acceptodds
Under review as a conference paper at ICLR 2027

TacitVLA: Reasoning Vision-Language-Action Models Without Reasoning Latency

Abstract

Textual reasoning before acting has improved vision-language-action (VLA) models, but at the cost of latency: every action conditioned on the reasoning text must wait for autoregressive decoding. We show that actions do not need to wait for or condition on the generated reasoning to benefit from it, because the reasoning supervision already improves the backbone representation. We introduce TacitVLA, a reasoning VLA whose action expert never attends to the reasoning it generates for the current call. On long-horizon tasks where the reasoning text also serves as episodic memory, it can generate each entry during action execution rather than before acting. We evaluate it in a controlled study on four benchmarks spanning manipulation and autonomous driving, memory and non-memory tasks: RoboMME, RMBench, LIBERO, and PhysicalAI-AV. On all four, TacitVLA matches the performance of its explicit-reasoning counterpart at up to 3.4× lower latency, and matches the latency of the same architecture trained without reasoning at up to +17.1% success. In absolute terms, it surpasses the best previously published results on the RoboMME Counting suite (82.7% vs. 66.8%) and RMBench (77.5% vs. 56.5%), and reaches 72.9% on LIBERO-Plus and a 52.5% scene score on the public split of the PhysicalAI-AV closed-loop driving challenge.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, embodied reasoning, robot learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SYEDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether VLA strategies must rely on reasoning text generated during the current inference process when executing actions. The proposed approach maintains supervision over the reasoning process but masks the current reasoning tokens when passing information to a flow-matching-based action execution module. During deployment, the system skips text generation for tasks that do not require memory; for tasks involving memory, it generates "ledger entries" alongside action execution, which are then provided to subsequent inference steps. The study evaluates TacitVLA against "Explicit Reasoning" and "No Reasoning" baselines using benchmarks such as RoboMME, RMBench, LIBERO/LIBERO-Plus, and PhysicalAI-AV. Results demonstrate that TacitVLA achieves overall performance comparable to "Explicit Reasoning" while reducing per-inference latency by up to 3.4 times.

Critical strengths
  1. The comparison between "Explicit Reasoning" and "Tacit VLA" employs identical training sequences, data, prompts, and loss function formulations; the sole distinction lies in whether the "action expert" has access to the current reasoning tokens. This approach is designed to investigate directly whether action generation must rely on previously generated reasoning content, rather than merely comparing two entirely disparate systems. The comparative analysis presented in Table 1 further reinforces this argument.
  2. A simple deployment strategy has enabled a significant reduction in latency. Latency per invocation decreased from 495 ms to 147 ms for LIBERO tasks, and from 220 ms to 127 ms for the RoboMME Counting task. This method achieves a tangible improvement in efficiency without requiring the sacrifice of essential textual memory.
Critical weaknesses
  1. Novelty is limited, and the additional empirical insights presented require more robust substantiation. As noted in Section 2, existing studies (e.g., ECoT-Lite, VLA-OS, and ZR-0) have already explored "reasoning-supervision-based strategies," which eliminate the text-generation phase during deployment. The authors distinguish their work from these predecessors through more rigorous experimental controls, an extension to memory-based tasks, and the inclusion of latency evaluations. While these contributions are valuable, the core advancement lies in the execution of controlled empirical studies and the proposal of deployment strategies, rather than the establishment of a fundamentally new principle for "decoupling reasoning from action." Consequently, the significance of this paper hinges largely on the new understanding derived from its experiments.
  2. Aggregated similarity metrics mask substantial differences in performance regarding robustness and safety. In the LIBERO-Plus benchmark (Table 12), TacitVLA achieves an overall success rate comparable to that of the "Explicit Reasoning" approach (72.9% versus 73.5%). However, when subjected to camera perturbations, TacitVLA’s performance drops by 9.3 percentage points (42.2% versus 51.5%), whereas it improves by 8.9 percentage points (60.0% versus 51.1%) when faced with sensor perturbations. Consequently, this aggregated similarity reflects, to some extent, a balancing of strengths and weaknesses rather than a robust retention of generalization capability.
  3. Based on comparisons involving memory tasks, it is not possible to isolate the representational gains specifically attributable to "reasoning supervision." Section 4.1 notes that the impact of the supervision path can be isolated by comparing TacitVLA against a "No Reasoning" version. However, Section 5 explicitly acknowledges that the "No Reasoning" version lacks both reasoning supervision and the "ledger" mechanism required for memory tasks; consequently, this comparison measures the combined benefits derived from both.
Questions for authors
  1. What new findings does this study offer that go beyond the existing conclusions in the literature discussed in Section 2?
  2. How can the discrepancy between the results for "Camera" and "Sensor" in Table 12 be explained? Furthermore, how robust are these findings when faced with changes in evaluation methodology or the inherent randomness of independent training processes?
  3. Which experimental results specifically isolate the impact of "reasoning supervision," and which ones measure its combined effect when paired with "ledger access"? Is there evidence to suggest that "reasoning supervision" provides independent benefits, even when memory inputs remain constant?
4 comments

We thank the reviewer for recognizing that our comparison changes only the action expert's access to the current reasoning, so it directly tests whether action generation must rely on the generated reasoning. We also thank the reviewer for noting that latency falls from 495 ms to 147 ms on LIBERO, and that the memory tasks keep their textual memory. We ran four new controls, summarized in the table under W1, and answer the camera and memory questions in our second comment (W2 and W3).

