Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Latent Activation Editing: Inference-Time Refinement of Learned Policies for Safer Multirobot Navigation

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a pre-trained multi-robot navigation policy can be made safer at inference time by monitoring and editing its internal latent activations, without retraining or changing its weights, cutting collisions by nearly 90% w

desk verdict A genuine new idea—inference-time latent editing for frozen robot policies—with a solid ablation study and a real-hardware demo, but the 89.6% collision-reduction headline is computed on a hand-picked hard set and needs a random-distribution follow-up. read the letter →

arxiv 2509.20623 v2 pith:JQBUEVM6 submitted 2025-09-24 cs.RO

Satyajeet Das , Darren Chiu , Zhehui Huang , Lars Lindemann , Gaurav S. Sukhatme This is my paper
classification cs.RO
keywords latentactivationeditinginference-timeinterventionmultirobotnavigationcollisionavoidanceworldmodelreinforcementlearningquadrotorswarm
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces Latent Activation Editing (LAE), a two-stage inference-time intervention for frozen reinforcement-learning policies. An online classifier monitors intermediate activations to flag states that are about to lead to collisions, and an editing module replaces those activations with surrogates predicted by a latent collision world model (LCWM). The central claim is that amplifying the policy's internal sense of risk, by injecting predicted future pre-collision activations, induces earlier and safer avoidance maneuvers. Across 2,600 simulated configurations, the paper reports a reduction from 5,623 to 583 collisions (an 89.6% drop), an increase in collision-free trajectories from 0 to 2,175, and a rise in average success rate from 0.58 to 0.64, with real-world Crazyflie flights confirming feasibility. The paper further argues that successful editing must preserve the latent components encoding the robot's own dynamics, since editing the full latent catastrophically degrades behavior.

What carries the argument

The central mechanism is the latent collision world model (LCWM), an action-free temporal predictor that operates directly on the frozen policy's intermediate activations. Given a short history buffer of n past latents (n=3 in the experiments), a GRU evolves hidden states h_i and outputs a predicted future latent via Z'_t = W h_t + b, clamped to at most the collision time. This surrogate activation replaces the flagged latent in the policy's forward pass. The LCWM is trained exclusively on collision-bearing trajectories and pre-collision windows (horizon H=250, prediction horizon m=10 by default). Its role is to 'amplify' the policy's risk signal by pushing the activation toward the state it

What would settle it

Measure the distribution of edited activations Z'_t relative to the natural latents the frozen policy produces during safe operation (e.g., via reconstruction error through the policy's encoder, Mahalanobis distance, or a density estimate). If Z'_t systematically falls far outside the safe-latent distribution, then the safety gains come from generic perturbation rather than from the claimed amplification of the policy's risk perception, and the method would likely fail on unseen environments.

Watch

Extended reading notes

Core claim

The core discovery is that steering a pre-trained policy's intermediate activations at inference time—without any weight updates—can substantially improve a narrow behavioral axis such as collision avoidance. The authors hypothesize that a latent collision world model, trained only on pre-collision windows from the policy's own rollouts, can predict how activations evolve toward a crash, and that injecting these predicted future activations amplifies the policy's internal risk perception. They show that a lightweight GRU-based LCWM, paired with a high-accuracy classifier, reduces cumulative collisions from 5,623 to 583 on 2,600 designated configurations, increases zero-collision trajectories

Load-bearing premise

The entire safety mechanism rests on the untested premise that the GRU's predicted latent activations, trained only on pre-collision windows, stay on the frozen policy's feasible latent manifold; if they drift off it, the policy produces infeasible or hazardous actions, and the authors' own ablations show that off-manifold full-latent edits cause catastrophic collisions.

