acceptodds
Under review as a conference paper at ICLR 2027

CTRL-Z: Code agent Trial robot actions and Reason with Latent world model

Abstract

Code-as-policy approaches translate language instructions into executable programs, enabling compositional task planning and reuse of robot skills. However, generating a valid program don't establish whether a candidate action will produce the desired physical change. Action-conditioned world models offer a way to anticipate such changes, but their latent predictions are not directly expressed in the language of executable programs generated by code agent. To address these questions, we propose a closed-loop system in which a code agent can read, reason, make prediction and update object physical parameters based on those predictions, and generate actions accordingly. To make it readable for the agent, we build Structured Object Program, a compositional program describing object primitive geometry, physical parameters, interaction properties and rendering object programs into geometric views. For the prediction system, we propose a dual-stream world model that accepts RGB images and rendered geometric views as input while predicting future visual and geometric latent variables. Physical parameters regarding the dynamic parts of the object are decoded from these predicted latents and fed back to the agent, enabling it to understand how its current actions affect the object's future properties. Experiments across manipulation tasks and object generalization settings show that our framework improves task success, cross-object generalization, and agent interaction efficiency

Primary area:
reinforcement learning
Keywords:
robot learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yRJpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a framework that equips a code-based robotic agent with a learned world model and Structured Object Programs (SOPs). These components enable the agent to predict the outcomes of candidate actions and revise control programs before execution. The framework also supports visuomotor policy post-training. Experiments on simulated manipulation tasks and two real-world tasks report higher task success rates, improved cross-object generalization, and fewer environment attempts.

Critical strengths
  1. The pipeline is interesting. This paper uses a world model to predict the possible outcomes of actions proposed by the coding agent. These predictions help the agent evaluate and revise its programs before execution.

  2. Ablation studies are useful. The comparisons with RGB-only and duplicated-RGB inputs show that SOP-rendered geometry improves the accuracy of future object-state predictions.

Critical weaknesses

The advantages of the learned world model over a physics simulator remain unclear for the evaluated tasks, which primarily involve rigid-body and articulated-object manipulation. The predicted information, such as part poses and configurations, could also be obtained from a physics simulator. Although Table 2 includes an oracle simulator reference, it does not provide a comparison against a practical simulator baseline initialized from the same perceptual inputs and SOP representation. A controlled comparison between simulator-based and world-model-based planning would help clarify the benefits of the proposed approach. This comparison should report task success rates and computational costs under matched LLM token budgets and candidate-action evaluation budgets.

Questions for authors

Please refer to the Critical Weaknesses section.

In line 033, please use the correct LaTeX quotation marks: ``Code as Policies''.

2 comments

We thank the reviewer for identifying the need for a perception-initialized simulator comparison. We have added an Estimated-SOP Simulator baseline to main-text Table 2, with a dedicated controlled comparison in Appendix B.6, Table 10.

To make the revisions easy to locate, the following index summarizes the added experiments, the points they support, and their manuscript locations. The requested simulator comparison is accompanied by controls examining the SOP interface, predictive evaluation, and downstream manipulation.

Added experimentSupported pointLocation
Controlled predictive-evaluation ablationSeparates the benefits of predictive selection, reward-guided revision, and future-state feedback under shared budgets.§5.6, Table 3; App. B.1, Table 5
SOP representation and geometry-input controlsExamines the complementary roles of executable object structure, rendered geometry, and agent-readable future states.App. B.2, Table 6
Direct future-state prediction baselineTests the benefit of latent rollout beyond directly regressing future states, under both RGB-only and RGB-plus-geometry inputs.App. B.3, Table 7
Fixed-candidate-bank evaluationMeasures ranking accuracy, false-positive selection, and bad-action rejection independently of candidate generation and code revision.§5.4, Figure 4; App. B.4, Table 8
Cumulative and independently initialized stage evaluationsIdentifies the execution stages where predictive feedback improves completion beyond current SOP priors.App. B.5, Table 9
Perception-initialized physics simulatorQuantifies proximity and remaining gaps to practical and oracle simulation in task success and completion cost under matched budgets.Main-text Table 2; App. B.6, Table 10
Four-task RoboTwin 2.0 evaluationTests the framework on an additional task suite beyond the primary custom ManiSkill3-based evaluation.App. B.7, Table 11
Expanded real-robot evaluationAdds a no-WM control and a cluttered long-horizon task to test predictive feedback during physical, sequential manipulation.§5.7, Table 4; App. A.1, Figures 5–7

[W1, Q1] Practical physics simulation from the same perceptual inputs.

