Edge labels

Label types. There are three main types of edge labels: Reasoning / Planning / Evaluation.

• Reasoning edges define the context (premise set) of a reasoning step, i.e., steps that actually include the contents of reasoning. One should be able to judge if a reasoning step is valid (errorless) and coherent (align with the problem) by only looking at the reasoning edges.
• Planning edges define the context of a planning step. It includes plan decomposition (backward chaining), sequences of plans, and the motivation of a plan. The network of planning edges shows the backbone of the entire reasoning trace.
• Evaluation edges show the self-reflection, i.e., LLM’s judgment on its own reasoning trace. This edge can be used to track the meta-reasoning ability of an LLM. Still, the model’s superficial attitude towards a step might not align with its internal certainty (perplexity, token probability entropy, etc.), which must be taken with caution. For instance, if the model evaluates a step to be wrong (evaluate:refute), the step still might have low perplexity/high confidence.

Connected node types. ReasoningFlow restricts node types that can be connected via a certain edge label. This is a deliberate design choice to reduce parser errors and enhance applicability.

• How to read the examples?

reason:premise-conclusion

This edge represents the basic premise-conclusion relationship. To apply this edge, both nodes should be asserted facts/propositions, as listed below.

• From: $\color{e0e0e0}{■}$ Context / $\color{FFD6A5}{■}$ Fact/$\color{FDFFB6}{■}$ Reasoning / $\color{CAFFBF}{■}$ Restatement / $\color{bdfbdf}{■}$ Assumption / $\color{9bf6ff}{■}$ Example / $\color{c3b1e1}{■}$ Conclusion
• To: $\color{FDFFB6}{■}$ Reasoning / $\color{bdfbdf}{■}$ Assumption / $\color{a0c4ff}{■}$ Reflection / $\color{c3b1e1}{■}$ Conclusion

(Deductive) Premises → Conclusion

This might also include commonsense-included deductive reasoning, where some premises are missing.

$\color{FFD6A5}{■}$ Sulferic acid is a strong acid, → $\color{FDFFB6}{■}$so it will fully dissociate, providing H+ ions.

$\color{FFD6A5}{■}$ I recall that the energy splitting due to the Zeeman effect is proportional to the magnetic field. → $\color{FDFFB6}{■}$ So, if you increase the magnetic field, the splitting should increase as well.

(Math) Inputs and equations → Calculation results

$\color{FDFFB6}{■}$r^2 h = 2r^3 → $\color{FDFFB6}{■}$h = 2r

$\color{FDFFB6}{■}$ Corrosion rate = (10 × 8.76 × 10³) / (2 × 96485 × 8960 × 7.1e-6) → $\color{FDFFB6}{■}$ So, corrosion rate ≒ 87600 / 12264 ≒ 7.14 mm/year

(Inductive) Observations → Inductive Hypothesis