{"id":"2dc1b902-751d-4ff9-aa35-1750d9b3a6de","arxiv_id":"2508.01005","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"MAO-ARAG uses an RL-trained planner to dynamically pick and combine query reformulation, document selection, and generation agents per question, claiming high answer quality while keeping cost and latency acceptable.","lead":"This paper describes MAO-ARAG, a question answering system in which a planner agent trained by reinforcement learning assembles a different mix of retrieval and generation modules for each question. The goal is to match a query's difficulty with the right amount of computation, keeping answers accurate while controlling cost and delay.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The adaptivity claim rests on an underspecified composite reward: without the cost-penalty weight and normalization, the RL planner could collapse to a fixed cheapest or highest-F1 workflow, and F1 serving as both reward and metric invites overfitting.","rationale":"The reader's verdict is UNVERDICTED because the full text is machine-unreadable, so the methods and experiments cannot be verified. My stress-test of the abstract's central claim reinforces that verdict rather than moving it: the single most load-bearing condition is that the composite reward F1 - cost gives the planner a stable and transferable learning signal. The reader's weakest assumption identifies exactly this premise, including the unspecified weight and normalization, and also flags F1's reliability as an answer-quality proxy. My concern adds two concrete failure modes: (1) a degenerate reward weighting collapses the planner to a fixed cheapest or most expensive workflow, which would make the 'adaptive' claim vacuous; (2) using F1 as both training reward and evaluation metric creates a direct reward-hacking channel, especially for paraphrase-rich open-domain QA. A third issue, latency not appearing in the reward despite being claimed as an outcome, makes the cost-latency claim underdetermined. None of these concerns by themselves prove the paper is wrong; they identify missing evidence that one focused ablation can supply. Since the reader's UNVERDICTED status reflects insufficient verifiability, and my concern does not change the direction of the verdict, I keep UNCHANGED. The proposed concrete test is intentionally minimal: a w-sweep plus fixed-policy baselines would distinguish genuine per-query adaptation from reward-weight collapse, and a cost-vs-latency plot would test whether the latency claim is actually enforced.","tokens_in":23410,"tokens_out":3602,"duration_ms":45956,"concrete_test":"Ask the authors for an ablation that sweeps the cost-penalty weight w, with reward normalization held fixed, and report (i) the distribution of selected workflows over test queries, (ii) F1 and cost for fixed cheapest-workflow and fixed most-expensive-workflow baselines, and (iii) the quality-cost frontier of the trained planner. The adaptivity claim is supported only if there exists a w where the planner's workflow distribution is not degenerate (e.g., not >95% one workflow), and the planner's frontier strictly dominates both fixed baselines. Additionally, plot latency against cost for the submitted experiments; if latency is not enforced by the cost penalty, state the actual latency constraint and how it is incorporated into training or evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the RL-trained planner 'intelligently selects' per-query workflows so that MAO-ARAG beats fixed RAG pipelines on quality, cost, and latency. That claim requires the training objective R = F1 - w·cost to be a stable, non-degenerate learning signal. The abstract reports no value or normalization for w, no reward scaling, and no credit-assignment detail. If w is too large, the optimal policy is always the cheapest workflow; if w is too small, it is always the highest-F1 workflow. In either case the planner is not adaptive, and the reported quality-cost frontier would be explained by executor choice rather than learned planning. The problem is compounded because F1 is both the training reward and the headline evaluation metric: in open-domain QA, where correct answers are often paraphrases of the reference, optimizing token-level F1 can favor terse or verbatim output over genuine answer quality. Finally, latency is claimed as a controlled outcome, yet the stated reward contains only a cost penalty; cost and latency need not be monotonically related. Because the supplied full text is encoding-corrupted and contains an unrelated arXiv header, no equations, tables, or baseline comparisons are available to check whether the planner's workflow distribution actually varies across queries. The load-bearing gap is therefore the unspecified training objective and the absence of evidence that learned workflow selection, rather than a fixed policy or executor