acceptodds
Under review as a conference paper at ICLR 2027

Reflective VLA: In-Context Action Consequences Make VLAs Generalize

Abstract

Most vision-language-action (VLA) models are reactive: they predict actions from the current instruction and observation. However, deployment-specific factors such as camera geometry, robot calibration, and actuation bias can be difficult to identify from a single observation, limiting the policy's ability to adapt its behavior based on observed action consequences. We propose Reflective VLA, which conditions each decision on observation–action–consequence triplets collected during the policy's own task execution. Each triplet records what the robot observed and executed, and how the scene changed afterward, providing action-aligned feedback for subsequent decisions without test-time parameter updates. Architecturally, Reflective VLA routes all observation modalities through the VLM under shared attention, allowing the action expert to attend directly to past triplets and the current observation. A block-causal mask enables parallel multi-frame training without leakage and supports KV-cached inference. Reflective VLA maintains strong performance on LIBERO and improves success on SimplerEnv-Bridge. Under deployment perturbations on LIBERO-Plus and LIBERO-Plus-Hard, it improves average success by 5.4 and 4.2 percentage points over a matched reactive baseline, respectively. Hard evaluates deployment configurations outside the task-specific training distribution within camera and calibration perturbation families represented during training. Controlled ablations support the joint use of historical actions and observed consequences beyond visual history alone, and real-robot studies provide complementary evidence under camera shifts.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision Language Action models, vIn-Context Learning, Robot Generalization

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_rN2COpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

To improve the generalization of Vision-Language-Action models to unseen tasks and environments, the paper proposes Reflective VLA, which explicitly incorporates historical observations, actions, and their consequences into the context. Specifically, the model uses previous observation–action–consequence tuples as in-context information, allowing it to reason not only about the current state but also about how previous actions affected the environment. Experiments show that this simple reflective context substantially improves VLA generalization across unseen tasks and environments.

Critical strengths

The main strength of the work lies in its use of observation–action–consequence tuples as in-context information for VLA models. Through controlled ablations and evaluations under distribution shifts, the authors provide empirical evidence that incorporating action consequences, beyond simply adding observation history, can improve generalization to unseen environments and configurations.

Critical weaknesses
  • W1. Additional computational cost. The method requires processing multiple historical observation–action–consequence tuples, increasing the context length and consequently the computational and inference costs.
  • W2. Limited analysis of the core mechanism. The proposed idea is relatively simple, and the paper could provide stronger evidence that the gains are specifically attributable to action consequences, rather than merely to additional temporal context.
Questions for authors
  • Q1. Can the authors provide additional analysis or qualitative case studies to clarify why action consequences improve generalization, and whether the gains are specifically attributable to action consequences rather than simply to additional temporal context?
  • Q2. Could the authors report the training time, GPU memory usage, and inference latency for different history lengths K, particularly relative to the shortest-context setting? This would clarify the computational trade-off introduced by longer reflective contexts.
  • Q3. How sensitive is Reflective VLA to the quality of the observed action consequences? It would be useful to evaluate its robustness under noisy, ambiguous, or inaccurate consequence observations.
3 comments
AuthorsOpenReview ↗

Response to Reviewer rN2C

We thank the reviewer for the positive assessment and the precise questions.

Q1. Computational cost for different K.

Thank you for the suggestion. The cost for each K is:

KInference latency (ms)Inference memory (GB)Train batch / GPUTotal training time (h)Train memory / GPU (GB)
11245.019322445
21305.073163664
41485.18684274
81785.39944679

Inference on one RTX 4090; training on H20 GPUs.

Inference. The overhead is small. Past triplets are KV-cached, so an 8× longer context costs 1.43× the latency and 0.38 GB more memory, and K=4 recovers most of the accuracy gain (76.7% vs. 77.8%) at 1.19×. Inference runs once per action chunk, not once per control step, so the extra 54 ms at K=8 is small relative to the chunk duration (one second on the real robot).

Training. All K frames of a sample are processed jointly, so we reduce the batch size per GPU as K grows and keep the number of supervised action targets per step constant (batch × K = 32). Even so, from K=1 to K=8 the total training time rises from 24 to 46 hours (1.9×) and the memory per GPU from 45 to 79 GB (1.8×). Memory is what limits us to K=8.

Q2. Are the gains due to action consequences or to more temporal context?

