{"id":"72bbfd2d-aafc-4116-9659-cfab0c83aab8","arxiv_id":"2506.10408","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A review that classifies Reasoning Agentic RAG into predefined (System 1-like) and agentic (System 2-like) workflows, surveying their designs and training strategies.","lead":"This survey groups recent Retrieval-Augmented Generation systems into two families: predefined reasoning pipelines that follow fixed steps, and agentic reasoning agents that autonomously decide when and how to search. It is a practical map for engineers and researchers choosing RAG architectures, not a new method or dataset.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The predefined/agentic split is internally inconsistent: Self-RAG and MCTS-RAG are placed in 'predefined reasoning' yet satisfy the paper's own agentic definition, so the central two-branch taxonomy is not a sound partition.","rationale":"I read the paper as a survey whose central contribution is the two-branch taxonomy: predefined reasoning versus agentic reasoning, mapped onto System 1 and System 2. The load-bearing condition is that these two branches are genuinely distinct and jointly cover the methods surveyed. That condition is not met internally. The paper's own descriptions cross the boundary: Section 3's Self-RAG description attributes to a 'predefined' method the very self-reflective, retrieval-deciding behavior that Section 4 uses to define 'agentic' reasoning, and Table 2 labels some predefined methods as 'Agentic' in the Control Type column. The reader's weakest assumption was about completeness and the possible existence of orthogonal axes; my concern is sharper — the binary partition is inconsistent with the paper's own definitions. This is a correctness issue for the organizing claim, not a stylistic preference about the System 1 / System 2 analogy. I would therefore move from the reader's UNVERDICTED status to CONDITIONAL: the survey can be accepted only if the authors either supply an unambiguous classification criterion and reclassify methods accordingly, or explicitly weaken the central claim from a two-branch taxonomy to a non-exclusive design-space map. A single consistency audit of Table 2 against Section 4's definition would settle the issue.","tokens_in":13276,"tokens_out":3824,"duration_ms":47483,"concrete_test":"Use the Section 4 definition of agentic reasoning as a binary checklist: does the LLM decide whether or when to retrieve during inference? does it self-reflect and modify queries? does it autonomously call external tools? Apply this checklist to every method in Table 2. If Self-RAG, MCTS-RAG, or Adaptive-RAG is classified as agentic, the predefined/agentic partition fails; publish the resulting classification matrix and, if it contains contradictions, revise the taxonomy in Sections 1 and 3 or explicitly define a non-binary organizational scheme.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 4 defines agentic reasoning as the LLM acting as an autonomous agent that 'identifies knowledge gaps, determines when and what to retrieve, and interacts with external tools.' Section 3 classifies Self-RAG as loop-based predefined reasoning, but its own description says the model 'can decide mid-generation to fetch additional information or to critique its own draft answer' — the same criterion. MCTS-RAG is likewise described as dynamically integrating retrieval and reasoning through an iterative decision-making process, yet is filed under tree-based predefined reasoning. Table 2 makes the contradiction explicit: Self-RAG, MCTS-RAG, and R2R carry control type 'Agentic' inside the Predefined Reasoning table. The boundary is further blurred by the 'hybrid-modular' subcategory, which is said to contain an agent that coordinates modules. So the central claim is not merely incomplete; the two categories are not mutually exclusive under the definitions given. The System 1 / System 2 framing inherits this problem because it rests on the same binary. The claim could be rescued by choosing an unambiguous criterion (e.g., trained vs. prompted decision-making, or fixed control graph vs. free-form model generation) and reclassifying methods consistently.