{"id":"a5005e30-7d39-467a-b392-03f57c12bf32","arxiv_id":"2607.18368","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A learned meta-policy that selects among named memory heuristics achieves the best held-out QA accuracy in the RoomKG benchmark while keeping memory operations symbolic and traceable.","lead":"The paper trains a controller that chooses, at each step, which simple memory rule (e.g. keep the most recently seen fact, drop the least used fact) to apply in a navigation-and-question-answering game. It reports small gains over fixed rules while keeping every memory decision inspectable.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract claims train/test room splits, but §4 describes the test set as 'same underlying dynamics with permuted question order'; new-room generalization is not evidenced by the reported protocol.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the claimed held-out generalization depends on the test split being a meaningful generalization test. My reading of §4 confirms the discrepancy: the abstract and contributions say 'train/test room splits,' while §4 says 'same underlying dynamics with permuted question order.' If the latter is the actual protocol, then the central claim of generalization to new rooms is not tested. The released code makes this directly verifiable, and the absence of a room-disjoint evaluation would not necessarily invalidate the more modest claim that learned heuristic selection adapts within a fixed layout, but it would invalidate the headline wording. The secondary statistical concern is also real: with 5 seeds and overlapping standard deviations, the 1.0-point advantage over the strongest fixed baseline is not shown to be reliable. However, the split mismatch is the more fundamental issue, since it concerns what the experiment is actually measuring. The paper has genuine strengths: it is built on an existing benchmark, the contribution is clearly scoped as a meta-policy layer, execution remains symbolic, and the released code supports reproducibility. The limitations section is unusually candid, acknowledging the single benchmark family, one primary capacity, deterministic dynamics, and dependence on the heuristic library. These strengths do not fix the split ambiguity. Since the reader already assigned CONDITIONAL and my concern reinforces that judgment rather than moving it, the verdict should remain unchanged; the authors should be asked to clarify the split protocol, and if the protocol is only question-order permutation, to either add a room-disjoint experiment or soften the claim.","tokens_in":14648,"tokens_out":2968,"duration_ms":26733,"concrete_test":"Open the released repository (github.com/humemai/roomkg-meta-policies, v0.1.0), locate the RoomKG environment's train/test construction, and confirm whether the train and test splits differ in room sets or only in question order. Then rerun the main comparison (StarE-GNN NS-TKG vs the best S-TKG baseline, K=512) with genuinely disjoint room sets, report per-seed paired scores, and run a paired significance test (e.g., Wilcoxon or bootstrap) on the difference. If the released split is only a question-order permutation and no room-disjoint experiment is added, the abstract's 'room splits' claim should be revised or the generalization claim downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—'strong train/test generalization' and 'best held-out performance on train/test room splits'—requires that test episodes come from rooms or layouts not seen in training. Section 4, however, describes the held-out environment as 'same underlying dynamics with permuted question order.' That wording is consistent with a fixed room graph and only shuffled queries. If that is what was run, the agent may have memorized the room layout during training, and Table 1's 47.28 vs 46.28 point estimate does not test generalization to new rooms. The conclusion's claim about 'this bounded-memory RoomKG setting' and the abstract's 'room splits' wording are therefore unsupported by the protocol as written. A secondary but related weakness: the 1.0-point edge over the best fixed S-TKG baseline (46.28) is within one standard deviation (StarE-GNN std 0.98; best S-TKG std 0.89–1.46) across 5 seeds, and no significance test is reported. Even under the intended split, the advantage over strong fixed symbolic controls is not established beyond noise. Both issues are checkable because the code is released; the split question is the more load-bearing because it determines whether the headline generalization claim is being tested at all.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies memory management in RoomKG, a partially observable gridworld whose observations and hidden state are represented as RDF graphs. It proposes a neuro-symbolic meta-policy that, given an encoded representation of the agent's bounded temporal knowledge-graph memory, selects among named symbolic heuristics for question answering (MRA/MRU/MFU), exploration (MRA/MRU/MFU), and forgetting (FIFO/LRU/LFU), and then executes the selected heuristics symbolically. The method is instantiated with GCN, R-GCN, or StarE-GNN encoders, head-specific attention pooling, and three DQN-style value heads (with a combinatorial 27-way head as an ablation). Experiments at memory capacity 512 compare fixed symbolic baselines, end-to-end LSTM/Transformer agents, and neuro-symbolic variants. The paper reports that StarE-GNN achieves the best held-out QA accuracy (47.28) over five seeds, and it provides qualitative traces showing interpretable switching among named heuristics.","tokens_in":14947,"tokens_out":5027,"duration_ms":46846,"significance":"If the empirical claims are supported, the paper offers a clean and timely formulation: learned selection over explicit symbolic memory operators rather than opaque latent memory policies, while preserving inspectability. The open-source code and the explicit limitations section are strengths. However, the empirical support is currently not conclusive: the described test protocol does not match the 'room split' claim, and the main advantage over strong fixed symbolic controls is within one standard deviation with no significance testing. The