{"id":"535b7376-1629-4bf1-bfaf-3c96f75b99e9","arxiv_id":"2606.03904","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"MAdam preconditions MOO solver directions with preference-conditioned curvature so that Adam's adaptive steps respect the intended metric instead of entangling it with gradient history.","lead":"MAdam is a wrapper for multi-objective optimizers that preconditions the direction given to Adam using the preference-conditioned curvature of the scalarized loss. If it works, it could make training more faithful to the intended trade-offs in multi-task and Pareto problems.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Preconditioning the reconciled direction r with scalarized curvature C does not imply Adam's EMA of g^2 (g = C^{-1/2} r) equals identity, because E[r r^T] generally ≠ C.","rationale":"The reader's weakest assumption correctly flags sufficiency of the fix. The more precise load-bearing point is the internal technical step that is supposed to deliver the fix: the claimed identity collapse. The mismatch between curvature and second-moment matrix of r is a concrete, falsifiable gap in that step. If the concrete_test passes, the claim is internally consistent; if it fails, the headline mechanism does not operate as stated, moving the verdict from UNVERDICTED to CONDITIONAL pending correction or clarification of the preconditioner.","tokens_in":1738,"tokens_out":491,"duration_ms":35042,"concrete_test":"In any reported experiment, record the per-coordinate EMA v_t of the preconditioned reconciled directions for the first 500 steps; compute ||v_t - 1||_∞. If the deviation exceeds 0.2 on average across runs, the collapse does not occur and the central mechanism fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim states that preconditioning the reconciled direction by the preference-conditioned curvature of the scalarized objective produces a whitened input on which Adam's second moment collapses to identity. Adam maintains a diagonal EMA v_t = β₂ v_{t-1} + (1-β₂) g_t ⊙ g_t. For v_t to collapse to the identity vector, the preconditioned vector g must satisfy E[g_i²] = 1 for each coordinate. This holds if the preconditioner is the inverse square root of the second-moment matrix of the input vector itself. The paper instead uses the curvature (Hessian or similar) of the scalarized loss L_λ. The reconciled direction r is the output of an arbitrary MOO solver (MGDA, linear scalarization, etc.) and is not in general distributed as ∇L_λ; hence E[r r^T] ≠ curvature matrix of L_λ. The resulting g therefore retains coordinate-wise second-moment variation, so Adam continues to apply a non-trivial diagonal adaptation on top of the supplied metric. This breaks the asserted collapse and leaves the realized update governed by a composite metric rather than the pure preference-conditioned metric.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.3","summary":"The paper identifies two systematic gaps when MOO solvers (loss-balancing, gradient-balancing, Pareto-based) hand reconciled directions to Adam: a weighting mismatch in which Adam's second-moment statistics entangle the time-varying preference vector, and a geometric mismatch in which Adam's adaptive metric distorts the Euclidean geometry assumed by the solvers. It introduces MAdam, a drop-in wrapper that preconditions the reconciled direction by the preference-conditioned curvature of the scalarized objective; the abstract asserts that this whitens the input so Adam's second moment collapses to identity and the realized update is governed by the preference-conditioned metric. Experiments across multi-task learning, Pareto-front recovery, PINNs, and medical imaging report consistent gains over Adam for every solver family.","tokens_in":2009,"tokens_out":437,"duration_ms":12898,"significance":"If the central preconditioning argument holds and the realized update is indeed governed purely by the supplied preference-conditioned metric, the method would supply a lightweight, solver-agnostic fix for a pervasive coupling problem in multi-objective machine learning, with direct applicability to the listed domains.","major_comments":[{"comment":"Abstract: the claim that preconditioning the reconciled direction r by the curvature C of the scalarized loss L_λ produces a whitened input on which 'Adam's second moment collapses to identity' is not supported by the stated construction. Adam's v_t is an EMA of g_t ⊙ g_t where g = C^{-1/2} r; for the EMA to become the identity vector it is necessary that E[g_i²] = 1 coordinate-wise, which holds only if C is (approximately) the second-moment matrix of r itself. Because r is the output of an arbitrary MOO solver and is not in general distributed as ∇L_λ, E[r r^T] ≠ C, so the second-moment adaptation remains non-trivial