Reference Seeds.Json

[
  {
    "generation": "reference",
    "iter_id": null,
    "name": "Chain-of-Thought",
    "thought": "By encouraging the LLM to think step by step rather than directly outputting an answer, chain-of-thought reasoning enables complex problem-solving through intermediate steps. This practice improves the model's ability to handle tasks that require deeper reasoning and provides insight into its decision-making process.",
    "graph_summary": "(reference pattern, no graph implementation provided — meta-LLM should adapt the pattern to AgentCanvas)",
    "diff_narrative": "Add an instruction that asks the planner LLM to think step by step before emitting an action; expose a 'thinking' field alongside the action so reasoning is visible.",
    "fitness": null
  },
  {
    "generation": "reference",
    "iter_id": null,
    "name": "Self-Consistency with Chain-of-Thought",
    "thought": "While an LLM can arrive at the correct answer, its reasoning may vary. By repeatedly asking the same question with high temperature settings, we can generate different reasoning paths. We then combine multiple answers from these Chain-of-Thought (CoT) agents to produce a more accurate final answer through ensembling.",
    "graph_summary": "(reference pattern, no graph implementation provided — meta-LLM should adapt the pattern to AgentCanvas)",
    "diff_narrative": "Run N planner LLM samples at high temperature per step; majority-vote (or score-aggregate) over the N proposed actions before issuing the env step.",
    "fitness": null
  },
  {
    "generation": "reference",
    "iter_id": null,
    "name": "Self-Refine (Reflexion)",
    "thought": "To enhance its performance, an LLM can iteratively improve its answer based on feedback. By reflecting on its previous attempts and incorporating feedback, the model can refine its reasoning and provide a more accurate solution.",
    "graph_summary": "(reference pattern, no graph implementation provided — meta-LLM should adapt the pattern to AgentCanvas)",
    "diff_narrative": "Add a critic step after the planner: critic reviews the proposed action against task + history, returns feedback; loop until critic approves or max-refine-rounds reached, then commit the action.",
    "fitness": null
  },
  {
    "generation": "reference",
    "iter_id": null,
    "name": "LLM Debate",
    "thought": "By letting different LLMs debate with each other, we can leverage their diverse perspectives to find better solutions for tasks.",
    "graph_summary": "(reference pattern, no graph implementation provided — meta-LLM should adapt the pattern to AgentCanvas)",
    "diff_narrative": "Spawn K role-conditioned planner agents per step (e.g., Navigator, Cartographer, Skeptic), run R debate rounds where each sees the others' reasoning, then a Final-Decision agent picks the action.",
    "fitness": null
  },
  {
    "generation": "reference",
    "iter_id": null,
    "name": "Step-back Abstraction",
    "thought": "Let LLM first think about the principles involved in solving this task which could be helpful. By understanding the underlying principles, the model can better reason through the problem and provide a more accurate solution.",
    "graph_summary": "(reference pattern, no graph implementation provided — meta-LLM should adapt the pattern to AgentCanvas)",
    "diff_narrative": "Add a 'principle' node that distills high-level navigation constraints (e.g., 'instruction mentions a door — look for door candidates first') from the instruction, then feed those principles into the planner alongside the per-step observation.",
    "fitness": null
  },
  {
    "generation": "reference",
    "iter_id": null,
    "name": "Quality-Diversity",
    "thought": "Similar to Quality-Diversity methods, let LLM generate multiple diverse interesting solutions could help. By encouraging the model to explore different reasoning paths, we can increase the chances of finding the best solution.",
    "graph_summary": "(reference pattern, no graph implementation provided — meta-LLM should adapt the pattern to AgentCanvas)",
    "diff_narrative": "Generate N diverse candidate actions per step under a 'try something different' prompt; pass all candidates into a Final-Decision agent that picks the best given task + history.",
    "fitness": null
  },
  {
    "generation": "reference",
    "iter_id": null,
    "name": "Dynamic Assignment of Roles",
    "thought": "Similar to Auto-GPT and expert prompting, we can use dynamic control flow in the design to let the agent decide what expert we should use.",
    "graph_summary": "(reference pattern, no graph implementation provided — meta-LLM should adapt the pattern to AgentCanvas)",
    "diff_narrative": "Add a routing agent that classifies the current sub-problem (e.g., 'is this a turn-and-look situation, a forward-progress situation, or a re-plan situation?') and dispatches to a role-specialised planner per sub-problem.",
    "fitness": null
  }
]