acceptodds
Under review as a conference paper at ICLR 2027

Heard but Not Heeded: Language Survives in a VLA's Representations but Not Its Actions

Abstract

A fine-tuned Vision-Language-Action (VLA) model that scores on the LIBERO manipulation benchmark stops following language instructions when given a scene outside its training distribution. We diagnose this failure through three measurements: (1) the model still internally represents which object was named ( accuracy vs baseline), (2) attention roughly doubles when the object is mentioned, yet (3) the policy grasps the wrong object. This dissociation understanding language and attending correctly, but choosing the wrong action shows that instruction-following failures need not stem from language misunderstanding or visual misidentification. We validate this dissociation with a within-distribution control experiment. Attention allocation to the target object is virtually identical in both settings ( vs of total attention), yet language-driven attention shifts are five times stronger when behavior fails. This suggests that insufficient attention magnitude cannot account for the behavioral failure. The evidence points to a breakdown at the interface where the model translates visual and language information into motor commands. We introduce object-level behavior metrics (recording which object was actually grasped, rather than only whether the task succeeded) and show that benchmark structural properties make instruction-following failures invisible to standard evaluation metrics.

Primary area:
interpretability and explainable AI
Keywords:
Vision-Language-Action Models, Out-of-Distribution Generalization, Mechanistic Interpretability, Instruction Following

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_yXmAOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a diagnostic study of instruction-following failures in a VLA model under distribution shift. Through in-distribution and out-of-distribution behavioral tests, linear probing, and attention analysis, the authors show that information about the instructed target remains recoverable and attention responds to the instruction, even when the model selects the wrong object. These findings suggest a gap between internal representations and action selection in the tested settings.

Critical strengths

The paper identifies an interesting gap between language-conditioned attention and action selection: in the tested OOD scenes, the model tends to select the same object despite changes in the instruction, even though attention shifts toward the named target. This finding motivates examining how language influences actions beyond attention alone and shows why task success may be insufficient to establish instruction following.

Critical weaknesses
  1. The behavioral evaluation does not sufficiently isolate the effect of distribution shift. Sections 4.1.1 and 4.1.2 compare settings that differ in scene, object configuration, and instruction content. A matched comparison using the same tasks and instructions while varying the visual environment would provide stronger evidence. The paired control in Section 4.1.3 is useful but covers only the pan–moka-pot pair, limiting the generalizability of its findings.
  2. The writing is overly conversational, particularly in the descriptions of the experiments, which makes the setup and findings harder to follow.
Questions for authors
  1. Have you evaluated additional OOD scenes, object pairs, or tasks to assess whether the observed behavior generalizes beyond the reported cases?
  2. Could you provide spatial attention maps for matched instructions on the same image to show where the reported attention changes occur?
Reviewer_81dYOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates why a fine-tuned vision-language-action model fails to follow language instructions in out-of-distribution manipulation scenarios despite strong benchmark performance. The authors examine this behavior through object-level grasp evaluation, linear probing of internal representations, and visual attention analysis, alongside in-distribution control experiments. Their findings show that information about the instructed object remains decodable and attention responds to language even when the model selects the wrong object, while an analysis of LIBERO suggests that standard task-success metrics can mask such instruction-following failures.

Critical strengths

The paper investigates an important failure mode in vision-language-action models through a combination of behavioral evaluation, representation probing, and attention analysis, showing that language information can remain accessible even when the model fails to act on it. Furthermore, the object-level evaluation and structural analysis of LIBERO reveal how standard task-success metrics can mask instruction-following failures, providing useful insights for designing more reliable VLA benchmarks.

Critical weaknesses
  1. Insufficient evidence for the central representation–action dissociation. The behavioral failures are demonstrated in Kitchen Scenes 2 and 3, whereas the representation probing uses a separate four-object permutation scene. It is therefore unclear whether the model retains the correct referent information in the same instances where action selection fails. Furthermore, probing and attention measurements are correlational, and without causal interventions or matched-instance analysis, the claimed localization of the failure to the vision-language-to-action interface remains insufficiently supported.

  2. Limited experimental validation and generalizability. The experiments focus on a single VLA model and a small number of out-of-distribution scenarios, with limited trials per condition. Some scenarios also introduce unfamiliar objects, making it difficult to isolate instruction-following failures from other generalization limitations. Additionally, the claim that approximately 75% of LIBERO's fine-tuning tasks do not require language is based primarily on task structure, without direct validation through language-removal or instruction-shuffling experiments. These limitations weaken the broader conclusions about VLA instruction following and benchmark reliability.

Questions for authors
  1. Were representation probing and attention measurements performed on the same instances where the model selected the wrong object? If not, how is the central dissociation established?

  2. What evidence rules out unfamiliar-object manipulation difficulties as an alternative explanation for the observed failures?

  3. Has the claim that approximately 75% of LIBERO tasks do not require language been validated through instruction removal or shuffling experiments?

