Should research agents verify answers before searching longer?

AREX starts from the observation that deep research "requires agents to find answers that jointly satisfy multiple constraints," and that discovering such answers "is costly, whereas verifying a candidate can often be decomposed into tractable constraint-wise checks." From this discovery–verification asymmetry the abstract draws its design claim: a research agent "should do more than simply search longer" and should "recursively improve its current answer by verifying intermediate results and using the partially verified state to guide subsequent refinement." The introduction names the target. Many existing systems extend a single search trajectory with more reasoning, tool calls or context, and the challenge is "not merely to search longer, but to identify which constraints remain unresolved and use that diagnosis to formulate a more targeted next research problem."

The mechanism is two nested loops. An inner research loop gathers evidence and constructs a provisional answer. An outer self-improvement loop audits that answer constraint-wise, identifies unresolved claims, and launches targeted follow-up research. The conclusion describes the result as turning provisional answers into "partially verified research states" that preserve supported evidence while isolating unresolved claims. The paper says a single longer trajectory does not guarantee systematic progress because "early errors may persist, exhausted directions may be revisited, and partially valid candidates may be accepted prematurely." To keep the loop running over long horizons, AREX learns a context-update tool that compresses interaction history into a compact improvement state holding verified evidence and unresolved constraints, "without relying on an external model." Training uses verified synthetic tasks and high-quality trajectories, through agentic mid-training and long-horizon reinforcement learning.

This qualifies Does search budget scale like reasoning tokens for answer quality?. That note treats search budget as the variable that sets answer quality. AREX agrees that inference-time computation "can broaden exploration" but argues that what the extra budget is spent on matters, and that spending it on constraint-directed follow-up beats spending it on a longer trajectory. It also differs in scale from Can recursive self-improvement speed up the research process itself?. There the recursion is over the research process itself. Here it is over the current answer to one query, though the excerpt uses the same label, "Recursively Self-Improving." The draft-then-revise shape resembles Can iterative revision cycles match how humans actually write?. The difference the excerpt supports is what steers the revision: unresolved constraints, not a noisy draft. My reading, not the paper's: the asymmetry is close kin to the generation-verification gap that the self-improvement literature leans on.

The excerpt does not establish how much the audit loop contributes. It reports "strong performance across deep search, wide search, agentic reasoning, and tool-use benchmarks" but names no benchmarks, scores, baselines or model sizes, and it gives no ablation separating the outer loop from the context-update tool or the training recipe. It does not say how the constraint-wise audit is carried out or how reliable it is. The premise is also hedged: verification "can often" be decomposed, so tasks whose constraints do not decompose cleanly are outside what it argues. The one training finding is called "preliminary": it favors reinforcing decision-critical steps over treating all actions uniformly, and estimating step utility is left to future work. What follows at this strength is a design argument backed by a stated result: spend inference effort on diagnosing unresolved constraints, not only on extending search.