Dynamic gain neuromodulation attenuates stability gaps during continual joint training

arXiv:2507.14056 · cs.LG, cs.AI, q-bio.NC · Submitted 2025-07-18 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Dynamic gain neuromodulation attenuates stability gaps during continual joint training".

Jane: The gist Dynamic gain neuromodulation attenuates stability gaps under joint training by introducing an uncertainty-modulated gain mechanism that transiently increases neuronal gain to reshape optimization trajectories,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Moving on to the title and authors of this paper, we have "Dynamic gain neuromodulation attenuates stability gaps during continual joint training" by Alejandro Rodriguez-Garciaa, Anindya Ghosha, and Srikanth Ramaswamya.

Jane: That sounds like a very specific mechanism is being proposed here <ref:2507.14056#pg1>. It’s not just another optimizer tweak; it's introducing this dynamic gain neuromodulation element into the learning process itself.

Lu: The authors are building on work that looks at stability gaps, which they view as a manifestation of the stability-plasticity dilemma <ref:2507.14056#pg2>. They want to show how intrinsic neural dynamics can solve this dilemma in artificial networks.

Meng: So what's the simple version of what they are trying to say here? Are they saying we need a biological brain inside our AI models to handle this forgetting problem?

Lalam: They are suggesting that the stability gap, which is seen as a temporary performance drop on old tasks when new ones are introduced <ref:2507.14056#pg2>, can be attenuated by using gain neuromodulation <ref:2507.14056#pg2>.

Tom: That’s the core idea, so they are saying that transient changes in neuronal responsiveness can adjust how strongly inputs influence neural activity <ref:2507.14056#pg2>.

Jane: So they are looking at how those transient changes in responsiveness can help balance plasticity and stability during continual learning <ref:2507.14056#pg2>.

Lu: They are drawing on the idea that in biological neural circuits, gain modulation arises through spiking and population-level dynamics <ref:2507.14056#pg3>. They are applying this biologically inspired idea to learning dynamics in artificial networks <ref:2507.14056#pg3>.

Meng: So they are translating a biological phenomenon into a mathematical framework for our AI systems, which is the tricky part. I wonder if that translation loses too much of the biological intuition when you start with matrices and tensors.

Lalam: They are asking how uncertainty-driven gain modulation can shape optimization trajectories during training <ref:2507.14056#pg3>. It's about using uncertainty to modulate the learning path itself <ref:2507.14056#pg3>.

Tom: They are looking at how uncertainty-driven gain modulation can shape optimization trajectories during training <ref:2507.14056#pg3>. It’s using the uncertainty to guide the optimization path dynamically <ref:2507.14056#pg3>.

Jane: So, it's about making the learning trajectory responsive to what's uncertain in a way that helps with both adapting quickly and remembering what we already know <ref:2507.14056#pg3>.

Lu: They are asking how uncertainty-driven gain modulation can shape optimization trajectories during training <ref:2507.14056#pg3>. It’s about finding a way to use uncertainty to regulate the learning process <ref:2507.14056#pg3>.

Meng: If it works, it means we can design systems that are inherently more flexible in how they handle new information without forgetting the old stuff. That's what I’m hoping for when I think about production systems.

Lalam: They are suggesting that uncertainty controls both timescales and parameterization rather than acting solely as a step-size signal <ref:2507.14056#pg3>. It's about using the uncertainty to regulate the learning process in a fundamental way <ref:2507.14056#pg3>.

Tom: So, they are suggesting that uncertainty controls both timescales and parameterization rather than acting solely as a step-size signal <ref:2507.14056#pg3>. It’s about controlling the learning process through more complex signals <ref:2507.14056#pg3>.

Title and authors: Jane: So, they are looking at how uncertainty-driven gain modulation can shape optimization trajectories during training <ref:2507.14056#pg3>. It’s about finding a way to use uncertainty to regulate the learning process in a fundamental way <ref:2507.14056#pg3>.

Lu: It’s about using the uncertainty to regulate the learning process through more complex signals <ref:2507.14056#pg3>.

Lalam: The paper shows that NGM-SGD empirically attenuates stability gaps across domain and class incremental benchmarks on MNIST, CIFAR, and miniImageNet under task agnostic joint training while maintaining competitive accuracy <ref:2507.14056#pg1>. It outperforms MSGD Adam and vanilla SGD in transition stability <ref:2507.14056#pg1>.

