Retrieval shifts from content to reasoning skeletons: PatternBloom lets a 7B agent compound gains as multi-hop depth grows
Synopsis
PatternBloom changes the retrieval unit of agentic retrieval-augmented generation from what content to read to how to reason, using typed graph skeletons distilled from high-scoring Stage-I reinforcement-learning trajectories into an external Graph Pattern Memory, then coupling the policy to that frozen memory with a pattern-coverage reward. Across seven QA benchmarks, the 7B version ranks first on EM and F1 on all seven datasets, leads R1-Searcher on average out-of-distribution EM, and the same frozen pattern memory transfers to a Qwen3-8B backbone.
Interpretation
A retrieval-augmented agent no longer retrieves only passages, knowledge-graph subgraphs or hyperedges; it retrieves a typed directed reasoning skeleton that prescribes how to traverse retrieved evidence rather than which evidence to read. Paragraph, knowledge-graph and structure-aware methods differ in retrieval unit, yet the reasoning route through the evidence stays inside the language model's hidden state, so queries of the same logical type are solved from scratch each time. On seven QA benchmarks, the 7B PatternBloom ranks first on EM and F1 across all seven datasets, leads R1-Searcher on average out-of-distribution EM, and stays first on the five out-of-distribution test sets.
Two reinforcement-learning stages make the evidence graph itself the supervised object, then lift reasoning skeletons from high-reward trajectories into a retrievable external memory. Earlier reinforcement learning rewarded only the final answer string, so a noisy graph with extra triples and a clean graph received the same reward whenever both yielded the matching answer. The Information-Density Reward normalizes a frozen scorer's information gain on the answer per triple and gates malformed trajectories to zero by format; the Pattern-Augmented Reward takes a convex combination of answer reward and trajectory coverage of retrieved skeletons, with coverage divided by skeleton size rather than by the union.
Removing any one component lowers the score, so all three act together. The full configuration reaches an average F1 of 55.9 across HotpotQA, 2Wiki, MuSiQue and Bamboogle. Removing the Pattern-Augmented Reward drops average F1 to 44.1, disabling the Graph Pattern Memory to 46.6, and skipping Information-Density Reward pre-training to 52.9.
Perspective
The demonstrated setting is multi-hop QA over a Wikipedia chunk hypergraph, trained on HotpotQA and 2WikiMultiHopQA and tested on seven benchmarks; the authors state that cross-domain transfer to non-Wikipedia corpora such as biomedical, legal or financial documents lies outside the present scope and name zero-shot cross-domain transfer as the next benchmark.
Experiments use a fixed random seed and report no error bars; the authors also note that the pattern memory is validated only on the Wikipedia hypergraph substrate, with cross-domain transfer untested.