(W1 + Q1) New findings beyond ECoT-Lite, VLA-OS, and ZR-0

As the reviewer notes, the more rigorous controls, the extension to memory tasks, and the latency evaluation are valuable contributions. We agree that the significance of the paper hinges on the new understanding derived from its experiments. We hope to clarify that this understanding goes beyond, and in part revises, the existing conclusions in the literature discussed in Section 2.

On the central question of whether a policy whose action never conditions on the current reasoning can match one that does, prior work disagrees. In ECoT-Lite [1], the arms whose actions never attend to the reasoning (reasoning pre-training and co-training) trail full ECoT by 3.7 to 8.1 points (their Tables 1 and 2), though these arms also change the training mixture. Its authors conclude that "full test-time reasoning has the best performance, as expected". Because ECoT-Lite's reasoning dropout deploys without the text but trains the action to attend to the reasoning whenever it is kept, it does not test whether an action that never conditions on the reasoning can match one that does. In a different architecture, VLA-OS [2] finds that implicit planning beats explicit planning. There, separate planning heads rather than the backbone write the plan, and the policy is a 0.5B model trained from scratch, where explicit planning scores below the action-only baseline. An arXiv preprint from June 2026, ZR-0 [3], claims "without any performance loss" yet reports no experiment testing this: it includes no explicit-reasoning or matched no-reasoning policy, removes its entire pretraining stage in its only ablation, and drops reasoning supervision on its benchmarks.

We vary only the action expert's attention within one training sequence, on pretrained backbones of the architecture recent reasoning VLAs adopt. The masked policy matches a strong explicit-reasoning policy in overall success on all four benchmarks at up to 3.4 times lower latency. When only the attention changes, conditioning the action on the current reasoning is therefore not needed for overall success.

We also study a regime that none of these prior works considers: memory tasks, where the reasoning text must persist as a ledger. Because TacitVLA generates each entry during action execution, generating the ledger adds no latency (Section 6.2), and it maintains a correct ledger in 99.7% of Press Button episodes (Table 14).

Our new experiments add findings that none of these works tests:

New findingEvidence
On memory tasks, a ledger alone does not reproduce the gainWith the ledger kept and reasoning supervision removed, a policy reaches 66.8 on RoboMME Counting even with ground-truth entries, against TacitVLA's 89.5 (W3 below)
The gain requires task-relevant reasoning contentSupervised but task-irrelevant text (another task's reasoning, unrelated English, or random words) recovers at most 1.6 of the 5.3-point LIBERO-Plus gain
Reasoning supervision improves the backbone without the action-expert gradientWith the action gradient detached from the backbone, TacitVLA still outperforms No Reasoning on LIBERO-Plus (69.4 vs. 67.6)
Omitting the text without the training-time mask loses performanceOmitting generation from the Explicit Reasoning checkpoint, at TacitVLA's latency, lowers LIBERO-Plus success to 70.5, below TacitVLA's 72.9

These results show where the benefit of reasoning enters the action. Reasoning reaches the action through the representation it trains, so the action need neither wait for nor attend to the reasoning text, and on memory tasks, its generation can move off the action's critical path. The core finding holds across policies of 2.5B to 3.9B parameters built on pretrained backbones, three reasoning formats, and four benchmarks in two domains, with 50 demonstrations per task to 3.8 million real driving events. The evaluation includes closed-loop driving, in which policies trained on real recordings act on out-of-distribution, simulator-rendered images.

In the revised paper, Section 2 now states these distinctions explicitly, and Section 6.1 and Appendix B add the new findings.

[1] Chen et al. Training Strategies for Efficient Embodied Reasoning (ECoT-Lite). CoRL 2025. https://arxiv.org/abs/2505.08243

[2] Gao et al. VLA-OS. NeurIPS 2025. https://arxiv.org/abs/2506.17561

[3] Li et al. ZR-0. arXiv 2026. https://arxiv.org/abs/2606.30552

(W2 + Q2) Camera and sensor axes on LIBERO-Plus

We agree that aggregate metrics can mask per-axis differences that matter for robustness and safety. We hope to clarify that our claim is that TacitVLA matches Explicit Reasoning at under a third of its latency (147 vs. 497 ms on LIBERO-Plus), so the question for each axis is whether TacitVLA matches. It matches within 1.4 points on five of seven axes, and the two exceptions point in opposite directions (differences computed from Table 12):

Difference (points)Robot init.SensorCameraLanguageObjectsBackgroundLightAll
TacitVLA − Explicit Reasoning−1.2+8.9−9.3−0.7−1.2+0.8−1.4−0.6
TacitVLA − No Reasoning+10.2+8.3+1.2+6.4+3.9+5.1+1.4+5.3

TacitVLA also scores above No Reasoning, which runs at the same latency, on all seven axes.

