Pith. sign in

REVIEW 3 major objections 4 minor 63 references

Suppression Sticks, Locality Is Fragile: A Closed-Loop Target-and-Control Audit of Task-Vector Negation in VLA Policies

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Subtracting a single skill vector from a multitask robot policy suppresses the target skill in every held-out closed-loop test, yet preserves unrelated control skills only partially—mean retention 52%, worst case 4%.

desk verdict A credible first closed-loop audit of task-vector negation in VLA policies, with an honest evaluation design; the headline numbers rest on one seed but the qualitative result should survive multi-seed replication. read the letter →

arxiv 2608.04692 v1 pith:IIZRRAB7 submitted 2026-08-05 cs.RO cs.LG

classification cs.ROcs.LG
keywords task-vectorarithmeticmodeleditingmachineunlearningvision-language-actionpoliciesclosed-loopevaluationbehaviorallocalityLIBERObenchmarkmasking
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

Task-vector subtraction is a one-line weight edit: take the difference between a single-skill expert and a shared base model, and subtract it from a multitask vision-language-action policy to remove that skill without retraining. This paper asks whether the edit is local—whether suppressing one behavior leaves the other nine intact—and answers by rolling out the target and every control skill after each edit, in closed loop, on the LIBERO-Goal benchmark. The target side is robust: five of ten skills drop to at most 15% success after subtraction and, at their chosen operating strengths, all five stay at 0/20 on held-out initial states. The control side is not: baseline-normalized control retention averages 52%, from 78% for the cleanest edit down to 4% for the worst, every edit materially harms at least one nominally unrelated skill, three skills resist suppression entirely, and two edits collapse every control. The authors conclude that closed-form subtraction is a fast but brittle intervention, and that only closed-loop target-and-control evaluation can measure locality in embodied model editing.

What carries the argument

The central object is the per-skill task vector $\tau_i = \theta_{\mathrm{exp},i} - \theta_{\mathrm{base}}$, the weight delta between a single-skill expert and a shared base, subtracted from the multitask goal checkpoint as $\theta_{\mathrm{edit}} = \theta_{\mathrm{Goal}} - \alpha \tau_i$. The load-bearing instrument is the closed-loop audit: a full 10×10 negation matrix on LIBERO-Goal in which each negated vector is rolled out against its own target and all nine controls at 20 trials per cell, plus a held-out state split (initial states 20–39, disjoint from the 0–19 used to select each skill's strength $\alpha^*$) that separates target suppression from control retention. A pre-specified taxonomy classifies each row as separation (target at or below 20% with mean controls at or above 40%), resistant, or global collapse. Supporting probes include the norm-matched sign comparison (the identical vector added versus subtracted at matched $\ell^2$ norm), a forward-only alignment score that ranks within-suite suppressibility without seeing any rollout outcome, and a relearning probe that fine-tunes the negated checkpoint against a task-naive base.

What would settle it

Retrain all ten LIBERO-Goal experts under five independent seeds and recompute the full 10×10 negation matrix with per-skill $\alpha^*$ operating points on held-out states; the quantitative claim weakens if any separation row becomes resistant or if mean held-out control retention moves well above 52%. A complementary check is running the cleanest edit (t0 at $\alpha^* = 0.75$) on a physical robot: if the target is not suppressed, or if control retention diverges sharply from the simulated 78%, the closed-loop audit does not survive contact with hardware.

Watch

Extended reading notes

Core claim

The paper's central claim is that single-vector task-vector subtraction in a vision-language-action policy delivers robust target suppression but fragile, heterogeneous locality. On LIBERO-Goal, subtracting one skill's vector produces three qualitatively different regimes: five skills show target–control separation (target at or below 15% success, controls above a 40% threshold), three skills resist suppression, and two edits suppress the target while collapsing every control. At per-skill operating strengths $\alpha^*$ re-evaluated on held-out initial states, all five suppressible targets stay at 0/20 success, but baseline-normalized mean control retention is 52%, ranging from 78% (t0) down to 4% (t6), with four of five rows zeroing at least one unrelated control and the fifth depressing its worst control from a ≥85% baseline to 35%. The separation pattern recurs across continuous-regression, discrete-token, and flow-matching action heads, while mean task-vector cosine (0.011–0.014 on the tested suites) predicts neither suppression nor control survival; multi-vector composition collapses the policy at $k \geq 2$ even at norm-matched scales. The authors read this as evidence that closed-loop target-and-control behavior, not static loss or weight-space geometry, is the grounded measure of locality for embodied model editing.