quality, drives the results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAO-ARAG, a multi-agent orchestration framework for adaptive retrieval-augmented generation. It defines several executor agents (query reformulation, document selection, generation) and a planner agent trained with reinforcement learning to select and compose a workflow for each query, using an outcome-based reward (F1 score) and a cost-based penalty. The abstract claims that experiments on multiple QA datasets show high answer quality while keeping cost and latency within acceptable limits, and the paper points to a public GitHub repository for code.","tokens_in":23573,"tokens_out":3453,"duration_ms":42631,"significance":"If the central claim is substantiated, the paper would offer a practical way to trade off answer quality, cost, and latency in RAG systems by learning per-query workflow selection rather than using a fixed pipeline. The multi-agent decomposition is a plausible and timely design, and the provision of a public code link is a concrete reproducibility-oriented strength. However, the submitted full text is machine-unreadable and contains no accessible experimental details, so the empirical contribution cannot currently be assessed. The contribution is incremental if the RL planner simply selects among known RAG variants, but it could still be useful for practitioners.","major_comments":[{"comment":"The supplied full text is machine-unreadable: it consists of encoding-corrupted characters and even contains an unrelated header 'arXiv:2508.00991v1 [astro-ph.GA]'. No equations, tables, dataset names, baseline scores, error bars, or significance tests are visible. The central empirical claim in the abstract ('Experiments conducted on multiple QA datasets...') is therefore unsupported in the submitted manuscript. A clean, readable version with the complete experimental protocol must be provided before the paper can be evaluated.","section":"Full text"},{"comment":"The planner's training signal is described only as an 'outcome-based reward (F1 score) and a cost-based penalty', with no relative weight, normalization, or reward scaling specified. This is load-bearing for the adaptivity claim: if the cost penalty dominates, the optimal policy collapses to always selecting the cheapest workflow; if F1 dominates, it collapses to always selecting the highest-F1 workflow, and neither behavior is adaptive. The manuscript needs to state the exact reward equation, the numerical weight and normalization used, and a sensitivity analysis over that weight.","section":"Abstract"},{"comment":"The abstract claims that latency is maintained within acceptable limits, but the stated reward function contains only a cost-based penalty; cost and latency are not necessarily monotonically related (for example, parallel executor invocations can increase token cost while reducing wall-clock latency). The paper should either add an explicit latency term to the reward or report latency measurements with the same transparency as cost, including how latency is defined and measured.","section":"Abstract"},{"comment":"Because F1 is used both as the RL training reward and as the headline evaluation metric, the reported answer-quality gains may partly reflect reward overfitting to token-level overlap. This is a particular concern in open-domain QA, where correct answers are often paraphrases of the reference. The evaluation should include at least one metric not used during training (e.g., LLM-based correctness judgment or normalized exact match) or demonstrate that the F1 improvements are robust across answer formulations.","section":"Abstract (evaluation)"}],"minor_comments":[{"comment":"The phrase 'multi-turn framework' is introduced without explanation; please define what constitutes a turn (for example, one planner-executor iteration) at first use.","section":"Abstract"},{"comment":"The GitHub repository link should be accompanied by a version tag or commit hash and a license statement to make reproducibility claims concrete.","section":"References/Code"},{"comment":"The garbled characters and the unrelated astronomy arXiv header indicate that the submission file was corrupted during preparation; the file should be regenerated and checked for readability before resubmission.","section":"Full text"}],"recommendation":"major_revision","confidential_remarks":"The manuscript as supplied is not reviewable because the full text is encoding-corrupted and contains an unrelated arXiv header from an astronomy paper. I recommend asking the authors to resubmit a clean, readable PDF before any substantive review. The major_revision recommendation is based on the readable abstract: the central claim is plausible, but the reward specification, evaluation validity, and empirical evidence are all load-bearing points that need to be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front. First, the PDF is mangled: encoding corruption throughout, plus a stray astro-ph header embedded in the body. Nobody can actually read the paper as submitted. Second, based on the abstract alone, this is a sensible, moderately novel systems idea: an RL-trained planner that selects among executor agents over multiple turns, balancing an F1-based answer quality signal against a cost penalty. That combination is plausibly new relative to the single-round, iterative, and reasoning RAG families the authors cite.