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript surveys Reasoning Agentic Retrieval-Augmented Generation (RAG), proposing a two-branch taxonomy: predefined reasoning (route-based, loop-based, tree-based, and hybrid-modular pipelines) versus agentic reasoning (prompt-based and training-based methods). The taxonomy is mapped onto the cognitive-science distinction between System 1 and System 2. The survey reviews representative methods in each branch, provides a summary table with code links, and concludes with future directions on reward design, tool control, retrieval efficiency, and robustness.","tokens_in":13566,"tokens_out":4288,"duration_ms":49484,"significance":"The survey addresses a timely topic with strong industrial relevance. Its strengths include a curated GitHub repository, a compact table of methods with code availability, and a detailed treatment of recent RL-based training approaches such as Search-R1, R1-Searcher, ReZero, DeepRetrieval, and DeepResearcher. If the taxonomy were consistently applied, the paper would be a useful organizing reference for practitioners. However, the central classification claim is currently not sound as stated: the paper's own definitions and Table 2 put the same methods into both predefined and agentic categories, and the System 1/System 2 framing is asserted rather than argued. The main contribution therefore needs substantial clarification before the survey can be relied upon.","major_comments":[{"comment":"The central predefined/agentic distinction is not mutually exclusive under the paper's own definitions. Section 4 defines agentic reasoning as the LLM acting as an autonomous agent that \"identifies knowledge gaps, determines when and what to retrieve, and interacts with external tools\". Section 3 describes Self-RAG as a loop-based method in which \"the model can decide mid-generation to fetch additional information or to critique its own draft answer\", which satisfies the agentic definition, and Table 2 lists Self-RAG's control type as \"Agentic\" while placing it in the Predefined Reasoning table. The same issue affects MCTS-RAG (tree-based predefined reasoning but control type \"Agentic\"), and the hybrid-modular category is described as containing \"an agent that coordinates modular components\". This overlap means the two categories do not form a sound partition, undermining the paper's central claim. A fix would be to choose a single unambiguous criterion, such as fixed control graph versus free-form model-generated action sequence, or trained versus prompted retrieval decisions, and then reclassify all methods consistently.","section":"§3, §4 (especially Table 2)"},{"comment":"The System 1/System 2 alignment is asserted without operational grounding. The manuscript does not define what makes a workflow \"fast\" or \"deliberative\" in a way that applies to the surveyed systems; for example, a route-based system that invokes a confidence estimator before retrieval is not obviously System 1, and a training-based agent using a learned policy is not obviously System 2. Because Table 1 merely restates the problematic predefined/agentic split, the cognitive-science framing inherits the same weakness. I recommend presenting this as a heuristic analogy and supporting it with explicit behavioral or architectural criteria, rather than as a categorical alignment.","section":"§1, Table 1"}],"minor_comments":[{"comment":"The sentence \"This synergy supports a predefined reasoning, where the agent iteratively queries external sources (e.g., a local database or web search) and refines its reasoning based on the retrieved evidence\" contradicts the intended meaning of predefined reasoning, since iterative querying and refinement are exactly what the paper later defines as agentic; this should be reworded.","section":"§3 (opening paragraph)"},{"comment":"There are several typographical errors: \"Funciton calling\" (Table 2) should be \"Function calling\", \"apporaches\" (§4 introduction) should be \"approaches\", and \"predefine reasoning\" (§3 final paragraph) should be \"predefined reasoning\". The Search-o1 paragraph also contains a line-broken \"dist ill\" that should be \"distill\".","section":"Table 2 and §4"},{"comment":"The legend uses \"Router-based\" while the text and Figure 3 use \"Route-based\"; the terminology should be made consistent.","section":"Figure 4"},{"comment":"The bibliography entry [Zhang et al., ] has no year; it should be completed (the citation appears to refer to an ICLR 2024 workshop paper).","section":"References"},{"comment":"The paper calls itself a comprehensive review, but the coverage is necessarily selective; I suggest explicitly stating the selection criteria or qualifying the scope as representative rather than comprehensive.","section":"Abstract and §1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The stress-test note is right: the paper's central taxonomy doesn't hold up under its own definitions. Self-RAG and MCTS-RAG are filed under predefined reasoning even though the paper's agentic-reasoning definition—LLM identifies knowledge gaps, decides when and what to retrieve, interacts with tools—describes exactly what those systems do. Table 2 even flags their control type as \"Agentic\" inside the Predefined Reasoning table. So the two-branch partition is not mutually exclusive as stated.