Load-bearing premise

The headline numbers—the 5/3/2 regime split and the 52% mean held-out retention—rest on task vectors trained from a single checkpoint with one expert-training seed, while all rollouts come from LIBERO MuJoCo simulation rather than physical hardware.

Editorial extensions

If this is right

  • Target-only success metrics are misleading: all five held-out edits score as perfect removals (0/20 targets), yet the control panel shows worst-case retention of 4% and four of five edits zeroing at least one unrelated control.
  • Single-skill subtraction does not compose: norm-matched cumulative negation of two or more suppressible skills drops control success to 10% at $k=2$ and to 0% for $k \geq 3$, so multi-skill removal needs a different mechanism.
  • Locality is a behavioral, closed-loop quantity: mean task-vector cosine and static-loss diagnostics do not track which controls survive, so embodied model editing must be audited by rollout, not by loss or weight geometry.
  • The edit is fast but reversible: subtraction needs no removal-time data or gradients (given precomputed expert deltas), but the relearning probe shows the suppressed skill returns to 85% within 250 fine-tuning steps while a task-naive base stays at 0%—behavioral masking, not durable erasure.
  • The qualitative pattern transfers across action-head families—continuous regression, discrete tokens, flow matching—but not across task families with shared motor primitives, where controls collapse with the target.

Reading between the lines

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

  • Read operationally, the 52% average hides a worse typical case: four of five separation rows zeroed at least one control, so a practitioner should plan for one or more controls at 0% success and treat the mean retention as the optimistic summary.
  • The paper's own three-seed probe shows the suppression threshold $\alpha^*$ is run-dependent (one independently retrained t0 expert kept 70% target success at $\alpha=1.0$); a real deployment would need per-seed calibration, partially eroding the 'closed-form, CPU-in-a-second' advantage at the point of use.
  • The component ablation (VLM-only negation reaches 0% target with better control retention than the full edit) suggests an untested extension: spreading the subtraction across components at per-component scales could improve the locality–suppression trade-off beyond what the paper reports.
  • The audit scores single episodes; because the closed-loop argument says errors compound through future observations, a natural next test is whether control damage grows over longer or repeated rollouts even where single-episode retention looks acceptable.
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 / 4 minor

Summary. The paper audits single-skill task-vector subtraction from MergeVLA-Goal and other VLA policies by closed-loop rollouts of the target skill and every control skill. It reports a 5/3/2 regime split on LIBERO-Goal (separation, resistance, global collapse), a held-out validation showing 0/20 target suppression but only 52% mean baseline-normalized control retention, cross-suite and cross-architecture transfer results, negative results for mean weight-space cosine as a predictor, a forward-only Alignment Score diagnostic, gradient-baseline comparisons, and a relearning probe consistent with behavioral masking. The central claim is that single-vector subtraction yields robust target suppression with fragile, heterogeneous locality, and that closed-loop target-and-control evaluation is the appropriate lens for embodied model editing.

Significance. If the quantitative results were seed-stable, this would be a valuable first closed-loop audit of task-vector negation in embodied policies, with a clean selection/evaluation split, raw x/n counts with asymmetric Wilson intervals, a sensible minimum reporting unit, and explicitly scoped negative results. The paper is also notable for shipping code and data, for being transparent about the limits of its diagnostics, and for distinguishing descriptive taxonomy from confirmatory inference. However, the headline numbers are single-seed, and the authors' own three-seed probe shows seed-dependent suppression thresholds; the contribution is therefore currently a well-executed case study rather than an established empirical law. The qualitative 'fragile locality' message may survive replication, but the specific 5/3/2 split and the 52% mean retention need either multi-seed support or a prominent single-checkpoint qualifier.