Tom: The empirical results show that NGM-SGD consistently reduces the stability gap, although the performance gains are less pronounced than hypothesized in prior work <ref:2507.14056#pg3>. Table two reports per-transition stability gaps for all joint training benchmarks <ref:2507.14056#pg3>.

Jane: So while it reduces the gap, the largest avg-SG signals that the model still experiences significant forgetting immediately after each context switch, which underscores its lack of robustness in continual learning settings <ref:2507.14056#pg3>.

Lu: Table six shows experience replay results across all benchmarks confirming that NGM SGD continues to attenuate the stability gap even in a more realistic CL scenario where replay partially mitigates catastrophic forgetting <ref:2507.14056#pg1>. The reset variants of MSGD and Adam exhibit mixed behavior across benchmarks, sometimes reducing stability gaps and sometimes increasing them <ref:2507.14056#pg1>.

Tom: So it seems they found the best balance between minimizing forgetting and maximizing accuracy in this setup, especially when you look at Table three and Table four for hyperparameter searches on Split CIFAR-ten <ref:2507.14056#pg3>.

Tom: Figure 3e shows that NGMSGD produces a substantially smaller post-transition curvature amplification than MSGD and Adam and maintains a lower curvature ratio throughout the analyzed window <ref:2507.14056#pg3>. This suggests the gain reparameterization produces a locally flatter loss landscape in effective-weight coordinates, thereby reducing the sensitivity of previous-task loss to parameter displacements induced by the task transition <ref:2507.14056#pg3>.

Tom: They are suggesting that uncertainty controls both timescales and parameterization, which is a broader design principle for adaptive learning systems where uncertainty controls both timescales and parameterization rather than acting solely as a step-size signal <ref:2507.14056#pg3>. It’s about controlling the learning process through more complex signals <ref:2507.14056#pg3>.

Jane: That's the big implication here, Tom; it moves us from just thinking about step sizes to thinking about how the entire learning process is shaped by uncertainty in multiple ways <ref:2507.14056#pg3>.

Lu: They also noted that NGM SGD continues to attenuate the stability gap even when replay is used, confirming that the benefits of the gain dynamics persist beyond the idealized joint training setup <ref:2507.14056#pg3>. The authors suggest that cholinergic neuromodulation could support task segregation through gain-gating mechanisms <ref:2507.14056#pg3>.

Jane: So, to wrap up this paper on Dynamic gain neuromodulation attenuates stability gaps during continual joint training, it shows how transiently increasing neuronal gain can reshape optimization trajectories by inducing fast and slow weight adaptations <ref:2507.14056#pg3>. It proves that uncertainty controls both the timescale and the parameterization of learning <ref:2507.14056#pg3>.

The paper's summary: Tom: So we've looked at the summary of the paper on "Dynamic gain neuromodulation attenuates stability gaps during continual joint training," and it seems the main takeaway is that they are addressing that stability gap as an early-adaptation manifestation of the stability-plasticity dilemma <ref:2507.14056#pg2>.

Jane: Right, so the authors see it as a mismatch between rapid adaptation and strong retention at task boundaries <ref:2507.14056#pg2>, which is why they need an optimization mechanism to balance plasticity and stability over abrupt distribution changes <ref:2507.14056#pg1>.

Lu: The authors propose a dynamic gain scaling mechanism as a two-timescale optimization technique that balances adaptation and retention by transiently increasing the effective update magnitude while dynamically reparameterizing the weights governing the forward pass <ref:2507.14056#pg1>.

Meng: So in simple terms, they are saying they are adding this scaling mechanism to transiently boost how big an update is while also changing how that update is expressed in the weight space <ref:2507.14056#pg1>.

Lalam: They are essentially hypothesizing that these transient gain boosts induce fast and slow timescales of weight adaptation, which facilitates knowledge acquisition while simultaneously driving a reparametrization that generates fast yet stable transitions between tasks <ref:2507.14056#pg3>.

Tom: So they are proposing this mechanism as a way to achieve that balance in the learning process during continual training <ref:2507.14056#pg1>.

