acceptodds
Under review as a conference paper at ICLR 2027

TARA: Temporal Action-Reasoning Alignment for Steerable Vision-Language-Action Policies

Abstract

Recent research increasingly explores generalist Vision-Language-Action (VLA) policies that leverage large-scale vision-language pretraining to generalize from only a few demonstrations. A promising direction is Chain-of-Thought reasoning, which enables models to reason step-by-step about the environment state and next subgoals before generating low-level actions. However, current VLAs tend to ignore intermediate thoughts when executing actions as they over-rely on vision. Prior work has tried to address this with information-dense intermediate representations, but such approaches struggle in low-data settings. We therefore propose a data-efficient framework TARA (Temporal Action-Reasoning Alignment) that improves policy steerability through precise temporal alignment between reasoning and actions. We heuristically decompose demonstrations into meaningful subgoals using implicit visual and behavioral information and train flow-matching policies to predict the actions required for each subgoal. We show that this temporal alignment increases attention to reasoning traces while preserving learned fine-grained behavioral knowledge. Moreover, aligned policies learn latent representations that better support compositional generalization and demonstrate robustness on LIBERO-Plus (+3.7%), generalizability through improved policy steerability on specifically constructed benchmark (+4.8%) across different subgoal representations, and significant improvements on a real-world Franka robot.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Chain-of-Thought, Embodied Reasoning, Policy Steerability

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

This paper proposes Temporal Action-Reasoning Alignment (TARA), a framework that improves VLA policy steerability by aligning action chunks with intermediate subgoals. The method uses heuristic trajectory segmentation to construct subtask and direction annotations and trains flow-matching policies to predict variable-length action sequences corresponding to subgoal completion. Experiments on LIBERO-Plus, OOD-90, and real-world manipulation demonstrate improvements in generalization and instruction following.

Critical strengths
  1. The temporal mismatch between high-level reasoning and low-level action execution is an important problem in hierarchical VLA policies.

  2. The proposed segmentation and alignment strategy is relatively straightforward and can leverage existing demonstration data.

Critical weaknesses
  1. Insufficient evidence for the independent contribution of temporal alignment. The current experiments do not clearly isolate the effects of subgoal annotation, variable-length action prediction, and their combination. This makes it difficult to determine whether the proposed temporal alignment mechanism is responsible for the reported improvements.

  2. Limited evaluation of generalization and robustness. The OOD evaluation protocol requires further clarification, and the real-world experiments contain very few trials. These limitations weaken the evidence supporting the broader generalization and robustness claims.

Questions for authors
  1. The proposed approach combines subgoal annotations with variable-length action chunking. Steerable Policies [1] explores fine-grained behavioral instructions, while Mixture of Horizons [2] investigates adaptive action horizons. Furthermore, Do What You Say [3] addresses reasoning-action alignment, and LaRA-VLA [4] explores integrating multimodal reasoning into action generation through latent representations. Could the authors clarify the methodological novelty of TARA compared with these approaches, particularly regarding the independent contribution of temporal action-reasoning alignment?

  2. The current experiments do not sufficiently isolate the contributions of fine-grained subgoal annotations and variable-length action chunking. Could the authors provide controlled ablations to demonstrate the independent benefits of temporal alignment?

  3. The paper mentions training high-level planners using LIBERO-90 demonstrations. Could the authors clarify whether the evaluated OOD tasks are seen by the high-level planner and distinguish low-level policy generalization from generalization of the entire hierarchical system?

  4. The real-world experiments include only five trials per task, making the reported improvements difficult to assess statistically. Could the authors conduct more extensive evaluations and investigate robustness to inaccurate subgoal predictions or segmentation errors?

References

[1] Steerable Vision-Language-Action Policies for Embodied Reasoning and Hierarchical Control. arXiv 2026.

[2] Mixture of Horizons in Action Chunking. ICML 2026.

[3] Do What You Say: Steering Vision-Language-Action Models via Runtime Reasoning-Action Alignment Verification. ICRA 2026.

[4] Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models. ICML 2026.

