Direct Metric Optimization: Methods and Insights

• Direct Metric Optimization (DMO) is a design principle that aligns model training with task-specific metrics, addressing the mismatch between conventional losses and final evaluation criteria.
• DMO techniques include exact reformulation, surrogate loss learning, and metric-derived preference construction to optimize metrics directly in domains like imbalanced classification and recommendation.
• Key challenges in DMO involve handling non-differentiability, conflicting gradients in multi-metric settings, and domain-specific constraints, which necessitate tailored optimization strategies.

Direct Metric Optimization (DMO) denotes a family of optimization strategies in which the training or search signal is derived from the task metric itself, or from an exact reformulation, differentiable surrogate, metric-induced preference relation, or jointly learned surrogate of that metric, rather than from a generic proxy such as cross-entropy, squared error, or an unstructured reward mixture. Across binary imbalanced classification, ranking-based recommendation, text-to-image alignment, dense video captioning, zero-shot speech synthesis, scientific simulation, geometry processing, and control, the common motivation is a mismatch between what models are optimized for during training and what they are ultimately evaluated on (Peng et al., 21 Jul 2025, Li et al., 2021, Tamboli et al., 16 Mar 2025, Lee et al., 12 Jan 2025, Li et al., 2024, Gressmann et al., 22 Mar 2026).

1. Conceptual scope and recurring technical obstacles

In the literature, DMO is not a single algorithm but a design principle: optimize the quantity that matters operationally. The target may be a classification metric such as precision, recall, or $F_\beta$; an IR metric such as RR, AP, $nDCG$, or RBP; a perceptual or alignment score such as CTC, speaker verification similarity, CLIP, HPS, PickScore, or Aesthetic; or an expensive scientific objective and feasibility profile returned by a simulator. This suggests a unifying interpretation of DMO as a response to objective mismatch rather than as a domain-specific recipe.

The main technical obstacle depends on the domain. In binary imbalanced classification, the difficulty comes from the indicator $1\{f_\theta(x_i)>t\}$, which makes precision, recall, and $F_\beta$ discontinuous and ties optimization to both model parameters $\theta$ and threshold $t$. In multi-metric generative alignment, different metrics may live on incompatible numerical scales and may produce conflicting gradients. In scientific optimization, the simulator may be expensive, non-differentiable, and constrained by a thin feasible/infeasible partition. In output-feedback control, direct policy optimization is hindered by coordinate redundancies that create degenerate stationary points. These obstacles motivate different DMO realizations: exact reformulation, preference construction, differentiable end-to-end metrics, surrogate gradients, or geometry-aware optimization (Peng et al., 21 Jul 2025, Tamboli et al., 16 Mar 2025, Gressmann et al., 22 Mar 2026, Kraisler et al., 2024).

2. Exact reformulation for discrete classification metrics

A particularly explicit formulation of DMO appears in binary imbalanced classification, where the aim is to optimize the metric of interest directly rather than optimize a surrogate and hope the target metric improves. The paper "Exact Reformulation and Optimization for Direct Metric Optimization in Binary Imbalanced Classification" studies three settings: fixed precision optimize recall (FPOR), fixed recall optimize precision (FROP), and optimize $F_\beta$-score (OFBS). With positive set $P$, negative set $N$, $N_+=|P|$, classifier $nDCG$0, and threshold $nDCG$1, it defines

$nDCG$2

$nDCG$3

$nDCG$4

The central contribution is an Exact Reformulation and Optimization (ERO) framework. It introduces auxiliary variables $nDCG$5, lifts the metric optimization problem into a constrained problem over $nDCG$6, and replaces the indicator equality by an exact piecewise-linear continuous constraint: $nDCG$7 for $nDCG$8. With

$nDCG$9

the lifted problem becomes amenable to first-order optimization, and the paper further converts the equality into exact inequalities. The resulting constrained program is optimized with an exact $1\{f_\theta(x_i)>t\}$0-type penalty method: $1\{f_\theta(x_i)>t\}$1 with multiplicative penalty updates $1\{f_\theta(x_i)>t\}$2, projected ADAM on the box constraints $1\{f_\theta(x_i)>t\}$3, $1\{f_\theta(x_i)>t\}$4, and a logit regularizer

$1\{f_\theta(x_i)>t\}$5

to avoid singular points $1\{f_\theta(x_i)>t\}$6. The paper proves exactness under a non-singularity condition, reports that ERO consistently gives the best optimization performance across all 8 constrained FPOR/FROP tasks, and shows that replacing exact reformulation by sigmoid smoothing hurts both feasibility and objective performance on ADE-v2 (Peng et al., 21 Jul 2025).

