Retrieval turns agentic: agents must now judge what to fetch, when to stop searching, and what to trust

Choosing context under a budget

In a preprint, SpectRAG addresses a familiar failure in multi-hop retrieval: standard systems fetch passages that look relevant individually but do not connect into an answer. The framework builds a weighted chunk graph that encodes semantic similarity, document adjacency, section hierarchy, shared entities, citation links, and questions about utility generated by a language model. Selection then follows the graph rather than ranking passages in isolation. m:…7650

The authors report that SpectRAG improves evidence recovery and generation quality in multi-hop reasoning tasks compared to dense, hybrid, HyDE, LightRAG, and NaiveRAG baselines. Graph-spectral structure is particularly effective for inference-oriented and cross-document retrieval, according to the paper, which makes sense: when an answer requires chaining facts across documents, the relationships between chunks matter as much as the chunks themselves. For builders managing long-document retrieval, the lesson is to budget tokens by structure, not just by similarity scores. m:…7650

Knowing when to search at all

A second preprint tackles the opposite waste: agents that search indiscriminately, paying latency and cost on every step whether or not outside evidence would help. The proposed fix is metacognitive reliability supervision, which estimates the reliability of the current reasoning state to decide whether external evidence is needed. Current-state reliability scores effectively distinguish whether a model can proceed without search, according to the authors. m:…9703

The supervision signal matters as much as the decision. Reliability-aware process rewards for GRPO training provide finer-grained supervision for search behavior than trajectory-level feedback, the paper reports, because they judge each step rather than only the final outcome. The overall result is an improved accuracy-efficiency trade-off in agentic search. In practical terms, an agent that knows when it is on solid ground searches less, answers faster, and spends less, while still reaching for evidence when its own reasoning is shaky. m:…9703

A harder test for long context

A third preprint asks whether current long-context evaluations are even measuring the right thing. GraphNIAH is a benchmark for long-context reasoning over real-world graphs, built on real citation relationships rather than synthetic needles planted in filler text. Using real-world citation relationships avoids benchmark contamination and synthetic data bias, according to the authors, which makes the test closer to how retrieval actually fails. m:…2804

The results are sobering for frontier models. GPT-5.6, Gemini-3, and Qwen3.5 show performance degradation in multi-hop reasoning as depth increases, and the models often stop searching citation relationships in context before reaching the answer, even when the information is present. That is a stopping problem, not a context-length problem: the answer is in the window, but the model gives up the search early. It connects directly to the other two preprints, since better chunk selection and better search-or-proceed judgment are both responses to exactly this failure. m:…2804

What changes for builders

Read together, the three preprints move retrieval from plumbing to judgment. The questions an agent system must now answer are which chunks form a connected argument, whether the current reasoning state is reliable enough to proceed, and how deep a citation chain to follow before stopping. None of these is solved by a larger context window alone, and GraphNIAH suggests that a larger window can even hide the problem by making early stopping less visible. Builders should therefore test retrieval the way these papers do: on multi-hop tasks with real structure, measuring both accuracy and how much searching actually happened. m:…7650 m:…9703 m:…2804