Camera and sensor are exactly the axes known to swing with changes to a policy's design. In the LIBERO-Plus evaluation ([1], Table 2), removing OpenVLA-OFT's wrist camera costs it 46.0 points on camera and 25.9 on sensor noise, and π0 and π0-FAST, which share a PaliGemma backbone, differ by 51.3 points on camera.

The paper's findings also hold under changes in evaluation methodology, and on RoboMME when each policy is retrained at other loss weights. Every per-axis result in Table 12 has a standard deviation of at most 1.1 points. The match holds on standard LIBERO (98.6 vs. 98.1) as it does under perturbations, and in both driving evaluations: open loop on held-out real recordings (minADE 1.116 vs. 1.140 m) and closed loop on out-of-distribution, simulator-rendered images (scene score 52.5 vs. 52.3). In Table 11, each policy is retrained at three cross-entropy weights in two RoboMME setups. TacitVLA exceeds No Reasoning in all six settings, ranks best among the three policies in five, and is within 1.2 points of Explicit Reasoning in the sixth.

The revised paper adds this per-axis analysis to Section 6.1 and qualifies the claim in the abstract and conclusion to overall success.

(W3 + Q3) Isolating reasoning supervision from ledger access

We agree that on memory tasks, TacitVLA against No Reasoning measures the combined benefit of reasoning supervision and ledger access, as Section 5 states. On the non-memory tasks, where no policy keeps a ledger, that comparison already isolates reasoning supervision (Section 4.1). On the memory tasks, the paper's central comparison holds ledger access constant: TacitVLA and Explicit Reasoning keep a ledger in the same way, and TacitVLA matches it (89.5 vs. 86.7 on RoboMME Counting, 77.5 vs. 77.0 on RMBench). We also trained a new control on RoboMME Counting under the same protocol as Table 2: TacitVLA's recipe with ledger access kept and only the supervision on the current reasoning text removed. We clarify that on these tasks, the reasoning text is the predicted next subtask, which becomes the next ledger entry, so a policy without this supervision never learns to write its own ledger. At test time, the simulator therefore supplies its memory inputs: the ground-truth subgoal at each replan, in the wording of the training labels. A deployed policy would not have these privileged subgoals, which is why the submission trained only deployable policies. As a diagnostic, however, the control tests whether ledger access with correct memory inputs can replace reasoning supervision.

Ledger access alone does not reproduce the gain, even with ground-truth memory inputs:

PolicyReasoning supervisionLedger at testSuccess (%)
No Reasoningnonone79.0 ± 1.0
Controlnosimulator's ground truth66.8 ± 12.2
TacitVLAyesits own entries89.5 ± 1.3

TacitVLA, which writes its own ledger, scores 22.7 points above the control (89.5 vs. 66.8). We believe the control falls even below No Reasoning (66.8 vs. 79.0) because predicting the next subtask is what teaches the backbone to track progress from the ledger. Evidence that reasoning supervision has an independent benefit with memory inputs held constant comes from the ledger-free tasks, where it adds 5.3 points on LIBERO-Plus and 6.1 in closed-loop driving (Tables 5 and 6), and from the stop-gradient and text controls (W1), which also run without a ledger. We have added this control to Table 2 and Section 6.1 of the revised paper, and the revised Sections 4.1 and 5 state which comparisons isolate reasoning supervision.

Overall

The reviewer noted that the significance of the paper hinges on the new understanding derived from its experiments. We hope the findings above, now in the revised paper, provide it, and would be grateful if the reviewer would consider updating their assessment.

[1] Fei et al. LIBERO-Plus. CVPR 2026. https://arxiv.org/abs/2510.13626

Thank you for the detailed responses and additional experiments, which strengthen the empirical analysis and address some of my concerns. However, my concerns about the novelty beyond prior work, robustness across independent training runs, and the separation of reasoning supervision from ledger access remain only partially resolved. I will therefore maintain my original rating

We thank the reviewer for the follow-up and for noting that the new experiments strengthen the analysis. We address each of the three remaining points below.

Novelty beyond prior work. The prior works the reviewer lists leave open the question of whether a reasoning VLA must pay the latency of reasoning at deployment. ECoT-Lite [1] concludes that "full test-time reasoning has the best performance, as expected," which our controlled comparison contradicts. VLA-OS [2] and ZR-0 [3] point the opposite way and are conducted in settings that do not isolate the action's access to the reasoning. TacitVLA provides the missing comparison by changing only that access, showing that the action does not need to wait for the reasoning on all four benchmarks, spanning manipulation and driving, and achieving up to 3.4 times lower latency at no cost in overall success.