Attribution. Table 3a in the main paper compares contexts with the same length K that differ only in content:

History contentCameraCamera†Rob. Calib†Avg
None (reactive)90.074.055.273.1
Observations only: (O), (O,O′)88.8, 88.573.2, 73.355.0, 55.172.3, 72.3
Actions without consequences: (O,A)89.174.957.473.8
Actions paired with consequences: (O,A,O′)95.776.361.377.8

Additional temporal context alone gives no gain: observation-only history stays at the reactive level on every split. The gain appears only when each executed action is paired with its observed consequence.

Why consequences help. The evaluated shifts change how commands map to observed motion, which a single frame does not reveal. Under robot calibration shift, for example, every command is offset by a hidden bias. A completed triplet exposes this bias as the difference between the commanded pose in A and the reached pose in O′, and the policy can then correct its subsequent commands. Observation-only history shows that the scene changed but not which command caused the change, so it cannot provide this evidence. This is consistent with (O,O′) staying at 55.1 on Rob. Calib† while the paired triplet reaches 61.3.

Q3. Sensitivity to the quality of consequence observations.

Noisy consequences. We add random LIBERO-Plus sensor noise (blur and fog) on top of a deployment shift (camera). The noise degrades all observations of the episode, including the consequence frames in the context:

MethodDeployment shift (camera)+ sensor noise
Reactive baseline90.088.6
Reflective VLA95.794.3

Both policies lose the same 1.4 points, so Reflective VLA keeps its 5.7-point advantage.

Ambiguous or inaccurate consequences. These can mislead the policy, as our Conclusion notes. A simple safeguard is to discard triplets whose consequences are known to be unreliable; with an empty context, the same model runs at the reactive level (73.1%, Table 3b in the main paper). Detecting such consequences automatically is future work.

We thank the reviewer again for the constructive feedback.

Reviewer_rN2COpenReview ↗

Reply to Authors

Thank you for the clear and comprehensive response, which has addressed my concerns. I will keep my current rating.

Thank you for your follow-up and positive assessment. We are glad that our response addressed your concerns, and we appreciate your thoughtful review and constructive feedback.

Reviewer_5NECOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper conditions a VLA on observation–action–consequence triplets accumulated during task execution. It improves over a matched reactive baseline.

Critical strengths

The method is interesting, and the improvement over limited benchmark is promising.

Critical weaknesses
  1. The compared benchmark and method is limited. More results on different benchmark ,such as robotwin or robodojo, and SOTA VLA/WAM method are necessary.
  2. The paper should report the overrall latency and computation cost.
Questions for authors

1.Could you report more results on different benchmarks with different method backbone? 2. Is this method general enough to be applied to WAMs, rather than being limited to VLAs? 3. Under the same backbone, data, and interaction budget, what advantage does rolling task-execution history offer over the cited ICWM pre-task probing context?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5NEC

We thank the reviewer for the review and suggestions.

Q1. More benchmarks, baselines, and backbones.

Thank you for the suggestion. We added RoboTwin 2.0 as a new benchmark and tested the method on two state-of-the-art VLA backbones. For each backbone, we add the reflective context (K=4) to the published reactive model and train it under the same protocol (50 tasks; 50 clean and 500 randomized demonstrations per task). Average success (%):

BackbonePolicyCleanRandomized
π0.5Reactive82.7476.76
π0.5+ Reflective context83.3280.54
LingBot-VLAReactive86.5085.34
LingBot-VLA+ Reflective context87.1286.83

Reactive numbers are those reported in the LingBot-VLA paper. Reflective context keeps clean performance (+0.6 on both backbones) and improves the randomized setting, by 3.8 points on π0.5 and 1.5 points on LingBot-VLA.

Q2. Overall latency and computation cost.

The submission reports latency in Fig. 4 in the main paper and Table 7 in Appendix C.4. We agree it should be more visible and will move it into the main text.

PolicyModel latency (ms)Relative
Reactive (K=1)1241.00×
Reflective VLA (K=4)1481.19×
Reflective VLA (K=8)1781.43×
MemoryVLA1981.60×

Inference. Measured on one RTX 4090 (bf16, batch 1, 10 flow-matching steps); preprocessing adds about 20 ms. Past triplets are KV-cached, so an 8× longer context costs only 1.43× the reactive latency, and K=4 keeps most of the gain (76.7% vs. 77.8%) at 1.19×. Each inference yields a whole action chunk, so this cost is paid once per chunk (one second of motion on the real robot).

Training. Reflective VLA and the reactive baseline use the same hardware and the same number of steps. Each step supervises the same number of frames per GPU: 4 sequences × K=8 for Reflective VLA versus 32 single frames for the baseline (Table 5 in the appendix).

Q3. Applicability to WAMs.