The practical simulator is instantiated from the same estimated SOP geometry and kinematics, with predefined collision models and physical settings fixed before testing. It executes candidate action chunks and applies the same SOP progress function to simulator-produced states. The Oracle Simulator remains a separate privileged reference initialized from restored ground-truth simulator states.

Learned and practical evaluators share the code agent, candidate-generation procedure, action interface, and online limits: 1M LLM tokens per episode, 12 candidate evaluations per decision, and 10 minutes per episode.

The added comparison reports:

EvaluatorTask SRUnseen-category SRTime (min)Tokens (M)
Dual-stream WM43.5 ± 1.1%33.8 ± 1.7%4.2 ± 0.50.15 ± 0.08
Estimated-SOP Simulator45.7 ± 0.8%39.5 ± 1.6%3.4 ± 0.40.15 ± 0.06
Oracle Simulator47.2 ± 0.7%42.6 ± 1.4%3.1 ± 0.40.13 ± 0.05

The learned evaluator is 2.2 percentage points below the practical simulator in overall success, and 3.7 points below the oracle. The practical simulator retains higher success and shorter successful-episode completion time. These results support the scoped claim that learned predictive evaluation approaches simulator-guided planning on the evaluated tasks; they do not establish superiority over physics simulation.

[W1] Additional evidence on evaluation quality.

Appendix B.4, Table 8 also compares these evaluators on shared fixed candidate banks. Top-1 agreement is 81.5% for the dual-stream WM and 93.4% for the practical simulator, with false-positive selection of 48.6% and 25.2%, respectively. This quantifies the remaining evaluator-quality gap, complementing the online success and cost comparison.

[Minor correction] Quotation marks.

The requested LaTeX quotation marks for “Code as Policies” have been corrected in the Introduction.

Dear Reviewer, thank you again for suggesting the perception-initialized simulator comparison. With approximately one day remaining in the rebuttal period, we would like to gently follow up on the added comparison under matched budgets (Table 2; Appendix B.6, Table 10) and our clarification of the scope of the results. If your schedule permits, we would greatly appreciate your thoughts on whether these additions address your concern. Thank you for your time and consideration.

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

The paper proposes CTRL-Z, a closed-loop framework that connects an LLM-based code agent with an action-conditioned latent world model through Structured Object Programs (SOPs) — compositional, parameterized object programs describing part geometry, kinematics, affordances, and manipulation templates, which can be rendered into geometric views. A dual-stream world model takes RGB history plus SOP-rendered geometry and autoregressively predicts future latents, from which a physical readout decodes dynamic part states back into the SOP so the agent can score and revise candidate action programs before execution. World-model rollouts are further used for VLA post-training and skill-library refinement. Experiments in simulation across articulated opening, pick-and-place, and precision tasks, plus a small real-robot study, show gains over code-based and visuomotor baselines and better generalization to unseen objects.

