acceptodds
Under review as a conference paper at ICLR 2027

The Action Doesn't End the Thought: Continuing Computation in Vision-Language-Action Models

Abstract

Most vision-language-action (VLA) models restart their action expert at every de- cision, discarding evidence the robot can no longer see and computation the next decision could reuse. We propose continuing computation: the expert carries its internal state across decisions and refines it with each new observation. We im- plement this in LycoReco, a VLA with a Continuous Thought Machine expert, and compare carry and reset experts trained identically except for state lifetime. When earlier cues differ but current inputs are identical, the carried state lets the robot act on information it no longer sees. It also saves expert computation: when replanning at every step on LIBERO, carry with one internal update per obser- vation outperforms reset with twelve. Carry’s successive action chunks are more consistent when nothing new is revealed, yet adapt when a remembered target moves. LycoReco is competitive with published VLAs on LIBERO (97.5%) and SIMPLER-WidowX (64.6%) and leads the compared systems on the MIKASA memory tasks (62.6%) without a separate memory module. State lifetime is there- fore a consequential architectural choice, and resetting the action expert at every decision should not be the default.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Robot Learning, Recurrent Computation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Bo5BOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the effect of retaining computation state across observations in VLA action experts and introduces LycoReco, based on a Continuous Thought Machine. Paired carry/reset experiments examine the use of past information, internal update budgets, and action-chunk consistency. The complete model reports mean success rates of 97.5% on LIBERO, 64.6% on SIMPLER-WidowX, and 62.6% across five MIKASA tasks.

Critical strengths

The controlled experiments are the strongest part of the paper. Carry and reset models use matched architectures and training settings, allowing a fairly direct study of state lifetime. The paired-cue experiment in Table 2 is particularly effective: current observations are held identical while earlier cues change. Carry reaches the correct target first in 135 of 150 replays, whereas a deterministic memoryless policy is bounded at 50.

The compute-budget analysis provides evidence beyond tasks that explicitly require memory. With replanning at every step on LIBERO, carry with one internal update outperforms reset with twelve, and this comparison is repeated across three training seeds. Chunk disagreement and target-movement experiments further examine whether consistency is compatible with responding to new observations.

The paper is clearly written. Comparisons with published systems are distinguished from matched experiments on state lifetime, and the appendices provide detailed training settings, state definitions, and computational accounting.

Critical weaknesses

The efficiency conclusion currently concerns mainly the action expert's internal computation. Appendix C.2 reports complete-decision latency at K=1, but the invocation rates differ when K>1: carry processes observations within an action chunk, whereas some reset policies run once per chunk. Since the VLM accounts for most computation, fewer internal updates do not directly establish a proportional improvement during policy execution. Total computation and inference time over the same number of control steps would clarify the practical benefit.

Most comparisons still use one training seed. Paired scene intervals capture variation over evaluation scenes but not training variance. In addition, the CTM and recurrent Transformer in Appendix E differ in both parameter count and state capacity. The evidence for carrying state is stronger than the evidence for CTM's advantage over other recurrent architectures.

A smaller point concerns the target-movement experiment. Section 4.5 already states that the 100 scenes are selected from those originally solved by carry. The 65% and 100% results should therefore be interpreted as post-perturbation success on that subset, rather than recovery rates over the full task distribution. This supports responsiveness to new evidence, but not a broader claim of perturbation robustness.

Questions for authors
  1. At the same K and number of control steps, what are the total inference calls, computation, and inference time for carry and reset?
  2. Beyond the three-seed budget comparison, are repeated MIKASA training results or analyses of the CTM/Transformer state-capacity difference available?
3 comments

We sincerely thank the reviewer for their thoughtful feedback and valuable suggestions. Below, we provide the requested efficiency measurements, three-seed evaluations, and matched-capacity architectural comparisons.

Total computation and inference time over the same number of control steps would clarify the practical benefit

We agree that practical efficiency should be compared over the same number of control steps, counting both VLM and action-expert computation.