Editorial extensions

If this is right

  • LAE reduces cumulative collisions from 5,623 to 583 (89.6%) on 2,600 configurations, increases zero-collision trajectories from 0 to 2,175, and raises average success rate from 0.58 to 0.64, with statistical significance.
  • The method transfers to real hardware: a classifier with ~2k parameters and a GRU editor with ~7k parameters run on Crazyflie 2.1's STM32 at 100 Hz with under 1 ms added latency, fitting within a 1 kHz control loop.
  • Effective latent editing must preserve self-dynamics components: editing the full Z_1 or the downstream Z_2 results in 67,951 and even more collisions, respectively, versus 583 for the partial Z_1 edit.
  • The editing horizon H and prediction horizon m are critical: H=50 leaves 1,124 collisions, H=250 yields 583, and m=30 causes 60,547 collisions, so the temporal targeting of edits is as important as the editor itself.
  • The framework is modular and behavior-agnostic: the same classifier/editor structure can, in principle, be re-targeted to other behavioral axes by changing the labeling scheme for the dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to add a projection or uncertainty check that keeps the predicted latent Z'_t on the frozen policy's feasible latent manifold; the paper's own full-latent ablations show that off-manifold edits are catastrophic, so a manifold-constraint could make LAE more robust and principled.
  • Because LAE leaves safe-region behavior identical to the base policy, it could be composed with other safety layers (e.g., reachability-based filters) without re-training, potentially layering complementary safety mechanisms.
  • The paper's framing suggests that pretrained policies encode a learnable 'risk axis' in their latent space; if that is true, the LCWM could be replaced by a simple linear probe or concept-direction intervention once the risk direction is identified, yielding an even lighter editor.
  • The dependence on labeled collision windows implies that LAE's effectiveness may degrade in environments where collision types or avoidance dynamics differ substantially from the training distribution; evaluating on out-of-distribution obstacle layouts would clarify the method's generalization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces Latent Activation Editing (LAE), an inference-time method that modifies intermediate activations of a frozen RL policy to reduce collisions in multi-quadrotor navigation. A behavior classifier flags unsafe latents, and a latent collision world model (LCWM), implemented as a GRU, predicts a future latent from a short history and replaces the flagged activation before the downstream policy layers. The method is evaluated on 2,600 deterministic simulator configurations selected because the base policy collides at least once: collisions drop from 5,623 to 583, zero-collision trajectories rise from 0 to 2,175, and average success rate rises from 0.58 to 0.64. Additional ablations compare editing strategies, horizons, and prediction lengths, and real-world Crazyflie experiments are reported qualitatively.

Significance. If the quantitative claims hold on a representative deployment distribution, LAE would be a genuinely useful post-deployment safety refinement: it requires no retraining, adds fewer than 10k parameters and under 1 ms latency on resource-constrained hardware, and the ablation showing that editing must preserve self-dynamics latents is a valuable design insight. The systematic comparison against KD-tree retrieval, sparse autoencoders, and encoder–decoder projections is also informative. However, the headline effect is currently established only on a collision-selected evaluation subset, and the central mechanism lacks a safeguard for staying on the policy's feasible latent manifold. The significance is therefore conditional on additional evaluation and analysis.

major comments (3)
  1. [Sec. V-A; Abstract; Sec. V-C] The evaluation set is constructed as 2,600 configurations 'where the base policy collides at least once,' so the baseline has zero zero-collision trajectories by construction and the 89.6% collision reduction is a conditional improvement on this hard subset, not an effect on the deployment distribution. The Abstract states 'nearly 90% fewer cumulative collisions' and 'substantially increases the fraction of collision-free trajectories' without this qualifier. No results are reported on a random or held-out set of configurations, no false-positive rate on baseline-safe runs is given, and no count of collisions introduced by LAE in scenarios the baseline would have completed safely is provided. The claim 'while preserving task completion' is therefore unverified outside the selected subset. Please add a full-distribution evaluation (random configurations, reporting total collisions per epi
  2. [Sec. IV-C, Eq. (5); Sec. V-E.1] The LCWM produces Z'_t = W h_t + b from a three-step history buffer and is trained only to match pre-collision latent trajectories. There is no constraint, projection, or uncertainty estimate ensuring that Z'_t lies on the frozen policy's feasible latent manifold. The paper's own ablation shows that editing the entire Z1 or Z2 latent causes catastrophic degradation (67,951 collisions), demonstrating that off-manifold edits are not benign. Since the safety claim rests on the edited latent being a valid input to the downstream policy, the paper should at least quantify the distance of predicted latents to the empirical latent manifold and/or the resulting action deviation, and discuss or implement a safeguard (e.g., trust region, rejection sampling, uncertainty threshold) before presenting LAE as a general safety layer.
  3. [Sec. V-E.2, V-E.3; Sec. V-C] The default hyperparameters H=250, m=10, and n=3 are selected using the same 2,600 configurations on which the headline result is reported, with no separate validation set described. The paired t-test and Cohen's d are computed on the same set used for model selection. This risks overfitting the tuning choices to the evaluation set and weakens the 'statistically significant' claim. Please add a held-out evaluation set or cross-validated hyperparameter selection, and report performance on configurations not used for tuning.
minor comments (5)
  1. [Sec. IV] Typo: 'mulit-quadrotor' should be 'multi-quadrotor'.
  2. [Sec. IV-C] Duplicate word in 'where where m defines the number of steps' and later 'we adopt it as the default in for all experiments'.
  3. [Sec. V-D.2] Duplicate phrase 'such as such as dictionary size'.
  4. [Sec. V-F] The real-world experiments are qualitative demonstrations (two crossing scenarios with no collision counts or statistical comparison). Please state clearly that these are illustrative feasibility results, not quantitative validation.
  5. [Sec. I] Contribution 1 claims 'the first activation-space intervention demonstrated on learned robot policies.' This is a strong claim; consider tempering it or adding a more thorough comparison with prior robotics latent-space editing work.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the safety improvement is an empirical result whose training objectives do not contain the evaluation metric; self-citations are infrastructure, not load-bearing.

