Adaptive Multi-Objective Loss Explained

• Adaptive Multi-Objective Loss is a framework that dynamically combines multiple objectives to address conflicts, scale differences, and gradient mismatches during training.
• It employs techniques like adaptive gradient combinations, dynamic scalarization, and feedback control to enhance multi-task learning, reinforcement learning, and physics-informed modeling.
• The approach aims to achieve Pareto optimality and common descent through real-time adjustments, ensuring balanced progress in complex optimization landscapes.

Adaptive multi-objective loss denotes a class of optimization constructions in which multiple training objectives are not collapsed once and for all into a fixed weighted sum, but are instead combined, normalized, scheduled, or otherwise mediated in a state-dependent manner during learning. Across multi-task learning, reinforcement learning, adversarial optimization, physics-informed learning, decision-focused learning, and generative alignment, the unifying motivation is that static scalarization often fails when objectives conflict, differ sharply in scale or variance, or induce different gradient geometries. The literature therefore replaces fixed coefficients with adaptive gradient combinations, adaptive scalarizations, dynamic multipliers, objective-conditioned surrogates, or optimizer-aware corrections, usually with Pareto optimality, common descent, or downstream decision quality as the underlying reference principle (Sener et al., 2018, Murillo-Gonzalez et al., 12 May 2026, Liu et al., 2 Jun 2026, Bischof et al., 2021).

1. Conceptual scope and basic formulations

In its most conventional form, multi-objective training begins from a vector of objective functions. For multi-task learning with shared parameters $\theta^{sh}$, task-specific parameters $\theta^1,\dots,\theta^T$, and empirical losses $\hat L^t(\theta^{sh},\theta^t)$, the standard surrogate is a weighted sum,

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$

where $c^t$ are fixed or heuristically chosen coefficients. The central criticism developed in the literature is that this surrogate only represents a principled compromise when tasks do not compete, whereas actual training often exhibits conflicting gradients, incompatible trade-offs, or strong heterogeneity across objectives (Sener et al., 2018).

A broader formulation writes the problem directly as vector optimization,

$$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$

with Pareto optimality replacing the search for one globally preferred scalar minimum. This view appears in gradient-based multi-task optimization, sparse deep multi-task learning, preference-conditioned policy optimization, and fractional-gradient methods, even though the operational mechanisms differ substantially (Sener et al., 2018, Hotegni et al., 2023, Murillo-Gonzalez et al., 12 May 2026, Shaw et al., 10 Jul 2025).

The term “adaptive” is not uniform across this literature. In some works it means recomputing task trade-offs online from current gradients, as in MGDA-style methods (Sener et al., 2018). In others it refers to adaptive scalarization geometry, such as dynamically modulating the smoothness of a Tchebycheff surrogate based on conflict signals (Murillo-Gonzalez et al., 12 May 2026). Elsewhere it denotes feedback control over multipliers in a scalar penalty loss (Sun et al., 2024), interval-based or RL-based reweighting of detector losses (Luo et al., 2021), stochastic online loss-shape updates (Raymond et al., 2023), or environment-conditioned decision-objective surrogates (Zhang et al., 2023). A recurring misconception is to treat all such methods as simple dynamic weight tuning. Several papers explicitly argue that the decisive object is not merely the coefficient vector but the optimization geometry induced by gradient combination, scalarization, or solver differentiation (Sener et al., 2018, Liu et al., 2 Jun 2026).

A useful organizing distinction is between methods that adapt weights, methods that adapt scalarization shape, methods that adapt optimizer geometry, and methods that adapt the loss function itself.

Family	Representative mechanism	Representative papers
Gradient-mediated	Min-norm common descent direction	(Sener et al., 2018, Shaw et al., 10 Jul 2025)
Scalarization-mediated	Modified Chebyshev, smooth Tchebycheff, weighted sum with adaptive priorities	(Hotegni et al., 2023, Murillo-Gonzalez et al., 12 May 2026, Chen et al., 10 Jan 2026)
Feedback/scheduling	PI-like multiplier control, interval weighting, RL action over weighting rules	(Sun et al., 2024, Luo et al., 2021)
Surrogate/objective learning	Learned decision-objective or learned loss network	(Zhang et al., 2023, Raymond et al., 2023)

This taxonomy suggests that “adaptive multi-objective loss” is best understood as a family of mechanisms for preserving or improving multi-objective trade-offs under nonstationary optimization conditions, rather than as one fixed algorithmic template.

