{"id":"b32c6b17-8e15-4dcf-8d07-17770722aab0","arxiv_id":"2505.14139","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"FlowQ uses energy-guided flow matching to learn an offline RL policy approximating π(a|s) ∝ πβ(a|s) exp(Q(s,a)) with guidance applied during training rather than at inference.","lead":"FlowQ trains a flow-based policy for offline reinforcement learning by shifting action samples toward high Q-values during training, so policy updates no longer need action sampling and inference needs no guidance. The authors report competitive D4RL scores, but the written algorithm contains a sign ambiguity that contradicts the stated objective.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's action target uses -∇Q, which under §4.1's energy-guided construction yields πβ exp(-Q), not the claimed πβ exp(Q); the central claim is not implemented as written.","rationale":"The central claim is that FlowQ implements π ∝ πβ exp(Q) with a constant-cost policy update. For that claim to hold, the sign of the Q-gradient in the flow target must be positive. The paper's own §4.1 construction is for p1 exp(-λE); with E = Q the endpoint is disadvantage-weighted, and with E = -Q the derived αc must contain +∇Q. Algorithm 1 contains -∇Q, so the implemented regression is inconsistent with the stated objective in either reading. I regard this as more load-bearing than the first-order Taylor assumption, because the sign error invalidates the central claim even in the ideal case where the Taylor expansion is exact; the Taylor approximation would only weaken an already mis-targeted objective. The reader's weakest_assumption (curvature of Q) is real but secondary. The invalid conditional-path step in A.1 is an additional independent problem, but the sign mismatch is sufficient by itself. The empirical timing claim is less central because the paper explicitly restricts it to policy updates. I therefore keep the reader's REJECT verdict: the manuscript as submitted does not support its central claim. A corrected sign, a valid derivation of the conditional path, and either code or fixed-hyperparameter results would be needed before the method's stated objective can be assessed.","tokens_in":15184,"tokens_out":9679,"duration_ms":98676,"concrete_test":"Train the Algorithm 1 regression on a 2D synthetic problem with known πβ and a linear Q (so the first-order Taylor expansion is exact), and compare the ODE endpoint distribution to πβ exp(Q) and πβ exp(-Q). If the endpoints match πβ exp(-Q), the negative-sign target in Algorithm 1 is confirmed as written; if they match πβ exp(Q), the paper needs to state its sign convention explicitly, since the text and algorithm disagree.","verdict_should_be":"REJECT","load_bearing_attack":"Section 4.2 states the objective π(a|s) ∝ πβ(a|s) exp(Q(s,a)). The energy-guided framework of §4.1 targets p1 exp(-λE), so matching exp(Q) requires E = -Q. Under the paper's first-order Gaussian approximation, the conditional mean is αc(t,x1) = t x1 - (1-t)^2 λ(t) ∇E(tx1). Substituting E = -Q yields αc = t x1 + (1-t)^2 λ(t) ∇Q(tx1). Algorithm 1 instead sets a_t = t a1 - (1-t)^2 λ(t) ∇Q(s, t a1) + (1-t)ε, which is the αc for E = +Q, i.e. for target πβ exp(-Q). The velocity target in Algorithm 1 inherits the same flipped sign. This is not a Taylor-accuracy issue: even with a linear Q, the specified regression fits a flow whose endpoint marginal is approximately πβ exp(-Q). Appendix A.1 does not repair the sign; it asserts a conditional path via an integral over the conditioning variable x1, which is not a valid conditional density, and therefore supplies no independent signed target. As written, the algorithm optimizes a different objective than the one claimed, so the D4RL results cannot be cited as evidence for the stated advantage-weighted policy.","agreement_with_reader":"partial"},"referee_report":null,"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-07T15:40:55.135205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}