Critical strengths
  • A clean solution to a real interface problem. Making latent world-model predictions readable and revisable by a code agent through a bidirectional program interface (rendered geometry in, decoded part states out) bridges two research lines that have developed largely separately, and the near-oracle performance of the learned predictor suggests the design captures most of the achievable gain.
  • Fairly controlled within-framework ablations. The dual-stream vs. RGB-only vs. oracle-geometry comparisons isolate the value of rendered geometry rather than "just another stream", and the multi-backbone and multi-setting generalization results make the trends more convincing.
Critical weaknesses
  • The decisive baseline is missing. The central claim is that latent rollout with SOP-rendered geometry yields accurate, program-usable prediction, but there is no comparison against directly predicting future part states (e.g., a state-decoding/object-centric head on the same encoder with the same supervision); if a direct predictor matches it, the SOP-geometry framing largely collapses.
  • Modest realized gains, fragile backbone dependence, thin real-world evidence. Against the same-backbone execution-feedback code agent the gain is only ~6–8 points, with other LLM backbones the framework nearly fails, and the real-robot study (two tasks, no sim-to-real description, admittedly illustrative figures) is too under-documented to support the physical-manipulation claims.
Questions for authors
  1. Please add a direct part-state prediction baseline (same encoder and supervision, no latent rollout / rendered geometry). Does the dual-stream model still win?
  2. Why does the "Current SOP" configuration succeed so rarely despite providing grasp poses, joint axes, and manipulation templates? A per-stage failure analysis (grounding, grasping, parameter choice) would clarify what the world model actually contributes.
  3. For the real-robot experiments: which configuration is evaluated, and how are SOP grounding and the (simulator-trained) world model transferred to real scenes? Please report grounding-failure rates.
  4. Since candidates are selected by reward rather than executed directly, please report selection accuracy — how often does the world model's misranking lead to a bad action? Also, what is the per-decision compute cost of long-horizon multi-candidate rollouts?
  5. The entire evaluation is on a self-built SAPIEN suite. I would encourage validating on at least one established manipulation benchmark (e.g., RLBench, ManiSkill, LIBERO, or the CaP-X benchmark the authors cite) to substantiate the generality claims beyond a platform where all priors and training data are self-annotated.
  6. Longer-horizon, multi-object or cluttered tasks would test whether the single-object template-based grounding and the 52-step rollout accuracy scale beyond the current task suite.
4 comments

We thank the reviewer for the specific suggestions. We have added the requested predictive controls and external-task evaluation. The controlled ablation and expanded real-robot results are now in the main text, Sections 5.6–5.7, Tables 3–4, and Figure 4 now includes candidate-selection quality.

ConcernEvidenceLocation
W1/Q1Direct stateApp. B.3, T7
W2/Q2Controlled/stage ablations§5.6, T3; App. B.1/B.5, T5/T9
Q4Selection/costFig. 4; T3; App. B.4, T8
W2/Q3/Q6Real robot§5.7, T4; App. A.1
Q5RoboTwinApp. B.7, T11
W2Backbones§5.2, T2

[W1, Q1] Direct future-part-state prediction.

Appendix B.3, Table 7 compares direct state prediction with latent rollout plus readout under matched observation histories, action chunks, training data, state supervision, and online budgets:

PredictorSeen SRUnseen-category SRGeometry/kinematics OOD SR
RGB-only: direct state20.4%1.6%0.0%
RGB-only: latent + readout27.8%10.1%9.3%
RGB + SOP geometry: direct state38.5%17.9%6.2%
RGB + SOP geometry: latent + readout42.1%29.3%25.5%

The RGB-only direct baseline requested by the reviewer is therefore included. Holding SOP geometry input fixed, latent prediction improves success by 3.6, 11.4, and 19.3 percentage points. At H = 52, position/rotation errors decrease from 14.2 cm / 12.6° to 6.1 cm / 6.8°. The larger gains under distribution shifts support the utility of latent prediction beyond directly regressing the program’s state variables.

[W2, Q2] Why current SOP priors are insufficient.

We added cumulative full-task diagnostics and independently initialized stage subtasks in Appendix B.5, Table 9. Relative to Current SOP, Full CTRL-Z increases cumulative pre-grasp completion from 21.5% to 65.2%, grasp completion from 10.7% to 45.6%, and manipulation completion from 5.2% to 44.1%. Structural knowledge alone therefore does not ensure successful execution.

Main-text Table 3 further separates predictive selection, reward-guided revision, and future-state feedback; the last adds 15.3 points on precision manipulation over reward-only revision. These stage results diagnose execution progress rather than uniquely assigning each failure to a single root cause.

[Q4] Selection reliability and computational cost.

Figure 4 and Appendix B.4, Table 8 evaluate shared fixed candidate banks: dual-stream Top-1 improves from 50.3% to 81.5%, and false-positive selection falls from 73.2% to 48.6%. Top-1 concerns SOP progress ranking; false positives concern actual failure among candidates predicted to reach the stage goal.

Online comparisons use caps of 1M tokens/episode, 10 minutes/episode, and 12 evaluations/decision, counting revised candidates. Table 3 reports successful-episode completion costs of 3.2 ± 0.3 minutes and 0.16 ± 0.08M tokens for M3, versus 9.7 ± 0.1 minutes and 0.98 ± 0.01M tokens for M0. These quantify end-to-end completion costs.