full rationale

The paper's derivation chain is self-contained. The behavior classifier (Sec. IV-B) is trained on latent activations labeled by a time-to-collision heuristic (Eq. 1), and the LCWM (Sec. IV-C) is trained to minimize reconstruction error to future latents from pre-collision windows (Eq. 4). No training objective contains the evaluation metric (total collisions) as a term, and the edited latent Z'_t = W h_t + b is not computed from the collision count. The observed 89.6% collision reduction therefore is not forced by construction: the same editing pipeline can degrade performance catastrophically when applied to full Z1 or Z2 latents (Sec. V-E.1, 67,951 collisions), demonstrating that the result depends on the specific choice of latent and is empirically contingent. The citations to Huang et al. [2] and QuadSwarm [40] are self-citations by overlapping authors, but they are used as the frozen policy and simulator infrastructure, not as evidence for the safety improvement; no uniqueness theorem or ansatz is imported from them. The evaluation is restricted to 2,600 configurations where the baseline collides, and the paper does not report full-distribution numbers, which is a validity limitation but not circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical entities are postulated. The introduced components, LAE and LCWM, are model artifacts (a classifier plus a GRU), not entities with independent falsifiable handles.

free parameters (6)
  • H (editing horizon) = 250
    Determines how many steps before a logged collision are labeled unsafe in Eq. 1; chosen in Sec. V-E.2 for lowest collisions and best success.
  • m (prediction horizon) = 10
    Number of steps the LCWM predicts ahead (Eq. 4); tuned in Sec. V-E.3, m=10 best, larger m destabilizes.
  • n (history buffer length) = 3
    Number of past latents fed to the GRU; stated as 'we found n=3 to suffice' in Sec. V-C, task-dependent.
  • GRU hidden size = 32
    Capacity choice for the editor; used for real-time deployment on Crazyflie.
  • Classifier hidden size = 64
    Capacity choice for the two-layer MLP behavior classifier; no systematic study reported.
  • Edited latent components (Z1_Partial) = neighbor and obstacle embeddings only, leaving self-dynamics untouched
    Hand-designed choice of which components of Z1 to edit; ablations show it is critical, since editing all of Z1 gives 67,951 collisions (Sec. V-E.1).
assumptions (4)
  • domain assumption Time-to-collision labeling (Eq. 1) is a valid safety surrogate.
    Labels all latents within H steps of a logged collision as unsafe; the classifier and editor are both trained on these labels. If the surrogate mislabels latents, the trigger is miscalibrated (Sec. IV-A).
  • domain assumption GRU-predicted activations remain on the policy's feasible latent manifold.
    No constraint ensures in-manifold validity; off-manifold edits are shown to be catastrophic in ablations (67,951 collisions for Z1_All, Sec. V-E.1).
  • ad hoc to paper Amplifying pre-collision latent evolution induces earlier and safer avoidance.
    The central hypothesis of the paper, stated in Sec. I and IV-C; not directly measured, only inferred from final safety metrics.
  • domain assumption Action-free latent prediction with n=3 history suffices to capture latent evolution toward collisions.
    The LCWM ignores actions and uses only a short latent history, relying on the recent action-free world model idea (ref [41]).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Activation Editing: Inference-Time Refinement of Learned Policies for Safer Multirobot Navigation." pith.science (2026). https://pith.science/paper/JQBUEVM6

