This edition covers three preprints listed on Hugging Face Daily Papers on October 23, 2026, the third Research Radar drawn from that list, because none was published on October 24 or 25. None is peer-reviewed, and the figures are the authors' own.
OpenDS-World: deep search on the live web (arXiv 2610.30181)
Most deep-search benchmarks freeze a corpus. This one, submitted October 19, asks agents to "autonomously explore the real Web" and combine evidence from several independent sources to reach "a unique factual answer". OpenDS-World has 200 tasks covering different modalities, heterogeneous file formats and interactive access, and every answer was independently verified by humans. Under a single tool environment, the best of 12 frontier models reached 39.0% accuracy. Humans, across 1,000 independent attempts, reached 21.1%. The authors report "persistent challenges in tool coordination and long-horizon search management" and large differences in how models end searches and recover from failures.
Two cautions. A live web drifts, so scores may not be reproducible over time. And the human baseline measures people under the benchmark's own conditions, not expert researchers with unlimited time.
SCOPE: policies compiled into predicate programs (arXiv 2610.31062)
Arnon Mazza and Elad Levi start from the point that agent policies now span "multi-turn transcripts and tool-call trajectories". They argue that fine-tuned guardrail models degrade on such compound policies, while LLM judges are expensive and "tend to produce unstable predictions". SCOPE compiles a policy into "an executable program of neural predicates executed by small models, deterministic extractors, and code that composes their outputs". It then optimises the program against adversarially generated counterexamples that target where it deviates from the policy. The authors report the highest average accuracy against prompt-optimisation, decomposition and workflow-generation baselines. They also report that programs optimised with the weakest executor transfer to other models unchanged and reach accuracy comparable to frontier reasoning models on inexpensive ones. Each verdict can be traced to the predicates that produced it.
RouteWeaver: choosing the execution mode (arXiv 2610.19705)
Routers usually fix the execution style in advance. RouteWeaver, first submitted October 14 and now in its third version, trains one policy to decide whether a query needs a single call, several rounds or an agentic loop, and which models to use within that mode. Its COMET-GRPO method splits credit between the mode decision and the execution inside it. The authors report an average accuracy of 61.4% across mathematics, coding, knowledge, logic and factual recall, an 11.03% relative gain over the strongest baseline. The telling ablation is that without progressive exploration, "the routing policy collapses entirely to single-round execution".
What remains uncertain. The abstracts give no dollar cost per task next to the accuracy figures; OpenDS-World analyses search cost but does not summarise it there. The SCOPE benchmark is described only as "real-world policy tasks".
Analysis: These papers are about deciding how much machinery a task needs. OpenDS-World shows the gap is still large when the web is real. SCOPE suggests a policy check can be cheap, stable and auditable when it is decomposed. RouteWeaver shows that a learned router left alone drifts to the cheapest mode. For production stacks, the useful pattern is to give small models narrow, checkable decisions and to measure whether the agentic path is chosen only when it pays.
Sources and evidence
Three October 23 Hugging Face listings: OpenDS-World's 200 live-web deep-search tasks top out at 39.0% for the best of 12 frontier models against 21.1% for humans; SCOPE compiles agent policies into programs of small-model predicates that transfer across models and give traceable verdicts; and RouteWeaver learns when to use single-call, multi-round or agentic execution, reaching 61.4% average accuracy but collapsing to single-round without progressive exploration.