major comments (3)
  1. [Abstract; Experiments (Complete LIBERO-Goal Matrix); Discussion and Limitations item 1] The paper's headline quantitative claims—the 5/3/2 regime split in Fig. 2/Table 3 and the held-out retention panel in Table 1 (mean 52%, t6 4%)—are computed from task vectors trained from a single MergeVLA-Goal checkpoint with one expert-training seed. The three-seed probe in Supp. Table S17 shows that at fixed α=1.0 an independently retrained t0 expert leaves the target at 70% (seed 2), which under the paper's own taxonomy would move that row from 'suppressed' to 'resistant'; the suppression threshold is run-dependent. The abstract and introduction state 'Across all ten LIBERO-Goal skills...' and 'all five suppressible targets remain at 0%' without a seed qualifier, whereas Discussion/Limitations item 1 disclaims seed robustness 'for the matrices.' Because the three-regime map and the 52% mean retention are the central quantitative contributions, the claims need either multi-seed replication of the full matrix and held-out panel (with seed-specific splits and retention distributions) or an explicit, prominent reframing of all headline numbers as single-checkpoint observations. This is a load-bearing point, not a routine caveat.
  2. [Experiments (Constrained α Selection); Supp. Tables S11–S13] The selection rule is stated as 'smallest α with target SR = 0%, then highest mean control SR' (Setup/Experiments), but for t5 and t6 the α⋆ values in Table 1 (1.0 and 1.25) are taken from the diagonal α-robustness grid without a development-panel control-retention measurement at the candidate α values; Supp. Table S12 labels these 'diagonal-grid selection' while the dev panel is measured only for t0, t2, t3. This matters for the headline held-out numbers: at α=1.0 the t6 row has 63.9% normalized control retention (Supp. Table S14), and an unsampled α between 1.0 and 1.25 might suppress the target with substantially higher control retention, changing the 52% mean and the '4% floor' result. The authors should either report the development-panel control SR at every candidate α for all five suppressible skills and re-run the held-out panel at the resulting α⋆, or justify why the coarser diagonal grid is sufficient for the stated rule.
  3. [Experiments (Complete LIBERO-Goal Matrix); Table 3] The 'three resistant' classification depends on a single 20-rollout cell: t4 has target SR 30% after negation, above the pre-specified ≤20% threshold but well within sampling noise of that threshold (a 6/20 count has a 95% Wilson interval straddling 20%). If t4 were reclassified as suppressed, its mean control SR of 10% (Supp. Table S4) would place it in global collapse, changing the taxonomy from 5/3/2 to 5/2/3. The manuscript reports threshold sensitivity for the separation cutoff (Supp. Table S5) but not for the target-suppression cutoff; I ask for a target-threshold sensitivity sweep and per-cell Wilson intervals for Table 3, or an explicit statement that the resistant class is robust to this cell's uncertainty.
minor comments (4)
  1. [Introduction vs. Table 1] The Introduction quotes 57.8% in-sample control retention while Table 1 reports 52% held-out retention; Supp. Table S14 is a helpful crosswalk, but a one-sentence note in the main text clarifying that the two numbers use different panels (fixed α on selection states vs. α⋆ on held-out states) would prevent confusion.
  2. [Figure 2 caption] The caption says 'Terracotta outlines mark the diagonal target cells,' but the regime labels are encoded by shape and color; consider adding a legend or explicit outline key inside the figure itself for accessibility.
  3. [Supp. Table S17] The parenthetical about the third run's step-4000 stdout surviving while the step-10000 checkpoint is used is confusing; clarify which checkpoint is actually evaluated and why the stdout mismatch does not affect the reported numbers.
  4. [Supp. Table S30] The relearning table shows the task-naive base floor eventually recovers by k=2000; a sentence noting this would help readers interpret the early divergence as a slope difference rather than a permanent gap.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the audit's central claims are genuine held-out transfer results, and its forward-only diagnostic is validated against outcomes it does not use.

full rationale

The central claim (robust target suppression with fragile, heterogeneous locality) rests on a held-out evaluation: α⋆ is selected on rollouts over initial states 0–19 and re-evaluated on disjoint states 20–39 (Table 1, Supp. Table S13), so the 0/20 target suppression and 52% mean normalized control retention are not fitted values. The 5/3/2 taxonomy is a descriptive classification of the full 10×10 matrix with a pre-specified threshold rule (target ≤20%, mean control ≥40%), and the paper explicitly labels post-hoc subgroup statistics as descriptive rather than confirmatory. The Alignment Score AS_i is forward-only and uses no negation outcome (Listing S1, Eq. 4); its within-suite ranking (Spearman 0.72–0.75) is a genuine non-circular prediction that the paper honestly demotes when its threshold fails to transfer across suites. Self-citations (Guo et al. 2025; Wang et al. 2026a,b; Chen et al. 2026; Li et al. 2026) appear only in Related Work as orthogonal input-side efficiency methods and are not load-bearing. The single-seed limitation and simulation-only scope are explicitly disclosed and affect external validity/robustness, not circularity. No equation or claim reduces to its own inputs by construction.