3. Ranking-based recommendation and the train–evaluation metric gap

In ranking-based recommendation, DMO has been studied as direct optimization of user-oriented IR metrics rather than generic losses. The paper "New Insights into Metric Optimization for Ranking-based Recommendation" compares RR, AP, $1\{f_\theta(x_i)>t\}$7, and RBP in both pairwise and listwise learning-to-rank settings, using LambdaRank-style pairwise optimization and a listwise MF model. For user $1\{f_\theta(x_i)>t\}$8, with relevance labels $1\{f_\theta(x_i)>t\}$9 and rank $F_\beta$0, it defines

$F_\beta$1

and normalized RBP

$F_\beta$2

The pairwise RBP-inspired update follows a LambdaRank construction: $F_\beta$3 while the listwise RBP-inspired objective simplifies to

$F_\beta$4

A notable empirical conclusion is that optimizing the same metric used for evaluation is not necessarily best. Optimizing RR is consistently worst, whereas AP-, $F_\beta$5-, and especially RBP-inspired objectives are more reliable; in the pairwise setting, $F_\beta$6 performs best overall, and in the listwise setting the proposed $F_\beta$7 loss is competitive across metrics and is independent of $F_\beta$8 during training. The thesis "Metric Optimization and Mainstream Bias Mitigation in Recommender Systems" sharpens the interpretation: the superiority of RBP-inspired optimization mainly benefits active users, so DMO can improve overall ranking accuracy while still amplifying a mainstreamness-related skew in who gains the most (Li et al., 2021, Li, 2023).

4. Generative-model alignment through metric-derived preferences and differentiable metric losses

In modern generative modeling, DMO frequently appears as metric-aligned preference learning or as end-to-end optimization of differentiable perceptual criteria. In dense video captioning, VidChain introduces Metric-based Direct Preference Optimization (M-DPO), which replaces human preference labels by evaluation-metric comparisons at each step of a Chain-of-Tasks decomposition. For sub-task $F_\beta$9, candidates $\theta$0 are sampled from a reference model, scored by a task metric $\theta$1, and converted to pairwise preferences by

$\theta$2

The final objective keeps only pairs whose metric gap exceeds $\theta$3: $\theta$4 This metric-derived preference construction improves both DVC and temporal video grounding, and the ablation reported for VideoLLaMA2 + CoTasks shows a progression from baseline $\theta$5 to $\theta$6 with ordinary DPO, to $\theta$7 with basic M-DPO, and to $\theta$8 SODA$\theta$9 with preference-gap-aware M-DPO (Lee et al., 12 Jan 2025).

BalancedDPO applies a related idea to text-to-image diffusion, but with multiple metrics. Instead of mixing raw rewards, it converts each metric-specific scorer into a binary vote

$t$0

then aggregates by majority vote,

$t$1

and runs a single DPO-style update on the consensus label. The paper emphasizes that this operates in preference distribution space rather than reward space, avoids normalization and weight-tuning across incompatible reward scales, and keeps gradient-update complexity at $t$2 while preference collection is $t$3. It also periodically refreshes the reference model $t$4. On Pick-a-Pic, PartiPrompt, and HPD, BalancedDPO improves average win rates by $t$5, $t$6, and $t$7, respectively, over DiffusionDPO (Tamboli et al., 16 Mar 2025).

DMOSpeech realizes a third variant: direct end-to-end optimization of differentiable metrics in zero-shot speech synthesis. After distilling a latent diffusion teacher into a 4-step student generator $t$8, it optimizes the generator with distillation, adversarial, CTC, and speaker verification losses: $t$9 The CTC term directly targets intelligibility,

$F_\beta$0

and the SV term directly targets speaker similarity through cosine similarity in embedding space. The resulting model uses only 4 sampling steps, reports $F_\beta$1 versus $F_\beta$2 for the $F_\beta$3 teacher, and reaches $F_\beta$4 and $F_\beta$5, with large ablation gains from adding CTC and SV losses (Li et al., 2024).

5. Surrogate, geometric, and structured-space formulations

In expensive scientific optimization, DMO often means learning a differentiable representation of the objectives and constraints themselves. DMOSOPT introduces a joint surrogate

$F_\beta$6

with a shared backbone, a regression head for objectives, and a classification head for binary constraints. The surrogate is used to produce a unified optimization gradient that pushes candidates toward better objective values and higher feasibility, and its partial derivatives define parameter sensitivities

$F_\beta$7

These sensitivities are then mapped into NSGA-II distribution indices. The framework is retrained iteratively as new expensive simulator evaluations arrive, supports optional gradient-based feasibility correction, and on the large CA1 network benchmark reaches comparable quality to the MEGP baseline with about $F_\beta$8 less compute and to the GPR baseline with about $F_\beta$9 less compute, while requiring about $P$0 fewer epochs than MEGP and $P$1 fewer than GPR to reach equivalent hypervolume (Gressmann et al., 22 Mar 2026).