@misc{pith2026250920623,
  author       = {Pith},
  title        = {Pith review of: Latent Activation Editing: Inference-Time Refinement of Learned Policies for Safer Multirobot Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JQBUEVM6}},
  note         = {Machine review of arXiv:2509.20623}
}
read the original abstract

Reinforcement learning has enabled significant progress in complex domains such as coordinating and navigating multiple quadrotors. However, even well-trained policies remain vulnerable to collisions in obstacle-rich environments. Addressing these infrequent but critical safety failures through retraining or fine-tuning is costly and risks degrading previously learned skills. Inspired by activation steering in large language models and latent editing in computer vision, we introduce a framework for inference-time Latent Activation Editing (LAE) that refines the behavior of pre-trained policies without modifying their weights or architecture. The framework operates in two stages: (i) an online classifier monitors intermediate activations to detect states associated with undesired behaviors, and (ii) an activation editing module that selectively modifies flagged activations to shift the policy towards safer regimes. In this work, we focus on improving safety in multi-quadrotor navigation. We hypothesize that amplifying a policy's internal perception of risk can induce safer behaviors. We instantiate this idea through a latent collision world model trained to predict future pre-collision activations, thereby prompting earlier and more cautious avoidance responses. Extensive simulations and real-world Crazyflie experiments demonstrate that LAE achieves statistically significant reduction in collisions (nearly 90% fewer cumulative collisions compared to the unedited baseline) and substantially increases the fraction of collision-free trajectories, while preserving task completion. More broadly, our results establish LAE as a lightweight paradigm, feasible on resource-constrained hardware, for post-deployment refinement of learned robot policies.

Figures

Figures reproduced from arXiv: 2509.20623 by the authors.

Figure 1
Figure 1. (a) Conceptual overview of LAE. An online behavior classifier monitors the intermediate latents activation Z of a frozen RL policy. Safe activations pass unchanged, while unsafe ones are replaced by edited surrogate activation Z ′ generated by the activation editing module, without modifying policy weights. (b) Real-world quadrotor navigation illustrating LAE behavior. Without LAE, the RL policy collides with an obs… view at source ↗
Figure 2
Figure 2. Overview of LAE integrated on a pre-trained multi-quadrotor [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Quantitative comparison of the base RL policy with and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Representative trajectory comparison. Left: base RL policy (no LAE), which collides with obstacles. Right: RL policy with LAE, which avoids collisions while remaining identical to the base policy in safe regions and still reaching the goals. Trajectories are coloured b…
Figure 6
Figure 6. Figure 6: Choice of Zt : Comparing Z 1 All, Z 2 All, and Z 1 Partial. Z 1 Partial offers the best safety–performance trade-off. time deployment on Crazyflie hardware. These results high￾light that while both architectures can realize the LCWM hypothesis, lightweight RNN such as …
Figure 8
Figure 8. Figure 8: Choice of prediction horizon m for LCWM. 3) Choice of Prediction Horizon m: The LCWM operates by predicting future latent activations over a fixed prediction horizon m, which specifies how many steps into the future the model predicts. A sufficient lookahead is require…
Figure 9
Figure 9. Figure 9: Real-world deployment with 4 Crazyflie quadrotors navigating among cylindrical obstacles (bilateral crossing). Left: with the baseline RL policy, Drone 1 collides with an obstacle, leading to task failure. Right: with LAE enabled, all drones avoid collisions and reach …
Figure 10
Figure 10. Figure 10: Real-world deployment with 4 Crazyflie quadrotors navigating among cylindrical obstacles (four-way crossing). Left: with the baseline RL policy, Drones 1, 2, and 4 collide with obstacles, leading to task failure. Right: with LAE enabled, all drones avoid collisions an…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. World-Task Factorization for Robot Learning

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    Introduces world-task factorization for robot policies using Bayesian evidence and AICON graph plus learned modulator, outperforming baselines with zero-shot generalization in heterogeneous robotics settings.