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

The central claim rests on three domain assumptions the paper states but does not prove: that LoRA expert deltas from a shared base are meaningful skill directions (violated by the t1 direction-misalignment anomaly), that LIBERO MuJoCo rollouts proxy real embodied behavior (unvalidated, explicitly listed as a gap), and that single-seed expert training is representative (partially contradicted by the paper's own three-seed probe, Supp. Table S17). Free parameters are limited to the per-skill edit scale alpha-star (outcome-fitted on rollouts, held-out validated) and pre-specified analysis thresholds. No invented entities: the Alignment Score and the three-regime taxonomy are diagnostic constructs with behavioral falsifiable handles, not postulated entities.

free parameters (4)
  • per-skill edit strength alpha-star = 0.75 (t0), 1.0 (t2/t3/t5), 1.25 (t6); alpha=1.0 for the full 10x10 matrix
    Selected by a fixed rule on rollouts (smallest alpha with target SR=0, then highest mean control SR) on states 0-19; outcome-fitted, validated on disjoint states 20-39, so the '0% target' headline is at a tuned operating point.
  • regime taxonomy thresholds = target <=20% suppressed; mean control >=40% separation
    Pre-specified decision rule for the three regimes; the paper shows the 5/3/2 split is invariant for cutoffs 30-40% (Supp. Table S5), so the split is not a knife-edge.
  • Alignment Score probe scale epsilon = epsilon in [0.1, 0.5], averaged
    Diagnostic hyperparameter of the forward-only AS_i; reported as a within-suite ranking (Spearman 0.72 at n=10), not a portable threshold.
  • interference-graph edge threshold = >=50 percentage-point control-SR drop
    Choice for the directed collateral graph (Fig. 4A); descriptive only and does not enter the headline numbers.
assumptions (4)
  • domain assumption Per-skill LoRA fine-tunes from a shared base (theta_base, fixed shared init seed 42) yield task vectors tau_i = theta_exp,i - theta_base whose subtraction from theta_Goal edits a single behavior
    Setup, Eq. (1); the whole audit treats expert deltas as skill directions. The t1 direction-misalignment anomaly (Supp. Sec. G) shows the assumption is not uniformly satisfied.
  • domain assumption LIBERO closed-loop success rate in MuJoCo is a valid proxy for the presence or absence of embodied skills
    All evaluation uses LIBERO rollouts; real-robot and non-LIBERO validation are explicitly listed as gaps (Discussion, Limitations item 2).
  • domain assumption A single expert-training seed is representative for the quantitative claims
    Full 10x10 matrices use one seed; the three-seed probe (Supp. Table S17) shows run-dependent suppression thresholds, so this assumption is partially violated for the numbers.
  • ad hoc to paper The Alignment Score AS_i, a forward-only hidden-state shift ratio at a small probe step, ranks target-suppression susceptibility
    Supp. Eq. (4), Supp. Sec. E; the paper demotes it to a within-suite ranking because its threshold inverts on LIBERO-Spatial (Supp. Fig. S7).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Suppression Sticks, Locality Is Fragile: A Closed-Loop Target-and-Control Audit of Task-Vector Negation in VLA Policies." pith.science (2026). https://pith.science/paper/IIZRRAB7

@misc{pith2026260804692,
  author       = {Pith},
  title        = {Pith review of: Suppression Sticks, Locality Is Fragile: A Closed-Loop Target-and-Control Audit of Task-Vector Negation in VLA Policies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IIZRRAB7}},
  note         = {Machine review of arXiv:2608.04692}
}
read the original abstract

Task-vector arithmetic offers a closed-form way to modify a model, yet its behavioral locality remains unclear in closed-loop robot control. We present a target-and-control audit of per-skill task-vector subtraction from multitask vision-language-action (VLA) policies. Across all ten LIBERO-Goal skills, subtraction produces three qualitatively different regimes: target-control separation for five skills, resistance for three, and global collapse for two. On held-out initial states, the five suppressible targets remain at 0% success; however, mean baseline-normalized control retention is only 52%, and each target-suppressing edit materially harms at least one nominally unrelated control. Additional Goal panels show separation across tested policies with continuous-regression, discrete-token, and flow-matching action heads, whereas we observe no clean separation on Spatial and control collapse on the tested Object and Long-horizon panels. Mean task-vector cosine does not account for this variation. A matched-norm control identifies a local sign asymmetry around one Goal anchor, while multi-vector outcomes vary with anchor and scale. Retain-aware gradient baselines provide data-dependent comparators but require removal-time data and optimization; subtraction is data- and gradient-free only at edit time, assuming precomputed expert deltas. Finally, a single-skill relearning probe is consistent with behavioral masking, not certified unlearning. These results characterize task-vector subtraction as a fast but brittle intervention and underscore the need for closed-loop target-and-control evaluation when assessing locality in embodied model editing.