[W2, Q3, Q6] Real-robot transfer and longer-horizon manipulation.

Section 5.7, Table 4 and Appendix A.1 now cover three tasks, with 100 trials per method per task. Deployment uses code-generated actions with frozen code-agent weights. The parameter estimator and world model are trained on mixed simulation and real-world data, and the resulting pipeline is directly deployed.

We added a cluttered long-horizon task: open the drawer, grasp the red block, and place it inside. Full CTRL-Z obtains 69%, 63%, and 37% on stacking, drawer opening, and the long task, versus 59%, 55%, and 26% without WM prediction. The matched comparison retains the same grounding, tools, and execution feedback. Recorded examples remain in Appendix A.1, Figures 5–7.

[Q5] Evaluation beyond the primary custom suite.

Appendix B.7, Table 11 adds single-arm evaluation on four RoboTwin 2.0 tasks: open_microwave, open_laptop, place_object_scale, and stamp_seal. We retain the original task success predicates under our shared online budgets. Mean success across the four tasks is 6.2% without WM, 28.1% with RGB WM, and 45.9% with Full CTRL-Z. These results extend the evaluation to articulated opening, object placement, and precision interaction outside our primary ManiSkill3-based suite.

[W2] Backbone dependence.

Table 2 shows within-backbone gains from RGB to dual-stream prediction for Gemini (5.5% → 8.3%) and Qwen (3.4% → 7.2%), but their absolute success remains low even with oracle evaluation (12.7% and 10.8%). We therefore interpret predictive evaluation as improving a given code agent, while recognizing that overall performance remains dependent on the underlying agent’s capabilities.

We thank the authors for the thorough and responsive rebuttal. The new experiments address most of our concerns: the direct part-state prediction baseline (App. B.3, Table 7) answers our Q1 — the dual-stream model retains a clear advantage, particularly under distribution shift — and the fixed-candidate-bank evaluation (Fig. 4, App. B.4), cost budgets, RoboTwin evaluation (App. B.7), and the new cluttered long-horizon real-robot task address Q4–Q6. We appreciate the scoping of the backbone-dependence claim.

Three clarifications would help us finalize our assessment:

  1. (Q3) The real-robot world model and parameter estimator are now described as trained on "mixed simulation and real-world data." Please specify the amount of real-world data, and confirm whether the new ablations (Tables 3, 5, 7) use models trained with the same data mixture.
  2. (Q2/Q3) Please report the grounding-failure rate in the real-robot experiments (and in simulation if available). Since the WM and no-WM conditions share the same grounding, the matched comparison remains fair, but the rate is needed to interpret residual failures.
  3. Given App. B.6 (Table 10), where the estimated-SOP simulator slightly outperforms the learned world model at lower cost, could the authors clarify when the learned world model is preferable to a perception-initialized simulator (e.g., deployment without a reconstructable simulator, contact-rich or unmodeled scenes)? This affects how we weigh the contribution of the learned predictor relative to the SOP interface itself.

We thank the reviewer for the positive assessment and these specific clarification requests.

1. Real-world training data and consistency across ablations.

The parameter estimator and world model use the same pool of 500 complete real-robot interaction trajectories, combined with simulation data. After training, the resulting models are used directly for evaluation, without additional evaluation-time updates.

Tables 3 and 5 present the same controlled ablation in the main text and appendix. Their WM-based configurations reuse the trained parameter estimator and world model, while varying how predictions are used for selection and revision. The direct-state and latent-prediction models in Table 7 use the same training-data partitions and simulation/real-world mixture. Thus, the reported differences are not attributable to different amounts of real-world training data.

2. Real-robot grounding-failure rates.

We defined an grounding failure as an episode in which, before physical execution, the grounding module misses or misidentifies a task-relevant object or part, assigns incorrect structural or kinematic relationships, or fails to instantiate the SOP. Grounding fails in 2/100 trials (2%) for block stacking, 13/100 (13%) for drawer opening, and 21/100 (21%) for long-horizon manipulation. These trials are included in the 100-trial denominator of Table 4, without test exclusion or replacement.