Carrying state does not require processing every observation within an action chunk: state lifetime and inference frequency are independent design choices. A carry expert can keep its state across chunk-level decisions without updating it at intermediate control steps, and our dual-rate experiment (Section 4.6) likewise runs the VLM less often than the expert. Appendix C.2 described only the every-step schedule, and we will revise it. We compared carry and reset under a matched schedule in which both the VLM and the expert run once every KK steps (experimental setup from Section 4.4). Over the same control steps, both make the same number of VLM and expert calls at identical total computation:

MethodKTicksVLM callsExpert callsVLM ms/refreshExpert ms/refreshWall ms/refreshTotal inference (s)LIBERO success
LycoReco carry111,0001,00051.4151.11356.05456.58942.25% ± 3.70 pp
LycoReco reset111,0001,00052.0681.08856.79356.94418.17% ± 5.86 pp
LycoReco carry1121,0001,00051.3964.77359.90760.15149.38% ± 3.37 pp
LycoReco reset1121,0001,00051.7404.82760.49461.25436.00% ± 2.75 pp
LycoReco carry2150050051.5341.12156.12329.07436.75% ± 0.50 pp
LycoReco reset2150050051.5771.10556.04428.92234.42% ± 1.77 pp
LycoReco carry21250050051.7494.88160.08031.02758.00% ± 4.24 pp
LycoReco reset21250050051.3014.82159.76230.91546.50% ± 2.83 pp

We acknowledge that end-to-end savings are modest in our implementation. With a ~3M-parameter expert and a ~3B-parameter VLM, the VLM dominates total computation, so reducing expert ticks does not yield proportional full-policy latency savings. Nevertheless, our budget experiment (Section 4.4) demonstrates a favorable performance–computation tradeoff: carry outperforms reset at every matched tick budget on the frozen PaliGemma. Reducing ticks from T=12T=12 to T=1T=1 lowers action-expert latency from 5.835 ms to 1.279 ms.

We additionally trained the following with the adapted encoder of the headline LIBERO system, while both VLM and action expert run once every KK steps:

LIBERO:

MethodK=1K=1K=10K=10
1-tick carry97.00%97.25%
12-tick carry97.50%97.50%
1-tick reset90.25%97.25%
12-tick reset90.75%97.00%

While both methods saturate near ~97% at K=10K=10, carry maintains this performance even under every-step replanning (K=1K=1), where reset drops by roughly 7 percentage points.

Practical implications under latency constraints. For latency-constrained robotic control, achieving stronger policy performance within the available inference budget is important. Model capacity and tick count offer two complementary scaling dimensions: larger experts provide greater representational capacity, while additional ticks enable further iterative refinement. Carrying state can improve this tradeoff rather than eliminate the need for either. In our LIBERO budget sweep, carry outperforms reset at every matched tick budget, while additional ticks can still improve carry's performance. Reducing the required tick budget may also create headroom for larger experts within the same latency constraint. While 1-tick carry does not universally surpass 12-tick reset across every evaluation regime (such as the saturated K=10K=10 setting above), carry consistently achieves higher performance at every matched tick budget.

Potential benefits in dual-system VLAs. When an action expert operates more frequently than the VLM, its inference latency becomes a key constraint on control frequency. Our dual-rate experiment (Section 4.6) shows that carry maintains its advantage over reset when the VLM refreshes less frequently, supporting its applicability to such architectures. At π0.5\pi_{0.5}'s expert scale (roughly 100 times larger), a hypothetical LycoReco-like expert would require approximately 0.86 GFLOPs at one tick versus 10.3 GFLOPs at twelve (Table 6). However, this larger model has not been trained due to compute budget. We will clarify these limitations and distinguish demonstrated expert-level efficiency from end-to-end benefits.

Most comparisons still use one training seed

We have expanded our evaluations across three independent training seeds per benchmark in the next revision. Success rates (%) are reported as mean ± sample standard deviation. The budget sweep results are provided in our response on practical efficiency above.

