are considered the same?"""# Adaptive reasoning with dynamic mode selection
reasoning_trace = reasoner.reason_adaptively(problem)
print(reasoning_trace)
print("\nMode sequence:", reasoner.mode_history)
final_answer = reasoner.extract_final_answer()
print("\nFinal Answer:", final_answer)
Practical Guidance
Parameter
Recommendation
Notes
Max iterations
8-15
More iterations explore thoroughly; cap to avoid repetition.
Mode diversity
Encourage all 4 modes
Don't let meta-agent favor one mode; track usage.
Context filtering strictness
Moderate
Balance focus with information retention.
Integration point
After problem statement
Let model choose modes rather than prescribing strategy.
Situations where single strategy hits local optimum
Benchmarks across diverse domains (mixed reasoning requirements)
Training-free inference (no retraining of base model)
When NOT to Use
Simple single-type problems (overhead not justified)
Models that struggle with complex instruction following
Domains with single dominant reasoning mode
Common Pitfalls
Meta-agent always selects same mode; use diversification prompts
Context gates too aggressive (removes useful information); test filtering
Not monitoring mode history; divergent mode should alternate with convergent
Context explosion; periodically summarize reasoning to avoid token bloat
Reference
See https://arxiv.org/abs/2602.10063 for full prompt templates, detailed mode descriptions, meta-agent implementation, and benchmarks across mathematics (AIME, MATH), coding (HumanEval), and spatial reasoning (Rotate3D).
