selection-inference
**Selection-Inference** is the **modular reasoning framework that decomposes multi-step reasoning into alternating phases of evidence selection (identifying relevant facts from context) and logical inference (deriving conclusions from selected facts) — enabling interpretable, verifiable, and more accurate multi-hop reasoning** — the structured approach that addresses the fundamental weakness of end-to-end reasoning by making each step's evidence and logic explicit and independently auditable.
**What Is Selection-Inference?**
- **Definition**: A two-module reasoning framework where a Selection module identifies the most relevant facts or premises from the available context, and an Inference module derives logical conclusions from exactly those selected facts — iterating these steps for multi-hop reasoning chains.
- **Separation of Concerns**: Rather than asking a single model call to simultaneously find relevant information and reason over it, Selection-Inference divides these cognitively distinct tasks into specialized steps.
- **Iterative Application**: For multi-hop reasoning, the framework alternates: Select → Infer → Select (with new derived fact added to context) → Infer → ... until the answer is reached.
- **Explicit Evidence Chain**: Each inference step produces a derived fact with explicit provenance — the set of selected facts used as premises — creating a verifiable reasoning trace.
**Why Selection-Inference Matters**
- **Interpretability**: Every reasoning step shows exactly which facts were selected and what conclusion was drawn — human reviewers can verify each step independently.
- **Error Isolation**: When reasoning fails, the framework makes it clear whether the failure was in selection (wrong facts retrieved) or inference (wrong conclusion from correct facts) — enabling targeted improvement.
- **Compositional Reasoning**: Complex questions requiring synthesis of 3–5 facts across a document are handled through iterative selection and inference — each step is simple even when the overall reasoning is complex.
- **Reduces Hallucination**: By grounding each inference in explicitly selected evidence, the model is less likely to fabricate facts — the selected premises constrain the inference space.
- **Modular Improvement**: Selection and inference modules can be independently improved — better retrievers improve selection, better reasoners improve inference, without coupling the two.
**Selection-Inference Architecture**
**Selection Module**:
- Input: context (document, passage, accumulated facts) + current question or sub-goal.
- Process: identify the 1–3 most relevant facts from context that bear on the current reasoning step.
- Output: selected fact set with relevance justification.
- Implementation: can be a separate prompt, fine-tuned retriever, or attention-based selector.
**Inference Module**:
- Input: selected facts + reasoning goal.
- Process: derive a logical conclusion or intermediate fact from the selected evidence.
- Output: derived conclusion with reasoning trace.
- Implementation: separate prompt instructed to reason only from provided premises.
**Iteration Controller**:
- Determines when reasoning is complete (answer derived) vs. when additional Selection-Inference cycles are needed.
- Adds derived facts to the context for subsequent selection steps.
- Terminates when the answer to the original question is produced or maximum steps reached.
**Selection-Inference vs. Alternatives**
| Approach | Evidence Handling | Interpretability | Multi-Hop Capability |
|----------|------------------|-----------------|---------------------|
| **Direct Prompting** | Implicit | Low | Limited (1–2 hops) |
| **Chain-of-Thought** | Mixed with reasoning | Medium | Moderate (2–4 hops) |
| **Selection-Inference** | Explicit per step | High | Strong (3–6+ hops) |
| **ReAct** | Tool-based retrieval | High | Strong (with tools) |
Selection-Inference is **the principled decomposition of reasoning into its fundamental cognitive operations** — demonstrating that separating "what information is relevant" from "what conclusion follows" produces more accurate, more interpretable, and more trustworthy multi-step reasoning than asking models to perform both tasks simultaneously.