  2. Inference-Time Policy Alignment for Fair Reinforcement Learning

    cs.LG 2026-07 reject novelty 5.0 of 10

    A frozen RL policy can be reweighted at test time by a learned generalized-Gini welfare critic to improve fairness metrics, though the central equivalence mixes up two different welfare objectives.

Reference graph

Works this paper leans on

44 extracted references · 19 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Recent advances in robot learning from demonstration,

    H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard, “Recent advances in robot learning from demonstration,”Annual review of control, robotics, and autonomous systems, vol. 3, no. 1, 2020

    2020

  2. [2]

    Collision avoidance and navigation for a quadrotor swarm using end-to-end deep reinforcement learning,

    Z. Huang, Z. Yang, R. Krupani, B. S ¸enbas ¸lar, S. Batra, and G. S. Sukhatme, “Collision avoidance and navigation for a quadrotor swarm using end-to-end deep reinforcement learning,” inIEEE Int. Conf. Robot. Autom. (ICRA), 2024

    2024

  3. [3]

    Decentralized control of quadrotor swarms with end-to- end deep reinforcement learning,

    S. Batra, Z. Huang, A. Petrenko, T. Kumar, A. Molchanov, and G. S. Sukhatme, “Decentralized control of quadrotor swarms with end-to- end deep reinforcement learning,” inConf. on robot learning, 2022

    2022

  4. [4]

    Deep reinforcement learning for robotics: A survey of real- world successes,

    C. Tang, B. Abbatematteo, J. Hu, R. Chandra, R. Mart ´ın-Mart´ın, and P. Stone, “Deep reinforcement learning for robotics: A survey of real- world successes,”arXiv preprint arXiv:2408.03539, 2024

    arXiv 2024

  5. [5]

    The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery

    Z. C. Lipton, “The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery.” Queue, vol. 16, no. 3, pp. 31–57, 2018

    2018

  6. [6]

    Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,

    A. Rajeswaran, V . Kumar, A. Gupta, G. Vezzani, J. Schulman, E. Todorov, and S. Levine, “Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,”arXiv preprint arXiv:1709.10087, 2017

    arXiv 2017

  7. [7]

    Learning to modulate pre-trained models in rl,

    T. Schmied, M. Hofmarcher, F. Paischer, R. Pascanu, and S. Hochre- iter, “Learning to modulate pre-trained models in rl,”Advances in Neural Information Processing Systems, vol. 36, 2023

    2023

  8. [8]

    Fine-tuning reinforcement learning models is secretly a forgetting mitigation problem,

    M. Wolczyk, B. Cupial, M. Ostaszewski, M. Bortkiewicz, M. Zajac, R. Pascanu, L. Kucinski, and P. Milos, “Fine-tuning reinforcement learning models is secretly a forgetting mitigation problem,”arXiv preprint arXiv:2402.02868, 2024

    arXiv 2024

Show all 44 references
  1. [9]

    Deep reinforcement learning that matters,

    P. Henderson, R. Islam, P. Bachman, J. Pineau, D. Precup, and D. Meger, “Deep reinforcement learning that matters,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

    2018

  2. [10]

    Challenges of real- world reinforcement learning,

    G. Dulac-Arnold, D. Mankowitz, and T. Hester, “Challenges of real- world reinforcement learning,”arXiv:1904.12901, 2019

    1904 arXiv

  3. [11]

    Steering language models with activation engi- neering,

    A. M. Turner, L. Thiergart, G. Leech, D. Udell, J. J. Vazquez, U. Mini, and M. MacDiarmid, “Steering language models with activation engi- neering,”arXiv preprint arXiv:2308.10248, 2023

    2023 arXiv

  4. [12]

    Representation engineering: A top-down approach to ai transparency,

    A. Zou, L. Phan, S. Chen, J. Campbell, P. Guo, R. Ren, A. Pan, X. Yin, M. Mazeika, A.-K. Dombrowskiet al., “Representation engineering: A top-down approach to ai transparency,”arXiv preprint arXiv:2310.01405, 2023

    2023 arXiv

  5. [13]

    Templeton,Scaling monosemanticity: Extracting interpretable fea- tures from claude 3 sonnet

    A. Templeton,Scaling monosemanticity: Extracting interpretable fea- tures from claude 3 sonnet. Anthropic, 2024

    2024

  6. [14]