\n\nWhat the paper does well is frame a real operational problem: fixed RAG pipelines trade off quality and cost poorly across query types. Routing at the workflow level, rather than just at retrieval, is a reasonable place to look for gains. The abstract is clear about the components and the training setup.\n\nThe soft spots are mostly reporting gaps rather than design flaws. The stress-test's worry about R = F1 - w·cost is legitimate but unresolved: we do not know the weight, normalization, or credit-assignment scheme, so we cannot tell whether the planner actually adapts per query or collapses to a fixed cheapest or highest-F1 workflow. That is a question the full paper should answer, but the abstract alone does not. The second concern, that F1 serves as both reward and headline metric, is a mild and common circularity risk in QA; it would matter more if the goal were factual correctness beyond the reference wording. Also, no numbers appear in the abstract: no dataset names, no F1 or cost values, no baselines. Without the full text, none of the empirical claims can be checked.\n\nI would not call the central idea incoherent or the authors unserious. The motivation is clear and the approach is standard reinforcement-learning practice over a well-defined action space. The biggest barrier is entirely mechanical: the corrupted source file cannot be reviewed.\n\nWho should read this: people working on cost-efficient RAG deployments and multi-agent LLM pipelines. If the experiments actually show that learned workflow selection beats fixed pipelines on the quality-cost-latency frontier, it is a modest but useful contribution. My recommendation: the authors must re-upload a clean PDF first. Once the text is readable, this deserves normal peer review rather than a desk reject. I would not cite it until I have seen the numbers.","headline":"Unreadable manuscript, but the abstract describes a plausible RL-trained adaptive RAG controller; the reward specification is a real question, not a demonstrated flaw.","tokens_in":24244,"tokens_out":3119,"would_cite":false,"duration_ms":40590,"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":"MAO-ARAG claims that a reinforcement-learned planner agent can pick and sequence RAG modules per query, keeping answer quality high while holding cost and latency in bounds.","keywords":["retrieval-augmented generation","multi-agent orchestration","reinforcement learning","question answering","adaptive workflow planning","cost-aware reward","RAG pipeline selection"],"falsifier":"On a held-out set with a known mix of simple and multi-hop questions, train two variants: one with the cost penalty set to zero and one with F1 removed. A planner that is truly adaptive should route the simple questions through short workflows and the hard questions through long ones, while the two degenerate variants should respectively route everything through the longest and shortest workflows; observing any other pattern would show the composite reward is not what produces adaptivity.","tokens_in":23086,"feed_emoji":"🤖","tokens_out":5911,"duration_ms":67303,"temperature":0.7,"pith_summary":"Fixed retrieval-augmented generation (RAG) pipelines choose one architecture—single-round, iterative, or reasoning—for every question, so they must compromise between accuracy and expense on queries of varying difficulty. This paper tries to establish that the compromise is unnecessary: a learned planner can compose the same RAG modules into a different workflow for each query. The proposed system, MAO-ARAG, treats adaptive RAG as a multi-turn decision problem in which a planner agent selects among executor agents (query reformulation, document selection, generation) and is trained by reinforcement learning with a reward of answer F1 minus a cost penalty. Experiments on multiple QA datasets are reported to show that the dynamic workflows reach high answer quality while keeping cost and latency within acceptable limits. If the paper is right, the practical upshot is that adaptivity lives in the planner's routing decisions rather than in hand-designing separate pipelines.","feed_headline":"One planner agent learns the right RAG workflow per query","feed_subtitle":"Trained on answer F1 minus cost, it keeps QA accurate without letting latency balloon.","key_machinery":"The load-bearing object is the planner agent: a multi-turn policy