Figures

Figures reproduced from arXiv: 2608.04692 by the authors.

Figure 1
Figure 1. A single edit-time subtraction removes the target skill, but its locality is fragile and heterogeneous. We derive an expert update from a single-skill expert and subtract it from a multitask VLA in one edit-time operation, then roll out the target and every control behavior. All frames are real MergeVLA LIBERO-Goal rollouts at α=0.75: the t0 target opens the drawer without the edit and fails after subtraction (same … view at source ↗
Figure 2
Figure 2. Full ten-skill audit reveals separation, resistance, and global collapse. Complete LIBERO-Goal negation matrix in natural task order (row = negated τ i , column = evaluated skill; closed-loop SR%, 20 rollouts/cell, α=1.0; no-edit baselines above). Terracotta outlines mark the diagonal target cells (a low value indicates suppression; the box does not imply success). The right strip labels each row’s regime by color a… view at source ↗
Figure 3
Figure 3. The tested t0 outcomes depend on sign, anchor, and aggregate scale. Closed-loop SR of target t0 and con￾trol t1 (20 rollouts/cell; relative perturbation norm ×∥τ 0∥ above each pair). Goal-anchored subtraction shows separation (0/100); the equal-norm addition is retained (50/90); the 0.71× goal-anchored mean is retained (90/100). The 1.4× goal-anchored sum nearly collapses the pair (0/5), whereas the base-anchored me… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Collateral damage is structured, whereas mean weight-space cosine does not track control survival. (A) Directed Goal interference graph on the 8-skill core (excluding the two global-collapse rows t8, t9): each edge i→j marks a ≥50-point control-SR drop from subtracting…
Figure 5
Figure 5. Figure 5: Target suppression recurs across three action-head families, but control retention remains model-dependent. Rows subtract τ i ; columns evaluate skills; red boxes mark target cells, and no-edit baselines appear above each matrix. MergeVLA t1 resists suppression; OpenVL…
Figure 6
Figure 6. Figure 6: Observed closed-loop collapse along the tested edit paths from the goal anchor. Closed-loop success rate of the edited Skill A (t0, blue circles) and the retained Skill B (t2, orange squares) as we interpolate from θGoal along the negation direction −τ A (left) and the…
Figure 7
Figure 7. Figure 7: Locality is a cliff, not a floor. Held-out control retention across the five suppressible Goal targets at their selected α ⋆ (bars: baseline-normalized nine-control mean; diamonds: worst single control); values match [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 48 canonical work pages

  1. [1]

    Arditi, A.; Obeso, O.; Syed, A.; Paleka, D.; Panickssery, N.; Gurnee, W.; and Nanda, N. 2024. Refusal in Language Models Is Mediated by a Single Direction. In Advances in Neural Information Processing Systems (NeurIPS)

  2. [2]

    Black, K.; Brown, N.; Driess, D.; et al. 2024. _0 : A Vision-Language-Action Flow Model for General Robot Control. arXiv preprint arXiv:2410.24164

  3. [3]

    A.; Jia, H.; Travers, A.; Zhang, B.; Lie, D.; and Papernot, N

    Bourtoule, L.; Chandrasekaran, V.; Choquette-Choo, C. A.; Jia, H.; Travers, A.; Zhang, B.; Lie, D.; and Papernot, N. 2021. Machine Unlearning. In IEEE Symposium on Security and Privacy

  4. [4]

    Brohan, A.; Brown, N.; Carbajal, J.; et al. 2023. RT-2 : Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. In Conference on Robot Learning (CoRL)

  5. [5]

    Chen, Z.; Wang, S.; Guo, W.; Cai, Q.; Zhang, H.; Li, P.; Zhao, Y.; and Guo, Y. 2026. PHASER: Phase-Aware and Semantic Experience Replay for Vision-Language-Action Models. arXiv preprint arXiv:2606.03598

  6. [6]