\n\nThat said, the survey has real value. It covers a fast-moving corner of RAG—reasoning agents that decide when to retrieve—and collects the recent 2024–2025 literature (Search-R1, R1-Searcher, ReZero, DeepResearcher, MCTS-RAG) in one place, with a GitHub repo of links. The prompt-based vs. training-based split within agentic reasoning is useful and accurate. The System 1/System 2 framing is more hook than analysis, but it's harmless if the categories were clear.\n\nThe soft spots go beyond the taxonomy. Section 3 says predefined reasoning \"empower[s] agents to determine what information to seek\" after defining it as fixed-module pipelines, and the hybrid-modular subcategory explicitly includes an agent coordinating modules—agentic by any reasonable reading. There are also minor editorial errors like \"Funciton calling\" in Table 2. These are fixable but suggest the manuscript wasn't polished.\n\nThe taxonomy could be rescued by choosing one unambiguous criterion, say fixed control graph vs. free-form model generation, or trained vs. prompted decision-making, and reclassifying methods consistently. As is, a careful reader can't fully trust the organizing claim, which is the paper's whole point.\n\nWho gets value? Practitioners wanting a quick map of recent agentic RAG methods, and researchers after a bibliography. No new results, and the taxonomy needs work. A serious editor should still send it to peer review: the area is important, the survey is substantive, and the inconsistencies are fixable in revision. I'd want that fixed before publication.","headline":"A useful survey of recent agentic RAG work, but its central predefined/agentic taxonomy is internally inconsistent and needs a clearer criterion before it can be trusted.","tokens_in":14028,"tokens_out":2913,"would_cite":false,"duration_ms":27041,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey organizes agentic RAG systems into two reasoning paradigms—predefined, fixed pipelines and autonomous, tool-orchestrating agents—mapped to System 1 and System 2 cognition.","keywords":["retrieval-augmented generation","agentic RAG","reasoning","System 1 / System 2","reinforcement learning","tool use","LLM agents","survey"],"falsifier":"A reader could test the taxonomy's completeness by coding recent agentic RAG papers on two independent rubrics: the paper's binary (predefined vs. agentic) and a set of orthogonal features (routing, looping, retraining, tool interface). If a substantial fraction of systems mix both branches—for example, an RL-trained retriever wrapped inside a fixed route-based workflow—or if the binary explains little variance in efficiency and adaptability across an evaluation benchmark, the central claim loses force.","tokens_in":13117,"feed_emoji":"🧠","tokens_out":5171,"duration_ms":51444,"temperature":0.7,"pith_summary":"This survey argues that the current landscape of reasoning agentic retrieval-augmented generation (RAG) is organized by two design families. Predefined reasoning uses fixed modular pipelines—routing, looping, tree-structured exploration, and hybrid modules—to control when and how retrieval happens. Agentic reasoning lets the language model itself decide when to retrieve and which tools to call, either through prompting or through reinforcement-learning training. The paper maps these families onto dual-process cognition: predefined reasoning as System 1 (fast, rule-based) and agentic reasoning as System 2 (slow, deliberative). A sympathetic reader comes away with a structured map of the field and a clear statement of the open problems—reward design, tool control, retrieval efficiency, and generalization—that will decide whether these systems work in industrial settings.","feed_headline":"RAG agents split into two reasoning styles","feed_subtitle":"A survey maps fixed pipelines to System 1 and autonomous tool use to System 2, and flags open industry problems.","key_machinery":"The