conceptual contribution can stand after a corrected evaluation or appropriately narrowed claims.","major_comments":[{"comment":"The abstract and §1 claim 'train/test room splits' and 'strong train/test generalization', but §4 defines the held-out environment as 'same underlying dynamics with permuted question order.' This does not hold out rooms or layouts; test performance could reflect memorization of a fixed room graph. The protocol as written therefore does not test the headline generalization claim. Please state exactly what differs between train and test. If new-room generalization is intended, add an evaluation on unseen room configurations; otherwise revise the abstract and conclusions to claim only robustness to question-order permutation.","section":"Abstract / §1 / §4 (Experimental Setup)"},{"comment":"The central quantitative claim is that StarE-GNN is 'best' among all systems. With 5 seeds, StarE-GNN gives 47.28 ± 0.98, while the strongest fixed symbolic controls give 46.28 ± 0.89 and 46.28 ± 1.46, and R-GCN gives 47.00 ± 1.67. The point differences are less than the reported standard deviations, and no paired significance test, confidence interval, or effect size is reported. A ~1.0-point gain over a strong baseline cannot support 'best held-out performance' without such an analysis. Report seed-paired tests or per-episode bootstrap intervals, or soften the comparative claim.","section":"Table 1 / §5.1"}],"minor_comments":[{"comment":"The StarE-GNN message function is written abstractly with ϕr and ψ; give the concrete instantiation used in the experiments, since the qualifier-awareness claim rests on this choice.","section":"§3.2, Eq. (6)"},{"comment":"The caption says 'BFS denotes breadth-first exploration in the symbolic baselines,' but no row or column is labeled BFS. Clarify or remove.","section":"Table 1 caption"},{"comment":"The phrase 'head-specific handling for QA as described in Section 4' appears inside Section 4 itself and should refer to Section 3.4.","section":"§4.2"},{"comment":"State whether the Q-value trace is from a single seed and how representative it is; the supplement may contain all heads, but the main text should give the selection criterion.","section":"Figure 3"},{"comment":"Minor typo: 'meta-policy that learnswhich' should be 'meta-policy that learns which.'","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the mismatch between the abstract's room-split claim and Section 4's described protocol. The released code should be checked for what the split actually is; if it is only a question-order permutation, the authors need either new-room experiments or a substantial revision of the central claim. The statistical weakness of the main comparative result is also load-bearing and should be addressed with proper tests."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the contribution is real but narrow — learned selection among named symbolic memory heuristics via three value heads — and the paper is refreshingly honest about its limitations. But the abstract says \"train/test room splits\" while Section 4 describes the test set as \"same underlying dynamics with permuted question order.\" That is a direct contradiction. If the test set contains no new rooms or layouts, the headline claim about generalization to held-out rooms is not supported by the protocol as written. The authors need to resolve this before the paper is taken at face value.\n\nWhat is actually new: the three-head meta-policy (QA/explore/forget) over a 3x3x3 symbolic action space, with encoder choice (GCN/R-GCN/StarE-GNN) as a proper ablation. The idea of keeping execution symbolic while learning which heuristic to fire is clean, and the qualitative traces make the inspectability claim concrete. Code and data are released, which is real evidence. The authors also explicitly state that the substrate is prior work, so the self-citation pattern is fine.\n\nSoft spots: (1) The split inconsistency is load-bearing. If the test is just a shuffled question order, the agent may have memorized the room layout, and Table 1 does not test the claimed \"train/test room splits.\" (2) The best result is 47.28 vs 46.28 for a strong fixed baseline; with standard deviations around 1.0 and no significance test, that gap is not established beyond noise. (3) Only capacity 512 is evaluated, and only one benchmark family. The authors list these limitations themselves, so this is not an oversight, but it does cap the strength of the claims.\n\nThe paper is coherent on its own terms. The math is standard DQN with separate heads; no fitting-as-prediction is happening. The method is sensible and the empirical question is well-posed. I would not call the 1-point difference statistically meaningful based on what is reported.\n\nRecommendation: send it to peer review, but with a firm request to resolve the split inconsistency and to report confidence intervals or a significance test. The contribution is legitimate and deserves referee time. I would not cite it in my own work until the split question is clarified; if clarified, it would be a useful reference for neuro-symbolic memory control.","headline":"A modest, honest paper whose central generalization claim is undercut by an internal inconsistency about what the held-out split actually is.","tokens_in":15435,"tokens_out":1635,"would_cite":false,"duration_ms":14792,"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":"A controller that learns which memory heuristic to apply at each step outperforms fixed symbolic rules in a partially observable knowledge-graph task while staying fully inspectable.","keywords":["neuro-symbolic reinforcement learning","partial observability","temporal knowledge-graph memory","meta-policy learning","symbolic memory heuristics","RDF graph annotations","RoomKG","graph neural networks"],"falsifier":"Run the best StarE-GNN meta-policy on held-out rooms whose layouts were never seen in training, across many seeds; if its test accuracy no longer beats the best fixed symbolic baseline by more than the noise level, the claim of improved held-out generalization from learned heuristic selection is falsified.","tokens_in":14516,"feed_emoji":"🧠","tokens_out":10136,"duration_ms":82831,"temperature":0.7,"pith_summary":"Partially