The WM and no-WM conditions share the initial grounding procedure. We therefore report these rates as diagnostics of the shared perception stage, separately from end-to-end task failure. They complement the execution-stage diagnostics in Appendix B.5. In simulation, episodes that terminate due to grounding failure are also counted as failures and retained in the task-success denominator, without exclusion or replacement.

3. Role of the learned world model relative to physics simulation.

Thank you for highlighting this distinction. Our simulator comparison is intended to examine how closely learned, program-readable prediction can approach simulator-guided planning under matched inputs and budgets. In Table 10, the learned WM achieves 43.5% overall success, compared with 45.7% for the estimated-SOP simulator and 47.2% for the oracle. We view physics simulation as a strong reference when an adequately matched simulator is available, the practical simulator also achieves shorter completion time in this setting.

The intended deployment value is the ability to reuse this learned predictor across objects and scenes after SOP grounding, particularly when constructing and calibrating an adequately matched executable physics model is difficult. This could reduce per-object or per-asset simulator-construction effort. To assess this cross-object reuse, we evaluate the pipeline on unseen object instances, novel part compositions, unseen categories, and geometry/kinematics OOD settings, covering diverse objects and structural variations beyond the world-model training distribution (Section 5.3; Appendix B.3, Table 7).

The learned predictor’s contribution is also examined with the SOP input held fixed. In Table 7, latent rollout improves success over direct state prediction by 11.4 points on unseen categories and 19.3 points on geometry/kinematics OOD. In Table 4, adding WM prediction to the same real-robot grounding and tool pipeline improves success by 10, 8, and 11 points across the three tasks. These results support complementary contributions from the SOP interface and learned prediction. Establishing comparative advantages over practical simulation in more complex contact-rich or unmodeled scenes remains an important next step.

Dear Reviewer, thank you again for your careful follow-up and constructive discussion. With approximately one day remaining in the rebuttal period, we would be grateful if, when your schedule permits, you could take a look at our latest clarifications regarding training data, grounding-failure rates, and the role of learned prediction relative to physics simulation. We would welcome any further feedback and remain happy to address any remaining questions. We sincerely appreciate your time and effort.

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

This paper proposes CTRL-Z, a framework that integrates code-based robotic agents with an action-conditioned world model for predictive evaluation and refinement of manipulation programs. It introduces Structured Object Programs (SOPs) to represent object geometry, semantic parts, kinematics, and affordances, enabling the world model to predict physical state changes and guide action selection. Experiments in simulation and real-world manipulation demonstrate improvements over existing baselines. It also shows promising results in generalization to unseen objects and configurations.

Critical strengths
  1. SOPs provide a shared representation of semantic parts, geometric parameters, kinematic structure, and manipulation primitives. This makes predicted object-state changes more interpretable and accessible to code-based agents than latent predictions alone.

  2. The appendix reports real-world success rates of 69% for block stacking and 63% for drawer opening, compared with 56% and 54% for π0.5, respectively. These results are encouraging, although broader testing is needed to establish robustness.

Critical weaknesses
  1. The methodological novelty and contribution need clearer justification. The proposed method can be categorized as model-based. The framework combines several existing ideas: code-as-policy agents, parameterized object representations, manipulation templates, action-conditioned latent world models, and model-based policy optimization. The paper cites related approaches such as ShapeAssembly, Analytic Concepts, DINO-WM, and V-JEPA 2-AC, but it remains unclear which technical components constitute the principal methodological advance beyond their integration.

  2. The experiments do not fully isolate the benefit of predictive action evaluation. The paper attributes its improvements to pre-execution physical prediction and the proposed SOP interface. However, the complete framework introduces several factors simultaneously, including structural grounding, geometric and manipulation priors, candidate action evaluation, program revision, and skill reuse.

Questions for authors
  1. What is the key technical contribution of SOPs beyond existing parameterized object representations and manipulation templates? Can the authors demonstrate that this representation is necessary for the observed improvements?

  2. How accurately do predicted progress rewards rank candidate actions according to their actual outcomes in the environment? How often does the agent select an action that the world model predicts to be successful but that fails during execution?

  3. What is the total inference cost of CTRL-Z relative to the baselines, including candidate generation, world-model rollouts, and code-agent revision? How many predictive evaluations are typically required before executing an action?