6 comments
AuthorsOpenReview ↗

Author Rebuttal

We thank the reviewer for their feedback. We are glad to hear that the reviewer finds (a) the temporal mismatch between high-level reasoning and low-level action execution an important problem in hierarchical VLA policies, and (b) our segmentation and alignment strategy straightforward and able to leverage existing demonstration data. To address your concerns, we have run an additional ablation to isolate the effect of temporal alignment. We have clarified the OOD-90 protocol, in particular the role of the planner and the distinction between low-level policy generalization and end-to-end generalization, and clarified the methodological novelty of TARA relative to related work, and added qualitative examples that demonstrate recovery from inaccurate subgoals. We are expanding our real-world experiments to 5-6 tasks with 10 trials each, and running an experiment with corrupted segmentation to test robustness to segmentation errors.

W1/Q2: “Isolate the effects of subgoal annotation, variable-length action prediction, and their combination.”

We ablate subgoal annotations and variable-length action prediction and find that both are important for the success of TARA, see below.

SetupOOD-90
0.525.6
+ Subgoal annotations (ECoT-Lite-like)26.1
+ Variable-length action prediction26.0
+ TARA30.9

These results indicate that neither subgoal annotations nor variable-length action prediction alone explains TARA's improvements.

W2/Q3 “Are the evaluated OOD tasks seen by the high-level planner?”

We include end-to-end zero-shot robustness and generalisation with LIBERO-Plus Figure 3 (right) and our real-world experiments (Figure 4). TARA outperforms ECoT-Lite-like by 3.7% on LIBERO-Plus and significantly outperforms baselines on unseen objects in real-world experiments. This demonstrates that TARA generalizes end-to-end/zero-shot. By training a planner on demonstrations for OOD tasks, we assess the generalization capabilities of the low-level policies in OOD scenarios, since our focus is not on improving high-level planning and reasoning. We will clarify this in the paper.

W2/Q4 “The real-world experiments include only five trials per task”

We will expand our real-world experiments to 5-6 tasks, each with 10 trials.

Q1: “Novelty of TARA”

By tying the action horizon to subgoal completion, we improve the causal effect of reasoning on action generation. Steerable Policies [1] investigates adaptive selection of subgoal representations, while we improve steerability of individual representations via TARA. Do What You Say [3] approaches alignment using computationally expensive test-time verification to select among sampled action chunks, while we directly improve the conditional action distribution π(action∣observation,reasoning)\pi(\text{action}|\text{observation},\text{reasoning}). Mixture of Horizons [2] fuses sampled fixed horizons to optimize the test-time trade-off between reactivity and compounding errors, while we train flow-matching policies on variable-length targets aligned with subgoals. We repurpose existing trajectory annotation strategies, such as directional end-effector motion in LaRA-VLA [4] or visual traces in ThinkAct [5], as segmentation heuristics such that any subsegment sampled during training does not contradict the annotation.

Q4: “Robustness to inaccurate subgoal prediction”

Given the identical (imperfect) planner, TARA outperforms ECoT-Lite-like on OOD-90, LIBERO-Plus, and real-world tasks (Figures 3 and 4, Table 1). We use PaliGemma as the planner, which might make mistakes. Nevertheless, TARA beats ECoT-Lite-like by 4.8% on OOD-90 and by 3.7% on LIBERO-Plus (end-to-end zero-shot). We aim to effectively leverage pretraining knowledge, generalizability, and reasoning capabilities of VLMs for embodied control; our goal is not to improve the reasoning or achieve a perfect planner. We include supplementary videos on OOD-90 that show TARA recovering from inaccurate subgoal predictions once the planner self-corrects, while ECoT-Lite-like does not (TARA: videos/tara_subgoal_recovery, videos/tara_subgoal_recovery_2; baseline: videos/baseline_subgoal_recovery, videos/baseline_subgoal_recovery_2).

Q4: “Robustness to segmentation errors”

We will run an experiment with corrupted segmentation.

