{"id":"a1a451a7-ca4e-4b0f-a0f1-da418c68ecdd","arxiv_id":"2606.00547","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"MERIT maintains episode-level and turn-level memories with RL-optimized retrieval and a process reward model, outperforming baselines on BIRD-Interact with positive transfer to Spider2-Snow.","lead":"The paper proposes MERIT, a dual-level memory system for text-to-SQL agents that learns retrieval policies via reinforcement learning at both episode and turn horizons, plus a process reward model for dense training signals. A smart generalist might read it to see how multi-scale memory can make conversational database agents more efficient by reusing past experiences at different decision levels.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Lightweight PRM accuracy as dense proxy for turn-level RL is unverified and central to local policy optimization","rationale":"The reader's weakest_assumption directly identifies the same dependency; the empirical claim on BIRD-Interact and transfer cannot be assessed until the PRM proxy quality is checked. No other internal inconsistency is visible from the given abstract.","tokens_in":1730,"tokens_out":286,"duration_ms":9553,"concrete_test":"On the BIRD-Interact validation interactions, compute Spearman correlation between PRM-assigned turn rewards and (a) human-annotated step utility or (b) final episode success; if correlation < 0.5, re-train the local policy with oracle rewards and measure whether the reported success-rate gain disappears.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"MERIT's turn-level retrieval policy is trained via RL using only a lightweight Process Reward Model to supply dense proxy rewards, because intermediate supervision is absent. For the local policy to improve state-conditioned memory selection (and thereby reduce turns while raising success rate), the PRM scores must be sufficiently correlated with actual downstream utility. The abstract provides no description of PRM training data, architecture, or any accuracy/ablation results, leaving open the possibility that noisy or misaligned proxies produce a retrieval policy no better than the dynamic-retrieval baseline it claims to beat.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes MERIT, a dual-level long-term memory framework for interactive text-to-SQL agents. It maintains episode-level memory for global strategic guidance and turn-level memory for local, state-conditioned decisions. Both levels use reinforcement learning to optimize retrieval policies; a lightweight Process Reward Model supplies dense proxy rewards to train the turn-level policy in the absence of intermediate supervision. Experiments claim that MERIT outperforms no-memory, static-retrieval, and dynamic-retrieval baselines on BIRD-Interact (higher success rate, fewer turns) and shows positive transfer to Spider2-Snow without benchmark-specific tuning.","tokens_in":1851,"tokens_out":372,"duration_ms":13876,"significance":"If the empirical gains are reproducible and the Process Reward Model proxies prove sufficiently accurate, the multi-horizon retrieval approach would advance experience reuse in multi-turn agents beyond single-horizon or heuristic methods, with potential impact on interactive database interfaces.","major_comments":[{"comment":"Abstract and the section describing the turn-level policy (likely §3.2 or equivalent): the central claim that the RL-trained local retrieval policy improves over dynamic-retrieval baselines rests on the assumption that the lightweight Process Reward Model supplies sufficiently accurate dense proxy rewards correlated with downstream utility. No training data, architecture, accuracy metrics, or ablation results for this PRM are provided, leaving open the possibility that noisy proxies yield no genuine improvement.","section":"Process Reward Model and turn-level RL training"}],"minor_comments":[{"comment":"The abstract states outperformance and transfer results but supplies no numerical values, error bars, baseline details, or dataset sizes, reducing immediate verifiability of the strongest_claim.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comment on the Process Reward Model. We agree that additional details are required to support the claims about the turn-level retrieval policy.","responses":[{"response":"We agree that the submitted manuscript does not provide the requested details on the Process Reward Model (PRM). The description mentions its use for dense proxy rewards but omits training data, architecture, accuracy metrics, and ablations, which weakens support for the claim of improvement over dynamic-retrieval baselines. In the revised version we will add a dedicated subsection (in §3.2 or an appendix) covering: the PRM architecture and training procedure, the dataset used to train it, quantitative accuracy metrics (including correlation with downstream task success), and ablation results isolating the PRM's contribution to turn-level policy performance. These additions will allow direct evaluation of whether the proxy rewards are sufficiently accurate and correlated with utility.","revision_made":"yes","referee_comment":"[Process Reward Model and turn-level RL training] Abstract and the section describing the turn-level policy (likely §3.2 or equivalent): the central claim that the RL-trained local retrieval policy improves over dynamic-retrieval baselines rests on the assumption that the lightweight Process Reward Model supplies sufficiently accurate dense proxy rewards correlated with downstream utility. No training data, architecture, accuracy metrics, or ablation results for this PRM are provided, leaving open the possibility that noisy proxies yield no genuine improvement."}],"tokens_in":1328,"tokens_out":316,"duration_ms":19909,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing here is a dual-horizon retrieval setup for interactive text-to-SQL: episode-level memory for strategy and turn-level memory for local decisions, both trained with RL policies. The turn level uses a lightweight Process Reward Model to generate dense signals because direct intermediate labels are missing. That combination is new relative to the static and single-horizon dynamic baselines they cite.