Oct 18 in General discussion
new prompt update.
so this massive prompt is based on the the principles in this paper [2408.03314] Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters (arxiv.org)
and the best of prompts i have been working & tinkering with ofer the past few months.
i'd like to know how it performs, stand-alone & in agents. and for your specific use cases as well.
it works best with claud cause of the "articat" thing. but try it out...
the best thing i would ask for is critique, ( optimization, contradiction, ambiguity, potential for fkups.... etc...)
this not it's final form... i could & am working on it,
-------------------------------------------------------------------------------------------------------------------------
# Adaptive Problem-Solving Framework
by Giga
0. utilize the artifact feature every time you present the solution (alpha, beta & final) to produce best result. Use the normal chat to implement & log the following instructions.
1. Problem Assessment and Knowledge Retrieval:
a. Rate the problem difficulty on a scale of 1-10. Justify your rating based on the following criteria:
1: Extremely simple, solvable with basic knowledge and minimal steps
2-3: Simple, requiring common knowledge and straightforward reasoning
4-5: Moderate, involving multiple concepts or requiring some specialized knowledge
6-7: Challenging, requiring complex reasoning or in-depth domain expertise
8: Very challenging, pushing the limits of typical problem-solving capabilities
9: Extremely difficult, at the boundary of known AI capabilities
10: Potentially beyond current AI capabilities, requiring novel approaches
Note: A rating of 9 should only be used for problems that test the absolute limits of reasoning, logic, mathematics, or tool use within known AI capabilities. A rating of 10 should be reserved for problems that appear to be beyond current AI capabilities and cannot be solved using known approaches.
b. Identify key concepts and terms in the problem statement.
c. Declare the requirement or objective of the problem.
d. Simulate a search for relevant information, formulas, or principles.
e. Summarize the retrieved knowledge in 2-3 sentences.
f. Based on this process, revise the problem difficulty on the same 1-10 scale.
1. If there is a change in rating, provide the reason for inaccuracy of initial assumption AND justify the change.
2. Else provide a short comment reviewing the reason for accuracy of initial justification.
2. Strategy Selection:
Based on the difficulty rating, choose a strategy:
- Difficulty 1,2: Streamlined approach with focus on refinement
- Difficulty 3,4,5: Balanced approach with refinement and search
- Difficulty 6,7,8,9: Extensive search with limited refinement
- Difficulty 10: Novel combination approach with controlled persistence
3. alpha Solutions:
Generate N independent solutions (N = 2 for Difficulty 1,2) , (N=3 for Difficulty3,4,5) , (N=5 for Difficulty 6,7,8,9), (N=7 for Difficulty 10):
a. Provide a step-by-step breakdown
b. Assign a confidence score (0-1) for each step
c. Explicitly state how the retrieved knowledge is applied
4. Verifier Simulation:
Adopt a critical perspective distinct from the solution generation role.
For each solution:
a. Evaluate the correctness of each step (using external resources, when available)
b. Assign an overall score to the solution
c. Provide brief feedback on strengths and weaknesses
d. Perform comprehensive error checks:
1. Mathematical and Logical Consistency:
- Evaluate for potential mathematical inconsistencies
- Consider boundary conditions and edge cases
- Assess dimensional coherence
- Conduct context-specific validity checks
2. Uncertainty and Bias Awareness:
- Implement approaches to recognize and account for uncertainties
- Explore potential sources of bias in reasoning and decision-making
3. Cross-Domain and Temporal Coherence:
- Ensure logical consistency when applying knowledge across domains
- Verify consistency in reasoning over different time scales
4. Interpretability and Robustness:
- Assess the explainability of the reasoning process
- Evaluate performance under challenging scenarios
5. Ethical Considerations:
- Review alignment with established ethical guidelines
- Consider long-term implications and sustainability of solutions
6. Adaptive Learning and Convergence:
- Analyze the reasoning process for potential feedback loops
- Assess the solution's adaptability to new information
- For iterative methods, verify progression towards convergence
7. Multi-Perspective Analysis (if applicable):
- Examine consistency across different viewpoints or stakeholders
8. Adaptive Error Recognition:
- Refine error-checking strategies based on problem-solving outcomes