Yes. The triplet interface is a prefix-conditioning scheme, so any model that predicts actions (or future frames and actions) from a token prefix can consume (O,A,O′) blocks under the same mask.

To verify this, we apply it to Cosmos-Policy, the strongest WAM on LIBERO-Plus in a recent robustness study of VLAs and WAMs (arXiv:2603.22078). Both versions are trained on the same data as in Table 2 in the main paper:

MethodCamera†Rob. Calib†
Cosmos-Policy, reactive61.338.2
Cosmos-Policy + reflective context (K=4)70.445.3

Reflective context improves both splits (+9.1 and +7.1 points), which shows that the benefit is not specific to VLA backbones.

Q4. Advantage over ICWM pre-task probing.

ICWM is concurrent work that was developed independently. Both methods condition the policy on interaction triplets; ICWM collects them in a probing phase before execution, while we take them from the task execution itself. This gives two advantages:

  • No separate probing phase. ICWM needs 20 random exploratory actions in a predefined safe workspace, repeated whenever the configuration changes. Rolling history needs no extra interaction or setup.
  • Consistent context. ICWM trains on segments of demonstration trajectories but deploys on random probing motions. Our context is the preceding chunks of the same episode in both cases.

The trade-off is that our first chunk is reactive, but one completed triplet already raises success from 73.1% to 75.3% (Table 3b in the main paper). The two designs are also compatible: probing could initialize the buffer that rolling history then updates.

We hope these clarifications address the concerns, and we would be grateful if the reviewer would reconsider the assessment.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer 5NEC, we would like to kindly follow up on our response above. We have added RoboTwin 2.0 results with two VLA backbones and a WAM experiment with Cosmos-Policy, and clarified computational costs and the comparison with ICWM. We would appreciate your feedback on whether these address your concerns and would be happy to discuss any remaining questions. Thank you for your time!

Thanks for the esponse. Given the additional computational overhead introduced by the reflective mechanism and the relatively minor improvements on LingBot-VLA, I will maintain my original score.

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

This paper proposes Reflective VLA, a vision-language-action framework that leverages interaction history to adapt to deployment-specific conditions. The key idea is to condition action prediction on observation–action–consequence triplets collected during the robot's own execution, allowing the policy to adjust subsequent actions based on historical actions and their observed effects without test-time parameter updates. Architecturally, the method combines a vision-language model with a flow-matching action expert through shared attention, and employs block-causal attention for joint multi-frame training and KV caching for efficient inference. Experiments on LIBERO, SimplerEnv-Bridge, LIBERO-Plus, LIBERO-Plus-Hard, and real-world robotic tasks demonstrate improved task success rates under camera viewpoint changes and robot calibration shifts.

Critical strengths
  1. Meaningful research problem. The paper addresses an important challenge in deploying VLA models: generalization to changes in camera configurations, robot calibration, and other deployment-specific conditions. This problem has clear practical relevance for real-world robotic applications.

  2. Intuitive design with practical implementation. The method incorporates historical actions and their observed consequences as context, enabling VLAs to leverage interaction feedback without test-time parameter updates. The overall design is relatively simple and can be integrated into existing VLA architectures. Moreover, block-causal attention enables joint multi-frame training, while KV caching reduces redundant computation over historical context, helping to mitigate the overhead of longer input sequences.

Critical weaknesses
  1. Limited conceptual novelty. The main contributions lie in continuously incorporating interaction history during task execution and the corresponding training and inference implementations. While these designs are practically useful, the underlying concept of leveraging interaction history for deployment adaptation offers limited novelty.

  2. Redundancy of consequence observations during inference. According to Appendix C.1, the consequence observation of each triplet coincides with the starting observation of the subsequent triplet during evaluation. Consequently, the full (O, A, O') and (O, A) variants contain the same unique observation information at test time. The improvement in Table 3a may therefore be partially attributed to repeated observations or differences in token organization, rather than additional interaction evidence.

  3. Insufficient evidence for the underlying adaptation mechanism. The paper primarily demonstrates the effectiveness of the proposed method through improvements in task success rates, without directly verifying whether the model learns to exploit action–consequence correspondences to infer deployment-specific conditions. The observed gains could also arise from richer temporal state information or task execution history, rather than genuine adaptation to the action–response dynamics.

  4. Limited scope of OOD generalization. LIBERO-Plus-Hard primarily evaluates larger perturbation magnitudes within camera and calibration perturbation families already represented during training, rather than entirely unseen types of perturbations. Therefore, the current experiments provide limited evidence for generalization to more diverse or fundamentally different deployment conditions.

