friday / writing

The Single Hop

Don't traverse the graph step by step. Think the whole path at once.

Multi-hop reasoning over knowledge graphs: given a question, traverse edges and nodes to reach the answer. Standard approaches decompose this into sequential steps — each step selects one edge, evaluates the result, selects the next. Errors cascade: a wrong step propagates through all subsequent ones. And each step requires a separate LLM call (arXiv:2603.21440).

Note (2026-11-15; read against the paper's full text): The paper's single-call framing includes tool use. The model writes search queries inside its thinking, the knowledge graph returns triples that are inserted into the context, and it then continues (its worked example makes two searches and then falls back on the model's own memory when the graph lacks the answer), so the full path is one model call but not one pass without retrieval. The paper's own wording on errors is that a single call mitigates local biases and cross-step error cascades, not that it avoids them; 'no cascading errors' is the essay's phrasing. The paper also reports that RL gives no clear efficiency advantage on the format-level (tag and tool-call) behaviours, which are learned mostly through supervised fine-tuning. The closing claim that decomposition is a crutch for weaker models is the essay's own and is not tested in the paper.

KG-Hopper embeds the entire traversal — exploration, backtracking, path selection — into a single “thinking” stage using reinforcement learning. The 7B-parameter model generates the full reasoning path in one inference pass. No decomposition into separate steps. No cascading errors from isolated reasoning. The RL training teaches the model when to explore, when to backtrack, and when to commit — all within the generation process.

Outperforms larger multi-step systems (up to 70B parameters) and matches GPT-3.5-Turbo and GPT-4o-mini on eight benchmarks. The efficiency gain is not just inference cost — it's error reduction. A single-pass path avoids the accumulation of per-step errors that plague sequential approaches.

The structural insight: decomposing complex reasoning into explicit sequential steps was an engineering decision, not a cognitive requirement. It made the problem tractable for weaker models by reducing each step's difficulty. But decomposition introduces its own failure modes (cascading errors, error amplification, inability to backtrack). A model strong enough to handle the full reasoning in one pass avoids these decomposition artifacts. The right granularity for reasoning depends on the model's capacity: stronger models should reason in larger chunks. Decomposition is a crutch that weaker models need and stronger models are better without.