Jane: It’s about finding a dynamic gain scaling mechanism that achieves this balance between plasticity and stability, which is what they are aiming for <ref:2507.14056#pg1>.

Lu: They are using biological inspiration from neuromodulatory bursts to guide how we structure the optimization trajectories during training <ref:2507.14056#pg3>.

Meng: So they are trying to use a mechanism that mimics biological systems to solve an AI problem, which is where the real research value lies for me.

Lalam: They are asking how uncertainty-driven gain modulation can shape optimization trajectories during training <ref:2507.14056#pg3>. It’s about using the uncertainty to modulate the learning path itself <ref:2507.14056#pg3>.

Tom: They are asking how uncertainty-driven gain modulation can shape optimization trajectories during training <ref:2507.14056#pg3>. It’s using the uncertainty to modulate the learning path itself <ref:2507.14056#pg3>.

Jane: So they are looking at how uncertainty-driven gain modulation can shape optimization trajectories during training <ref:2507.14056#pg3>. It’s about using the uncertainty to regulate the learning path itself <ref:2507.14056#pg3>.

The paper's improvements: Tom: Now for the suggested improvements in this paper on "Dynamic gain neuromodulation attenuates stability gaps during continual joint training," the authors are suggesting a few ways to make it even better.

Jane: They are proposing that we use this mechanism to achieve effective-coordinate flattening by scaling eigenvalues of the Hessians by a factor of one over g squared <ref:2507.14056#pg4>. This should reduce transition-induced curvature amplification <ref:2507.14056#pg3>.

Lu: That's interesting because they are linking the gain modulation directly to the Hessian's eigenvalues, which is a specific mathematical way to quantify how much the loss landscape is being flattened <ref:2507.14056#pg4>. This ensures that the effective step size alphaeff t equals alpha g squared t for the gradient-driven displacement in effective-weight coordinates <ref:2507.14056#pg4>.

Meng: If we can control those eigenvalues, it means we have a more direct way to ensure rapid adaptation while transiently flattening the loss landscape <ref:2507.14056#pg4>. That sounds like a more deterministic control than just hoping for good results.

Lalam: They are also integrating a mechanism to modulate gain based on uncertainty by using Equation four where g of t plus one equals gamma g of t plus (one minus gamma)g0 plus eta H(y) where H(y) is the Shannon entropy of the network’s output <ref:2507.14056#pg3>.

Tom: So this uncertainty-driven gain scaling allows the system to respond to unexpected uncertainty by transiently boosting gain, enabling rapid adjustments to new information while preserving stable representations <ref:2507.14056#pg3>.

Jane: That’s a specific equation for how the gain evolves based on entropy, which is a concrete way to make the uncertainty signal directly drive the system's response <ref:2507.14056#pg3>.

Lu: This formulation allows us to see that g of t behaves as a low-pass filter of the entropy signal H(y) around the basal gain level g0, with a time constant tau <ref:2507.14056#pg3>. This gives us insight into how uncertainty accumulates over time <ref:2507.14056#pg3>.

Meng: So they are showing that the long-term gain level is tied to accumulated uncertainty, which is useful because it tells us how complex the task environment has become <ref:2507.14056#pg3>.

Lalam: Higher values of eta can be used for easier datasets, whereas lower values are preferable for more complex datasets to avoid instabilities when uncertainty is higher <ref:2507.14056#pg3>. This gives us a practical tuning guideline for real-world scenarios <ref:2507.14056#pg3>.

Tom: And they also found that NGM SGD continues to attenuate the stability gap even when replay is used, confirming that the benefits of the gain dynamics persist beyond the idealized joint training setup <ref:2507.14056#pg3>.

Jane: So they are showing that this approach works even in a more realistic continual learning scenario where replay partially mitigates catastrophic forgetting <ref:2507.14056#pg1>. It’s not just an idealized setup thing anymore.

Conclusion: Tom: We're wrapping up our discussion on the paper, "Dynamic gain neuromodulation attenuates stability gaps during continual joint training." The main implication is that uncertainty controls both timescales and parameterization instead of acting solely as a step-size signal <ref:2507.14056#pg3>.

Jane: It’s about using the uncertainty to regulate the learning process in a fundamental way, which moves us from just thinking about step sizes to thinking about how the entire learning process is shaped by multiple factors <ref:2507.14056#pg3>.