    S.; Tarun, A

    Chundawat, V. S.; Tarun, A. K.; Mandal, M.; and Kankanhalli, M. 2023. Zero-Shot Machine Unlearning. IEEE Transactions on Information Forensics and Security

  7. [7]

    Deeb, A.; and Roger, F. 2024. Do Unlearning Methods Remove Information from Language Model Weights? arXiv preprint arXiv:2410.08827

  8. [8]

    Dinh, L.; Pascanu, R.; Bengio, S.; and Bengio, Y. 2017. Sharp Minima Can Generalize for Deep Nets. In International Conference on Machine Learning, 1019--1028. PMLR

Show all 63 references
  1. [9]

    Eldan, R.; and Russinovich, M. 2023. Who's Harry Potter? Approximate Unlearning in LLMs . arXiv preprint arXiv:2310.02238

  2. [10]

    Fan, C.; Liu, J.; Zhang, Y.; Wong, E.; Wei, D.; and Liu, S. 2024. SalUn : Empowering Machine Unlearning via Gradient-Based Weight Saliency in Both Image Classification and Generation. In International Conference on Learning Representations (ICLR)

  3. [11]

    Foret, P.; Kleiner, A.; Mobahi, H.; and Neyshabur, B. 2021. Sharpness-Aware Minimization for Efficiently Improving Generalization. In International Conference on Learning Representations (ICLR)

  4. [12]

    K.; Roy, D.; and Carbin, M

    Frankle, J.; Dziugaite, G. K.; Roy, D.; and Carbin, M. 2020. Linear Mode Connectivity and the Lottery Ticket Hypothesis. In International Conference on Machine Learning (ICML)

  5. [13]

    Fu, Y.; Zhang, Z.; Zhang, Y.; Wang, Z.; Huang, Z.; and Luo, Y. 2026. MergeVLA : Cross-Skill Model Merging Toward a Generalist Vision-Language-Action Agent. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). a rXiv:2511.18810

  6. [14]

    Garipov, T.; Izmailov, P.; Podoprikhin, D.; Vetrov, D.; and Wilson, A. G. 2018. Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNN s. In Advances in Neural Information Processing Systems (NeurIPS)

  7. [15]

    Ginart, A.; Guan, M.; Valiant, G.; and Zou, J. 2019. Making AI Forget You: Data Deletion in Machine Learning. In NeurIPS

  8. [16]

    Golatkar, A.; Achille, A.; and Soatto, S. 2020. Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In CVPR

  9. [17]

    Graves, L.; Nagisetty, V.; and Ganesh, V. 2021. Amnesiac Machine Learning. In AAAI Conference on Artificial Intelligence

  10. [18]

    Guo, C.; Goldstein, T.; Hannun, A.; and van der Maaten, L. 2020. Certified Data Removal from Machine Learning Models. In ICML

  11. [19]

    Guo, W.; Chen, Z.; Wang, S.; He, J.; Xu, Y.; Ye, J.; Sun, Y.; and Xiong, H. 2025. Logic-in-Frames: Dynamic Keyframe Search via Visual Semantic-Logical Verification for Long Video Understanding. In Advances in Neural Information Processing Systems, volume 38, 124389--124422

  12. [20]

    Guo, W.; Zhang, H.; Li, P.; Cai, T.; Chen, Z.; Guo, Y.; He, X.; Yang, Y.; Sun, Y.; and Xiong, H. 2026. A Brain-inspired Embodied Intelligence for Fluid and Fast Reflexive Robotics Control. arXiv preprint arXiv:2601.14628

  13. [21]

    J.; Shen, Y.; Wallis, P.; et al

    Hu, E. J.; Shen, Y.; Wallis, P.; et al. 2022. LoRA : Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations (ICLR)

  14. [22]

    T.; Wortsman, M.; Gururangan, S.; Schmidt, L.; Hajishirzi, H.; and Farhadi, A

    Ilharco, G.; Ribeiro, M. T.; Wortsman, M.; Gururangan, S.; Schmidt, L.; Hajishirzi, H.; and Farhadi, A. 2023. Editing Models with Task Arithmetic. In International Conference on Learning Representations

  15. [23]

    Jin, X.; Ren, X.; Preotiuc-Pietro, D.; and Cheng, P. 2023. Dataless Knowledge Fusion by Merging Weights of Language Models. In International Conference on Learning Representations (ICLR)

  16. [24]

    Karamcheti, S.; Nair, S.; Balakrishna, A.; et al. 2024. Prismatic VLM s: Investigating the Design Space of Visually-Conditioned Language Models. In International Conference on Machine Learning (ICML)