observable agents must decide what to record, retrieve, and discard over time; this paper argues that those memory decisions can be learned without giving up symbolic control. The proposed meta-policy reads a memory made of RDF fact triples annotated with timestamps and access counts, and at every step chooses among three named retrieval rules (MRA, MRU, MFU), three exploration rules of the same family, and three forgetting rules (FIFO, LRU, LFU). The paper reports that, in the RoomKG navigation-and-query environment with a long-term memory cap of 512 facts, learned selection—particularly with the qualifier-aware StarE-GNN encoder—achieves the best held-out question-answering accuracy among the symbolic, neural, and neuro-symbolic systems compared, while every decision remains traceable to a named heuristic. A sympathetic reader would care because the result suggests that adaptive memory management does not require an opaque neural memory, and that picking among existing symbolic strategies is itself a learnable skill.","feed_headline":"Learned memory-rule switching beats fixed policies","feed_subtitle":"At a 512-fact memory cap, a learned rule-picker answers location queries best while keeping every memory decision traceable.","key_machinery":"The load-bearing object is the annotated RDF memory graph: every stored fact is a base subject-predicate-object triple plus statement-level qualifiers—time_added, last_accessed, num_recalled—so each memory item carries exactly the metadata that the heuristics rank. The symbolic action space is finite and named: MRA (most recently added), MRU (most recently used), and MFU (most frequently used) for question answering and exploration, and FIFO, LRU, and LFU for forgetting. The meta-policy encodes memory with a graph neural network (GCN, R-GCN, or StarE-GNN), pools node embeddings with head-specific attention into three value heads, and picks the highest-valued heuristic for each category; a de","core_discovery":"The paper's central claim is that memory management can be cast as a meta-decision problem: instead of learning what to remember, the agent learns which existing symbolic memory heuristic to apply at each step. In RoomKG at capacity 512, the qualifier-aware StarE-GNN configuration reaches 47.28 held-out test QA accuracy, ahead of the strongest fixed symbolic combination (46.28) and far ahead of end-to-end LSTM/Transformer baselines (around 9–11). The authors interpret the pattern as showing that adaptivity and inspectability need not be traded off: the controller's choices are named heuristics over explicit annotations, and qualifier-aware encoding is what lets the value heads exploit the re","pith_inferences":["Editorial: the same meta-selection mechanism should transfer to any memory system whose operations are named and ranked over metadata—buffer-pool replacement, cache eviction, episodic-memory stores—because the method's assumptions are only that memory is an annotated graph and operations are named heuristics.","Editorial: the abstract claims generalization to train/test room splits, while the experimental section describes the held-out test as the same dynamics with permuted question order; showing the result on genuinely unseen layouts would be the direct way to substantiate that stronger reading of the claim.","Editorial: a straightforward extension would be to vary the heuristic library and measure test accuracy as a function of library size; the authors expect a poor library would cap learned selection, so the curve should rise and plateau."],"forward_implications":["Adaptive memory control can stay inspectable: every QA, exploration, and forgetting choice is a named heuristic over explicit metadata, so failures can be traced to a concrete rule.","Encoder choice matters for heuristic selection: qualifier- and relation-aware encoders (StarE-GNN, R-GCN) outperform topology-only GCN, matching the information the heuristics depend on.","Modular three-head selection beats a single 27-way combinatorial head, suggesting that separate value heads with denser supervision are better for this decision problem.","End-to-end prediction over the full 245-action combinatorial space is poorly suited to bounded-memory temporal KG tasks; the symbolic substrate carries most of the performance.","Because fixed symbolic baselines themselves differ, which heuristic is active is a consequential variable even without learning—reinforcing the case for learned meta-selection."],"fun_headline_variants":["Meta-policy picks memory heuristics, tops fixed rules","Learn which memory rule to use, beat all fixed ones","Neuro-symbolic memory controller wins at 512 cap","Adaptive rule-switching memory beats static strategies","Symbolic memory meta-learning tops fixed policies in QA"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim depends on the held-out test being a genuinely different generalization target and the roughly one-point lead over the best fixed baseline being more than seed noise; the reported protocol permutes question order rather than introducing new rooms, and the standard deviations are the same size as the margin.","fun_headline_variants_meta":{"raw":{"variants":["Meta-policy picks memory heuristics, tops fixed rules","Learn which memory rule to use, beat all fixed ones","Neuro-symbolic memory controller wins at 512 cap","Adaptive rule-switching memory beats static strategies","Symbolic memory meta-learning tops fixed policies in QA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1108,"prompt_tokens":709,"completion_tokens":399,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":335}},"tokens_in":453,"tokens_out":399,"duration_ms":4133,"temperature":1.0,"reasoning_tokens":335,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:37:17.525973+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the best StarE-GNN meta-policy on held-out rooms whose layouts were never seen in training, across many seeds; if its test accuracy no longer beats the best fixed symbolic baseline by more than the noise level, the claim of improved held-out generalization from learned heuristic selection is falsified.","supporting_citations":[],"review_version":1}