Robustness across independent training runs. To directly answer this, we independently trained each policy three times on RoboMME Counting with different seeds:

PolicyRun 1Run 2Run 3Mean ± sd
No Reasoning79.078.579.579.0 ± 0.5
Explicit Reasoning87.587.586.087.0 ± 0.9
TacitVLA89.589.089.589.3 ± 0.3

Each policy varies by at most 1.5 points across runs, and the ordering of Table 2 in our paper holds in every run, particularly with TacitVLA matching Explicit Reasoning's performance and outperforming No Reasoning.

Separating reasoning supervision from ledger access. We agree that the separation should be studied, and our results do so directly. On RoboMME, the control keeps ledger access, removes only reasoning supervision, and even receives the simulator's ground-truth ledger entries while TacitVLA must write its own. Despite this advantage, the control trails TacitVLA by 22.7 points (66.8 vs. 89.5, Table 2 in our paper). On LIBERO-Plus and in closed-loop driving, where no policy keeps a ledger, reasoning supervision alone adds 5.3 and 6.1 points over No Reasoning (Tables 5 and 6 in our paper). Therefore, reasoning supervision brings a benefit of its own, both when ledger access favors the control and when no ledger is used at all.

[1] Chen et al. Training Strategies for Efficient Embodied Reasoning (ECoT-Lite). CoRL 2025. https://proceedings.mlr.press/v305/chen25a.html

[2] Gao et al. VLA-OS. NeurIPS 2025. https://arxiv.org/abs/2506.17561

[3] Li et al. ZR-0. arXiv 2026. https://arxiv.org/abs/2606.30552

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

This paper asks whether, in reasoning VLAs, the benefit of textual reasoning reaches the action through test-time conditioning, or is already carried by the supervision that trains the backbone. The authors separate three roles of the reasoning text — training target, episodic memory, inference-time input — and argue only the third is dispensable. Three policies sharing all training ingredients differ by a single attention-mask cell: Explicit Reasoning attends to the reasoning tokens, TacitVLA does not. Across four benchmarks covering manipulation and driving, memory and non-memory tasks, TacitVLA matches Explicit Reasoning in success at no-reasoning latency, 1.4×–3.4× lower.

Critical strengths
  1. The problem is real and the framing is clean. Explicit reasoning imposes a serial cost that is difficult to hide: every action chunk conditioned on generated text must wait for autoregressive decoding. Inference acceleration and latent/implicit reasoning reduce the constant but do not remove the dependency.
  2. The experimental design is simple and reproducible. TacitVLA achieves the decoupling through attention masking, without introducing an additional reasoning model or a complicated inference pipeline. which makes the result easy to adopt and easy to check.
  3. The empirical coverage is broad and latency is measured. Four benchmarks, both manipulation and driving, both memory and non-memory regimes, with per-stage timing alongside success rates, is substantially more than the claim strictly requires. The absolute numbers (RoboMME Counting 89.5, RMBench 77.5, LIBERO-Plus 72.9) are competitive on their own terms, and reporting the deferred-generation schedule and its execution-window budget is good practice.
Critical weaknesses

1.Does explicit reasoning actually help the action? This deserves a more careful discussion, since prior work reports that explicit reasoning brings little performance gain. My own view is that reasoning helps generalization only once it genuinely scales: on small domains or datasets the model is still essentially fitting, and the reasoning adds no information beyond a retrieval augmentation over the visual input. Experiments with a single CoT format and a narrow domain therefore seem insufficient to strongly support the authors' conclusion.

2.The claim that reasoning improves the backbone is plausible, but as written it remains unexplained — how is it improved, and what exactly is improved? Reasoning carries an inherent logical organization (object grounding, motion direction, etc.), so could the authors analyze this further? For instance, would the benefit persist if the reasoning were unrelated to the task?or what will happen ifthe resoning cot is meaningless?I think this could help to know whether the model learns reasoning or just retrieval augmentation over the visual input

3.Is there a stop-gradient ablation, i.e., one in which the action gradient cannot reach the VLM backbone? If the co-trained model still performs better under that setting, it would in fact strengthen the authors' conclusion.

4.Memory control: On memory tasks, No Reasoning keeps the observation but drops the ledger entirely, so the reported gains conflate reasoning supervision with symbolic memory — Press Button moving from 7.7 to 73.3 is largely a statement about having a ledger. Can the authors add a control that keeps the ledger but removes reasoning supervision, to isolate the contribution the paper actually argues for?

Questions for authors

Generalization: Does the benefit of reasoning supervision increase with training data scale and task complexity? Can the authors demonstrate improvements beyond in-distribution task performance?

Gradient pathways: Can the authors add an ablation in which the action gradient is detached from the backbone?perfomance drop is expected for all models,but this result would strengen the author's conclusion

Semantic content versus logical structure: Would ordinary scene descriptions, or task-relevant text without explicit reasoning structure, yield similar results? What happens when reasoning is replaced with carefully controlled task-irrelevant text or non-semantic tokens?