organizing device is a two-branch taxonomy of Reasoning Agentic RAG, carried by the System 1/System 2 analogy from cognitive science. Predefined reasoning is defined as structured, rule-based execution with fixed modular pipelines, subdivided into route-, loop-, tree-, and hybrid-modular designs; agentic reasoning is defined as autonomous, model-driven decision-making about tool interaction, subdivided into prompt-based and training-based implementations. The taxonomy does the analytical work: it groups representative systems, aligns each group with a cognitive mode, and frames the central efficiency-adaptability trade-off.","core_discovery":"The paper's central claim is that Reasoning Agentic RAG methods are not a random assortment of tricks but fall into two coherent families distinguished by who controls the retrieval process. In predefined reasoning, control sits in a fixed workflow: route-based methods (RAGate, Self-Route) turn retrieval on conditionally; loop-based methods (Self-RAG, CRAG) iterate retrieval-feedback cycles; tree-based methods (RAPTOR, MCTS-RAG) organize hierarchical search; hybrid-modular systems (Adaptive-RAG, Modular-RAG) compose specialized modules under an orchestrator. In agentic reasoning, control sits in the model: prompt-based approaches (ReAct, Self-Ask, function calling, Search-o1) let instruction-following capabilities drive tool use, while training-based approaches (Search-R1, R1-Searcher, ReZero, DeepRetrieval, DeepResearcher) use reinforcement learning to teach the model when and how to search. The claimed payoff is that this two-branch taxonomy gives practitioners a principled way to trade efficiency against adaptability, and gives researchers a shared vocabulary for comparing systems.","pith_inferences":["The System 1/System 2 mapping is an analogy, not a proven architectural law; a learned router inside a predefined pipeline or a routinized agentic policy would blur the boundary in a way the taxonomy does not address.","An independent benchmark could test whether the predefined/agentic split predicts performance on multi-hop, multi-source tasks better than orthogonal dimensions such as degree of tool access, ability to retrain, or number of reasoning steps.","The field may converge on hybrid controllers—predefined at the top, agentic at the leaves—so the two categories may be endpoints of a spectrum rather than a true dichotomy."],"forward_implications":["For well-structured queries with available evidence, predefined reasoning gives predictable, efficient answers; for open-ended, multi-step tasks, agentic reasoning becomes necessary.","Prompt-based agentic methods provide lightweight adaptability on frozen LLMs, while RL-trained methods trade training cost for more reliable search policies.","The efficiency-adaptability trade-off is the main axis for system choice: predefined systems are cheaper but rigid, agentic systems are flexible but require careful reward design and risk unproductive searches.","Progress in the field will come from finer-grained reward functions, more advanced tool configuration, retrieval-efficiency constraints, and training in realistic environments rather than static corpora."],"supporting_citations":[{"why":"RAGate is the central example of route-based predefined reasoning, showing conditional retrieval triggered by context or model confidence.","marker":"[Wang et al., 2024a]"},{"why":"Self-RAG is the foundational loop-based predefined reasoning method, using self-reflection to decide mid-generation whether to retrieve.","marker":"[Asai et al., 2023]"},{"why":"RAPTOR provides the tree-based predefined workflow, recursively organizing retrieved content into a summary tree for multi-hop reasoning.","marker":"[Sarthi et al., 2024]"},{"why":"Adaptive-RAG exemplifies hybrid-modular predefined reasoning by combining routing with retrieval adaptation based on question complexity.","marker":"[Jeong et al., 2024]"},{"why":"Modular-RAG is the most advanced hybrid-modular example, breaking the pipeline into composable, swappable modules.","marker":"[Gao et al., 2024]"},{"why":"ReAct is the foundational prompt-based agentic method, interleaving thought, action, and observation to drive retrieval.","marker":"[Yao et al., 2023]"},{"why":"Self-Ask demonstrates prompt-based agentic reasoning through step-by-step question decomposition and follow-up search actions.","marker":"[Press et al., 2023]"},{"why":"Search-o1 is the key large-reasoning-model example, triggering search on