-
Agentic Retrieval-Augmented Generation: A Survey — arXiv, Jan 2025 Why it matters: formalizes “agentic RAG” patterns (reflection, planning, tool use, multi-agent) and maps implementation choices you already teach. Great for framing why orchestration beats “just a better model.” (summarizepaper.com)
-
Reasoning↔RAG Synergy (Survey): Toward Deeper RAG-Reasoning Systems — arXiv, Jul 2025 Why it matters: unifies “reasoning-enhanced RAG” and “RAG-enhanced reasoning,” then spotlights agentic interleaving (search ↔ think loops). Solid taxonomy + dataset links you can fold into eval curricula. (summarizepaper.com)
-
LLM-based Agents in Medicine (Survey) — ACL Findings 2025 Why it matters: a rigorous vertical survey (healthcare) with evaluation tables, safety constraints, and workflow patterns (routing, oversight, audit). Use it as a model for domain-specific agent governance sections in your posts. (ACL Anthology)
-
TreeRL / ToTRL & contemporaries (LLM-RL roundups, 2025) — multiple arXiv 2025 entries aggregated Why it matters: captures the 2025 turn toward on-policy tree search and cheap post-training (e.g., critique fine-tuning) to elicit reasoning—directly relevant to agent planning quality vs. cost. Use as supportive citations when you discuss “don’t just swap models; improve the loop.” (lxl-sword.github.io)
-
Multi-Agent Safety & Oversight (2025) — assorted 2025 arXiv preprints in agent lists Why it matters: emerging patterns (tiered oversight, debate vulnerabilities, equilibrium-style coordination). Handy for a short “risks & mitigations” sidebar in your posts about multi-agent orchestration. (lxl-sword.github.io)
-
Versus “Fundamentals of Building Autonomous LLM Agents” (architecture pillars): #1 and #2 extend that base into deployable agent loops—where reasoning and retrieval co-evolve, and tools get scheduled intelligently. They’re closer to the problems you coach (planner quality, retrieval hygiene, evals).
-
Evaluation lens you can quote: While not 2025, OSWorld remains the most cited “computer-use” anchor (human ~72% vs SOTA much lower); pair it as context any time you temper claims about autonomy. It’s acceptable to include as a baseline and note it’s 2024. (arXiv)
- “For a 2025 view on agentic RAG patterns and how reflection/planning/tool-use actually change retrieval quality, see Agentic RAG: A Survey.” (summarizepaper.com)
- “If you want a single map of reasoning↔retrieval co-design in 2025 (including interleaved agent loops), the July ‘RAG-Reasoning’ survey is the best organizing reference.” (summarizepaper.com)
- “In regulated domains, the ACL 2025 survey on LLM agents in medicine shows how agent workflows, routing, and audit interact with safety and evaluation.” (ACL Anthology)
- “Recent 2025 LLM-RL works (TreeRL/ToTRL; critique fine-tuning) show you can improve planning/efficiency without swapping to a bigger model—ideal for agent loops under budget.” (lxl-sword.github.io)
Want me to compile PDFs + BibTeX for these and generate a one-page “What to cite in 2025 for agentic systems” handout?