[1] Steerable Vision-Language-Action Policies for Embodied Reasoning and Hierarchical Control. arXiv 2026

[2] Mixture of Horizons in Action Chunking. ICML 2026

[3] Do What You Say: Steering Vision-Language-Action Models via Runtime Reasoning-Action Alignment Verification. ICRA 2026

[4] Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models. ICML 2026

[5] ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning. NeurIPS 2025

AuthorsOpenReview ↗

Author Rebuttal (Addendum)

Q4: “Robustness to segmentation errors”

As mentioned in our initial rebuttal, we conducted an additional experiment to evaluate TARA's robustness to segmentation errors. Even under substantial segmentation errors, TARA outperforms the uncorrupted baselines. Introducing random boundary jitter to our LIBERO training data, with mean displacements of 25% and 50% of the original segment length (approximately 0.45s and 0.9s, respectively, given a mean segment length of 1.8s), reduces TARA's OOD-90 success rate by only 2.5% and 1.6%, respectively. This demonstrates that TARA remains effective even with substantial segmentation noise.

AuthorsOpenReview ↗

Author Rebuttal (Addendum 2)

W2/Q4 "real-world experiments contain very few trials"

We have expanded real-world experiments from 3 to 5 tasks and from 15 to 25 trials per method. The two new tasks test generalization to unseen goals and spatial reasoning. Full success is shown, with partial success in parentheses:

TaskECoT-Lite-likeTARA
put the apple on the plate0.40 (1.00)0.60 (1.00)
put the apple in the basket0.40 (1.00)0.80 (1.00)
put the pear on the plate0.00 (0.00)0.40 (0.60)
put the apple on the left plate (NEW)0.20 (0.80)0.40 (0.60)
put the apple next to the banana (NEW)0.20 (0.20)0.20 (0.80)
Average0.24 (0.60)0.48 (0.80)

TARA’s average full success rate is 48%, double that of ECoT-Lite-like (24%), and outperforms or ties ECoT-Lite-like on all tasks. Due to robot availability, we could not run the 10 trials per task we proposed earlier. We will extend the real-world evaluation to 10 trials per task in the final version.

Reviewer_yHbAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript presents Temporal Action-Reasoning Alignment (TARA), a framework designed to improve policy steerability in Vision-Language-Action (VLA) models. The authors argue that conventional fixed-horizon action chunking can introduce inconsistencies between intermediate reasoning subgoals and the actions executed by the policy, potentially causing the policy to ignore reasoning signals. TARA addresses this issue by heuristically segmenting demonstrations into meaningful subgoals and training flow-matching policies to predict variable-length action chunks aligned with subgoal completion. Experiments on OOD-90, LIBERO-Plus, and real-world Franka manipulation tasks report improvements in generalization and robustness. Additional analyses investigate direction-following behavior, attention allocation, and latent representations.

Critical strengths

Strengths

  1. TARA identifies a relevant limitation of hierarchical VLA policies: fixed-horizon action chunks may extend beyond the completion of the current reasoning subgoal, introducing inconsistencies between intermediate instructions and executed actions. Aligning action horizons with subgoal boundaries is an intuitive and potentially useful approach.

  2. Experiments are reasonably conducted across multiple settings.

  3. Additional behavior analysis is acknowledged and appreciated for understanding the policy steerability.

Critical weaknesses
  1. Insufficient isolation of the contribution of temporal alignment. A central concern is whether the reported improvements can be attributed primarily to temporal action-reasoning alignment. Appendix A.4.1 indicates that TARA additionally incorporates next-subgoal supervision for conquering the OOD-90 benchmark, whereas the baselines are trained using only the current subgoal. Moreover, Appendix Table 4 reports a substantial performance drop between TARA with and without next-subgoal chaining (97.3% versus 87.0% on LIBERO), while the temporal action-reasoning alignment mechanism is identical for both settings.