2 comments

We thank the reviewer for the constructive feedback. We have added targeted controls and brought the predictive-evaluation ablation into Section 5.6, Table 3, real-robot results into Section 5.7, Table 4, and candidate-selection results into Figure 4.

To help the reviewer locate the revisions, we summarize the added evidence and its connection to the concerns below:

  • W1 and Q1 — SOP contribution: We added separate controls for the agent-facing representation, geometric input, and future-state feedback in Appendix B.2, Table 6, with an accompanying discussion of their complementary roles.
  • W2 — Attribution of improvements: We added the controlled predictive-evaluation ablation to main-text Section 5.6, Table 3, with the supplementary presentation in Appendix B.1, Table 5, separating predictive selection, reward-guided revision, and explicit future-state feedback.
  • Q2 — Evaluation reliability: We expanded main-text Section 5.4, Figure 4 with candidate-selection results and added the complete fixed-candidate-bank evaluation and metric definitions in Appendix B.4, Table 8.
  • Q3 — Computational cost: We added completion-time and token-consumption comparisons to main-text Table 3 and specified the shared episode and candidate-evaluation budgets in the experimental setup of Appendix B.
  • Expanded real-robot evidence: We added a cluttered long-horizon task and a no-WM control. Quantitative results now appear in main-text Section 5.7, Table 4, while deployment details and qualitative examples are provided in Appendix A.1.

[W1, Q1] Contribution of the SOP interface.

Our contribution is the bidirectional interface through which object programs provide geometric inputs to the world model and receive predicted part states for program evaluation and revision. We test these functions separately in Appendix B.2, Table 6. On unseen categories, success rises from 5.4% to 12.8% when natural-language descriptions are replaced by SOPs under RGB-only prediction, while internal SOP scoring is retained. Adding rendered geometry raises success to 15.3%, and returning future part states to the agent raises it to 29.8%. These controls support the complementary roles of executable structure, geometric prediction, and program-readable future states.

[W2] Separating prediction, selection, and program revision.

Main-text Table 3 and Appendix B.1, Table 5 use the same frozen agent, observations, SOP priors, and online budgets:

ConfigurationOpening SRPick-and-place SRPrecision SR
M0: current SOP + self-check5.6%14.3%3.1%
M1: + predictive selection24.8%31.9%21.3%
M2: + SOP-reward-guided revision32.4%43.6%30.5%
M3: + future SOP part states44.5%53.1%45.8%

M1 selects an existing candidate using predicted progress; M2 additionally returns SOP progress reward for revision; M3 also returns numerical future SOP object states. Future-state feedback adds 12.1, 9.5, and 15.3 percentage points over SOP reward-only revision. These comparisons separate the benefit of predictive feedback from merely supplying current structural priors.

[Q2] Candidate ranking and false-positive selection.

We added shared, fixed banks of 20 action chunks per initial state, executing every candidate from the same restored simulator state. Figure 4 and Appendix B.4, Table 8 show that dual-stream prediction improves pairwise ranking from 64.9% to 85.9%, Top-1 from 50.3% to 81.5%, and bad-action rejection from 31.7% to 58.2%, while reducing false-positive selection from 73.2% to 48.6%.

Top-1 measures agreement with the highest actual SOP progress score. False-positive selection is the actual failure fraction among selected candidates predicted to satisfy the stage goal.

[Q3] Inference budgets and completion costs.

The supplementary online comparisons share a 1M-token episode cap, 10-minute wall-clock cap, and at most 12 candidate evaluations per decision, including revised-candidate re-evaluations (Appendix B, experimental setup). In Table 3, successful-episode completion time decreases from 9.7 ± 0.1 to 3.2 ± 0.3 minutes, and LLM usage decreases from 0.98 ± 0.01M to 0.16 ± 0.08M tokens, from M0 to M3. These quantify completion costs across repeated generation, evaluation, and revision.

Dear Reviewer, thank you again for your thoughtful feedback. With approximately one day remaining in the rebuttal period, we would like to gently follow up on our responses and revised manuscript. We have added controls examining the SOP contribution, predictive selection, and program revision, together with candidate-selection diagnostics. If your schedule permits, we would greatly appreciate any further thoughts on these additions and would be happy to clarify any remaining questions. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.