Skip to content

Gradient Routing: Localizing Computation in Neural Networks via Masked Gradients

Gradient Routing: Masking Gradients to Localize Computation in Neural Networks

arXiv· arXiv:2410.04332· Alex Cloud, Jacob Goldman-Wetzler, Evžen Wybitul, Joseph Miller, Alexander Matt TurnerML Alignment and Theory Scholars (MATS)Indexed by this site · 2026-10-14· Originally publishedAlignment

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

The authors use gradient routing to localize capabilities; with partial labeling they claim forgetting is better than filtering, and after retraining 0.7B, WMDP-bio increases by 0.182 and FineWeb-Edu by 0.032.

Questions
Networks are mainly trained on input-output pairs, yet internal mechanisms relate to transparency, harmful capabilities, and out-of-distribution generalization. The authors use data to localize specified computations to predetermined subregions, especially when labels do not cover all relevant data.
Methods
Forward pass is unchanged; backward pass uses a data-dependent weighted mask to decide which subregions are updated by which data. Representation experiments partition digits across the two encoding halves; unlearning uses Expand-Route-Ablate then removes the expanded neurons; reinforcement learning partitions updates for the two terminal cell types across two experts.
Experiments and results
In partial labeling, the authors state that ERA forgetting outperforms data filtering. After retraining 0.7B, WMDP-bio loss increases by 0.182, FineWeb-Edu by 0.032 (Table 1). Under 1% supervision, the authors state there is still non-trivial performance.
Limitations and future work
The authors note hyperparameter sensitivity, token-by-token routing ignoring adjacent context, scalable supervision limited to simple environments, and post-ablation behavior not necessarily coherent. Experiments include MNIST, CIFAR100, TinyStories-28M, 303M and 0.7B, and a 5×5 grid; the paper does not report the number of 0.7B repetitions.
Experimental setupMLP autoencoder、ResNet et al. · MNIST、CIFAR100 et al. · MAE、validation accuracy et al.
Evaluated models
MLP autoencoderResNet20-layer, 303M parameter TransformerTinyStories-28M0.7B parameter TransformerMixture of Experts (MoE) network
Benchmarks
MNISTCIFAR100FineWeb-EduTinyStoriesWMDP-bio5×5 gridworld
Metrics
MAEvalidation accuracyforget lossretain lossunlearningrobust unlearningcross-entropy lossground truth return
AI summaryThe researchers propose Gradient Routing, a training method that applies a data-dependent weighted mask to gradients during backpropagation, allowing users to configure which parameters are updated by which data points, thereby isolating capabilities into specific subregions of the neural network. The authors claim the method can learn representations partitioned in an interpretable way, achieve robust forgetting by ablating pre-specified network subregions, and localize modules responsible for different behaviors, enabling scalable oversight of reinforcement learners. They also find that gradient routing localizes capabilities even when applied only to a limited, temporarily selected data subset, and consider the method promising for real-world applications where high-quality data is scarce.

The researchers propose Gradient Routing, a training method that applies a data-dependent weighted mask to gradients during backpropagation, allowing users to configure which parameters are updated by which data points, thereby isolating capabilities into specific subregions of the neural network. The authors claim the method can learn representations partitioned in an interpretable way, achieve robust forgetting by ablating pre-specified network subregions, and localize modules responsible for different behaviors, enabling scalable oversight of reinforcement learners. They also find that gradient routing localizes capabilities even when applied only to a limited, temporarily selected data subset, and consider the method promising for real-world applications where high-quality data is scarce.

Related events and news

1 events · 2 news mentions

News that mentions this paper

梯度路由缓解RL奖励作弊及扩展实验

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    Use data-dependent gradient masks to localize safety-relevant computation to user-specified subregions of the network.

    Core question: Can we use data during training to localize safety-related internal computations to specified subregions, rather than only fitting inputs to outputs?