2. Pareto descent and adaptive gradient combination

A major line of work defines adaptivity through the current geometry of task gradients. In “Multi-Task Learning as Multi-Objective Optimization,” multi-task training is recast as

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$

with Pareto optimality as the target notion of solution quality. The method uses the multiple gradient descent algorithm (MGDA), solving at each step

$$\min_{\alpha^1,\ldots,\alpha^T} \left\{ \left\| \sum_{t=1}^T \alpha^t \nabla_{\theta^{sh}}\hat L^t \right\|_2^2 \ \middle|\ \sum_t \alpha^t=1,\ \alpha^t\ge 0 \right\},$$

thereby finding the minimum-norm point in the convex hull of the task gradients. The coefficients $\alpha^t$ are recomputed online and are explicitly distinguished from fixed user-specified weights $c^t$: their role is to produce a common descent direction for all objectives, not to encode exogenous preferences (Sener et al., 2018).

Because exact MGDA requires one backward pass per task through shared parameters, the same work introduces MGDA-UB, an upper-bound surrogate operating in representation space $\theta^1,\dots,\theta^T$0,

$\theta^1,\dots,\theta^T$1

Under the assumption that $\theta^1,\dots,\theta^T$2 is full rank, the solution either yields Pareto stationarity or a direction that decreases all objectives. This made adaptive Pareto-consistent loss construction practical for encoder-decoder architectures and many-task settings such as CelebA with 40 attributes (Sener et al., 2018).

A formally related but mathematically distinct approach appears in the multi-objective adaptive-order Caputo fractional gradient descent method. There the common descent direction is obtained from the convex subproblem

$\theta^1,\dots,\theta^T$3

or equivalently its epigraph form. The KKT conditions imply

$\theta^1,\dots,\theta^T$4

so the effective coefficients $\theta^1,\dots,\theta^T$5 again emerge adaptively from a per-iteration optimization problem rather than from manual scalarization. Here, however, the underlying gradients are adaptive-order Caputo fractional gradients, and the method is analyzed in terms of weak efficiency, criticality, and convergence toward Tikhonov-regularized solutions in quadratic settings (Shaw et al., 10 Jul 2025).

The common principle across these methods is that fixed scalar weights are replaced by coefficients induced by local multi-objective geometry. This suggests a stronger form of adaptivity than heuristic reweighting: the loss combination is not chosen because one objective is “important” in the abstract, but because the current first-order structure admits a jointly improving direction.

3. Adaptive scalarization and changing optimization geometry

A second line of work retains scalarization but makes its form itself adaptive. In sparse deep multi-task learning, a modified Weighted Chebyshev scalarization is used for joint optimization of task losses and a sparsity objective,

$\theta^1,\dots,\theta^T$6

subject to

$\theta^1,\dots,\theta^T$7

The importance vector $\theta^1,\dots,\theta^T$8 is externally chosen, not learned online, so the method is “adaptive” primarily across runs: by varying $\theta^1,\dots,\theta^T$9, one explores properly Pareto-optimal trade-offs between prediction performance and sparsity. The paper emphasizes that this differs fundamentally from a linear weighted sum because the modified Chebyshev form targets the worst weighted deviation from the reference point and, according to the cited theorem, can recover all properly Pareto-optimal solutions under non-convexity (Hotegni et al., 2023).

In “Adaptive Smooth Tchebycheff Attention for Multi-Objective Policy Optimization,” the scalarization parameter rather than the preference vector is adapted online. The smooth Tchebycheff objective is

$\hat L^t(\theta^{sh},\theta^t)$0

where $\hat L^t(\theta^{sh},\theta^t)$1 controls curvature. Small $\hat L^t(\theta^{sh},\theta^t)$2 approaches hard Tchebycheff, giving access to non-convex Pareto regions; large $\hat L^t(\theta^{sh},\theta^t)$3 approaches linear weighted sums, improving stability. The method computes objective-wise PPO gradients, estimates a conflict ratio

$\hat L^t(\theta^{sh},\theta^t)$4

and uses this signal to regulate $\hat L^t(\theta^{sh},\theta^t)$5 through a decay schedule, a conflict-dependent boost, and an EMA controller. The same scalarization induces an exact soft attention

$\hat L^t(\theta^{sh},\theta^t)$6

which is then mixed with a maintenance floor to prevent objective starvation (Murillo-Gonzalez et al., 12 May 2026).