  2. Overstated Generalization claim. The OOD-90 benchmark result requires careful interpretation. On page 18, the authors state that "While the low-level policy is trained only on the standard LIBERO dataset, the high-level planner is finetuned on demonstrations for LIBERO-90." Therefore, this is a reasonable setting for investigating low-level policy generalization under task-informed subgoals, but it does not establish end-to-end zero-shot generalization of the proposed architecture as a system. The claims should be qualified accordingly.

  3. Inconsistent results between Figure 3 in the main body and Table 3 in the appendix section. In the main result (Figure 3), the average SR seems above 30%, whereas the author reports 29.1& in Table 3 (appendix).

  4. Reliability of the real-world results. Figure 4 shows an unseen-layout result between 81 and 82%, while the described protocol involves two tasks with five trials each and discrete partial-success credits. Could the authors provide an explanation of how this percentage is calculated?

Questions for authors

Please check the weakness. Moreover, could the authors discuss and analyse the failure recovery behavior of TARA?

4 comments
AuthorsOpenReview ↗

Author Rebuttal

We thank the reviewer for their feedback. We are glad to hear that the reviewer finds that our work (a) addresses a relevant limitation of hierarchical VLA policies, with aligning action horizons to subgoal boundaries being an intuitive and potentially useful approach, (b) is evaluated reasonably across multiple settings, and (c) is accompanied by behavior analyses that help in understanding policy steerability.

To address your concerns, we ran an additional experiment that isolates temporal alignment from next-subgoal chaining, clarified the scope of our OOD-90 claims and the replanning-frequency difference behind Figure 3 and Table 3, explained how we compute real-world success rates, and added a qualitative analysis of failure recovery. We will expand our real-world evaluation to 5-6 tasks with 10 trials each and revise our paper accordingly.

W1: “Contribution of temporal alignment”

We trained TARA without chained subgoals and it scores 30.9% on OOD-90, equal to our previous results. We first clarify that, for all our experiments (Figures 3 and 4, Table 1), ECoT-Lite-like and TARA share the identical planner trained on chained subgoals. However, policies differ in how they receive chained subgoals: TARA is trained to receive chained subgoals directly, ECoT-Lite-like is trained on individual subgoals, and during inference, chained subgoals are split and passed individually. The reason for this difference is replanning frequency: during inference, ECoT-Lite-like replans every 10 steps, while TARA replans every 1-10 steps. To achieve a comparable replanning frequency of 5-10 steps, in the paper, we train on chained subgoals when subgoal completion is within 5 steps. To demonstrate that this only impacts replanning frequency, we trained TARA on individual subgoals as ECoT-Lite-like. By requeuing the policy without updating subgoals during inference until at least 5 steps are executed, we receive a replanning frequency of 5-10 steps and overall achieve 30.9% on OOD-90, equal to TARA in Figure 3/Table 1; without this mechanism, we score 29.1%, still above ECoT-Lite-like (26.1%). This isolates the contribution of temporal alignment; our final policy is trained on chained subgoals, as it provides lower inference latency.

W2: “OOD-90 does not establish end-to-end zero-shot generalization”

We include end-to-end zero-shot robustness and generalisation with LIBERO-Plus Figure 3 (right) and our real-world experiments (Figure 4). TARA outperforms ECoT-Lite-like by 3.7% on LIBERO-Plus and significantly outperforms baselines in real-world experiments. This shows that TARA generalizes end-to-end/zero-shot. By training a planner on demonstrations for OOD tasks, we assess the generalization capabilities of the low-level policies in OOD scenarios, since our focus is not on improving high-level planning and reasoning. We will clarify this.

W3: “Inconsistency between Figure 3 and Table 3”

Figure 3/Table 1 reports TARA with a replanning frequency in [5,10] (trained), while Table 3 uses a replanning frequency in [1,10] (see W1 for more information). We will add clarifications in the Appendix.

W4: “Reliability of the real-world results”

We will expand our real-world experiments to 5-6 tasks, each with 10 trials.

W4: “The percentage in Fig 4”

Thank you for pointing this out. The value shown in Figure 4 is slightly off, due to a miscalculation. TARA achieves 85% on unseen layouts by scoring 4/5 on “put the apple on plate” (3/5 complete successes, 2 partial successes = 1/5) and 4.5/5 on “put the apple in basket” (4/5 complete successes, 1 partial success = 0.5/5).

