AGENTIC · 2026-10-23
Agentic — Paper Daily, 2026-10-23
Community search over short user-generated posts, questions, answers and comments. Key metrics are search relevance (nDCG) and CTR. Focus: semantic matching, query understanding, short-text retrieval, ranking models, and LLM applications in search.
4.50
LEAD STORY ·
KAIST /
Korea UniversityKAISTKorea University Korea University and KAIST have Claude Code (Claude Opus 5.5) design retrieval tools for a search agent: it reads each question's execution log and score, jointly edits pre-generated index files, executable search scripts, and usage instructions, and after 10 rounds keeps the best version on the optimization set. With GPT-5.4-nano as the search agent, BrowseComp-Plus accuracy rises from DCI's 71.6% to 78.4%, and mean nDCG@10 on four BRIGHT subsets from 45.7 to 55.0. Prompt-only optimization reaches 75.6% and 52.0, so the rest of the gain comes from the generated index files and search scripts; handing the interface unchanged to Gemini 3.5 Flash-Lite still beats DCI (73.9% vs 60.9%).
4.50
Northeastern University
TsinghuaNortheastern University and Tsinghua find that in retrieve-while-reasoning RAG, later sub-queries keep drifting away from the original question along the direction earlier sub-queries already moved, and the effect grows with step count. MAPS therefore makes two changes: each round only passes distilled knowledge and a compressed reasoning summary to the model, never historical sub-queries; before writing sub-queries it lists several complementary directions, writes one sub-query per direction, retrieves in parallel, and merges. On Qwen3-32B, mean Cover-EM over six QA sets is 56.0 versus 49.2 for ParallelSearch and 43.9 for Search-o1; removing any of the three components hurts.
4.50
PParlant
Parlant used to ask an LLM whether each customer-service rule is relevant every turn—about 69,000 tokens for 36 rules. S&Q instead has an LLM offline write 10 triggering user messages and 10 easily confused counterexamples per rule and trains a logistic regressor; online each turn needs one embedding and 37 µs of CPU. On τ²-bench airline support at 90% recall, precision is 32%–34% versus 31% for a cosine threshold; after mining examples from real dialogues precision rises to 43.5%, and almost all of that lift comes from the counterexamples.
4.40
IBMIBM builds ClariEdge, a ~3,900-sample multi-turn RAG clarification set that specially includes hard negatives that "look ambiguous but the referent is already fixed by context," then LoRA-tunes small Granite models to decide whether to clarify and to list candidates. Granite-4.1-3B LoRA reaches 96.6% judgment accuracy versus 79.4% for prompt-based GPT-4o and 73.1% for GPT-5.6 Sol (which tends to clarify everything). After training without hard negatives, hard-negative-class accuracy falls from 87.3% to 35.4%.
4.36
Huawei
Southeast University
Sun Yat-sen University
Shandong Normal UniversityHuawei Noah and collaborators propose Uni-Mem: one Qwen3-4B-Instruct jointly learns what to write into a memory store and what to fetch at query time, both with online RL. Training alternates write and retrieve optimization, holding the other part fixed inside each contrastive group so reward differences come only from the part being trained; rewards are assigned per "event" (a new dialogue span or a question) to the related memory ops. On LoCoMo the LLM-judge score is 88.80%, 10.93 points above Memory-R2 (OSS), and 80.99% without SFT.
4.20
SSapienza University of Rome
Georgetown University
University of Glasgow
University of Pisa The University of Pisa and collaborators release ASQ, about 1.02M full retrieval traces from open search agents (Search-R1, AutoRefine, gpt-oss-20b) on MS MARCO, HotpotQA, and Researchy Questions, including every sub-query, retrieval result, and reasoning text. A question can trigger up to 186 retrievals while people usually search 2–3 times; agents often revisit earlier queries and fetch the same results again; adding a reranker or deepening retrieval does not reliably shorten the chain, and effects vary by agent and setup.
4.01
HKU and collaborators propose BIRD-RL, RL that trains small models to debug SQL on real databases. The hard part is that SQL mutates database state, so each sampled trajectory would otherwise need its own database container; BIRD-RL registers and reuses DB replicas per trajectory, raising space efficiency 12.8× on SQLite without Kubernetes. Training has four stages: imitate reasoning, imitate multi-turn tool use, reinforce reasoning, then interactive RL on real DBs. 7B and 14B models reach 44.40% and 48.00% success on BIRD-CRITIC-SQLite versus 48.20% for Claude-Opus-4.6 Agent.
4.01
IIQuest Research
Beihang University
Renmin University of China Renmin University and collaborators find that RL inside one fixed agent harness hurts when the model is moved to another harness. Harness Scaling splits the harness into five swappable parts—system prompt, tool interface, observation format, context management, workflow—picks a small orthogonal set of combinations for training, and pairs each task with two nearby combinations. Training Qwen3-4B-Instruct on ClawGym-Task raises GRPO mean score on unseen harness combinations from 29.81 to 39.04; transfer to Claude Code averages 37.22, above SFT (35.08) or RL (33.38) done directly on OpenClaw.
3.45
Microsoft
UCLAMicrosoft Research releases Flint-interactive so agents can generate interactive charts with click, filter, zoom, and cross-plot linking. Interaction is split in two: a compiler turns raw clicks and drags into semantic events such as "which data the user selected" and applies render updates; the agent only writes which event triggers which update (highlight, annotate, pan/zoom) and never low-level rendering. Use it via Flint-MCP in chat or via a programming API in dashboards and tables. The blog gives no reliability or cost numbers.
3.15
NVIDIANVIDIA describes Sentry, an agent security framework that runs on a BlueField-4 DPU and is independent of host software; all agent traffic to models and the public internet must pass through it. DOCA Tower is a transparent proxy that decrypts traffic and approves requests with Open Policy Agent rules; DOCA Argus reads host memory to check that the requesting process and container hashes are on an allowlist. On anomalies, hardware can isolate the agent in milliseconds, and logs live on the DPU where the agent cannot change them. Launch is planned for early next year; no performance numbers are given.
NeurIPS 2026
HKU
NUS