  17. [25]

    B.; and Fitzgerald, T

    Karli, U. B.; and Fitzgerald, T. 2026. RECALL: Recovery Experience Collection for Active Lifelong Learning in Vision-Language-Action Models. arXiv preprint arXiv:2606.23617

  18. [26]

    S.; Han, D.; and Choe, J

    Kim, H. S.; Han, D.; and Choe, J. 2025. NegMerge : Consensually Negating Weights for Machine Unlearning. In ICML

  19. [27]

    J.; Finn, C.; and Liang, P

    Kim, M. J.; Finn, C.; and Liang, P. 2025. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. In Robotics: Science and Systems (RSS)

  20. [28]

    J.; Pertsch, K.; Karamcheti, S.; et al

    Kim, M. J.; Pertsch, K.; Karamcheti, S.; et al. 2024. OpenVLA : An Open-Source Vision-Language-Action Model. In Conference on Robot Learning (CoRL). a rXiv:2406.09246

  21. [29]

    Kim, Y.; Hwang, D.; Cha, S.; and Choe, J. 2026. Knowledge Vector Weakening: Efficient Training-free Unlearning for Large Vision-Language Models. arXiv preprint arXiv:2601.21794

  22. [30]

    Kurmanji, M.; Triantafillou, P.; Hayes, J.; and Triantafillou, E. 2023. Towards Unbounded Machine Unlearning. In Advances in Neural Information Processing Systems (NeurIPS)

  23. [31]

    Kwon, J.; Kim, J.; Park, H.; and Choi, I. K. 2021. ASAM : Adaptive Sharpness-Aware Minimization for Scale-Invariant Learning of Deep Neural Networks. In International Conference on Machine Learning (ICML)

  24. [32]

    Li, H.; Xu, Z.; Taylor, G.; Studer, C.; and Goldstein, T. 2018. Visualizing the Loss Landscape of Neural Nets. In Advances in Neural Information Processing Systems (NeurIPS)

  25. [33]

    Li, H.; Zhang, Y.; Zhang, S.; Wang, M.; Liu, S.; and Chen, P.-Y. 2025. When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers. International Conference on Learning Representations (ICLR)

  26. [34]

    Li, N.; Pan, A.; Gopal, A.; et al. 2024. The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning. In ICML

  27. [35]

    Li, P.; Guo, W.; Zhang, H.; Cai, T.; He, X.; Guo, Y.; and Xiong, H. 2026. Spatial Memory for Out-of-Vision Manipulation in Vision-Language-Action. arXiv preprint arXiv:2605.22283

  28. [36]

    Li, Y.; Sun, C.-E.; and Weng, T.-W. 2025. Effective Skill Unlearning through Intervention and Abstention. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)

  29. [37]

    Lin, S.; Dong, J.; Chen, R.; Zhang, X.; Xu, L.; and Chen, X. 2026. CATA : Continual Machine Unlearning via Conflict-Averse Task Arithmetic. arXiv preprint arXiv:2605.18610

  30. [38]

    Liu, B.; Liu, Q.; and Stone, P. 2022. Continual Learning and Private Unlearning. In Conference on Lifelong Learning Agents (CoLLAs)

  31. [39]

    Liu, B.; Zhu, Y.; Gao, C.; et al. 2023. LIBERO : Benchmarking Knowledge Transfer for Lifelong Robot Learning. In NeurIPS

  32. [40]

    Liu, H.; Kim, C.; Liu, B.; Liu, M.; and Zhu, Y. 2026. Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning. arXiv preprint arXiv:2603.03818

  33. [41]

    ucki, J.; Wei, B.; Huang, Y.; Henderson, P.; Tram \`e r, F.; and Rando, J. 2024. An Adversarial Perspective on Machine Unlearning for AI Safety. arXiv preprint arXiv:2409.18025

  34. [42]

    Lynch, A.; Guo, P.; Ewart, A.; Casper, S.; and Hadfield-Menell, D. 2024. Eight Methods to Evaluate Robust Unlearning in LLM s. arXiv preprint arXiv:2402.16835

  35. [43]

    C.; and Kolter, J

    Maini, P.; Feng, Z.; Schwarzschild, A.; Lipton, Z. C.; and Kolter, J. Z. 2024. TOFU : A Task of Fictitious Unlearning for LLMs . arXiv preprint arXiv:2401.06121