Memory control: On memory tasks, No Reasoning drops the ledger entirely, so it measures the joint value of reasoning supervision and symbolic memory. Can the authors add a control that keeps the ledger but writes its entries without reasoning supervision, to isolate the contribution the paper argues for?

4 comments

We thank the reviewer for recognizing that the problem is real and the framing is clean, and that the design is simple and reproducible. We also thank the reviewer for noting the broad empirical coverage, with latency measured and absolute numbers competitive on their own terms. We ran all four analyses the reviewer requested, including the stop-gradient and ledger controls in our second comment (W3 and W4).

(W1 + Q1) Does reasoning help the action, and beyond in-distribution performance?

We agree that this deserves a careful discussion. We hope to clarify that this concern, together with our central claim that TacitVLA keeps the benefit of reasoning at the latency of a policy without it, is exactly why the paper spans three reasoning formats, four benchmarks, three pretrained backbones, and two domains (Table 1). As a result, the paper's central conclusion does not rest on a single CoT format or a narrow domain. The same breadth answers the reviewer's question: with the architecture and data held fixed, TacitVLA improves on No Reasoning in every setting:

Benchmark (evaluation)Training dataReasoning textNo Reasoning (%)TacitVLA (%)Gain
RMBench (memory)50 demonstrations per tasknext subtask60.477.5+17.1
RoboMME Counting (memory)400 demonstrationsnext subtask79.089.5+10.5
LIBERO-Plus (zero-shot perturbations)1,817 demonstrationsscene description and plan67.672.9+5.3
PhysicalAI-AV closed loop (trained on real video, tested on rendered scenes)3.8 million driving eventschain of causation46.452.5+6.1

On data scale, the gain is present at every training-set size we study, from 50 demonstrations per task to 3.8 million driving events. At the largest scale, where Explicit Reasoning lifts the closed-loop scene score from 46.4 to 52.3 (Table 6), TacitVLA matches it without generating any text (52.5). On task complexity, the gain is largest on the long-horizon memory tasks, where the reasoning also writes the ledger (W4). Beyond in-distribution performance, the standard LIBERO suites are near saturation, with all three policies within 0.5 points (98.5, 98.1, and 98.6), while LIBERO-Plus's zero-shot perturbations show a 5.3-point gain. Closed-loop driving is out of distribution twice over: the policies are trained only on real camera recordings but act on images rendered by AlpaSim [1], in held-out scenes where their own actions take them beyond the logged trajectories. There, the gain is 6.1 points. Evaluations in distribution may therefore show little gain from reasoning, which could explain the prior reports the reviewer mentions.

In W2 to W4 below, we analyze how this benefit arises. The gain requires reasoning that is correct for the current task (W2), reasoning supervision improves the backbone without the action gradient (W3), and its benefit holds beyond the ledger on memory tasks (W4). We have added this summary to Section 6.1 of the revised paper.

(W2 + Q3) Semantic content versus logical structure

We thank the reviewer for suggesting these controls, which directly test what reasoning supervision teaches the backbone. As for how the backbone improves, the cross-entropy on the reasoning text trains the backbone, and the action expert conditions on the backbone's representation of the context (Section 4.1), so what improves is that representation. We clarify that LIBERO's reasoning text already pairs an object-grounded scene description with the subtask plan (Table 1), so the ordinary scene description the reviewer mentions is part of the reasoning text itself.

To test whether its content or its logical structure matters, we retrained TacitVLA under the same protocol as Table 5 with this text replaced in three ways, each matched in length: another task's reasoning (same structure, wrong scene and goal), unrelated English (Wikipedia prose, fluent but task-irrelevant), and random words from the LIBERO reasoning vocabulary (meaningless). Task-relevant reasoning carries almost all of the benefit:

Reasoning text in trainingStructureTask contentLIBERO-Plus (%)Gain over No Reasoning
None (No Reasoning)––67.6–
Random wordsnono68.0+0.4
Unrelated Englishfluentno68.2+0.6
Another task's reasoningyeswrong task69.2+1.6
The task's own reasoning (TacitVLA)yesyes72.9+5.3

Another task's reasoning, with the same structure and vocabulary but the wrong content, recovers only 1.6 of the 5.3-point gain, and fluent or meaningless text recovers almost none. The benefit therefore comes from semantic content that is correct for the current scene and goal, rather than from the text's structure alone or from generic language modeling. That benefit also reaches the action without the text ever being generated. We have added these controls to Section 6.1 of the revised paper, with the full results in Appendix B.