    • Scenario: The authors argue that networks are mainly trained on input-output pairs, and internal mechanisms are neglected; these mechanisms relate to transparency, whether they contain sensitive information or harmful capabilities, and whether goals generalize reliably outside the training distribution. As systems become more capable, these mechanisms need to be explained and controlled.
    • Existing shortcomings: The authors argue that fully reverse-engineering conventional models is unnecessary, and switching to inherently interpretable architectures is also unnecessary; manually specifying internals is mostly infeasible, so data-driven targeted influence on internals (mechanistic supervision) is needed. When labels do not cover all harmful sources, filtering may be insufficient; typical forgetting can be recovered with a small amount of fine-tuning.
    • Proposed approach: Apply user-provided, data-dependent weighted masks to backpropagation, specifying which parameters, activations, or modules are updated by which tokens, documents, or labels. The authors report that even when routing only on a limited, temporarily selected subset, broader related capabilities still enter the target region.
    • Proposed method: gradient routing. Demonstrates three uses: partitioning representations into interpretable regions; ablating predetermined subregions to remove capabilities; and localizing different behaviors in reinforcement learning to modules, so that policies can be supervised when reliable labels are scarce.
    • Supervision conditions: The user specifies the masks. For language forgetting, data are divided into forget set and retain set; unlabeled forget data are treated as retain during training. In the grid world, terminal cell types are visible only in some episodes, and the policy input includes a supervision indicator, so it can condition on whether it is supervised.
  2. What related research exists?

    The authors place this paper in the lines of modular training, concept erasure, robust forgetting, and limits of data filtering.

    Modularity and pre-specified capabilities

    • Pfeiffer et al. (2023) survey modular learning: forward pass selects modules via routing functions. Some routing uses metadata (Waibel & II, 1992), such as modality, language, or robot high/low-level control and task type.
    • Gururangan et al. (2021) DEMix partitions data by domain, with one expert per layer per domain; turning off a domain expert approximates a model not trained on that domain. This paper adds ablation to it as a localization baseline.
    • Another line freezes pretrained weights and trains new modules (Rebuffi et al., 2017; 2018; Houlsby et al., 2019; Bapna & Firat, 2019). Zhang et al. (2024) learn weight masks, transfer subnetworks to randomly initialized models, and retrain from scratch, e.g., making vision models recognize ImageNet classes by shape rather than texture.

    Adversarial representation learning and concept erasure

    • Using discriminators to adversarially train feature extractors to control information in representations (Goodfellow et al., 2014; Schmidhuber, 1992; Ganin & Lempitsky, 2015; Ganin et al., 2016; Edwards & Storkey, 2015).
    • Another line modifies activations at inference to remove concepts (Ravfogel et al., 2020; Belrose et al., 2023; Elazar et al., 2020; Bolukbasi et al., 2016).
    • Authors' comparison: gradient routing localizes capabilities during training, with optional ablation afterward, rather than erasing at inference.

    Robust forgetting and post-hoc pruning

    • Machine unlearning aims to remove unwanted knowledge or capabilities from trained networks (Cao & Yang, 2015; Li et al., 2024). The authors cite multiple works claiming typical methods are fragile and can be recovered with a small amount of fine-tuning. Lee et al. (2024) and Łucki et al. (2024) argue that unwanted concepts are more easily bypassed than completely removed from weights.
    • TAR (Tamirisa et al., 2024) also targets robust forgetting for LLMs, but via fine-tuning; this paper does gradient routing during pretraining.
    • Pruning approaches include SISA (Bourtoule et al., 2021) which trains multiple models on data partitions and discards one, and deleting neurons or parameter subspaces related to harmful behavior. Lo et al. (2024) find that models pruned to remove concepts can quickly relearn them; the authors relate this to the limited success of post-hoc circuit discovery.

    Limits of data filtering and experimental baselines

    • The authors cite Anwar et al. (2024): limited or imperfect labeling is common. Filtering often relies on heuristics and blacklists (Albalak et al., 2024); rules and narrow classifiers struggle to effectively find harmful content (Welbl et al., 2021). Toxicity filtering may also exclude content from marginalized groups (Dodge et al., 2021; Chowdhery et al., 2023).
    • Baseline methods: Language forgetting compared data filtering, RMU (Li et al., 2024), DEMix + ablation. Reinforcement learning compared data filtering and naive training; the appendix also includes oracle filtering that sees all terminal states but downsamples according to supervision ratio. Benchmarks/datasets: MNIST, CIFAR100, FineWeb-Edu, TinyStories (Delphi), WMDP-bio forget set, 5×5 grid world.