We also provide the headline benchmark results:

ModelLIBEROSIMPLER-WidowXMIKASA (-VLA-v0 release)MIKASA (-v0 release)
LycoReco97.25 ± 0.2565.16 ± 5.6259.33 ± 5.6660.50 ± 4.10

In addition, the CTM and recurrent Transformer in Appendix E differ in both parameter count and state capacity

To isolate architectural differences from state capacity in Appendix E, we trained a Recurrent Transformer (RT) control matched to CTM's exact retained-state size (82 KiB) and trainable parameter count (2,996,301 for RT vs. 2,995,790 for CTM) using the same frozen PaliGemma backbone and training recipe on MIKASA. The per-task success rates (%) are reported below:

ExpertLifetimeSGTIMRC3RC5RC9Mean
LycoRecocarry612489521247.6
Matched Recurrent Transformercarry452563311034.8

A smaller point concerns the target-movement experiment

We agree. This experiment demonstrates responsiveness to new visual evidence within the scenes solved without perturbation. We do not claim general perturbation robustness, and we will clarify this distinction in Section 4.5 and the abstract/introduction to avoid the ambiguity raised.

Thank you to the authors for their point-by-point response to the review comments. The supplementary experiments and clarifications are reasonably thorough and have addressed most of my concerns. However, some issues remain only partially resolved, so I will maintain my score of Weak Accept.

Reviewer_GBSkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether a VLA action expert should preserve its computational state across observations. LycoReco couples PaliGemma with a compact Continuous Thought Machine whose activations, per-neuron histories, and synchronization accumulators continue across decisions. Separately trained carry/reset controls investigate hidden-cue retention, success versus ticks per observation, action-chunk consistency, and response to a moved target. The strongest evidence is a paired-cue experiment with identical current inputs and a three-seed comparison in which a one-tick carry expert outperforms a twelve-tick reset expert in the specified frozen-feature LIBERO setting.

Critical strengths
  1. This paper studies a useful architectural question: whether a VLA action expert should preserve its internal state across decisions. Separating state lifetime from computation per observation and action chunking gives the work a clear focus.
  2. The experimental design supports the central findings. Carry and reset experts are trained under matched conditions, while the paired-cue experiment holds current inputs fixed and changes only earlier cues, directly testing whether retained information affects action selection. The recurrent Transformer controls further broaden the evidence.
  3. The results suggest benefits for both computation and control. In the frozen-backbone LIBERO setting, one-tick carry outperforms twelve-tick reset across three seeds. The target-motion experiment also shows that the model can update its actions when a remembered target moves, addressing a potential drawback of persistent state.
Critical weaknesses

The paper does not establish a sufficiently distinct contribution beyond existing recurrent policies. The hidden-cue experiment demonstrates memory, while the LIBERO study does not isolate computation reuse from historical-information benefits.

The computational advantage is not validated in the configuration producing the headline performance. The budget sweep uses an unadapted frozen encoder and K=1, whereas the strong LIBERO system uses an adapted encoder and K=10.

The comparative evaluation does not establish superiority over existing memory-enabled VLA systems. The MIKASA ranking combines different environment releases and protocols, while the internal CTM–Transformer comparison leaves state capacity unmatched. These limitations weaken the broader design recommendation.

Questions for authors

Can the two MIKASA environment (line 886-VLA-v0 and -v0) releases be separated in Table 1 and the unsupported direct ranking removed, or can one relevant external baseline be evaluated under the same release and protocol?

What evidence distinguishes computation reuse from the benefit of historical information in the LIBERO sweep? Could an equal-information control resolve this, or should reuse be presented as an interpretation of the observed tradeoff?

Does the key one-tick carry versus twelve-tick reset advantage remain with the adapted encoder used by the high-performing LIBERO system, with the two experts otherwise matched? Please also clarify the scope implied by the different reset recipe in Appendix F.

Can the main-text consistency and responsiveness statements explicitly preserve the trajectory and scene-selection conditions of Appendix D.6? A shared-input trace analysis would be a useful supplementary diagnostic.

3 comments

We sincerely thank the reviewer for their thoughtful feedback and valuable suggestions. We agree with several of the scope concerns and will narrow several claims in the revision. We’d be happy to provide further clarifications during the discussion period.

The paper does not establish a sufficiently distinct contribution beyond existing recurrent policies. The hidden-cue experiment demonstrates memory, while the LIBERO study does not isolate computation reuse from historical-information benefits.

We agree that recurrence itself is not novel, and we do not claim it as a contribution. Our contributions are:

We formulate action generation as continuing computation. The expert iterates within each observation and continues the same iterations across observations. One transition rule updates the full computational state, which includes activations, per-neuron histories and synchronization accumulators. New observations refresh its context, and action readout leaves the state unchanged. To our knowledge, LycoReco is the first VLA which iterates its states within each observation and continues the same iteration across observations. Our experiments show that the number of iterations within each observation serves as an effective compute-scaling axis, and that carrying the state across observations shifts this entire scaling curve upward.