Reviewer_LF9EOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies failures of language following under distribution shift in a π05 policy fine-tuned on LIBERO. By overriding the instruction sent to the policy and recording the first object grasped, the authors find that the policy repeatedly selects a default object in some conditions despite instructions naming another target. In-distribution overrides select the requested object in 19/20 trials. To explain this behavior, the authors probe target-location information in prefix K/V representations and measure instruction-dependent attention changes using object segmentation masks. They report decodable target locations and increased attention, interpreting these results as a representation-attention-behavior dissociation. The paper also examines visual shortcuts in LIBERO's task structure and a negative result in which the controller compensates for simulated dynamics perturbations.

Critical strengths
  1. The author raised an interesting question, although this issue has already been a consensus within the field, but exploring its intrinsic mechanism is still a very important and interesting scientific question.
  2. Recording the object actually grasped is more diagnostic than task success alone because it helps distinguish target-selection errors from subsequent manipulation failures. In §4.1.1, keeping the image unchanged and transmitting another training instruction produces correct target selection in 19/20 trials, demonstrating language-dependent selection under those in-distribution conditions. Together with the grasp identities in Tables 1-2, this allows a concrete comparison of instruction sensitivity and task competence. (§4.1, Tables 1-2.)
  3. Section 4.3.3 identifies the anti-correlated object positions in the initial two-object design and introduces four-object permutations. Section 4.3 and Appendix B also include swapped-label, never-named, shuffled-label, and frame-held-out controls. Comparing attention to the same object within the same frame when it is named versus unnamed also controls pixels and mask area. These designs support the narrower but useful conclusion that the measured representations and attention are sensitive to the instruction. (§4.3-4.4, Appendix B.)
Critical weaknesses
  1. Although the author raised an important question, mere mechanism analysis does not constitute a strong scientific contribution to this issue, because it merely elaborates on the current consensus within the field at the level of phenomena. What is more anticipated is the corresponding solution. The author claims that the diagnosis can guide component-level repairs, but this judgment has not been verified through targeted retraining or intervention experiments, therefore, the guiding value of the diagnosis for repairs still lacks empirical support.
  2. Behavioral failures mainly involve KITCHEN_SCENE2/3, but §4.3.3, lines 304-309, explicitly switches the probe scene, and Table 3 measures attention to bowl/pudding. Table 5 combines these internal measurements with OOD 0/10 behavior without sample correspondence, leaving information retention during the same failures unestablished. There is also a specific inconsistency: Appendix B.4 (lines 688-690) retains two upstream explanations: information in a subspace inaccessible to the action expert, and attended token values lacking task-relevant information, while the conclusion (lines 468-470) states that controls "exclude weak visual grounding".
  3. LIBERO-Plus §4.2 and Appendix E already observe persistent manipulation of the original object after changing instructions and goals within a scene, overlapping with the default-object bias here. LangGap §III.A-B vary four semantic dimensions under fixed layouts and evaluate π05, sharing the instruction-change diagnostic logic.
  4. The author checked whether the model retained the information of "which object the instruction required" and also measured the model's attention to the object. However, Table 5 does not clearly state whether these internal measurements came from the same set of experiments where the model actually grabbed the wrong object. Without this correspondence, it is difficult to determine what exactly happened internally when the model made a mistake in grabbing.
  5. Section 4.6 uses the failure of the familiar term "moka pot" on task 18 to rule out vocabulary as an explanation. In Table 2, however, the authors report that the policy grasps the moka pot in all six trials, correctly selecting the object specified by the overridden instruction. The zero task success rate follows from retaining the original pan-based scoring goal, as explained in Appendix A. This result therefore provides no evidence that the policy fails to follow the familiar noun "moka pot". I suggest removing this example from the vocabulary-exclusion argument.
Questions for authors
  1. Please identify the scenes, instructions, frames, checkpoint, and rollouts underlying each entry in Table 5. Can target-location decoding, attention, and actual grasping be jointly reported on the same failing tasks and compared with correct behavior in the same scene?
  2. What evidence distinguishes the action expert failing to use an adequate representation from upstream information that is readable by an external probe but lacks task-relevant geometry or affordances? Can targeted activation patching/ablation or behavioral readout controls support it? Otherwise, will the conclusion be restricted to a correlational dissociation?
  3. Does "OOD" refer to the scene, layout, object, or task composition? How is overlap between KITCHEN_SCENE3 and the training suites handled? Is there a positive control establishing that the policy can grasp the pan under matched visual conditions? Please also explain why the task-18 moka-pot override is treated as a language failure when Table 2 shows the requested object being grasped.
  4. Can the authors provide the task-level BDDL audit, actual training-task list, and sampling weights supporting the 75% estimate, and explain the scene counts in Table 6? Is this estimate supported by language-free or mismatched-language evaluation of the policy, or does it describe potential shortcuts only?
  5. How were the 90 probe/attention frames generated, and do they come from correlated trajectories? Are cross-validation splits grouped by independent configuration or trajectory, and how were probe layers and regularization selected? Can attention sign tests account for clustering by frame/configuration or trajectory, with intervals for the behavioral conditions? These details affect the interpretation of significance and generalization.
  6. Can at least one of the repair methods proposed in §4.8 be actually implemented, and is the accuracy of object selection and the success rate of tasks improved in unseen distribution scenarios? Please compare the results with those obtained by fine-tuning with an equivalent amount of conventional data to test whether the diagnosis presented in this paper can provide specific guidance for effective repair.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.