    The authors position this paper as training-time mechanistic supervision: using data-dependent gradient masks to specify where to learn, without changing the loss. Appendix J also compares IIT, PackNet, PiggyBack; the authors claim the difference is that localization is specified by user masks, and the localization of learning updates is decoupled from the forward computation.

  3. How does the paper address this problem?

    Forward pass unchanged, backward pass reweights gradients by data masks; ERA first expands dimensions, then routes, then ablates; RL assigns behaviors to experts.

    gradient routing does not change the forward pass; it only uses user-specified, data-dependent edge weights during backpropagation to decide which data updates which subregions. The authors use "route X to Y" to mean restricting the learning updates of data X to region Y.

    Routing derivatives

    • Each data point can assign real-valued weights to the edges of the computation graph. The weights act only on the backward pass; when the weights are all 1, ordinary backpropagation is recovered. Any gradient-based optimizer, such as SGD or Adam, can be used with these modified gradients.
    • At terminal nodes, the routing derivative is 1. For non-terminal nodes, the routing derivative is the sum of the contributions from each child: first multiply by the edge weight, then by the ordinary local derivative.
    • This definition is written as ∂ L̃(zi)/∂ v=∑u∈child(v)αi(v,u)·∂ L̃(zi)/∂ u·∂ u(zi)/∂ v. The meaning is that child gradients are scaled by the edge weight before being propagated back.
    • In practice, it is not necessary to set a separate mask for every edge and every point. The authors usually mask only the outputs of a few layers and assign weights according to coarse partitions such as forget/retain.

    Implementation and masks

    • Algorithm 1 multiplies the activation by a mask at the layer output, then adds (1-mask) times the same activation after detach: the value is unchanged, and the masked-out path does not propagate gradients.
    • Most masks in this paper are 0/1. In the forgetting experiments, the weights on the original dimensions of the target layer can be less than 0. Appendix K lists the masks by task; the authors compare the choice to selecting an architecture, considering the capacity of the target region, the inductive bias, and whether the model still needs to use the region after ablation.

    Representation partitioning

    • MNIST uses a three-layer MLP: the encoder maps 28×28 to R32, and the decoder reconstructs all digits. Each of the two certificates looks at half of the encoding (R16), and its updates do not change the encoding. The authors' criterion is: if the full decoder can reconstruct, but a given certificate cannot, then the information is in the other half.
    • During training, gradients for 0–4 are restricted to the upper half, and 5–9 to the lower half. The loss is pixel MAE, plus an encoding L1 (weight 0.003) and a penalty on the sum of the absolute correlations between the upper and lower halves (weight 0.1). Appendix B: 200 epochs, batch 2048, Adam learning rate 1e-3.
    • Appendix B.1 applies the same idea to CIFAR100 classification: the convolutional layers of a 34-layer ResNet serve as the encoder, split in half along the channel dimension, routing by classes 0–49 and 50–99, with an encoding L1 added.

    Language models: scalar steering and ERA

    • Steering scalar: when predicting the location of California, the gradients in the residual stream of layers 6–18 are set to 0 except for dimension 0, so that updates for these positions are concentrated in the weights that write to dimension 0. Afterwards, the model can be steered by changing only a single scalar in this dimension.
    • Expand, Route, Ablate (ERA) has three steps. Expand: add randomly initialized neurons to the target layer. Route: train from scratch on next-token prediction; for selected forget tokens, lower the learning rate on the original dimensions of the target layer, and leave retain gradients unchanged. Figure 3 assumes a learning rate of 0 on the original dimensions of the target layer and draws the gradient path for forget tokens. Ablate: remove the added neurons, then do a few steps of fine-tuning on a small amount of retain data to repair the damage caused by the ablation.
    • For TinyStories, the target layers are 0–4, with 64 MLP neurons added per layer, and the mask weight on the original dimensions is −0.75; each token is then interpolated between the all-pass mask and the strong-routing mask according to its relative frequency in forget versus retain. An L1 with weight 1e-4 is added to the target-layer MLP activations.
    • The 0.7B experiment routes 20 virology-related tokens to dimensions 0–79 of the MLP in layers 0–7; these tokens have a mask weight of −5×10^−8 on the original dimensions, and the target-layer MLP activation L1 is 1×10^−7. The footnote lists these 20 tokens.