trained with reinforcement learning to select and sequence executor agents. Executor agents are the standard RAG modules—query reformulation, document selection, generation—so the planner's job is routing, not retrieval or generation itself. The mechanism that makes the routing adaptive is the composite reward, F1 score minus a cost-based penalty, evaluated after the answer is produced; it is what causes expensive workflow steps to be chosen only when they improve the outcome enough to pay for themselves. The multi-turn loop lets the planner revise its plan after seeing intermediate results, which is how the same module toolbox can serve both simple and multi-hop questions.","core_discovery":"The paper's central claim is that per-query workflow planning is itself a learnable policy. Given a question, the planner agent chooses and sequences executor agents from a fixed toolbox of RAG modules—query reformulation, document selection, generation—over multiple turns, and it improves those choices through reinforcement learning. The training signal is outcome-based: token-level F1 between the generated answer and the reference, minus a penalty proportional to execution cost, which is meant to force the planner to spend more compute only when it buys accuracy. Across several QA datasets, the paper reports that this dynamic planning achieves high answer quality and keeps cost and latency within acceptable limits, meaning the learned workflows sit on a better quality-cost-latency frontier than a fixed pipeline.","pith_inferences":["A direct experimental check the paper leaves implicit: vary the cost-penalty coefficient and measure whether the planner's workflow-length distribution shifts monotonically; a non-monotone shift would indicate the reward is not the driving signal.","The same planner-plus-executors formulation transfers to tool-using agents beyond RAG, where executor agents are external tools and the reward is task success minus tool cost.","Because F1 rewards surface word matches, the planner may be biased toward verbose or paraphrased answers; a semantic-similarity reward would likely change which workflows are judged successful.","A training-time cost penalty bounds expected cost, not worst-case latency, so a production deployment would likely need an explicit stop condition to cap tail latency on hard queries."],"forward_implications":["A deployed QA system can reuse existing retrieval and generation modules and gain adaptivity purely by swapping the hand-coded pipeline chooser for this learned planner.","Adjusting the cost-penalty weight at training time becomes a single control knob for the quality-cost trade-off, instead of a re-architecture decision.","Queries that are simple should be routed through short, cheap workflows while hard, multi-hop queries get longer ones, so gains over fixed pipelines should be largest on mixed-difficulty datasets.","Because the planner's output is an explicit sequence of agent calls, the system's per-query behavior remains inspectable as a workflow rather than being a black-box generation."],"supporting_citations":[],"fun_headline_variants":["Planner learns per-query RAG workflows via RL","Multi-agent planner picks RAG steps per query","RL-trained planner adapts RAG pipeline per query","Adaptive RAG: one planner, many workflows","Planner agent tunes RAG workflow to each query"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that answer F1 minus a cost penalty is a trustworthy, balanced training signal: F1 must be a reasonable stand-in for answer quality, and the penalty weight must be set so the planner does not collapse to always-cheapest or always-most-expensive workflows.","fun_headline_variants_meta":{"raw":{"variants":["Planner learns per-query RAG workflows via RL","Multi-agent planner picks RAG steps per query","RL-trained planner adapts RAG pipeline per query","Adaptive RAG: one planner, many workflows","Planner agent tunes RAG workflow to each query"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00038,"raw_usage":{"total_tokens":2022,"prompt_tokens":955,"completion_tokens":1067,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":1005}},"tokens_in":571,"tokens_out":1067,"duration_ms":10141,"temperature":1.0,"reasoning_tokens":1005,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:55:22.946017+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set with a known mix of simple and multi-hop questions, train two variants: one with the cost penalty set to zero and one with F1 removed. A planner that is truly adaptive should route the simple questions through short workflows and the hard questions through long ones, while the two degenerate variants should respectively route everything through the longest and shortest workflows; observing any other pattern would show the composite reward is not what produces adaptivity.","supporting_citations":[],"review_version":1}