APEX, for text-to-image alignment, again keeps scalarization but argues that fixed weights fail because of “variance hijacking” and gradient conflicts. It therefore separates normalization from prioritization. Dual-Stage Adaptive Normalization first standardizes each reward per group,

$\hat L^t(\theta^{sh},\theta^t)$7

then aggregates with adaptive weights and renormalizes: $\hat L^t(\theta^{sh},\theta^t)$8 The weights come from $\hat L^t(\theta^{sh},\theta^t)$9 Adaptive Priorities,

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$0

The three factors are respectively derived from gradient norms, negative cosine conflicts, and distance to empirical upper bounds. This is adaptive scalarization in a strong sense: not only do weights vary, but reward normalization is structured so that scale heterogeneity does not corrupt the semantic meaning of priorities (Chen et al., 10 Jan 2026).

These works show that adaptive multi-objective loss need not mean abandoning scalarization. An alternative is to preserve scalarization while making its geometry, normalization, or emphasis state-dependent.

4. Feedback control, stochastic scheduling, and loss-balancing heuristics

Other methods formulate adaptive multi-objective loss as a feedback-control or scheduling problem rather than a Pareto descent or scalarization-geometry problem. In M-HOF-Opt, the training loss has the structural-risk form

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$1

but the multiplier vector $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$2 is updated online from observed losses rather than selected by outer-loop hyperparameter search. The high-level controller maintains a shrinking setpoint $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$3 for the regularization vector, while the low-level multiplier controller uses the tracking error

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$4

to update multipliers through a PI-like exponential rule,

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$5

The resulting interpretation is “multiplier induced loss landscape scheduling”: the scalar penalty loss remains a weighted sum, but the weights become closed-loop control inputs that reshape the optimization landscape over epochs (Sun et al., 2024).

In multi-scale detector training, the same scheduling intuition appears in a purely statistical form. The total loss is

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$6

but the weights $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$7 over pyramid levels are updated online. Adaptive Variance Weighting computes interval losses $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$8, interval variances $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T}\quad \sum_{t=1}^T c^t \hat L^t(\theta^{sh},\theta^t),$$9, and variance reduction rates

$c^t$0

then upweights the two scales with maximal $c^t$1. Reinforcement Learning Optimization generalizes this by choosing among four weighting actions—AVW, favoring small-loss scales, favoring large-loss scales, or no enhancement—with action probabilities updated from whether the total interval loss decreased: $c^t$2 The paper interprets this as phase-dependent weighting: different heuristics are useful at different stages of detector training (Luo et al., 2021).

Physics-informed learning yields another variant. PINNs combine PDE residuals, boundary conditions, initial conditions, and sometimes data terms in a scalarized objective

$c^t$3

ReLoBRaLo adapts $c^t$4 using relative progress rather than raw magnitude or gradient norms. Its balancing core is

$c^t$5

augmented with exponential history and a Bernoulli random lookback to the initialization losses. The method is positioned as loss-based rather than gradient-based balancing, with lower overhead than GradNorm or learning-rate annealing while still targeting fairness of relative progress across PINN objectives (Bischof et al., 2021).

A closely related adversarial-optimization example is UniAda, where one universal perturbation $c^t$6 is optimized against steering and acceleration losses,

$c^t$7

Its Adaptive Weighting Scheme computes weighted gradient norms $c^t$8, normalized losses $c^t$9, relative inverse training rates $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$0, and mismatch losses

$$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$1

then updates the objective weights $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$2 by gradient descent on $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$3, followed by positivity enforcement and normalization. This is a direct transfer of training-rate balancing to shared-parameter adversarial optimization (Zhang et al., 25 Apr 2026).

These methods share an important feature: they do not necessarily seek exact Pareto stationarity, yet they still treat imbalance as a dynamical systems problem. The loss weights become controlled variables whose trajectories matter as much as their instantaneous values.