    Behavior modules in reinforcement learning

    • The environment is a 5×5 grid, with two terminal cell types, diamond and ghost, four-directional movement per step, and a maximum of 32 steps. Supervisory instructions are sampled independently with probability p. The policy observes the full state, including these instructions. A trial has a type label only if it reaches a terminal cell whose instruction is TRUE.
    • The policy is a mixture of experts with two MLP experts plus a gating network, mapping a convex combination of outputs to actions. When a type label is present, gradients enter only the expert of the reached type, and the gate is trained to activate only that module; unsupervised trials use ordinary REINFORCE, with gradients not masked by type.
    • In Appendix F, the training reward rMoE is 1 when the label is reached diamond or ghost, and 0 otherwise; it does not give a negative reward for ghost. Evaluation uses a separate ground truth return. The gating loss is used only when a full label is present.
  4. What experiments did the paper conduct?

    The authors claim ERA is better than filtering with partial labeling; after retraining 0.7B, WMDP-bio +0.182, FineWeb-Edu +0.032.

    Experimental Setup

    • System under test: MNIST three-layer MLP autoencoder; appendix uses a 34-layer ResNet on CIFAR100. Language side uses TinyStories-28M (8 layers, hidden 512), and modified nanoGPT: 20-layer 303M (GPT-2 tokenizer) and 0.7B (Qwen-2 tokenizer, embedding 1536). Grid policy is a two-expert MoE.
    • Data: MNIST trains on 60,000 images for 200 epochs; CIFAR100 trains on 50,000, tests on 10,000, for 200 epochs. TinyStories uses the Delphi version with 400,000 stories, 1 epoch, sequences truncated to 256; stories containing forest, tree, woodland and their plurals form the forget set, about 20% of training data. 303M trains on FineWeb-Edu for 10,000 steps. 0.7B trains on about 13B tokens, 10,000 steps, with about half of the WMDP-bio forget set added. Grid policy trains for 20,000 algorithm steps (328M environment steps), 512 parallel environments, 32 steps per rollout.
    • Annotation and baselines: TinyStories is trained with the proportion of forget stories labeled; unlabeled ones are treated as retain; validation set is always correctly labeled. Comparisons include data filtering (discarding the full-data model and retraining from scratch on retain), RMU, DEMix + ablation. Grid compares data filtering and naive training; appendix also has oracle filtering.
    • Metrics: pixel MAE; CIFAR validation accuracy. Unlearning is the difference in forget loss before and after the method; robust unlearning is the difference before the method and after fine-tuning on 64 forget stories post-method, taking the minimum validation forget loss. Retain is the retain loss after the method. WMDP uses cross-entropy, and additionally cross-entropy after removing 20 routed tokens. Grid evaluation is stepwise ground truth return: diamond is 1, ghost is −1, otherwise 0, discount 0.97.
    • Review and repetition: No LLM judge; language tasks look at loss, grid looks at environment return. MNIST 20 runs; TinyStories at least N=5 per setting, plots show 95% CI of the mean; CIFAR gradient routing metrics N=4. The paper does not report the number of runs for 0.7B and grid experiments, nor human agreement.
    • 0.7B retraining protocol: After ablating dimensions 0–79 of the MLP in layers 0–7, fine-tune on FineWeb-Edu for 32 steps (gradients do not flow into ablated dimensions), then train on 2 WMDP-bio forget samples for 20 steps, reporting the step with the lowest validation loss.

    Main Results

    Table 1 is the 0.7B model with 20 virology tokens routed. Loss is before ablation; retraining column follows the protocol in the previous section.

    Swipe horizontally to read the table

    DatasetLossAfter ablation (Δ)After retraining (Δ)
    WMDP-bio forget set2.5964.283 (+1.687)2.778 (+0.182)
    WMDP-bio (routed tokens removed)2.5674.205 (+1.638)2.738 (+0.171)
    FineWeb-Edu2.9254.864 (+1.939)2.957 (+0.032)
    • The authors claim that after retraining, the increase on the forget set is still higher than on FineWeb-Edu, so localization and robust removal hold. This remains true after removing the routed tokens, and the authors claim this is not merely lowering the probability of these 20 tokens.
    • Ablation itself raises loss on both datasets, and the ablation increase on FineWeb-Edu is not lower; the subsequent FineWeb-Edu fine-tuning is to recover some performance.
    • This table has only this one 0.7B model, with no side-by-side numbers against data filtering or RMU.