and the update is governed by a composite rather than a pure preference-conditioned metric. This directly undermines the stated resolution of both mismatches.","section":"Abstract"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for identifying this subtlety in the abstract's central claim. The comment correctly notes that exact collapse of Adam's second-moment vector to the identity requires the preconditioned direction to match the second-moment structure of the scalarized gradient, which does not hold for arbitrary MOO solvers. We address this point directly below and will revise the manuscript accordingly.","responses":[{"response":"We agree that the abstract's wording is imprecise. The preconditioning g = C^{-1/2} r aligns the input direction with the preference-conditioned metric of L_λ, so that the geometry of the update is governed by that metric rather than the original Euclidean geometry assumed by the MOO solver. However, Adam's subsequent EMA of g ⊙ g will generally remain non-identity because r is a reconciled direction whose second-moment structure need not coincide with C. Consequently the realized step is a composite of the preference-conditioned metric and Adam's adaptive normalization in the whitened coordinates. We will revise the abstract to remove the claim that the second moment 'collapses to identity' and instead state that MAdam ensures the effective metric of the update is the preference-conditioned curvature (with Adam providing coordinate-wise normalization within that metric). This preserves the resolution of the geometric mismatch while acknowledging that the weighting mismatch is mitigated rather than eliminated.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the claim that preconditioning the reconciled direction r by the curvature C of the scalarized loss L_λ produces a whitened input on which 'Adam's second moment collapses to identity' is not supported by the stated construction. Adam's v_t is an EMA of g_t ⊙ g_t where g = C^{-1/2} r; for the EMA to become the identity vector it is necessary that E[g_i²] = 1 coordinate-wise, which holds only if C is (approximately) the second-moment matrix of r itself. Because r is the output of an arbitrary MOO solver and is not in general distributed as ∇L_λ, E[r r^T] ≠ C, so the second-moment adaptation remains non-trivial and the update is governed by a composite rather than a pure preference-conditioned metric. This directly undermines the stated resolution of both mismatches."}],"tokens_in":1439,"tokens_out":480,"duration_ms":15175,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper identifies a weighting mismatch where Adam's second-moment history averages over changing preferences and a geometric mismatch where the adaptive metric warps the space the solver expects. It then builds MAdam as a wrapper that preconditions the reconciled direction with the curvature of the scalarized objective.\n\nWhat is new is the joint framing of those two gaps and the minimal wrapper that leaves both the solver and Adam untouched. The paper does well to focus on a practical, widely used pairing and to test the idea across multi-task learning, Pareto recovery, PINNs, and medical imaging, with gains reported for every solver family.\n\nThe central claim is that the preconditioned input makes Adam's EMA collapse to identity so the update is governed purely by the preference-conditioned metric. That does not hold in general. The reconciled direction r comes from the MOO solver and is not distributed as the gradient of the scalarized loss whose curvature is used for preconditioning. Therefore E[g g^T] after g = C^{-1/2} r is not identity, and Adam continues to apply a non-trivial diagonal adaptation. The abstract gives no derivation that resolves this, so the realized update remains a composite rather than the clean metric described.\n\nExperiments may still show net improvement, but they would need tight controls on tuning and implementation details to isolate the effect. The paper is aimed at people already using Adam inside multi-objective pipelines. It has enough of a concrete idea and some empirical signal to merit referee time, even though the theoretical justification needs tightening.","headline":"MAdam flags real coupling issues between MOO solvers and Adam and offers a preconditioner, but the claim that this makes the second moment collapse to identity does not follow from the given construction.","tokens_in":2506,"tokens_out":389,"would_cite":false,"duration_ms":18675,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"MAdam preconditions the reconciled direction with the preference-conditioned curvature of the scalarized objective to align Adam's updates with the multi-objective solver's intent.","keywords":["multi-objective optimization","Adam