5. Decision-focused, policy-conditioned, and optimizer-aware extensions

Adaptive multi-objective loss has also expanded beyond classical multi-task learning into end-to-end decision pipelines, policy optimization, and optimizer corrections. In multi-objective decision-focused learning, the issue is not merely how to weight losses but how to align prediction with the geometry of a downstream Pareto problem. MoDFL therefore combines three losses: $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$4 where $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$5 is a landscape loss in objective space based on sRMMD, $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$6 is the distance from a predicted representative decision to the true Pareto set, and $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$7 is a decision loss evaluating a scalarized representative Pareto point under true coefficients. This design is “adaptive” in the sense that the loss family is matched to objective-space geometry, solution-space geometry, and end-task decision quality simultaneously, although the coefficients $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$8 are fixed hyperparameters rather than online variables (Li et al., 2024).

In forecast-then-optimize power systems, Adaptive Decision-Objective Loss adopts a different route: it learns a surrogate $$\min_{x \in X} F(x)=\min_{x\in X}[f_1(x),\ldots,f_m(x)]^T,$$9 for the final downstream objective value across forecast errors and changing environment parameters $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$0, then trains the predictor by

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$1

The paper’s explicit rationale is that in a multi-stage setting, minimizing a decision loss relative to the “perfect-forecast” decision can fail to achieve global final optimality. ADOL therefore learns the objective utility itself rather than a deviation from a nominal target decision, yielding a context-conditioned scalar surrogate for a composite downstream decision process (Zhang et al., 2023).

Policy optimization provides yet another notion of adaptivity. Preference-conditioned MORL methods such as MOPPO and MOA2C sample a scalarization vector $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$2, compute vector returns and vector advantages, and only then scalarize: $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$3 With PopArt normalization, this becomes

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$4

The objective is adaptive because the loss changes with sampled preferences, while entropy regularization is itself adapted by an MDMM-style target-tracking rule in MOPPO (Terekhov et al., 2024).

A more radical extension appears in “Online Loss Function Learning,” where the loss function $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$5 is updated after every base-model update: $$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$6

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$7

Here adaptivity no longer concerns coefficients on known objectives; the learned loss itself changes online in response to training dynamics. Although this is not explicit Pareto balancing, it can be interpreted as adaptive objective composition across time, optimization regime, and train-versus-validation performance (Raymond et al., 2023).

Finally, MAdam shifts attention from the loss to the optimizer. If a multi-objective solver produces a reconciled direction

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$8

standard Adam can create a weighting mismatch and a geometric mismatch because its second-moment denominator marginalizes time-varying preferences and distorts Euclidean trade-off geometry. MAdam therefore estimates the preference-conditioned curvature

$$\min_{\theta^{sh},\theta^1,\ldots,\theta^T} \big(\hat L^1(\theta^{sh},\theta^1),\ldots,\hat L^T(\theta^{sh},\theta^T)\big)^\top,$$9

constructs the metric

$$\min_{\alpha^1,\ldots,\alpha^T} \left\{ \left\| \sum_{t=1}^T \alpha^t \nabla_{\theta^{sh}}\hat L^t \right\|_2^2 \ \middle|\ \sum_t \alpha^t=1,\ \alpha^t\ge 0 \right\},$$0

and whitens the direction before Adam: $$\min_{\alpha^1,\ldots,\alpha^T} \left\{ \left\| \sum_{t=1}^T \alpha^t \nabla_{\theta^{sh}}\hat L^t \right\|_2^2 \ \middle|\ \sum_t \alpha^t=1,\ \alpha^t\ge 0 \right\},$$1 This is adaptive multi-objective loss in a solver–optimizer sense: the scalarized objective may be unchanged, but the executed update is corrected so that Adam no longer overrides the intended multi-objective trade-off (Liu et al., 2 Jun 2026).