[1] NVIDIA et al. AlpaSim. 2025. https://github.com/NVlabs/alpasim

(W3 + Q2) Stop-gradient: detaching the action gradient from the backbone

We thank the reviewer for proposing this ablation, and we ran it. We retrained TacitVLA under the same protocol as Table 5 with the action expert's gradient detached from the backbone for the whole run. The co-trained model still performs better than No Reasoning, whose backbone does receive the action gradient, which the reviewer notes would strengthen our conclusion:

PolicyReasoning supervisionAction-expert gradient into the backboneLIBERO-Plus (%)
No Reasoningnoyes67.6
TacitVLA, action gradient detachedyesno69.4
TacitVLAyesyes72.9

The detached TacitVLA's backbone receives reasoning supervision but no gradient from the action expert, while No Reasoning's receives that gradient but no reasoning supervision, and TacitVLA still scores higher (69.4 vs. 67.6). Detaching lowers TacitVLA from 72.9 to 69.4, the drop the reviewer expected, so the action gradient improves the backbone further. We have added this ablation to Section 6.1 of the revised paper, with the full results in Appendix B.

(W4 + Q4) Memory control: keeping the ledger without reasoning supervision

We agree that on memory tasks, No Reasoning drops both reasoning supervision and the ledger, the symbolic memory the reviewer describes, so it measures their joint value. We hope to clarify that the contribution the paper argues for is tested against Explicit Reasoning, which keeps a ledger in the same way. On Press Button, TacitVLA matches it (74.3 vs. 73.3, Table 4), and both write a correct ledger in 99.7% of episodes (Table 14), so that comparison is not a statement about having a ledger. To separate reasoning supervision from the ledger, we ran the control the reviewer suggested: TacitVLA's recipe with the ledger kept and only the supervision on the current reasoning text removed.

We clarify that on memory tasks, the reasoning text is the predicted next subtask, which becomes the next ledger entry, so without reasoning supervision, the policy never learns to write its own entries. At test time, the simulator writes the control's entries instead, adding the ground-truth subgoal at each replan in the wording of the training labels. Such a control relies on privileged information that a deployed policy would not have, which is why the policies the submission trained either write their own entries or keep no ledger. As a diagnostic, however, it tests whether a ledger alone can stand in for reasoning supervision.

The control therefore needs a simulator that reports subgoals, which is why we ran it on RoboMME Counting rather than RMBench Press Button. RoboMME's simulator reports each subgoal online from the same task checker that labeled its training data, while RMBench's simulator provides no such subgoals at test time. We trained the control under the same protocol as Table 2.

Without reasoning supervision, keeping the ledger does not recover the gain, even with ground-truth entries:

PolicyReasoning supervisionLedger entries written bySuccess (%)
No Reasoningnono ledger79.0 ± 1.0
Controlnothe simulator (ground truth)66.8 ± 12.2
TacitVLAyesthe policy itself89.5 ± 1.3

TacitVLA, which writes its own ledger, scores 22.7 points above the control (89.5 vs. 66.8). Symbolic memory alone therefore does not account for the gain over No Reasoning. We believe the control falls even below No Reasoning (66.8 vs. 79.0) because predicting the next subtask is what teaches the backbone to track progress from the ledger. We have added this control to Table 2, and the revised Section 6.1 now credits TacitVLA's 10.5-point gain over No Reasoning to reasoning supervision and the ledger together.

Overall

The reviewer's questions led to four analyses, now in the revised paper. TacitVLA improves on No Reasoning across three reasoning formats, four benchmarks, and training sets from 50 demonstrations per task to 3.8 million events (W1). The gain requires reasoning that is correct for the current task (W2). With the action gradient detached, the co-trained model still outperforms No Reasoning (W3). Keeping the ledger without reasoning supervision does not recover the gain (W4). We hope these results address the concerns raised, and would be grateful if the reviewer would consider updating their assessment. We are happy to answer further questions during the discussion period.

These additional results substantially strengthen the empirical evidence and address most of my original concerns. I still believe that the paper offers a relatively limited understanding of reasoning-action integration, particularly regarding whether the model genuinely utilizes the logical structure of reasoning to guide action generation, rather than benefiting primarily from task-relevant semantic supervision. Nevertheless, I consider this a limitation of the scope of the current work rather than a reason to dismiss its contribution.I would like to raise my score to 3

We sincerely thank the reviewer for the follow-up and for raising the score. We are very glad that the additional results substantially strengthen the empirical evidence, and we are grateful for the reviewer's suggestions that made the revised paper stronger.

Reviewer_rS5POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TacitVLA aims to retain the benefits of reasoning supervision in vision-language-action (VLA) models while reducing the latency incurred before action execution. During training, the model learns both to generate reasoning text and to predict actions, but the action expert does not attend to the reasoning text for the current call. Instead, it predicts actions directly from the context, including visual observations and instructions. At deployment, text generation is omitted on non-memory tasks. On memory tasks, the model retains a ledger of previously generated subtasks that records what it has attempted. While the robot executes the current action chunk, the model generates a new subtask description, such as “pick up the second red cube,” for subsequent policy calls. Across the four benchmarks evaluated in the paper, TacitVLA and Explicit Reasoning exhibit broadly comparable aggregate performance, with differences across tasks. On LIBERO-Plus, TacitVLA achieves a success rate of 72.9%, compared with 73.5% for Explicit Reasoning and 67.6% for No Reasoning. Its mean latency per call on a single H100 GPU is 147 ms, compared with 497 ms for Explicit Reasoning and 147 ms for No Reasoning. Under this evaluation setting, TacitVLA therefore achieves substantially lower action latency with a small difference in aggregate success rate relative to Explicit Reasoning.