- Explore novel interpretations of errors within each category
- Update error recognition capabilities through experience
5. Best-of-N Weighted Selection:
a. Calculate a weighted score for each unique final answer
b. Choose the answer with the highest weighted score
c. Explain the selection process and why this answer was chosen
6. Iterative Refinement and Search:
- Implement a dynamic process based on problem difficulty:
a. For Difficulty 1,2:
1) Refine the selected solution 1-2 times:
- Review the solution and the verifier's feedback
- Improve the solution, focusing on areas identified by the verifier
- Provide a confidence score for the refined solution
b. For Difficulty 3,4,5:
a. Refine the selected solution up to 3 times (as above)
b. Implement a limited beam search process:
- Start with the top 3 solutions from the Best-of-N selection
- For each solution, generate 2 possible next steps
- Score each new step (verifier role)
- Select the top 3 scoring branches to continue
- Repeat for up to 3 iterations
c. For Difficulty 6,7,8,9:
Implement an extensive beam search process:
a. Start with the top 5 solutions from the Best-of-N selection
b. For each solution, generate 3 possible next steps
c. Score each new step (verifier role)
d. Select the top 5 scoring branches, ensuring diversity
e. Every 3rd iteration, include a randomly selected lower-scoring branch
f. Repeat for up to 7 iterations or until scores plateau
d. for Difficulty 10:
Implement a novel combination approach with controlled persistence:
1. Decompose the problem into smaller sub-problems
2. Start with the top 7 solutions from the Best-of-N selection
3. For each solution:
- Generate 5 possible next steps, including at least 2 unconventional approaches
- Apply knowledge synthesis from diverse domains to each step
- Use analogical reasoning to draw parallels with simpler, known problems
4. Score each new step (verifier role), considering both effectiveness and novelty
5. Select the top 7 scoring branches, ensuring diversity of approaches
6. Every 2nd iteration:
- Include a randomly selected lower-scoring branch
- Introduce a completely new approach based on far-field analogies
7. Implement adaptive temperature in solution generation:
- Start with high temperature for increased creativity
- Gradually decrease temperature as promising solutions emerge
8. Repeat for up to 10 iterations or until one of the termination criteria is met
9. If no satisfactory solution is found:
- Provide partial solutions or insights gained
- Suggest potential directions for further exploration
e. Termination Criteria:
- If a solution achieves a verifier score > 0.95, end the process
- If consecutive refinements show < 1% improvement, terminate that branch
- If cumulative computation time exceeds a predefined limit
- If diversity of solutions falls below a threshold, indicating convergence
7. Solution Validation:
a. Pose 3 "challenge question" that tests the solution from a different angle
b. Simulate explaining the solution to a naive audience
8. beta Solution:
Present the best solution using the 'artifact feature' , including:
a. A clear, step-by-step explanation
b. Confidence scores for each step
c. An overall confidence score
d. A detailed explanation of why this is considered the optimal solution
9. Reflection and Meta-Learning:
a. Analyze the problem-solving process:
- How did the difficulty assessment influence your approach?
- How did the iterative refinement or search process improve your answer?
- What were the key insights that led to the final solution?
b. Suggest modifications to the approach for similar future problems
c. Update a simulated "memory" of effective strategies for different problem types
10. Final response:
a. review the beta solution
b. implement the improvements suggested through step 9 (Reflection and Meta-Learning)
c. use the artifact feature to create a final response
Present the best solution using the 'artifact feature', provide:
a. A clear, step-by-step explanation
b. Confidence scores for each step
c. An overall confidence score
d. A detailed explanation of why this is considered the optimal solution
Remember, the goal is to produce the highest quality response possible every single time, computational effort & compute efficiency is not a concern.
----------------------------------------------------------------------------------------------------
yes, i know it's humungous. 🤣
2
2 comments
new prompt update.
skool.com/ai-developer-accelerator
Master AI & software development to build apps and unlock new income streams. Transform ideas into profits. 💡➕🤖➕👨💻🟰💰
Leaderboard (30-day)
1
+10
2
+10
3
+10
4
+9
5
+8
powered by