6. Limitations, misconceptions, and open directions

A recurring limitation is that many methods remove one class of manual choices only to introduce another. MGDA-UB avoids manual scalarization but assumes an encoder–decoder structure and, for its strongest guarantee, full rank of $$\min_{\alpha^1,\ldots,\alpha^T} \left\{ \left\| \sum_{t=1}^T \alpha^t \nabla_{\theta^{sh}}\hat L^t \right\|_2^2 \ \middle|\ \sum_t \alpha^t=1,\ \alpha^t\ge 0 \right\},$$2 (Sener et al., 2018). Modified Weighted Chebyshev sparse multi-task learning recovers richer Pareto trade-offs than linear sums, but it still requires choosing the preference vector $$\min_{\alpha^1,\ldots,\alpha^T} \left\{ \left\| \sum_{t=1}^T \alpha^t \nabla_{\theta^{sh}}\hat L^t \right\|_2^2 \ \middle|\ \sum_t \alpha^t=1,\ \alpha^t\ge 0 \right\},$$3, the reference point $$\min_{\alpha^1,\ldots,\alpha^T} \left\{ \left\| \sum_{t=1}^T \alpha^t \nabla_{\theta^{sh}}\hat L^t \right\|_2^2 \ \middle|\ \sum_t \alpha^t=1,\ \alpha^t\ge 0 \right\},$$4, and several sparsification hyperparameters externally (Hotegni et al., 2023). PASTA adapts scalarization curvature rather than preference weights, yet depends on a conflict threshold, a maintenance rate, and a decay schedule whose interaction remains an acknowledged limitation (Murillo-Gonzalez et al., 12 May 2026).

Another misconception is that any adaptive weighting scheme automatically solves multi-objective optimization. Several papers explicitly distinguish between dynamic weighting and genuine Pareto methods. A3M, for example, is adaptive in policy learning and opponent modeling, but its multi-objective reward remains a fixed weighted sum with static $$\min_{\alpha^1,\ldots,\alpha^T} \left\{ \left\| \sum_{t=1}^T \alpha^t \nabla_{\theta^{sh}}\hat L^t \right\|_2^2 \ \middle|\ \sum_t \alpha^t=1,\ \alpha^t\ge 0 \right\},$$5; the paper states that it does not learn or dynamically update these weights online (Li et al., 27 Jun 2026). Similarly, MOODY uses time-varying scalarized rewards driven by changing private preference vectors and adaptive few-shot retraining, but it remains a scalarization-based RL framework rather than an online Pareto optimizer (Tan et al., 2024).

Scalability is a persistent fault line. Gradient-based balancing methods may require multiple backward passes or pairwise gradient statistics. This is why MGDA-UB, ReLoBRaLo, APEX’s micro-batch gradient estimates, and M-HOF-Opt’s loss-only feedback rules all emphasize computational tractability (Sener et al., 2018, Bischof et al., 2021, Chen et al., 10 Jan 2026, Sun et al., 2024). By contrast, some decision-focused methods require Pareto-set approximations or differentiable solver surrogates, which can be expensive or unavailable in larger domains (Li et al., 2024).

A plausible synthesis suggested by these works is that future adaptive multi-objective loss design will increasingly combine four ingredients: calibrated per-objective normalization, geometry-aware conflict sensing, explicit preference or context conditioning, and optimizer-aware execution. This suggestion is directly motivated by the convergence of ideas across DSAN and $$\min_{\alpha^1,\ldots,\alpha^T} \left\{ \left\| \sum_{t=1}^T \alpha^t \nabla_{\theta^{sh}}\hat L^t \right\|_2^2 \ \middle|\ \sum_t \alpha^t=1,\ \alpha^t\ge 0 \right\},$$6 priorities (Chen et al., 10 Jan 2026), smooth Tchebycheff curvature control (Murillo-Gonzalez et al., 12 May 2026), preference-conditioned critic and actor losses (Terekhov et al., 2024), and metric-aware optimizer correction (Liu et al., 2 Jun 2026). At the same time, the literature consistently cautions that adaptive mechanisms remain local and gradient-based, usually target Pareto stationarity rather than global front recovery, and often leave preference elicitation unresolved (Sener et al., 2018, Hotegni et al., 2023, Shaw et al., 10 Jul 2025).

Taken together, the field defines adaptive multi-objective loss not as one algorithm, but as a design problem: how to ensure that multiple competing objectives retain meaningful influence throughout learning despite non-convexity, scale heterogeneity, delayed effects, and optimizer-induced distortions. The answer has evolved from static coefficient tuning toward dynamic, geometry-aware, and task-structured constructions, but no single mechanism has yet eliminated the underlying tension between expressivity, stability, computational cost, and controllable trade-offs.