\n\nWhat works is the motivation. Usefulness of past experience really does shift between planning and execution steps, so separating the horizons makes sense on paper. The transfer experiment on Spider2-Snow without retuning is also a reasonable check.\n\nThe soft spot is exactly the one the stress-test flags. The central empirical claim is that MERIT beats no-memory, static, and dynamic baselines on success rate and turns. The abstract states this but supplies zero numbers, error bars, baseline descriptions, or ablation results. More critically, the turn-level policy depends on the PRM being a good enough proxy; there is no mention of how the PRM was trained, what its accuracy is, or any correlation check against downstream utility. Without that, it is possible the local policy is not actually better than the dynamic baseline it claims to beat.\n\nThis is an empirical methods paper aimed at the text-to-SQL and agent-memory crowd. A reader working on interactive database agents might pick up the dual-level framing, but anyone needing reproducible numbers or verified reward-model quality will find the current write-up thin. The work shows coherent thinking about the problem but does not yet supply the evidence needed to trust the gains.\n\nI would not send it to peer review in this form; the missing quantitative support and PRM validation are load-bearing.","headline":"MERIT's dual-level RL memory with a PRM proxy for turn-level training is a plausible idea but the abstract gives no numbers, ablations, or PRM validation, so the performance claims are hard to evaluate.","tokens_in":2351,"tokens_out":431,"would_cite":false,"duration_ms":12488,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"MERIT learns dual-level retrieval policies to reuse past experiences in multi-turn text-to-SQL tasks.","keywords":["text-to-SQL","memory retrieval","reinforcement learning","interactive agents","process reward model","multi-horizon memory"],"falsifier":"If removing the Process Reward Model or using only sparse final rewards leads to no improvement or worse performance than single-level dynamic retrieval on BIRD-Interact, the central claim would be falsified.","tokens_in":2634,"feed_emoji":"🧠","tokens_out":556,"duration_ms":17759,"temperature":0.7,"pith_summary":"The paper introduces MERIT, a framework that maintains two types of memory for text-to-SQL agents: episode-level for overall strategy and turn-level for immediate decisions. Retrieval policies at both levels are trained with reinforcement learning, and a process reward model provides dense signals for the turn-level policy. This approach aims to overcome limitations of static and single-horizon dynamic retrieval methods. If successful, it would allow agents to more effectively reuse experiences across different stages of database query interactions, leading to higher success rates and fewer turns.","feed_headline":"Dual memory levels boost text-to-SQL agent success","feed_subtitle":"MERIT learns retrieval policies for episode and turn horizons, cutting interaction turns on BIRD-Interact with cross-benchmark transfer.","key_machinery":"Dual-level memory retrieval with reinforcement learning policies and a Process Reward Model for proxy rewards.","core_discovery":"MERIT maintains episode-level memory for global strategic guidance and turn-level memory for local decision support, with both levels using learned retrieval policies optimized via reinforcement learning. A lightweight Process Reward Model supplies dense proxy rewards to train the turn-level policy despite sparse intermediate supervision. On BIRD-Interact, this yields higher success rates and fewer interaction turns than baselines, with positive transfer to Spider2-Snow.","pith_inferences":["Similar dual-level memory could apply to other interactive agent tasks beyond text-to-SQL.","The Process Reward Model approach might generalize to other RL settings with sparse rewards.","Testing on more diverse databases could reveal limits of the transfer."],"forward_implications":["Agents can reuse memories useful for initial planning separately from those for local execution.","RL-optimized retrieval improves over fixed similarity heuristics.","Cross-benchmark transfer is possible without specific tuning.","Multi-horizon retrieval reduces average interaction turns while increasing success."],"fun_headline_variants":["Dual-level memory in interactive text-to-SQL agents","Episode-level and turn-level memory retrieval policies","Reinforcement learning for multi-horizon memory in SQL agents","Lightweight process reward model for local memory support"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The lightweight Process Reward Model provides accurate enough dense proxy rewards for training the turn-level retrieval policy without direct intermediate supervision.","fun_headline_variants_meta":{"raw":{"variants":["Dual-level memory in interactive text-to-SQL agents","Episode-level and turn-level memory retrieval policies","Reinforcement learning for multi-horizon memory in SQL agents","Lightweight process reward model for local memory support"]},"model":"grok-4.3","cost_usd":0.006894,"raw_usage":{"total_tokens":3202,"prompt_tokens":674,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":68937000,"prompt_tokens_details":{"text_tokens":674,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2470,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":674,"tokens_out":58,"duration_ms":17570,"temperature":1.0,"reasoning_tokens":2470,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T19:03:43.305565+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"If removing the Process Reward Model or using only sparse final rewards leads to no improvement or worse performance than single-level dynamic retrieval on BIRD-Interact, the central claim would be falsified.","supporting_citations":[],"review_version":1}