    Partial-Label Unlearning on TinyStories

    • Tested ERA, data filtering, RMU, DEMix+ablate at different forget labeling proportions for three types of loss (Fig 4, at least N=5, 95% CI). When labeling is below 100%, the authors claim ERA outperforms baselines including data filtering on both unlearning and robust unlearning.
    • The cost is worse retain, which the authors claim is proportional to the amount of data routed. DEMix+ablate has negative unlearning except at 100% labeling; the authors explain that the forget expert is trained only on labeled forget data, while the retain expert also sees unlabeled forget. The text does not give specific values for these differences.
    • At 100% labeling, the authors claim RMU has the highest forget loss before retraining, data filtering is most robust to retraining, and most of RMU's unlearning is undone by retraining. ERA's forget loss after retraining is higher than RMU (Fig 9), but after subtracting overall performance degradation, robust unlearning is comparable to RMU; ERA+RMU ("+" in the figure) is higher than either alone. Fig 4 also has a point for RMU at labeling proportion 0.95.

    Scalable Supervision in a Grid World

    • At the supervision proportions listed on the x-axis of Fig 5a, compares gradient-routed MoE, data filtering, and naive training on ground truth return; evaluation is on 512 random environments, 32 steps per environment. An example of a policy trained under 10% supervision is in Fig 5b, whose caption states both terminal cells are unsupervised, and turning still elicits the target behavior.
    • The authors claim routing separates modules that go toward diamond and ghost; using the diamond expert outperforms baselines at low supervision, with nontrivial performance at 1%, and at 5% achieves comparable performance to other methods at six times the supervision. The authors also claim that at high supervision, baselines improve largely because a higher proportion of evaluation environments are supervised, not because they perform well when unsupervised.
    • The authors claim naive training exploits the supervision gap: avoiding ghosts when supervised, and heading toward nearby ghosts when unsupervised. Below 40% supervision, data filtering cannot learn even when trained to the limit (Fig 12). Oracle filtering eventually converges; the appendix states that at 10% supervision it takes about 3–4 times more training steps to reach the routing method's performance, while the routing method's highest episode return is about 0.2 lower because ghosts on the shortest path are still walked into.

    Representation Partitioning and Steering Scalar

    • MNIST: In Fig 2 and Table 2 setting 1, the lower-half certificate has MAE 0.32 (±0.02) for 0–4 and 0.08 (±0.00) for 5–9, mean of 20 runs; the authors claim predictions for 0–4 are nearly constant, indicating the lower half has almost no information about these digits. Across-run 5/95 percentiles are (0.31, 0.33) for 0–4 and (0.08, 0.09) for 5–9.
    • CIFAR100: The authors claim gradient routing with L1 leads to localized certificates while validation accuracy drops; at N=4, the 95% CI widths for gradient routing metrics range from 0.03 to 0.07. The text does not give accuracy point estimates.
    • 303M: After routing only the California token to dimension 0 of the residual stream in layers 6–18, the dimension with the highest cosine similarity to the unembedding is California, followed by Oregon, Colorado, Texas, etc. (top 10 in the text, top 300 in Appendix D). The paper does not give similarity values. The authors claim this localizes a broader semantic concept and can be steered with a single scalar; Appendix D gives generation samples.

    Other Ablations and Analyses

    • Table 2 setting 4, gradient routing without regularization: lower-half certificate is 0.32 (±0.02) for 0–4 and 0.32 (±0.01) for 5–9.
    • Table 2 setting 8, no routing, no regularization, trained only on 5–9: both groups are 0.08 (±0.01 / ±0.00), and the authors claim unseen digits can still be reconstructed.
    • Table 2 setting 6, lower half also receives gradients for 0–9 with increased L1 on the lower half: 0–4 is 0.23 (±0.02), 5–9 is 0.13 (±0.01).
    • Fig 11: control with only expansion, ablation, and L1, without routing, has forget validation loss 1.47 (±0.02), same as base; the authors claim ERA's effect comes from routing, not ablation itself.
    • Fig 13: The authors claim that routing alone or gating alone cannot stably maintain expert specialization; specialization from early routing on average decays with training.
    • Appendix F states that the gradient-routed MoE's highest episode return is about 0.2 lower than a fully trained oracle filtering.
  5. What are some points that can be further explored?