  5. Lack of analysis of the online adaptation process. The paper does not investigate how the policy progressively adapts as more interaction history becomes available, nor does it evaluate changes in deployment conditions during task execution. The current experiments mainly consider perturbations that remain fixed throughout each episode, leaving the model's ability to handle dynamic deployment shifts unclear.

  6. Limited coverage of real-world experiments. The real-world evaluation includes only two relatively simple pick-and-place tasks, with 50 rollouts for each held-out camera-shift condition. Moreover, the lack of per-task results and uncertainty estimates makes it difficult to assess the generalizability of the observed improvements to more complex manipulation scenarios.

Questions for authors
  1. Necessity of explicit consequence observations. Appendix C.1 indicates that, during inference, each consequence observation already appears as the starting observation of the subsequent triplet. Could the authors compare representations with and without explicitly repeated consequence observations while preserving the same unique observation information? This would help disentangle the benefits of additional information from those of structured token organization.

  2. Role of action–consequence correspondence. Could the authors introduce controlled ablations involving shuffled actions, mismatched consequences, or cross-deployment histories? Such experiments would help verify whether the model genuinely relies on the correct correspondence between executed actions and their observed consequences, rather than simply benefiting from richer temporal context.

  3. Evidence for the adaptation mechanism. Could the authors analyze how action prediction errors or task execution performance evolve as the policy accumulates interactions under different deployment perturbations? Such an analysis would help distinguish genuine deployment adaptation from general temporal context modeling.

4 comments
AuthorsOpenReview ↗

Response to Reviewer LkPR (1/2)

We thank the reviewer for the detailed and constructive review.

Q1. Limited conceptual novelty.

We thank the reviewer for recognizing the practical value of the design. We would like to clarify the conceptual contribution, which the submission did not state clearly enough.

In-context adaptation is established in language models but little explored for VLAs. Most existing VLAs that use history treat it as memory for state and task tracking, not as evidence about the current deployment. The difference matters: MemoryVLA does not improve over the reactive baseline on Rob. Calib† (49.2 vs. 55.2), while Reflective VLA does (61.3). The key is pairing executed actions with their observed consequences. Visual history alone gives no gain ((O), (O,O′): 72.3 vs. reactive 73.1); only the paired triplet does (77.8; Table 3a in the main paper).

Block-causal training and KV-cached inference make this practical in a VLA (1.43× reactive latency at K=8) and the same interface also applies to world-action models. We will revise the contribution statement accordingly.

Q2. Consequence observations are redundant at inference.

Thank you for catching this. Appendix C.1 misstated the evaluation gap as zero; we apologize and will correct it. During training, we insert a random gap between neighboring triplets to avoid causal confusion. The gap also raises the information density of the context: the same K triplets cover a longer span of the episode with less redundancy between neighbors. Inference keeps a fixed gap (8 environment steps on LIBERO, 3 on SimplerEnv-Bridge). O′ = O_{i+C} and the next start O_{i+C+8} are therefore different frames, and (O,A) contains no frame showing each action's chunk-aligned outcome.

The gain is also not from extra frames alone: adding the same consequence frames without actions gives +0.0 ((O) 72.3 → (O,O′) 72.3), versus +4.0 with actions ((O,A) 73.8 → (O,A,O′) 77.8).

Q3. Insufficient evidence for the adaptation mechanism.

We agree that success rates alone are indirect, and we add an offline measurement of action prediction error.

Error and success as interactions accumulate. On held-out validation trajectories collected under the deployment perturbations, we give one fixed model k completed triplets from the same episode and measure the mean absolute error (MAE) of the predicted end-effector xyz position against the recorded action chunk, in meters. The table reports it next to the closed-loop success rate at the same k (average in Table 3b in the main paper):

Completed triplets0137
Action prediction error (xyz MAE, m)0.00450.00370.00350.0035
Success rate (%)73.175.376.777.8

As the number of completed triplets grows from 0 to 7, the action MAE falls from 0.0045 m to 0.0035 m (4.5 mm to 3.5 mm) and the average success rate rises from 73.1% to 77.8%. The gain holds under each perturbation (Table 3b in the main paper): Camera 90.0 → 95.7, Camera† 74.0 → 76.3, Rob. Calib† 55.2 → 61.3.

Richer temporal context. Table 3a in the main paper tests this alternative: visual history alone does not help ((O), (O,O′): 72.3% vs. reactive 73.1%), and only the paired triplet does (77.8%).

Q4. Controlled ablations with shuffled actions, mismatched consequences, or cross-deployment histories.