    Ganspace: Dis- covering interpretable gan controls,

    E. H ¨ark¨onen, A. Hertzmann, J. Lehtinen, and S. Paris, “Ganspace: Dis- covering interpretable gan controls,”Advances in neural information processing systems, vol. 33, pp. 9841–9850, 2020

    2020

  7. [15]

    Sdedit: Guided image synthesis and editing with stochastic differen- tial equations,

    C. Meng, Y . He, Y . Song, J. Song, J. Wu, J.-Y . Zhu, and S. Ermon, “Sdedit: Guided image synthesis and editing with stochastic differen- tial equations,”arXiv preprint arXiv:2108.01073, 2021

    2021 arXiv

  8. [16]

    World models,

    D. Ha and J. Schmidhuber, “World models,”arXiv:1803.10122, 2018

    2018 arXiv

  9. [17]

    Learning latent dynamics for planning from pixels,

    D. Hafner, T. Lillicrap, I. Fischer, R. Villegas, D. Ha, H. Lee, and J. Davidson, “Learning latent dynamics for planning from pixels,” in Proc. ICML, 2019

    2019

  10. [18]

    Dream to control: Learning behaviors by latent imagination,

    D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi, “Dream to control: Learning behaviors by latent imagination,”arXiv:1912.01603, 2019

    1912 arXiv

  11. [19]

    Efficient estimation of word representations in vector space,

    T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of word representations in vector space,”arXiv preprint arXiv:1301.3781, 2013

    2013 arXiv

  12. [20]

    Learning to generate reviews and discovering sentiment,

    A. Radford, R. Jozefowicz, and I. Sutskever, “Learning to generate reviews and discovering sentiment,”arXiv:1704.01444, 2017

    2017 arXiv

  13. [21]

    Emerging properties in self-supervised vision trans- formers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision trans- formers,” inProc. IEEE/CVF int. conf. on computer vision, 2021

    2021

  14. [22]

    Understanding intermediate layers using linear classifier probes,

    G. Alain and Y . Bengio, “Understanding intermediate layers using linear classifier probes,”arXiv preprint arXiv:1610.01644, 2016

    2016 arXiv

  15. [23]

    Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav),

    B. Kim, M. Wattenberg, J. Gilmer, C. Cai, J. Wexler, F. Viegaset al., “Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav),” inProc. ICML, 2018

    2018

  16. [24]

    Interfacegan: Interpreting the disentangled face representation learned by gans,

    Y . Shen, C. Yang, X. Tang, and B. Zhou, “Interfacegan: Interpreting the disentangled face representation learned by gans,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 4, 2020

    2020

  17. [25]

    Prompt-to-prompt image editing with cross attention control.(2022),

    A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y . Pritch, and D. Cohen-Or, “Prompt-to-prompt image editing with cross attention control.(2022),”URL https://arxiv. org/abs/2208.01626, vol. 3, 2022

    2022 arXiv

  18. [26]

    P- tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks,

    X. Liu, K. Ji, Y . Fu, W. L. Tam, Z. Du, Z. Yang, and J. Tang, “P- tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks,”arXiv:2110.07602, 2021

    2021 arXiv

  19. [27]

    Plug and play language models: A simple approach to controlled text generation,

    S. Dathathri, A. Madotto, J. Lan, J. Hung, E. Frank, P. Molino, J. Yosinski, and R. Liu, “Plug and play language models: A simple approach to controlled text generation,” 2020

    2020

  20. [28]

    Locating and editing factual associations in gpt,

    K. Meng, D. Bau, A. Andonian, and Y . Belinkov, “Locating and editing factual associations in gpt,” 2023

    2023

  21. [29]

    Elad,Sparse and redundant representations: from theory to appli- cations in signal and image processing

    M. Elad,Sparse and redundant representations: from theory to appli- cations in signal and image processing. Springer, 2010

    2010

  22. [30]

    Toy models of superposition,

    N. Elhage, T. Hume, C. Olsson, N. Schiefer, T. Henighan, S. Kravec, Z. Hatfield-Dodds, R. Lasenby, D. Drain, C. Chenet al., “Toy models of superposition,”arXiv preprint arXiv:2209.10652, 2022

    2022 arXiv

  23. [31]