We analyze what changes when expert computation continues instead of restarting: information retention, the computation–success tradeoff, and chunk consistency and responsiveness, all using matched carry/reset experts.

The table places LycoReco among recurrent and iterative action experts.

ArchitectureCarries state across observationsIterates within an observationIterated state continues to the next observationCarried state shapes how the current observation is readCarried-state size
BC-RNN / RoboFlamingo LSTM✓✗n/a✗O(1)
AVA-VLA✓✗n/a✓ (reweights visual tokens in the VLM)O(H)
RD-VLA✗✓✗n/a—
LycoReco✓✓✓✓ (queries VLM K/V in the expert)O(1)

H: action-chunk length. Outside VLAs, CTM's RL agents and adaptive-computation RNNs already carry and iterate. LycoReco is, to our knowledge, the first VLA action expert to do so.

Can the two MIKASA environment (line 886-VLA-v0 and -v0) releases be separated in Table 1 and the unsupported direct ranking removed, or can one relevant external baseline be evaluated under the same release and protocol?

Thank you for pointing this out. We ran three additional independent training runs on the -v0 environment release so both protocols can be evaluated and compared directly. We use the official episode step limits (matched with the baselines) for each environment:

Environment ReleaseLycoReco (3 seeds, %)
MIKASA-VLA-v059.33 ± 5.66
MIKASA-v0 (New)60.50 ± 4.10

The relative ranking still holds on the matched -v0 environment release. We will revise Table 1 to separate the two environment releases and remove cross-release rankings.

What evidence distinguishes computation reuse from the benefit of historical information in the LIBERO sweep? Could an equal-information control resolve this, or should reuse be presented as an interpretation of the observed tradeoff?

We agree that the LIBERO tick-budget comparison alone does not isolate computation reuse from access to historical information. It establishes that carrying the expert state improves success at a smaller per-observation update budget in the tested setting. We will present computation reuse as an interpretation consistent with this result, rather than a separately established causal explanation. Our closest equal-information evidence is on MIKASA (Appendix E): a reset expert that also reads the VLM features of every earlier frame reaches 38.4%, compared to 47.6% for carry and 20.4% for plain reset.

Does the key one-tick carry versus twelve-tick reset advantage remain with the adapted encoder used by the high-performing LIBERO system, with the two experts otherwise matched? Please also clarify the scope implied by the different reset recipe in Appendix F.

We trained the following matched carry and reset experts on the adapted encoder of the high-performing systems:

LIBERO:

MethodK=1K=1K=10K=10
1-tick carry97.00%97.25%
12-tick carry97.50%97.50%
1-tick reset90.25%97.25%
12-tick reset90.75%97.00%

At K=10K=10, all systems saturated LIBERO, therefore 1-tick carry doesn’t show a clear advantage over 12-tick reset. 1-tick carry may not outperform 12-tick reset, but our experiments show a systematic advantage of carry over reset on matched tick budgets on LIBERO. We will revise the manuscript to include the high-performing system comparison and the updated claim.

Regarding Appendix F: that table uses a larger reset reference expert with learned VLM feature projections (3.9M vs. 3.1M parameters), higher expert computation per observation (2.07 GFLOPs at T=8T = 8 vs. 0.10 GFLOPs at T=12T = 12), and a higher training budget. Both our headline benchmark results and the matched carry/reset budget sweeps use the smaller, compute-efficient expert that reads the VLM's K/V cache directly. We will make this distinction explicit in the revision.

Can the main-text consistency and responsiveness statements explicitly preserve the trajectory and scene-selection conditions of Appendix D.6? A shared-input trace analysis would be a useful supplementary diagnostic.

Thank you also for suggesting the shared-input trace diagnostic. We replayed the same 125 held-out MIKASA demonstrations (25 per task, 2,754 observations) through both MIKASA policies (T=12T = 12, action horizon H=4H = 4) without executing their predicted actions. On identical observation streams, carry exhibits 22.7% lower chunk disagreement than reset (95% CI: 15.1–29.5%). We will revise the manuscript to include this result.

