More context is not better evidence: retrieval now wins by pruning under a budget

Choose under a budget

Multi-hop questions punish the naive retrieval strategy. Answering them means gathering evidence across documents, and every extra chunk costs tokens, latency and attention. In a preprint, a framework called SpectRAG is reported to improve evidence recovery and generation quality in multi-hop retrieval-augmented generation compared with dense, hybrid, HyDE, LightRAG and NaiveRAG baselines. The comparison set matters: the gain is reported against several established approaches, not against a straw man. m:…6815

The reported mechanism treats retrieval as a graph problem rather than a ranking problem. In a preprint, the framework is reported to utilise a weighted chunk graph encoding semantic similarity, document adjacency, section hierarchy, shared entities and citation links. In plain terms, a chunk is judged not only by how similar it looks to the question, but by where it sits, what it connects to, and what cites it. And in a preprint, graph-spectral structure is reported to be particularly effective for inference-oriented and cross-document retrieval tasks. m:…6815

Prune by quality, not by count

A second preprint makes the pruning step explicit. In a preprint, DuQUAD is reported to improve accuracy by 3.0% on average over the strongest competing baseline, and to improve gold evidence recall by 9.4% compared with existing baselines. Recall here means recovering the evidence that actually supports the answer, so the gain is about finding the right material, not just producing a plausible answer from whatever was retrieved. m:…1280

In a preprint, DuQUAD is also reported to reduce computational cost compared with existing multi-agent frameworks through quality-aware evidence pruning. The pairing of higher recall with lower cost is the point: pruning is not a compromise made to save money, but a quality step in its own right, because removing weak evidence leaves the model less to be misled by. m:…1280

Embeddings need context, not just labels

The third preprint attacks a labelling bottleneck. Training an embedding model to retrieve the right chunk usually needs labelled examples of the right chunk, which are expensive to produce. In a preprint, contextual embedding models trained via distillation are reported to eliminate the need for gold-chunk labels. Distillation here means a larger or more capable setup teaches the embedding model, so the training signal comes from the teacher rather than from hand-labelled chunks. m:…1614

In a preprint, the 9B model is reported to achieve state-of-the-art performance on Context-bench, Q2D-Chunk and ConTEB, and visual retrieval performance is reported to improve with granular image chunking. The second finding generalises the point beyond text: how finely you cut the material up changes what retrieval can find, whether the material is a document or an image. m:…1614

Stored is not retrievable

The fourth preprint delivers the cautionary half of the story. Storing knowledge inside a model's weights is not the same as being able to get it back out on demand. In a preprint, entity-specific adapter modules are reported to encode context-free factual knowledge that generalises to unseen questions, which shows the storage side can work. m:…3837

But in a preprint, weight-space geometry is reported to correlate with subgraph semantics, with a correlation of positive 0.329, while not correlating with functional retrievability. And in a preprint, parametric knowledge graph memory is reported to be feasible for storing knowledge but to currently lack a retrieval mechanism that functions beyond semantic similarity. The gap between knowing and retrieving is the story of all four preprints: retrieval systems should be judged on what they can actually bring back, under a budget, not on what they contain. m:…3837