    Sparse autoencoders find highly interpretable features in language models,

    H. Cunningham, A. Ewart, L. Riggs, R. Huben, and L. Sharkey, “Sparse autoencoders find highly interpretable features in language models,”arXiv preprint arXiv:2309.08600, 2023

    2023 arXiv

  24. [32]

    Ls3: Latent space safe sets for long-horizon visuomotor control of sparse reward iterative tasks,

    A. Wilcox, A. Balakrishna, B. Thananjeyan, J. E. Gonzalez, and K. Goldberg, “Ls3: Latent space safe sets for long-horizon visuomotor control of sparse reward iterative tasks,” inConf. on Rob. Learn., 2022

    2022

  25. [33]

    Latent safety- constrained policy approach for safe offline reinforcement learning,

    P. Koirala, Z. Jiang, S. Sarkar, and C. Fleming, “Latent safety- constrained policy approach for safe offline reinforcement learning,” arXiv preprint arXiv:2412.08794, 2024

    2024

  26. [34]

    Safe reinforcement learning from pixels using a stochastic latent represen- tation,

    Y . Hogewind, T. D. Simao, T. Kachman, and N. Jansen, “Safe reinforcement learning from pixels using a stochastic latent represen- tation,”arXiv preprint arXiv:2210.01801, 2022

    2022 arXiv

  27. [35]

    Generalizing safety beyond collision-avoidance via latent-space reachability analysis,

    K. Nakamura, L. Peters, and A. Bajcsy, “Generalizing safety beyond collision-avoidance via latent-space reachability analysis,” arXiv:2502.00935, 2025

    2025

  28. [36]

    A scalable dis- tributed collision avoidance scheme for multi-agent uav systems,

    B. Lindqvist, P. Sopasakis, and G. Nikolakopoulos, “A scalable dis- tributed collision avoidance scheme for multi-agent uav systems,” in IEEE/RSJ int. conf. on int. robots and systems (IROS), 2021

    2021

  29. [37]

    Ego-swarm: A fully autonomous and decentralized quadrotor swarm system in cluttered environments,

    X. Zhou, J. Zhu, H. Zhou, C. Xu, and F. Gao, “Ego-swarm: A fully autonomous and decentralized quadrotor swarm system in cluttered environments,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), 2021

    2021

  30. [38]

    Safety barrier certificates for heterogeneous multi-robot systems,

    L. Wang, A. Ames, and M. Egerstedt, “Safety barrier certificates for heterogeneous multi-robot systems,” inAmer. cont. conf. (ACC), 2016

    2016

  31. [39]

    Gcbf+: A neural graph control barrier function framework for distributed safe multi-agent control,

    S. Zhang, O. So, K. Garg, and C. Fan, “Gcbf+: A neural graph control barrier function framework for distributed safe multi-agent control,” IEEE Transactions on Robotics, 2025

    2025

  32. [40]

    Quadswarm: A modular multi-quadrotor simulator for deep reinforcement learning with direct thrust control,

    Z. Huang, S. Batra, T. Chen, R. Krupani, T. Kumar, A. Molchanov, A. Petrenko, J. A. Preiss, Z. Yang, and G. S. Sukhatme, “Quadswarm: A modular multi-quadrotor simulator for deep reinforcement learning with direct thrust control,”arXiv preprint arXiv:2306.09537, 2023

    2023 arXiv

  33. [41]

    Learning to act without actions,

    D. Schmidt and M. Jiang, “Learning to act without actions,”arXiv preprint arXiv:2312.10812, 2023

    2023 arXiv

  34. [42]

    Sample factory: Egocentric 3d control from pixels at 100000 fps with asynchronous reinforcement learning,

    A. Petrenko, Z. Huang, T. Kumar, G. Sukhatme, and V . Koltun, “Sample factory: Egocentric 3d control from pixels at 100000 fps with asynchronous reinforcement learning,” inProc. ICML, 2020

    2020

  35. [43]

    Parametric umap embeddings for representation and semisupervised learning,

    T. Sainburg, L. McInnes, and T. Q. Gentner, “Parametric umap embeddings for representation and semisupervised learning,”Neural Computation, vol. 33, no. 11, pp. 2881–2907, 2021

    2021

  36. [44]

    Barlow twins: Self-supervised learning via redundancy reduction,

    J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny, “Barlow twins: Self-supervised learning via redundancy reduction,” inProc. ICML, 2021

    2021

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.