Q1: “Failure recovery behavior of TARA”

Since TARA and the baseline policies share the identical planner, we analyse and classify the behaviors that enable recovery from baseline failures on OOD-90. The table below shows the percentage of TARA-only successes attributable to each behavior.

TARA behavior% of tasks
Improved steerability54.5
Recovery from a wrong subgoal27.3
OOD behavior transfer18.2

Examples in supplementary

  • Steerability (“put the red mug on the left plate”): The ECoT-Lite-like baseline (videos/baseline_steerability) ignores the instruction and the planner’s subgoal and executes the most familiar task instead; TARA (videos/tara_steerability) uses subgoals to generalize to the unseen task.
  • Subgoal recovery (“pick up the tomato sauce and put it in the tray”): In both cases, the planner first steers left, then corrects itself. TARA (videos/tara_subgoal_recovery) follows the wrong subgoal but recovers with the next one; the baseline (videos/baseline_subgoal_recovery) does not.
  • OOD behavior transfer (“open the top drawer of the cabinet”): TARA (videos/tara_behavior) applies fine-grained behaviors such as opening drawers in unseen settings; the baseline (videos/baseline_behavior) fails to transfer.
AuthorsOpenReview ↗

Author Rebuttal (Addendum)

W4: “Reliability of the real-world results”

We have expanded real-world experiments from 3 to 5 tasks and from 15 to 25 trials per method. The two new tasks test generalization to unseen goals and spatial reasoning. Full success is shown, with partial success in parentheses:

TaskECoT-Lite-likeTARA
put the apple on the plate0.40 (1.00)0.60 (1.00)
put the apple in the basket0.40 (1.00)0.80 (1.00)
put the pear on the plate0.00 (0.00)0.40 (0.60)
put the apple on the left plate (NEW)0.20 (0.80)0.40 (0.60)
put the apple next to the banana (NEW)0.20 (0.20)0.20 (0.80)
Average0.24 (0.60)0.48 (0.80)

TARA’s average full success rate is 48%, double that of ECoT-Lite-like (24%), and outperforms or ties ECoT-Lite-like on all tasks. Due to robot availability, we could not run the 10 trials per task we proposed earlier. We will extend the real-world evaluation to 10 trials per task in the final version.

Dear Reviewer yHbA,

We appreciate your thoughtful comments on our paper. We hope that our rebuttal has addressed your concerns. If you have any further comments or questions, we’d be happy to clarify.

Best,

Authors of Submission51100

The authors have conducted additional experiments that provide further support for their claims and have committed to clarifying the remaining ambiguities in the revised manuscript. Based on these responses, I will raise my score to weak accept.

Reviewer_BaTQOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TARA (Temporal Action-Reasoning Alignment), improving generalization performance and policy steerability in VLA models. Current VLA models tend to ignore intermediate CoT reasoning when generating actions, relying on visual information. To address this issue, TARA segments demonstration trajectories into meaningful subgoals using heuristics (e.g., zero-crossings of translational velocities). It trains flow-matching policies to predict variable action chunks that cover the remaining steps required to complete the subgoal. Experiments using LIBERO-Plus, the OOD-90 benchmark, and a real-world Franka robot showed that temporal alignment increased attention to reasoning traces and enhanced both compositional generalization and policy steerability.

Critical strengths
  • The approach of dynamically adjusting the chunk length to match the remaining steps of the subgoal is a straightforward solution to the misalignment between reasoning and action chunks.
  • The paper writing is easy to follow; especially, Figure 1 is intuitive.
Critical weaknesses

The dataset segmentation relies heavily on a rule-based heuristic, such as zero crossings in translational velocities. This might not work well for tasks dominated by rotational manipulations or highly dynamic, complex manipulations. Reasoning errors from the high-level planner can directly propagate to the actions, which seems critical but is not evaluated much in the paper.