  36. [44]

    Matena, M.; and Raffel, C. 2022. Merging Models with Fisher-Weighted Averaging. In Advances in Neural Information Processing Systems (NeurIPS)

  37. [45]

    Meng, K.; Bau, D.; Andonian, A.; and Belinkov, Y. 2022. Locating and Editing Factual Associations in GPT . In Advances in Neural Information Processing Systems (NeurIPS)

  38. [46]

    Meng, K.; Sen Sharma, A.; Andonian, A.; Belinkov, Y.; and Bau, D. 2023. Mass-Editing Memory in a Transformer. In International Conference on Learning Representations (ICLR)

  39. [47]

    Octo Model Team ; Ghosh, D.; Walke, H.; Pertsch, K.; et al. 2024. Octo: An Open-Source Generalist Robot Policy. In Robotics: Science and Systems (RSS)

  40. [48]

    Ortiz-Jimenez, G.; Favero, A.; and Frossard, P. 2023. Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Models. In Advances in Neural Information Processing Systems

  41. [49]

    Physical Intelligence ; Black, K.; Brown, N.; Darpinian, J.; Driess, D.; Finn, C.; Levine, S.; Pertsch, K.; et al. 2025. _ 0.5 : a Vision-Language-Action Model with Open-World Generalization. arXiv preprint arXiv:2504.16054

  42. [50]

    Qu, D.; Song, H.; Chen, Q.; et al. 2025. SpatialVLA : Exploring Spatial Representations for Visual-Language-Action Model. In Robotics: Science and Systems (RSS)

  43. [51]

    Ranjan, R.; and Polyzou, A. 2026. VLA-Forget : Vision-Language-Action Unlearning for Embodied Foundation Models. In ACL 2026 Workshop on Knowledge and Foundation Models (KnowFM). a rXiv:2604.03956

  44. [52]

    Shi, W.; Lee, J.; Huang, Y.; et al. 2024. MUSE : Machine Unlearning Six-Way Evaluation for Language Models. arXiv preprint arXiv:2407.06460

  45. [53]

    Shukor, M.; Aubakirova, D.; Capuano, F.; Kooijmans, P.; Palma, S.; Zouitine, A.; et al. 2025. SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics. arXiv preprint arXiv:2506.01844

  46. [54]

    K.; Chundawat, V

    Tarun, A. K.; Chundawat, V. S.; Mandal, M.; and Kankanhalli, M. S. 2023. Fast Yet Effective Machine Unlearning. IEEE Transactions on Neural Networks and Learning Systems

  47. [55]

    Wang, S.; Guo, W.; Chen, Z.; Hu, X.; and Xiong, H. 2026 a . Where to Focus: Query-Modulated Multimodal Keyframe Selection for Long Video Understanding. arXiv preprint arXiv:2604.17422

  48. [56]

    Wang, S.; Guo, W.; Chen, Z.; Xu, Y.; Hu, X.; and Xiong, H. 2026 b . Less is More: Token-Efficient Video-QA via Adaptive Frame-Pruning and Semantic Graph Integration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 9856--9866

  49. [57]

    Wilson, E. B. 1927. Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association, 22(158): 209--212

  50. [58]

    Y.; et al

    Wortsman, M.; Ilharco, G.; Gadre, S. Y.; et al. 2022. Model Soups: Averaging Weights of Multiple Fine-Tuned Models Improves Accuracy Without Increasing Inference Time. In International Conference on Machine Learning (ICML)

  51. [59]

    Yadav, P.; Tam, D.; Choshen, L.; Raffel, C.; and Bansal, M. 2023. TIES -Merging: Resolving Interference When Merging Models. In Advances in Neural Information Processing Systems

  52. [60]

    Yang, E.; Wang, Z.; Shen, L.; et al. 2024. AdaMerging : Adaptive Model Merging for Multi-Task Learning. In International Conference on Learning Representations (ICLR)

  53. [61]

    Yao, J.; Chien, E.; Du, M.; Niu, X.; Wang, T.; Cheng, Z.; and Yue, X. 2024. Machine Unlearning of Pre-trained Large Language Models. In ACL

  54. [62]

    Yu, L.; Yu, B.; Yu, H.; Huang, F.; and Li, Y. 2024. Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch. In International Conference on Machine Learning

  55. [63]

    Zhang, R.; Lin, L.; Bai, Y.; and Mei, S. 2024. Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning. In Conference on Language Modeling (COLM)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.