Critical strengths
  1. The research question is clearly formulated, and the central comparison is well targeted. The paper distinguishes three roles of reasoning text: a training target, episodic memory, and an input conditioning the current action. By controlling whether the action expert can attend to the reasoning text for the current call, the study examines the dependence of action prediction on explicit textual reasoning. This design helps distinguish the effects of reasoning supervision from those of conditioning actions on generated reasoning, providing relatively direct empirical evidence about how reasoning benefits VLA policies.

  2. The design extends the decoupling of reasoning generation and action prediction to memory tasks. The policy retains a ledger of past subtasks as part of its context while deferring the generation of the current ledger entry to the action execution window. Consequently, the current action does not need to wait for text generation, while subsequent calls can still use the updated memory. This design preserves historical information while improving action responsiveness. The paper also explicitly characterizes the additional waiting time when text generation exceeds the execution window, demonstrating consideration of deployment constraints.

  3. The evaluation covers diverse settings and jointly examines task performance and inference latency. The paper compares No Reasoning, Explicit Reasoning, and TacitVLA across manipulation and autonomous driving, including both memory and non-memory tasks. Beyond standard benchmark evaluation, LIBERO-Plus assesses generalization under environmental perturbations. The loss-weight ablations, ledger correctness measurements, and latency breakdowns provide additional evidence for assessing the trade-off between performance and efficiency.

Critical weaknesses
  1. The contributions of reasoning supervision and textual memory are not fully disentangled on memory tasks. The No Reasoning baseline lacks both reasoning supervision and the subtask ledger. Consequently, this comparison measures their combined benefit rather than quantifying their individual contributions, limiting the attribution of the observed performance gains. An additional control using the same ledger-generation mechanism but without supervision on the current reasoning text would help clarify the source of these gains.

  2. Comparable aggregate performance does not establish that the benefits of explicit reasoning are preserved across perturbation conditions. On LIBERO-Plus, TacitVLA and Explicit Reasoning achieve overall success rates of 72.9% and 73.5%, respectively. However, under camera viewpoint perturbations, their success rates are 42.2% and 51.5%, a gap of 9.3 percentage points. This suggests that aggregate results can obscure substantial performance degradation under specific conditions. The authors should further analyze this discrepancy and qualify the conclusion that learning to generate reasoning is sufficient according to the tasks and metrics supported by the experiments.

  3. The additional value of the training-time attention mask has not been directly established. The current evaluation compares Explicit Reasoning, which generates reasoning text before predicting actions, with TacitVLA, whose action expert is trained not to attend to the current reasoning text. However, it does not report the performance of the Explicit Reasoning checkpoint when reasoning generation is omitted at deployment. Evaluating this checkpoint on LIBERO and LIBERO-Plus using only the context keys and values for action prediction, and comparing it with TacitVLA, would help determine whether training-time masking is necessary to retain performance or whether omitting text generation at deployment already provides comparable results.

Questions for authors

The paper addresses the action latency introduced by reasoning generation in VLA models. TacitVLA retains reasoning supervision while masking the action expert’s attention to the current reasoning text and, on memory tasks, generates ledger entries during action execution. Its main contribution lies in distinguishing the roles of reasoning text, examining the dependence of action prediction on current reasoning through controlled comparisons, and extending the design to textual memory. The experiments broadly support a favorable performance–latency trade-off. However, the contributions of reasoning supervision and memory remain insufficiently disentangled, substantial performance gaps occur under certain perturbations, and the additional benefit of training-time masking over simply omitting reasoning generation at deployment requires further validation.

2 comments

We thank the reviewer for the positive assessment and for recognizing the clearly formulated research question, the well-targeted central comparison, the extension to memory tasks, and the joint evaluation of performance and latency. We ran all three analyses the reviewer suggested, including the Explicit Reasoning checkpoint with generation omitted (W3, in our second comment), which falls below TacitVLA at the same latency (70.5 vs. 72.9 on LIBERO-Plus).

(W1) Disentangling reasoning supervision and textual memory

We agree that on memory tasks, the No Reasoning baseline lacks both reasoning supervision and the subtask ledger, so its gap to TacitVLA measures their combined benefit. We hope to clarify that the paper's central claim on memory tasks compares TacitVLA with Explicit Reasoning, which keeps a ledger in the same way, so the ledger cannot explain that match (89.5 vs. 86.7 on RoboMME Counting in Table 2, and 77.5 vs. 77.0 on RMBench in Table 4). To disentangle reasoning supervision from the ledger, we ran the control the reviewer suggested: TacitVLA's recipe with the ledger kept and only the supervision on the current reasoning text removed, trained on RoboMME Counting under the same protocol as Table 2.