Questions for authors
  • The proposed segmentation method heavily relies on rule-based heuristics. How robust are these heuristics when applied to massive, unstructured datasets like OXE, where human teleoperation noise is significant?
  • Does predicting the exact sequence length of a subgoal reduce the closed-loop reactivity of the system in highly dynamic environments?
5 comments
AuthorsOpenReview ↗

Author Rebuttal

We thank the reviewer for their feedback. To address your concerns, we have clarified that TARA’s efficacy does not depend on a specific segmentation scheme and that segmentation is not our main contribution. We have also clarified how TARA performs better than baselines even with an imperfect planner and how it retains closed-loop reactivity. We will add details on our segmentation setup to the Appendix. We are also running an experiment that trains TARA on corrupted segmentations to evaluate robustness to segmentation errors. We answer your questions below.

W1/Q1: “Rule-based heuristics might not work well for tasks dominated by rotational manipulations or highly dynamic, complex manipulations; applicability to datasets like OXE”

TARA’s efficacy does not depend on a specific segmentation scheme (Section 5.5); the segmentation scheme is not our main contribution. In addition to the proposed segmentation scheme, we evaluate TARA with the visual traces. As shown in Table 1, TARA (Visual Traces) outperforms ThinkAct-like (a respective baseline) by 6.6% on OOD-90 and 4.4% on LIBERO-Plus (spatial suite). Further, visual trace planners are pretrained on OXE and have been shown to be effective for various manipulation tasks [1,2]. This shows that TARA is robust to the choice of segmentation heuristics, and it can be used with unstructured datasets like OXE. Additionally, since both LIBERO and our real-world data are human teleoperated, we smooth velocities and merge nearby boundaries or non-contradictory segments to reduce noise. We will add more details to the appendix.

W2: “Reasoning errors from the high-level planner can directly propagate to the actions”

Given the identical (imperfect) planner, TARA outperforms ECoT-Lite-like on OOD-90, LIBERO-Plus, and real-world tasks (Figures 3 and 4, Table 1). We use PaliGemma as the planner, which might make mistakes. Nevertheless, TARA beats ECoT-Lite-like by 4.8% on OOD-90 and by 3.7% on LIBERO-Plus (end-to-end zero-shot). We aim to effectively leverage pretraining knowledge, generalizability, and reasoning capabilities of VLMs for embodied control; our goal is not to improve the reasoning or achieve a perfect planner.

We additionally provide supplementary videos on OOD-90 that show TARA recovering from inaccurate subgoal predictions once the planner self-corrects, while the ECoT-Lite-like baseline cannot (TARA: videos/tara_subgoal_recovery, videos/tara_subgoal_recovery_2; baseline: videos/baseline_subgoal_recovery, videos/baseline_subgoal_recovery_2).

Q2: “Does predicting the exact sequence length of a subgoal reduce the closed-loop reactivity?”

The benefit of TARA comes from training, not inference: our policies learn better representations (Section 5.4) and learn to ground action generation in subgoals (Section 5.3). In practice, we already truncate execution to the first H=10 actions (Lines 303-304).

[1] ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning. NeurIPS 2025.

[2] Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation. ICLR 2026.

AuthorsOpenReview ↗

Author Rebuttal (Addendum)

Q1: "Robustness to significant human teleoperation noise"

As mentioned in our initial rebuttal, we conducted an additional experiment to evaluate TARA's robustness to segmentation errors. Even under substantial segmentation errors, TARA outperforms the uncorrupted baselines. Introducing random boundary jitter to our LIBERO training data, with mean displacements of 25% and 50% of the original segment length (approximately 0.45s and 0.9s, respectively, given a mean segment length of 1.8s), reduces TARA's OOD-90 success rate by only 2.5% and 1.6%, respectively. This demonstrates that TARA remains effective even with substantial segmentation noise.

Dear Reviewer BaTQ,

We are grateful for your review. Please let us know if any points in our rebuttal need further clarification. We are happy to discuss them.

Best,

Authors of Submission51100

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.