optimizer","preconditioning","multi-task learning","Pareto optimization","gradient balancing","loss balancing","metric-aware optimization"],"falsifier":"Experiments on the same multi-task, Pareto, and physics-informed tasks where MAdam produces no consistent improvement or worse results than standard Adam would falsify the claim that the preconditioner resolves the identified mismatches.","tokens_in":2659,"feed_emoji":"","tokens_out":679,"duration_ms":20385,"temperature":0.7,"pith_summary":"The paper shows that handing reconciled directions from multi-objective solvers to Adam creates a weighting mismatch, where the preference vector becomes entangled in the second-moment history and collapses distinct trade-offs, plus a geometric mismatch, where Adam's metric distorts the Euclidean space the solvers assume. MAdam fixes both as a wrapper by preconditioning the direction with the curvature of the scalarized objective conditioned on the current preference. On this whitened input Adam's second moment becomes the identity matrix, so the actual step is controlled by the intended metric instead. Experiments across multi-task learning, Pareto recovery, physics-informed networks, and medical imaging show consistent gains over plain Adam for every solver family.","feed_headline":"MAdam preconditions Adam with scalarized curvature to fix MOO gaps","feed_subtitle":"Whitening the input with preference-conditioned curvature makes the update follow the solver's metric instead of entangling preferences or d","key_machinery":"The preference-conditioned curvature preconditioner applied to the reconciled direction before it enters Adam.","core_discovery":"MAdam resolves the weighting and geometric mismatches by preconditioning the reconciled direction by the preference-conditioned curvature of the scalarized objective; on this whitened input, Adam's second moment collapses to identity, so the realized update is governed by the preference-conditioned metric.","pith_inferences":["The same preconditioning idea could be tested on other adaptive methods such as RMSProp or Adagrad to check whether similar preference-entanglement effects appear.","Dynamic preference vectors that change during training might become easier to track once the curvature term isolates the current preference from historical statistics.","If the curvature computation stays cheap, the wrapper pattern suggests a general route for making any adaptive optimizer respect an externally supplied scalarization metric."],"forward_implications":["Adam's second-moment denominator no longer marginalizes the time-varying preference into a history average.","The adaptive metric stops turning aligned objectives into apparent conflicts by respecting the Euclidean geometry assumed by the solver.","Distinct Pareto trade-offs remain separated instead of collapsing toward a uniform mixture.","MAdam improves performance over Adam for loss-balancing, gradient-balancing, and Pareto-based solver families.","Gains appear across multi-task learning, Pareto-front recovery, physics-informed neural networks, and medical imaging."],"fun_headline_variants":["MAdam preconditions reconciled direction by scalarized curvature","Whitening inputs collapses Adam second moment to identity in MAdam","MAdam resolves MOO weighting mismatch with preference curvature","Geometric and weighting gaps fixed by MAdam curvature preconditioning"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That the weighting and geometric mismatches are the dominant sources of sub-optimality and that preconditioning with the preference-conditioned curvature removes them without new side effects or prohibitive cost.","fun_headline_variants_meta":{"raw":{"variants":["MAdam preconditions reconciled direction by scalarized curvature","Whitening inputs collapses Adam second moment to identity in MAdam","MAdam resolves MOO weighting mismatch with preference curvature","Geometric and weighting gaps fixed by MAdam curvature preconditioning"]},"model":"grok-4.3","cost_usd":0.010576,"raw_usage":{"total_tokens":4663,"prompt_tokens":650,"num_sources_used":0,"completion_tokens":63,"cost_in_usd_ticks":105762000,"prompt_tokens_details":{"text_tokens":650,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3950,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":650,"tokens_out":63,"duration_ms":26128,"temperature":1.0,"reasoning_tokens":3950,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T11:19:25.730299+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Experiments on the same multi-task, Pareto, and physics-informed tasks where MAdam produces no consistent improvement or worse results than standard Adam would falsify the claim that the preconditioner resolves the identified mismatches.","supporting_citations":[],"review_version":1}