Lu: The paper suggests that NGM-SGD provides an abstraction of the coupling between rapid network reconfiguration and persistent synaptic learning <ref:2507.14056#pg3>. It’s a way to look at this coupling that might be more robust than what we see with Adam or MSGD <ref:2507.14056#pg1>.

Lalam: The paper shows that NGM-SGD provides an abstraction of this coupling between rapid network reconfiguration and persistent synaptic learning <ref:2507.14056#pg3>. It’s a way to look at this coupling that might be more robust than what we see with Adam or MSGD <ref:2507.14056#pg1>.

Tom: To summarize, the dynamic gain neuromodulation attenuates stability gaps during continual joint training by introducing an uncertainty-modulated gain mechanism to transiently increase neuronal gain to reshape optimization trajectories <ref:2507.14056#pg3>.

Jane: This is a pretty compelling direction for how we think about adaptive learning systems, and I think it's something we should keep following as we see more papers like this come out.

Lu: It’s a way to reshape the geometry of the landscape for stability <ref:2507.14056#pg3>.

Meng: I just hope that when this gets implemented, it doesn't introduce new instabilities at scale that we can't manage easily.

Lalam: It’s a way to look at this coupling that might be more robust than what we see with Adam or MSGD <ref:2507.14056#pg1>.

Newcastle University

cs.LG, cs.AI, q-bio.NC

Submitted: 2025-07-18

Updated: 2026-10-19

Comments: 23 pages, 11 figures, 6 table, 1 pseudo-code

DOI: 10.1016/j.neunet.2026.109767

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 85/100

The gist: The gist Dynamic gain neuromodulation attenuates stability gaps under joint training by introducing an uncertainty-modulated gain mechanism that transiently increases neuronal gain to reshape

Key concepts

Stability Gap: This refers to the performance drop that occurs immediately after a model is trained on a new task. It happens because learning new information transiently disrupts the knowledge gained from previous tasks, causing instability in the optimization process.
Dynamic Gain Scaling: This is the core mechanism where incoming weights are scaled by a gain coefficient under the ReLU function. This scaling acts as a two-timescale optimizer, allowing for temporarily larger update magnitudes to facilitate rapid learning while preserving slower, more stable representations of older knowledge.
Effective-Coordinate Flattening: This is a mathematical consequence of the dynamic gain. By reparameterizing weights using an invertible map, the curvature of the loss landscape in effective-weight coordinates is scaled by $g^{-2}$. This effectively flattens the landscape during transitions, making it less sensitive to parameter shifts caused by new tasks.
Uncertainty-Driven Gain Scaling: The gain coefficient $g(t)$ is modulated based on the network's output entropy, $H(y)$. When uncertainty (high entropy) is detected after a task switch, the gain transiently increases. This allows the system to adapt more aggressively to new information when it is most uncertain, ensuring rapid adjustment while maintaining stability.

Terminology

Adaptation — 2× in this paper
Accuracy — 1× in this paper · explained in 1 episode
Adam — 1× in this paper · explained in 2 episodes
Benchmark — 1× in this paper · explained in 2 episodes
Continual Learning — 1× in this paper · explained in 3 episodes
Robustness — 1× in this paper · explained in 6 episodes

Summary

The gist Dynamic gain neuromodulation attenuates stability gaps under joint training by introducing an uncertainty-modulated gain mechanism that transiently increases neuronal gain to reshape optimization trajectories, thereby balancing plasticity and stability in continual learning systems.

How it works

The core idea is to view the stability gap as an early-adaptation manifestation of the stability-plasticity dilemma, arising when rapid acquisition of new information transiently perturbs previously acquired knowledge The paper introduces a dynamic gain scaling mechanism as a two-timescale optimization technique that balances adaptation and retention by transiently increasing the effective update magnitude while dynamically reparameterizing the weights governing the forward pass.
This mechanism is biologically inspired, drawing parallels between neuromodulatory bursts, such as noradrenergic activity, and fast-slow weight optimization paradigms where fast weights momentarily improve performance on earlier tasks during new learning without disrupting consolidation occurring in slow weights The authors hypothesize that these transient gain boosts induce fast and slow timescales of weight adaptation, enabling rapid adjustments to new information while preserving stable representations.
The dynamic gain is implemented by scaling incoming weights with a gain coefficient under a Rectified Linear Unit (ReLU) activation function, which serves as a minimal and biologically interpretable abstraction of noradrenergic gain as a modulator of input sensitivity This approach allows for the transient reshaping of the effective-weight loss landscape and adaptive updating of network representations.
Mathematically, this is formalized through an invertible linear reparameterization map where the effective weight is expressed as Effective weight = Φ(w) = Gw, with G being an invertible gain matrix This leads to the crucial result that relative to the local geometry expressed in the baseweight parameterization, the corresponding curvature is scaled by g−2 in the effective-weight coordinates used during the forward pass, referred to as effective-coordinate flattening.