The internal CTM–Transformer comparison leaves state capacity unmatched

Thanks for pointing this out, we trained a Recurrent Transformer (RT) control matched to CTM's exact retained-state size (82 KiB) and trainable parameter count (2,996,301 for RT vs. 2,995,790 for CTM) using the same frozen PaliGemma backbone and training recipe on MIKASA. The results below:

ExpertLifetimeSGTIMRC3RC5RC9Mean
LycoRecocarry612489521247.6
Matched Recurrent Transformercarry452563311034.8

We thank the reviewer again for their constructive feedback. We hope that our additional experiments and clarifications have addressed the concerns.

We would greatly appreciate your assessment of whether these new results sufficiently address your concerns. If any questions remain, we would be happy to provide further clarification during the discussion period.

Reviewer_L8sbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes that Vision-Language-Action (VLA) models should retain their internal state when making decisions, rather than starting from scratch each time. Based on the PaliGemma model, LycoReco implements this mechanism through a "Continuous Thought Machine" with 3 million parameters. On the LIBERO benchmark, the "carry" mechanism leverages information, even when it is no longer visible within the current field of view, while requiring fewer internal state updates. It produces more coherent and consistent action sequences and is capable of adjusting subsequent actions if a remembered target moves.

Critical strengths

The comparison between "carry" and "reset" is an appropriate experimental design, and the results demonstrate that memory plays a significant role. Under the same training protocol, the "carry" strategy achieved a score of 62.6% across five MIKASA tasks, compared to 25.8% for the "reset" strategy.

Critical weaknesses

None

Questions for authors
  1. Does carrying state still help in a flow or diffusion action expert (like pi0), or only in this recurrent CTM?

  2. Similar RNN designs do not appear to be novel in the VLA field; the authors might consider discussing their work in relation to AVA-VLA (Ava-vla: improving vision-language-action models with active visual attention, CVPR 2025) , a paper that likewise demonstrates the advantages of such non-reconstructive systems.

  3. The evaluation appears to lack multiple experimental run. Could you please report the average success rate and variance based on multiple independent runs?

2 comments

We sincerely thank the reviewer for their thoughtful feedback and valuable suggestions. We’d be happy to provide further clarifications during the discussion period.

Q1. Does carrying state also help flow-based action experts?

Carrying state also helps a flow-matching expert. We tested a recurrent Transformer expert that we trained with flow matching. Across three paired training runs, carry raises mean success from 14.47% to 21.53%, and carry beats reset in every pair. The expert updates its retained state from each observation and holds it fixed during flow integration.

For context, we include our direct-L1 CTM control below. Both experts use frozen PaliGemma features and the same retained-state size on MIKASA (82 KiB).

Action expertObjectiveCarry (%)Reset (%)
LycoReco CTML147.620.4
Recurrent TransformerFlow matching21.5314.47

LycoReco CTM has 3M parameters, and the Flow matching recurrent transformer has 3.79M parameters. However, carrying state in this flow-matching expert is not equivalent to our proposed continuing computation, since the retained state is updated across observations but held fixed during flow integration. How to extend continuing computation to the iterative generation process of flow-based action experts remains an open question for future work. We will add these results and state the limitation to the revision.

Q2. Relationship to AVA-VLA and recurrent policies.

We agree that recurrence itself is not novel, and we do not claim it as a contribution. We will expand our comparisons to include AVA-VLA and clarify our contributions more precisely:

  1. We formulate action generation as continuing computation. The expert iterates within each observation and continues the same iterations across observations, and successive action chunks are readouts of this ongoing process. LycoReco realizes this design by adapting a CTM into a compact 3M-parameter action expert. One transition rule updates the full computational state within and across observations. New observations refresh its context, and reading out actions leaves the state unchanged. To our knowledge, LycoReco is the first VLA which iterates its states within each observation and continues the same iteration across observations. Our experiments show that the number of iterations within each observation serves as an effective compute-scaling axis, and that carrying the state across observations shifts this entire scaling curve upward.

  2. We analyze what changes when expert computation continues instead of restarting. Matched carry/reset experiments and targeted diagnostics examine information retention, the computation–success tradeoff, and chunk consistency and responsiveness.