Thank you for suggesting a test of whether the policy relies on the correct action–consequence correspondence. We added the mismatched-consequence control. With the same model, K, and token layout, we keep O and A unchanged and replace each O′ in the context with a frame randomly selected from other episodes, so that it no longer corresponds to its action:

History given to the policyCameraCamera†Rob. Calib†Avg
No history (K=1)90.074.055.273.1
Correct consequences95.776.361.377.8
Mismatched consequences88.453.243.161.6

Average success drops from 77.8% to 61.6%, below the same model without history (73.1%), so the policy relies on the consequence that corresponds to each action. This mismatch is a deliberate intervention: in normal operation, each O′ is recorded right after its action is executed and always matches it.

Table 3a in the main paper provides a training-time counterpart: the (O,O′) variant removes the action from each triplet while keeping the same observations, K, and token slots, and the gain disappears (77.8% → 72.3%).

AuthorsOpenReview ↗

Response to Reviewer LkPR (2/2)

Q5. OOD is limited to larger magnitudes within training families.

We agree that our OOD evaluation covers larger magnitudes within the perturbation families seen in training. Our method needs training data with diverse deployment configurations, which is not hard to collect, so that the policy learns to infer the deployment shift from action consequences. It can then handle configurations outside the training range, such as larger camera shifts and calibration biases. Diverse data alone is not enough, because these factors are hard to identify from a single observation: the reactive baseline is trained on the same data and reaches 64.6% on LIBERO-Plus-Hard versus 68.8% for Reflective VLA, and 24% versus 70% under the Large real-world camera shift. We will state this scope more prominently in the introduction.

Q6. No analysis of online adaptation or dynamic shifts.

Progressive adaptation. We agree that it is important to analyze how the policy adapts as interaction history accumulates. In Q3, we provide the action prediction error and the success rate of one fixed model given 0, 1, 3, and 7 completed triplets.

Dynamic shifts. Our evaluation keeps each perturbation fixed within an episode because this is the more common form of deployment shift: camera placement and robot calibration are properties of a deployment and rarely change within a single episode. If the configuration does change mid-episode, the context adapts once the triplets collected after the change have replaced the older ones, i.e., within K chunks. We have not evaluated this case and will state the scope explicitly.

Q7. Limited real-world coverage.

Thank you for pointing this out. We add per-task results and uncertainty estimates.

Per-task results. Success rate (%) for Box / Bowl. Each cell has 25 rollouts per task (50 for Nominal):

MethodNominalSmall shiftMedium shiftLarge shift
Reactive88 / 6872 / 5248 / 2832 / 16
MemoryVLA88 / 6868 / 5248 / 3236 / 20
OC-VLA86 / 6676 / 6068 / 4856 / 40
Reflective VLA92 / 7288 / 7284 / 6880 / 60

Uncertainty. Success rate (%) pooled over both tasks, with 95% Wilson intervals:

MethodNominal (n=100)Small (n=50)Medium (n=50)Large (n=50)
Reactive78 [68.9, 85.0]62 [48.2, 74.1]38 [25.9, 51.8]24 [14.3, 37.4]
MemoryVLA78 [68.9, 85.0]60 [46.2, 72.4]40 [27.6, 53.8]28 [17.5, 41.7]
OC-VLA76 [66.8, 83.3]68 [54.2, 79.2]58 [44.2, 70.6]48 [34.8, 61.5]
Reflective VLA82 [73.3, 88.3]80 [67.0, 88.8]76 [62.6, 85.7]70 [56.2, 80.9]

Under the Large shift, Reflective VLA gains +46.0 [26.8, 60.6] points over Reactive, +42.0 [22.6, 57.2] over MemoryVLA, and +22.0 [2.7, 39.1] over OC-VLA, which uses calibrated camera extrinsics (95% Newcombe intervals). All three intervals exclude zero.

We hope these clarifications address the concerns, and we would be grateful if the reviewer would reconsider the assessment.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer LkPR, we would like to kindly follow up on our responses above. We have clarified the inference gap and added controlled ablations, adaptation analysis, and per-task real-world results with uncertainty estimates. We would appreciate your feedback on whether these address your concerns, and would be happy to clarify any remaining questions. Thank you for your time!

The additional experiments provide stronger evidence for the usefulness of interaction history, but do not convincingly establish deployment-specific adaptation rather than general temporal modeling. The mismatched-consequence experiment also introduces substantial distribution shifts, making its conclusions less conclusive. Moreover, the conceptual novelty relative to ICWM remains limited, and the revised inference-gap description contradicts the original evaluation protocol.

Overall, these concerns remain insufficiently addressed, and I maintain my recommendation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.