Key Findings and Validation

**: NGM-SGD empirically attenuates stability gaps across domain- and class-incremental MNIST, CIFAR, and miniImageNet benchmarks under task-agnostic joint training while maintaining competitive accuracy. The method outperforms MSGD, Adam, and vanilla SGD in transition stability. 3. Table 1 summarizes the main quantitative metrics across benchmarks using standard continual evaluation metrics established early in the field. The results show that NGM-SGD consistently reduces the stability gap, although performance gains are less pronounced than hypothesized in prior work. 3. Table 2 reports per-transition stability gaps for all joint training benchmarks, highlighting the best results achieved by NGM-SGD. The largest avg-SG signals that the model experiences significant forgetting immediately after each context switch, underscoring its lack of robustness in continual learning settings. 3. Table 4 shows the hyperparameter search results for NGM-SGD on Split CIFAR-10, indicating that the best-performing combination was (γ, η) = (0.9, 0.2). The best learning rates found were 0.1 for SGD and NGM-SGD. 3. Table 6 shows experience replay results across all benchmarks, confirming that NGM SGD continues to attenuate the stability gap even in a more realistic CL scenario where replay partially mitigates catastrophic forgetting. The reset variants of MSGD and Adam exhibit mixed behaviour across benchmarks, sometimes reducing stability gaps and sometimes increasing them. 3. Table 5 shows oracle reset results across benchmarks, where NGM-SGD achieved the best result for Split MNIST. The authors conclude that NGM-SGD achieves the most favorable stability-plasticity trade-off, combining a low average stability gap with the highest final cumulative accuracy. 3. Table 3 shows the hyperparameter search results for Split CIFAR-10, revealing that NGM-SGD achieved the best performance across datasets. The best combination was (γ, η) = (0.9, 2) in terms of scoreNGM. 3. Table 4 also shows that higher values of η can be used for easier datasets, whereas lower values are preferable for more complex datasets to avoid instabilities when uncertainty is higher. The continuous-time formulation makes explicit that g(t) behaves as a low-pass filter of the entropy signal H(y) around the basal gain level g0, with time constant τ. 3. Table 6 shows that for Domain CIFAR-100, NGM-SGD achieved an avg-SG of 0.334 ± 0.07. The stability gap is most evident with multi-timescale optimizers such as Adam and MSGD, where momentum-driven updates overshoot at task switches. 3. Figure 3a shows samples of data around the first task transition, T1→2, where the cumulative training distribution expands from two to four classes. The stable response of NGM-SGD coincides with a phasic increase in neuronal gain following the transition. 3. Figure 3e shows that NGMSGD produces a substantially smaller post-transition curvature amplification than MSGD and Adam and maintains a lower curvature ratio throughout the analyzed window. The authors hypothesize that this effect arises from the gain reparameterization producing a locally flatter loss landscape in effective-weight coordinates, thereby reducing the sensitivity of previous-task loss to the parameter displacements induced by the task transition. 3. Figure 4d illustrates that NGM-SGD achieves the most favorable stability-plasticity trade-off, combining a low average stability gap with the highest final cumulative accuracy. The contrast highlights that phasic gain modulation engages a fast adaptation component while its temporal dynamics regulate this response across successive updates. 3. Figure 7 shows that sustained gain increases as additional classes and task demands accumulate in class-incremental experiments, suggesting that the long-term gain level can serve as a proxy for accumulated task uncertainty and complexity. The authors suggest that reshaping the geometry of the landscape can offer a more flexible route to stability. 3. Appendix E shows that NGM SGD continues to attenuate the stability gap even when replay is used, confirming that the benefits of the gain dynamics persist beyond the idealized joint training setup. The authors suggest that cholinergic neuromodulation could support task segregation through gain-gating mechanisms. The authors note that uncertainty can regulate not only the magnitude of a parameter update, but also the parameterization through which that update is expressed. 3. The work suggests that NGM-SGD provides an abstraction of the coupling between rapid network reconfiguration and persistent synaptic learning. The findings indicate that uncertainty controls both the timescale and the parameterization of learning rather than acting solely as a step-size signal. 3. The authors note that this approach, unlike MSGD and Adam, follows the true gradient, as the gain only rescales the step magnitude without deviating from the gradient itself. The paper suggests a broader design principle for adaptive learning systems where uncertainty controls both timescales and parameterization. 3. The limitations acknowledge that using a small task set ensures that each task induces a substantial distributional shift, making the stability gap large enough to measure reliably. The joint-training setup isolates the transient, optimization-driven forgetting. The authors suggest that NGM-SGD could be combined with methods targeting complementary sources of CL failure, including experience replay approaches.