We clarify that on these tasks, the reasoning text is the predicted next subtask, which becomes the next ledger entry, so without supervision on that text, the policy never learns to write its ledger. At test time, the simulator writes the control's ledger instead, adding the ground-truth subgoal at each replan in the wording of the training labels. This control depends on privileged subgoals that a deployed policy would not have, which is why the submission trained only deployable policies. As a diagnostic, however, it addresses the question the reviewer raises.

A ledger without reasoning supervision does not reproduce TacitVLA's gain, even with ground-truth subgoals:

PolicyReasoning supervisionLedger at testSuccess (%)
No Reasoningnonone79.0 ± 1.0
Controlnosimulator's ground-truth subgoals66.8 ± 12.2
TacitVLAyesits own entries89.5 ± 1.3

TacitVLA, which writes its own ledger, scores 22.7 points above the control (89.5 vs. 66.8). The ledger alone therefore does not account for the gain over No Reasoning. We believe the control falls even below No Reasoning (66.8 vs. 79.0) because predicting the next subtask is what teaches the backbone to track progress from the ledger. Without any ledger, reasoning supervision alone adds 5.3 points on LIBERO-Plus and 6.1 in closed-loop driving, where no policy keeps one (Tables 5 and 6). We have added this control to Table 2 and Section 6.1 of the revised paper, which now credits the 10.5-point gain over No Reasoning to reasoning supervision and the ledger together.

(W2) Camera viewpoints and qualifying the conclusion

We agree that aggregate results can obscure per-axis differences. We hope to clarify that our claim is that TacitVLA matches Explicit Reasoning at under a third of its latency (147 vs. 497 ms per call on LIBERO-Plus). We therefore analyzed whether that match holds on each perturbation axis, with differences computed from Table 12:

Difference in success (points)Robot init.SensorCameraLanguageObjectsBackgroundLightAll
TacitVLA − Explicit Reasoning−1.2+8.9−9.3−0.7−1.2+0.8−1.4−0.6
TacitVLA − No Reasoning+10.2+8.3+1.2+6.4+3.9+5.1+1.4+5.3

The match holds within 1.4 points on five of seven axes. The two exceptions point in opposite directions: +8.9 on sensor noise and −9.3 on camera viewpoints. TacitVLA also scores above No Reasoning, which runs at the same latency, on all seven axes.

On the camera discrepancy, this axis is known to swing with changes to a policy's design. The LIBERO-Plus authors find models "most vulnerable to changes in camera viewpoint", and in their evaluation ([1], Table 2), π0 and π0-FAST share a PaliGemma backbone yet differ by 51.3 points on this axis. Removing only OpenVLA-OFT's wrist camera lowers its camera score by 46.0 points.

We have qualified the conclusion accordingly. The revised abstract, introduction, and conclusion now state that learning to generate the reasoning is enough to match Explicit Reasoning's overall success on each of the four benchmarks. The revised Section 6.1 presents this per-axis analysis, including the camera and sensor axes.

[1] Fei et al. LIBERO-Plus. CVPR 2026. https://arxiv.org/abs/2510.13626

(W3) Explicit Reasoning with generation omitted at deployment

We thank the reviewer for proposing this direct test of the training-time mask. We evaluated the Explicit Reasoning checkpoint of Tables 5 and 6 with generation omitted, so its action expert reads only the context keys and values, on the same inference path and at the same latency as TacitVLA. We ran it on LIBERO and LIBERO-Plus, as suggested, and also on open-loop driving, under the same protocol as Tables 5 and 6.

Omitting generation at deployment falls short of TacitVLA in all three evaluations:

PolicyStandard LIBERO (%)LIBERO-Plus (%)Driving minADE (m) ↓
Explicit Reasoning, with its text98.173.51.140
Explicit Reasoning, generation omitted96.670.51.321
TacitVLA98.672.91.116

At the same latency, TacitVLA leads the generation-omitted checkpoint by 2.0 points on standard LIBERO and 2.4 on LIBERO-Plus, and its driving error is 0.205 m lower. With its direct-action prompt instead, the Explicit Reasoning checkpoint scores lower still (95.7 on standard LIBERO and 69.5 on LIBERO-Plus). An action expert trained to read the reasoning text depends on it at deployment, so omitting generation alone does not suffice, and the training-time mask is what retains the performance of Explicit Reasoning without the text. We have added this comparison to Section 6.1 of the revised paper, with the full results in Appendix B.

Overall

Each of the reviewer's three suggestions is now answered in the revised paper. On memory tasks, the paper's central comparison keeps a ledger in the same way on both sides, and a ledger without reasoning supervision does not reproduce TacitVLA's gain (W1). TacitVLA matches Explicit Reasoning within 1.4 points on five of seven perturbation axes, and the claim is qualified to overall success (W2). Omitting generation from Explicit Reasoning falls below TacitVLA at the same latency, so the training-time mask matters (W3). We hope these results address the concerns raised, and would be grateful if the reviewer would consider updating their assessment. We are happy to answer further questions during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.