AGENTIC · 2026-10-20
Agentic — Paper Daily, 2026-10-20
Community search over UGC posts, questions, answers, and comments. Primary metrics: relevance (nDCG) and click-through rate (CTR). Focus: semantic matching, query understanding, short-text retrieval, ranking, and LLM-assisted search.
4.51
A search agent can answer correctly without using retrieved evidence. Grounded Credit Assignment separates credit for finding evidence from credit for answering. On the larger evaluated model, it improves average exact match over Search-R1 by 6.3 percentage points.
Key insight: Reward evidence acquisition separately from answer correctness. Otherwise, a lucky answer from memory credits the entire retrieval trace and can improve accuracy while weakening grounding. The proposed correction uses verifiable sub-outcomes without an extra reward model.
4.30
Munin indexes a local literature corpus and citation graph, then equips an open-weight model with reading, search, and computation tools. On paired literature questions, accuracy rises from 38.7% without the harness to 87.4% with it.
Key insight: A fixed retrieve-then-answer pipeline underperforms the unaided model on every tested backbone. The useful unit is the whole reading harness, including citation tracing and abstention when evidence is missing. The released evaluation suite lets adopters test their own corpus.
4.05
Model confidence tracks unsupported answers only when retrieved evidence constrains the response. Across controlled unsupported edits, perplexity achieves AUROC between 0.81 and 0.91 after answer-length matching.
Key insight: Interpret confidence together with retrieval coverage. Missing passages let the model answer confidently from parametric knowledge, changing what the score measures. The study finds that adding a second passage can restore discrimination lost with top-one retrieval.
4.05
GLOBE stores product-editing experience as explicit rules scoped to listing fields. Feedback updates the handbook, while moderation removes redundancy and deactivates harmful rules. Ranking-improvement success rises by 17.5 percentage points over prompt optimization.
Key insight: An inspectable behavior handbook is easier to revise than a monolithic prompt. Field scope and outcome counts make rules reusable and removable. These are experimental ranker outcomes; shared models in the feedback loop can still introduce preference bias.
4.05
SAFE-Mem labels individual memory claims by kind, authority, and scope, then checks those permissions during retrieval. Attack success is at least 77% lower than the strongest evaluated defense that preserves legitimate memory use.
Key insight: Remembering a statement does not authorize acting on it. Claim-level governance lets retrieval retain useful facts while blocking unauthorized influence. A provenance graph connects source sentences, claims, and entities for traceable decisions.
4.05
BRIA tests candidate tools in isolated branches and lets the task agent adopt only useful execution evidence. On MCP-Atlas, its average claim score exceeds MCP-Zero by 14.37 percentage points.
Key insight: Keep failed tool investigations out of the main reasoning trajectory. Isolated trials preserve alternatives and let routing depend on execution evidence rather than tool descriptions alone. Evaluation uses a shared library of 1,897 tools.
3.91
TreeOPSD branches continuations from shared interaction prefixes, backs terminal rewards up the tree, and distills toward a bounded teacher distribution. It improves average Deep Research accuracy over GRPO by 4.97 percentage points on the reported model.
Key insight: Alternative futures from the same state provide finer credit for earlier actions. A finite teacher target also limits overshooting during repeated updates on old rollouts. The method avoids an added critic, step annotations, or a stronger external teacher.
3.91
MultiAgent-OPD trains a Leader–Searcher–Answerer team with role-specific feedback and token-level supervision from a frozen teacher. Across seven datasets, macro avg@4 exceeds Prompted Critic by 4.06 percentage points.
Key insight: Use different feedback for different roles: team success for the leader, local process supervision for searchers, and correctness for the answerer. Scoring student-generated tokens shortens the credit path from a final answer to the search actions that supported it.
3.85
A study of production agent workloads examines how long KV state occupies GPU memory between model calls. Trace-driven replay with a public KV layout finds that a one-hour idle cap reduces the estimated residency ceiling by 3.6 times.
Key insight: Retention policy depends on the next call’s timing and deployment. The call traces are real; memory figures use a public model layout as a proxy. Representation changes and deployment-aware routing complement eviction when memory is tight.
3.40
AWSAWS provides a deployable reference for iterative retrieval across managed knowledge bases, an MCP gateway, and an instrumented agent runtime. Seven telemetry layers cover retrieval, traces, token usage, and evaluation. The post reports no comparative answer-quality uplift.
Key insight: Provision evaluation alongside the retrieval loop so routing failures, retrieval gaps, and token costs can be traced together. Cross-knowledge-base routing and within-base iteration are separate layers. The small demonstration corpora limit the strength of its empirical evidence.
EMNLP 2026 Workshop