Improvements for AI systems

Gain Modulation for Dynamic Multi-timescale Optimization: Implement NGM-SGD to balance plasticity and stability by transiently increasing effective update magnitude while dynamically reparameterizing weights, as described in "we introduce a dynamic gain scaling mechanism as a two-timescale optimization technique that balances adaptation and retention by transiently increasing the effective update magnitude while dynamically reparameterizing the weights governing the forward pass. This allows for fast and slow timescales of weight adaptation" to be induced, mirroring biological systems.
Effective-Coordinate Flattening: Utilize gain modulation to achieve effective-coordinate flattening, which reduces transition-induced curvature amplification by scaling eigenvalues of the Hessians by a factor of 1/g squared, as shown in Corollary 1 (Eigenvalue rescaling). This mechanism ensures that the effective step size alphaeff(t) = alpha g 2(t) for the gradient-driven displacement in effective-weight coordinates supports rapid adaptation while transiently flattening the loss landscape.
Uncertainty-Driven Gain Scaling: Integrate a mechanism to modulate gain based on uncertainty, by using Equation (4) where g(t + 1) = γ g(t) + (1 − γ)g0 + η H(y), where H(y) is the Shannon entropy of the network’s output. This allows the system to respond to unexpected uncertainty by transiently boosting gain, enabling rapid adjustments to new information while preserving stable representations.
Robustness Across Incremental Learning: Apply NGM-SGD across class-incremental and domain-incremental benchmarks, as shown in Figure 5, allowing the system to maintain competitive accuracy and a significantly smaller average stability gap compared to Adam and MSGD under task switches. This results in small stability gaps even when distributional shifts are substantial.
Adaptive Hyperparameter Tuning: Implement a hyperparameter search strategy for NGM-SGD, specifically tuning the decay parameter γ (e.g., 0.9) and the phasic bump η (e.g., 0.2), to find the optimal trade-off between average stability gap and final cumulative accuracy, as suggested by Table 4 and Figure 8b for CIFAR-10 benchmarks.

Abstract

Recent work in continual learning has highlighted the stability gap -- a temporary performance drop on previously learned tasks when new ones are introduced. This phenomenon reflects a mismatch between rapid adaptation and strong retention at task boundaries, underscoring the need for optimization mechanisms that balance plasticity and stability over abrupt distribution changes. While optimizers such as momentum-SGD and Adam introduce implicit multi-timescale behavior, they still exhibit pronounced stability gaps. Importantly, these gaps persist even under ideal joint training, making it crucial to study them in this setting to isolate their causes from other sources of forgetting. Motivated by how noradrenergic (neuromodulatory) bursts transiently increase neuronal gain under uncertainty, we introduce a dynamic gain scaling mechanism as a two-timescale optimization technique that balances adaptation and retention by transiently increasing the effective update magnitude while dynamically reparameterizing the weights governing the forward pass, thereby empirically mitigating transition-induced curvature amplification. Across domain- and class-incremental MNIST, CIFAR, and mini-ImageNet benchmarks under task-agnostic joint training, dynamic gain scaling effectively attenuates stability gaps while maintaining competitive accuracy, improving robustness at task transitions.