    The authors note that routing hyperparameters are sensitive, token-by-token context is ignored, and the scalable supervision experiments are limited to simple environments.

    Limitations and future directions noted by the authors

    • Performance is sensitive to which data are routed, where they are localized, and the mask weight magnitude, making it difficult to balance retain and forgetting; the authors speculate that methodological improvements would reduce this sensitivity (Section 5).
    • Language model experiments route by token, ignoring adjacent tokens. The authors find this unexpectedly effective, but some problems may require whole sequences; how to choose routing data to localize broad capabilities is an open question (Section 5).
    • Empirical results for scalable supervision are in simple, narrow settings; the method requires that the policy after ablation still produces coherent behavior, which is not generally true, and scaling supervision via localization may require new ideas (Section 5).
    • Appendix A adds: the toy environment lacks the complexity and asymmetry of real problems; ablating a previously active module still yields coherent behavior, which may not hold in practice (Appendices A, G).
    • On MNIST, the clean separation shown depends on the problem setting (e.g., no data augmentation like rotation) and strong L1. The authors say digit feature separation is "unnatural," and outside the chosen hyperparameters, visual differences are not so distinct (Appendix A).
    • The learning dynamics of gradient routing remain an open question. The authors speculate that localizing many concepts to separate regions would increase training loss (Appendix A).

    Experimental coverage

    • Tested systems: MNIST MLP autoencoder, 34-layer ResNet on CIFAR100, TinyStories-28M, 303M, and 0.7B Transformers, and a two-expert MoE on a 5×5 grid.
    • Language forgetting compared data filtering, RMU, and DEMix+ablate; TinyStories labeling ratios ranged from partial to 100%, robustness tested by fine-tuning on 64 forget stories. The 0.7B recovery test was FineWeb-Edu fine-tuning for 32 steps, then training on 2 WMDP-bio samples for 20 steps.
    • 0.7B routed only the 20 virology tokens from the footnote, with regions being dimensions 0–79 of the MLP in layers 0–7; training corpus was FineWeb-Edu plus about half of the WMDP-bio forget set. Language results are reported as loss.
    • Grid supervision ratios are the levels on the x-axis of Fig 5a, evaluated on 512 environments with 32 steps each; the appendix compares removing routing, removing gating, and oracle filtering that violates the environment definition.
    • The paper does not report repetitions for 0.7B and grid experiments, LLM evaluation, human agreement, or the number of WMDP-bio entries.
  6. Summarize the main content of the paper

    Gradient routing localizes capabilities during training: supports representation partitioning, ablation-based forgetting, and policy steering when labels are scarce.

    gradient routing uses data-dependent masks during training to restrict specified computations to predetermined subregions, to address transparency, harmful capabilities, and out-of-distribution goals, rather than only fitting inputs to outputs.

    • Problem: The authors argue that internal mechanisms are neglected by conventional training; labels often do not cover all harmful data, and filtering and post-hoc forgetting may both be insufficient.
    • Method: Forward pass unchanged, backward pass reweights gradients. MNIST splits 0–4 and 5–9 into two halves of the encoder. ERA first adds neurons to the target layer, then pushes updates from forget tokens into the new dimensions, and finally removes these dimensions. In reinforcement learning, only the expert corresponding to the type label is updated in episodes with type labels; training reward does not penalize ghosts.
    • Results: Table 2 setting 1, lower-half certificate MAE for 0–4 is 0.32 (±0.02) and for 5–9 is 0.08 (±0.00). With labeling below 100%, the authors claim ERA is better than data filtering, but retain is worse; at 100%, data filtering is more stable for retraining. Table 1: after retraining 0.7B, WMDP-bio +0.182, after removing routing tokens +0.171, FineWeb-Edu +0.032; the ablation itself does not make the FineWeb-Edu increase lower. The authors claim the grid policy still shows non-trivial performance with 1% supervision.
    • Authors' conclusion: Routing on narrow data draws broader capabilities into the target region; mechanistic supervision can be done without changing the loss, thus avoiding speculation when training only on observable outcomes. The authors see potential for high-risk deployments with scarce labels, while acknowledging hyperparameter sensitivity and that scalable supervision remains a narrow setting.
Read originalarxiv.org