A geometrically different DMO appears in continuous-time LQG output-feedback control. There the optimization target is the LQG performance index $P$2, but the parameterization of dynamic controllers contains internal-coordinate symmetries. The paper constructs a coordinate-invariant Krishnaprasad–Martin Riemannian metric on the manifold of minimal stabilizing controllers, quotients out the $P$3 action induced by controller-state changes, and performs Riemannian gradient descent on the quotient manifold. The quotient removes exactly $P$4 redundant directions for full-order controllers, and under the stated Hessian nondegeneracy assumption the method has a local linear convergence guarantee. This is DMO in the sense of direct policy optimization with a metric matched to the intrinsic orbit geometry of controller realizations (Kraisler et al., 2024).

In geometry processing, the paper "Metric Optimization in Penner Coordinates" reformulates metric optimization on meshes by changing variables from fixed-connectivity edge lengths to Penner coordinates on the space $P$5 of cone metrics. The key global statement is that logarithmic Penner coordinates yield a bijection

$P$6

so the optimization is unconstrained in a Euclidean coordinate system even when connectivity changes are needed. Angle constraints define a manifold parameterized by shear coordinates, and the paper studies objectives including log-length distortion, area distortion via best-fit scale factors, and a quadratic symmetric-Dirichlet-like energy. This formulation supports direct optimization over the full metric space rather than within a single triangulation chart (Capouellez et al., 2022).

A narrower usage occurs in approximate Frank–Wolfe on graph-structured support sets, where DMO denotes the support-selection oracle inside the outer optimization loop. Given a direction $P$7, the method chooses a graph-consistent support $P$8 of size $P$9, forms a sparse atom supported on $N$0, and updates the iterate by a Frank–Wolfe step. The proposed “Top-$N$1 + optimal visiting” heuristic starts from the top $N$2 coordinates, expands connected components by random edge visits, and scores candidate supports by $N$3. Here DMO is not a global learning principle but a graph-aware heuristic oracle, and the reported experiments conclude that the new DMO did not improve over the original heuristic (Pan et al., 2024).

6. Limitations, misconceptions, and acronym ambiguity

Several recurrent misconceptions are explicitly challenged in the literature. First, optimizing the evaluation metric itself is not always the best choice: in recommendation, same-metric train/test matching fails clearly for RR, while RBP-inspired losses are often stronger optimization targets even when evaluation uses AP or $N$4 (Li et al., 2021). Second, smooth approximations are not necessarily adequate for discontinuous metrics: in binary imbalanced classification, sigmoid-style smoothing can produce numerical mismatch, infeasible constraints, and vanishing gradients when sharpened, which motivates exact reformulation instead (Peng et al., 21 Jul 2025). Third, direct reward mixing is not equivalent to robust multi-metric alignment: BalancedDPO argues that reward rescaling issues and conflicting gradients make label aggregation in preference space preferable to weighted reward sums (Tamboli et al., 16 Mar 2025).

The limitations are correspondingly heterogeneous. ERO is currently specialized to binary classification, assumes non-singularity $N$5, and uses a deterministic finite-sum solver that may not scale well to very large datasets. RBP-based DMO can improve overall recommendation quality while concentrating larger gains on more active users, which exposes a trade-off between aggregate metric optimization and user-level distribution of benefit. Surrogate-based DMO can overfit early data, oversmooth sharp feasibility boundaries, scale poorly in hypervolume computation when the number of objectives is large, and lacks the calibrated uncertainty estimates typical of GP-based Bayesian optimization (Peng et al., 21 Jul 2025, Li, 2023, Gressmann et al., 22 Mar 2026).

The acronym itself is also ambiguous. In robotics, DMO may denote Deep Motion Optimizer, a learned recurrent refinement operator paired with an image-conditioned EBM for long-horizon motion prediction, rather than Direct Metric Optimization (Oba et al., 2023). In elliptic and parabolic PDE theory, “DMO coefficients” denotes Dini mean oscillation coefficients, as in the boundary Schauder and interface regularity results of "Schauder type estimates for degenerate or singular elliptic equations with DMO coefficients" and "Higher order parabolic systems with piecewise DMO and Hölder continuous coefficients"; these uses are terminologically unrelated to metric-directed training or search (Dong et al., 2023, Dong et al., 2024).

Taken together, these works show that Direct Metric Optimization is best understood as a broad optimization doctrine: model fitting or search is organized around the actual metric, constraint, or decision criterion of interest, with the technical mechanism chosen to match the structure of that criterion. The mechanism may be exact reformulation, pairwise preference induction, end-to-end differentiable loss design, joint surrogate learning, geometry-aware gradient descent, or a structured oracle, but the defining feature remains the same—alignment of optimization with the metric that governs final judgment.