self-assessed knowledge gaps and distilling retrieved documents.","marker":"[Li et al., 2025a]"},{"why":"Search-R1 is the central training-based agentic method, using outcome-based reinforcement learning to teach when and what to search.","marker":"[Jin et al., 2025]"},{"why":"DeepResearcher extends training-based agentic reasoning to real-world web environments, showing emergent planning and cross-validation behaviors.","marker":"[Zheng et al., 2025]"}],"fun_headline_variants":["RAG agents split into System 1 and System 2 reasoning","Fixed pipelines vs self-driven agents: RAG's two reasoning paths","RAG reasoning: predefined steps or autonomous tool use","Survey maps RAG agents to fast and slow thinking","System 1 vs System 2: how RAG agents decide to search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's organizing claim depends on the assumption that the binary split between predefined and agentic reasoning is a complete and meaningful partition of existing Reasoning Agentic RAG designs, and that the System 1/System 2 analogy is a valid analytical lens for them.","fun_headline_variants_meta":{"raw":{"variants":["RAG agents split into System 1 and System 2 reasoning","Fixed pipelines vs self-driven agents: RAG's two reasoning paths","RAG reasoning: predefined steps or autonomous tool use","Survey maps RAG agents to fast and slow thinking","System 1 vs System 2: how RAG agents decide to search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2653,"prompt_tokens":988,"completion_tokens":1665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":1578}},"tokens_in":604,"tokens_out":1665,"duration_ms":15872,"temperature":1.0,"reasoning_tokens":1578,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:26:18.452844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could test the taxonomy's completeness by coding recent agentic RAG papers on two independent rubrics: the paper's binary (predefined vs. agentic) and a set of orthogonal features (routing, looping, retraining, tool interface). If a substantial fraction of systems mix both branches—for example, an RL-trained retriever wrapped inside a fixed route-based workflow—or if the binary explains little variance in efficiency and adaptability across an evaluation benchmark, the central claim loses force.","supporting_citations":[{"cited_title":"Self-rag: Learning to retrieve, generate, and critique through self- reflection,","cited_arxiv_id":null,"evidence_quote":"Self-RAG is the foundational loop-based predefined reasoning method, using self-reflection to decide mid-generation whether to retrieve."},{"cited_title":"Rap- tor: Recursive abstractive processing for tree-organized retrieval,","cited_arxiv_id":null,"evidence_quote":"RAPTOR provides the tree-based predefined workflow, recursively organizing retrieved content into a summary tree for multi-hop reasoning."},{"cited_title":"Adaptive-rag: Learning to adapt retrieval-augmented large language models through question complexity,","cited_arxiv_id":null,"evidence_quote":"Adaptive-RAG exemplifies hybrid-modular predefined reasoning by combining routing with retrieval adaptation based on question complexity."},{"cited_title":"Modular rag: Transforming rag systems into lego-like reconfigurable frameworks,","cited_arxiv_id":null,"evidence_quote":"Modular-RAG is the most advanced hybrid-modular example, breaking the pipeline into composable, swappable modules."},{"cited_title":"React: Synergizing reasoning and acting in language models,","cited_arxiv_id":null,"evidence_quote":"ReAct is the foundational prompt-based agentic method, interleaving thought, action, and observation to drive retrieval."},{"cited_title":"Measuring and narrowing the compositionality gap in language models,","cited_arxiv_id":null,"evidence_quote":"Self-Ask demonstrates prompt-based agentic reasoning through step-by-step question decomposition and follow-up search actions."},{"cited_title":"Search-r1: Training llms to reason and leverage search engines with reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Search-R1 is the central training-based agentic method, using outcome-based reinforcement learning to teach when and what to search."},{"cited_title":"Deep- researcher: Scaling deep research via reinforcement learn- ing in real-world environments,","cited_arxiv_id":null,"evidence_quote":"DeepResearcher extends training-based agentic reasoning to real-world web environments, showing emergent planning and cross-validation behaviors."}],"review_version":1}