AVA-VLA carries action-related representations to initialize action placeholders and guide backbone visual attention, producing each chunk through one backbone forward pass. LycoReco instead applies the same update rule to its expert state both within and across observations, using the evolving state to query current VLM features and read out successive action chunks.

ArchitectureCarries state across observationsIterates within an observationIterated state continues to the next observationCarried state shapes how the current observation is readCarried-state size
BC-RNN / RoboFlamingo LSTM✓✗n/a✗O(1)
AVA-VLA✓✗n/a✓ (reweights visual tokens in the VLM)O(H)
RD-VLA✗✓✗n/a—
LycoReco✓✓✓✓ (queries VLM K/V in the expert)O(1)

Q3. Results across independent training runs.

We trained three independent runs on each benchmark to update the manuscript. The table reports mean success (%) ± sample standard deviation across training runs.

ModelLIBEROSIMPLER-WidowXMIKASA (-VLA-v0 release)MIKASA (-v0 release)
LycoReco97.25 ± 0.2565.16 ± 5.6259.33 ± 5.6660.50 ± 4.10

We thank the reviewer again for their constructive feedback. We hope that our additional experiments and clarifications have addressed the concerns.

We would greatly appreciate your assessment of whether these new results sufficiently address your concerns. If any questions remain, we would be happy to provide further clarification during the discussion period.

Comments on the paper

1 comment

We sincerely thank all reviewers for their constructive feedback and recognition of our work. The table below summarizes the key feedback and our responses.

DimensionReviewer L8sbReviewer GBSkReviewer Bo5BSummary
Formulation & Significance"studies a useful architectural question... gives the work a clear focus.""The paper is clearly written. Comparisons with published systems are distinguished from matched experiments"Strength: Reviewers highlight the significance of continuing computation in VLAs and the formulation of our research question
Experimental Design"The comparison between 'carry' and 'reset' is an appropriate experimental design""The experimental design supports the central findings""The controlled experiments are the strongest part of the paper"Strength: Reviewers highlight our experiment design: matched carry/reset controls, paired hidden-cue diagnostic, and compute-scaling sweeps as rigorous evidence.
Novelty vs. Recurrent VLAs"Similar RNN designs do not appear to be novel in the VLA field""does not establish a sufficiently distinct contribution beyond existing recurrent policies."—Contribution: We clarified our distinctions compared to other recurrent policies, among them, LycoReco is the first VLA expert to both iterate within an observation and continue that iteration across observations in state.
Model Architecture"Does carrying state still help in a flow or diffusion action expert (like pi0)?""CTM–Transformer comparison leaves state capacity unmatched.""CTM and recurrent Transformer in Appendix E differ in both parameter count and state capacity."Results: On frozen PaliGemma features with state and parameters matched, CTM carry beats matched RT carry on MIKASA. In a similarly matched flow-matching RT, carry improves MIKASA success confirming applicability across both direct-L1L_1 and flow-based experts.
Practical Benefits & Efficiency"advantages of such non-reconstructive systems""What evidence distinguishes computation reuse from... historical information?" / "Does the... advantage remain with the adapted encoder""Total computation and inference time over the same number of control steps would clarify the practical benefit."Results: We showed carry's advantage on both frozen and adapted encoder at matched budget. Carry also outperforms reset with full history info on MIKASA. We profiled end to end latency to be competitive and clarified practical benefits in scaling and dual-system VLAs.
Evaluation Protocol"report the average success rate and variance based on multiple independent runs""Can the two MIKASA environment (-VLA-v0 and -v0) releases""Most comparisons still use one training seed..."Results: We expanded our experiments to three seed runs. LIBERO (97.25 ± 0.25%), WidowX (65.16 ± 5.62%), MIKASA-VLA-v0 (59.33 ± 5.66%), and matched MIKASA-v0 (60.50 ± 4.10%).
Consistency & Responsiveness"is capable of adjusting subsequent actions if a remembered target moves.""A shared-input trace analysis would be a useful supplementary diagnostic.""This supports responsiveness to new evidence, but not a broader claim of perturbation robustness."Results: The new shared-input trace analysis reinfoces our consistency claim. We clarified the scope of the responsiveness claim to distinguish it from perturbation robustness.

We will incorporate new results and clarifications into a revised manuscript and welcome any further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.