AGENTIC · 2026-10-19
Agentic — Paper Daily, 2026-10-19
Community search over short user-generated posts, questions, answers and comments. Primary metrics: relevance (nDCG) and CTR. Focus on semantic matching, query understanding, retrieval, ranking and LLM-assisted search.
4.25
LEAD STORY ·
Huawei /
Leiden UniversityHuaweiLeiden University SearchJev directly scores legal search decisions and delegates uncertain cases to a reasoning model. The dual-system agents achieve a 3.7–4.7× speedup in active search time on BrowseComp-Plus.
Key insight: Separate short relevance and sufficiency decisions from planning and answer generation. Calibrated confidence controls delegation; track handoff frequency and end-to-end tail latency alongside decision accuracy.
4.15
University of EdinburghHuaweiThe paper couples intermediate retrieval supervision with action-local credit assignment. Under matched Qwen3-4B training conditions, evidence-coverage local credit raises average answer F1 by 3.09 points over outcome-only training.
Key insight: Design the retrieval signal and its credit location together. Evidence feedback is more useful when attached to the search action that produced it than when spread indiscriminately across a trajectory.
3.95
Apple
CMUSmall search agents can choose among reusable reasoning candidates instead of generating a fresh chain before every action. Shared-context likelihood scoring reduces total per-question model inference latency by 28–54% across the reported multimodal search tests.
Key insight: Turn recurring reasoning into scored options. A compact candidate library makes shared-cache scoring possible, but its coverage becomes a design constraint; retain an escape route for unfamiliar tasks.
3.60
MAST evaluates evidence retrieval and answering across 21 languages against a shared corpus. Its September 24 date refers to a project update opening working-notes submission, not a paper publication.
Key insight: Separate query language from evidence language and version the answer judge. Measure evidence recall and final correctness by language to distinguish retrieval gaps from generation errors.
3.35
StanfordCMU
University of Washington
SambaNovaThe harness stores candidate solutions and measured outcomes while each agent invocation receives a fresh, role-specific context. On kernel optimization, SLA matches the strongest baseline’s final performance with over 84% fewer tokens.
Key insight: Persist experiment evidence rather than entire conversations. Structured candidate records and explicit assignments can prevent duplicate work; the direct evaluations concern software, kernels and algorithms rather than document research.
3.35
Periodically discard agent sessions and candidate history, then restart from the best solution found so far. Across four open-ended optimization tasks, the reported improvement reaches 15% over parallel agents.
Key insight: When search plateaus, preserve the best artifact while resetting accumulated search history. This is a stronger reset than rebuilding conversational context, and its evidence comes from open-ended optimization rather than factual retrieval.
3.35
ParallelParallel breaks production web research into at least nine components and discusses their maintenance costs. It offers a build-versus-integrate framework rather than a controlled performance comparison.
Key insight: Decide whether research quality is the product’s differentiator. Teams focused on domain workflows can buy generic retrieval while retaining evidence acceptance and domain reasoning.
3.15
PProgress
Progress’s Smart Agent combines indexed documents, live business records through MCP, and web search in an iterative retrieval loop. The announcement describes two new access paths but reports no controlled accuracy or latency evaluation.
Key insight: Keep live records at their source and retrieve them when needed. Evaluate freshness, permissions and claim-level provenance when combining business APIs with static documents.
3.00
Thomson ReutersCoCounsel Legal combines research planning, retrieval and cited reports with maintained legal content and verification. The article cites roughly 1,500 attorney-editors supporting the content layer, but provides no controlled accuracy result.
Key insight: Traceable reasoning depends on maintained sources as well as citations. The transferable pattern is checking conclusions against source passages; the proprietary legal corpus and editorial resources are distinct prerequisites.