acceptodds
Under review as a conference paper at ICLR 2027

MIND: Mechanistic Interpretability for Neural Diagnosis of Visuomotor Policies

Abstract

Visuomotor policies can handle many robot manipulation tasks, but understanding their failures remains difficult. Unfamiliar conditions and small action errors can lead to failure, yet watching the robot alone tells us little about what went wrong inside the policy. We introduce MIND (Mechanistic Interpretability for Neural Diagnosis), a method that looks inside a trained policy to detect failures during execution and identify the internal concepts involved. MIND uses a sparse autoencoder to break down the policy’s internal activations into features that can be studied individually. We give meaning to these features by linking their activity to the robot’s position, movement, and surroundings, and by steering their activations to examine how they affect the robot’s actions. From successful and failed executions, we learn how the features behave on the way to a failure. During execution, MIND flags deviations from these patterns and uses the meaningful features contributing most to the change to describe the problem. Experiments in simulation and on a physical robot show that changes in these internal features help detect and explain failures. By using the same features for detection and diagnosis, MIND connects failure signals to understandable changes inside the policy.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Sparse Auto Encoder, Imitation Learning, Failure Detection, Interpretability

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_n844OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies failure detection and diagnosis for frozen vision-language-action policies. It trains SAEs on token-level internal activations, assigns features names from predefined state and decision concepts, and tests these names on held-out episodes. Temporal statistics of the named features feed a linear failure detector, whose concept-level contributions and activation histories are then verbalized by a language model.

The main contribution is using the same interpretable features for detection and diagnosis. Experiments cover several VLA models in simulation and real-world.

Critical strengths
  1. I appreciate that the explanation is tied to the features used by the detector, rather than generated independently of the alarm. The linear score decomposition makes the contribution of each named concept inspectable. This provides a good basis for understanding why the monitor flagged an execution, without modifying the deployed policy.

  2. The naming procedure uses quantitative criteria and evaluates whether the selected names remain valid on held-out episodes. This is better evidence than only inspecting a feature’s highest-activation examples. Closed-loop interventions further examine the behavioral relevance of selected decision features, complementing the observational concept-alignment results

Critical weaknesses
  1. Limited Novelty. Paper [1] already use SAEs to interpret VLA features and validate them through closed-loop steering. Paper [2] also employ per-token SAEs for concept identification, while paper [3] detect failures from VLA activations. MIND introduces a more systematic naming procedure and integrates these components for failure diagnosis. However, I find the methodological advance limited, and the added value of this integration is not sufficiently established.

  2. Insufficient Justification for SAE. Since MIND relies on predefined concepts and their labels, a natural alternative is to train linear concept probes on the same frozen activations and use their predictions in the same temporal monitoring and diagnosis pipeline. I notice the paper reports that linear probes recover all tested concepts on held-out robot attempts, but does not provide this end-to-end comparison. Without it, the necessity of SAE-based decomposition remains unclear.

  3. Limited Predefined Concept Vocabulary. Only features aligned with predefined concepts enter the monitor, potentially excluding informative but unnamed features. Held-out naming accuracy does not guarantee that the vocabulary captures the distinctions needed for failure diagnosis. This concern is particularly relevant given the limited failure-category accuracy: 15/48 for OpenVLA and approximately one-third for π0.5. I would appreciate further analysis of how vocabulary coverage affects diagnostic performance.

[1] Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models, 2026

[2] Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models, 2026

[3] SAFE: Multitask Failure Detection for Vision-Language-Action Models, 2025

I would consider raising my score if the response establishes a substantive benefit over direct concept probing and provides stronger evidence for the reliability and scope of the diagnoses.

Questions for authors
  1. How does MIND compare with linear concept probes followed by the same temporal monitor and diagnosis module, using the same activations, concept labels, and data splits? A comparison of failure detection and diagnosis would directly clarify the added value of the SAE representation.

  2. For the reported failure-category errors, can you distinguish cases where the necessary concept is absent from the vocabulary from cases where an available concept is incorrectly detected or interpreted? This would clarify whether the main limitation lies in concept coverage, feature alignment, or the language diagnosis.

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

This paper proposes MIND, a mechanistic interpretability framework that uses Sparse Autoencoders (SAEs) to map Vision-Language-Action (VLA) model activations into human-understandable robot state and motion concepts. By validating causal influence through feature "steering" (clamping activations) and monitoring feature dynamics during execution, MIND effectively detects runtime failures and diagnoses their root causes. Evaluated on OpenVLA, π0.5, and LingBot-VLA in both simulated and real-world environments, the framework achieves up to 0.991 AUROC in failure detection—halving false alarms—and demonstrates precise causal steering across 12 out of 14 motion concepts.

Critical strengths
  1. While conventional methods only identify when failures occur, this paper achieves both failure detection and root-cause diagnosis within a unified framework using SAE-extracted internal features. By explaining errors through token-level policy states, it significantly advances VLA model interpretability and reliability.
  2. The authors prove feature causality through closed-loop steering experiments, successfully inducing intended behavior across 12 out of 14 motion concepts in OpenVLA. Extensive evaluations across OpenVLA, π0.5, and LingBot-VLA in both simulation and real-world environments demonstrate exceptional detection accuracy (up to 0.991 AUROC) alongside reduced false alarms.
Critical weaknesses
  1. MIND maps basis features extracted via Sparse Autoencoders (SAEs) to a predefined set of state and decision-making concepts. Consequently, it remains unclear whether the method can effectively diagnose internal states or failure causes that fall outside this predefined vocabulary. Furthermore, while semantic consistency is verified across novel tasks and initial states, the framework’s ability to diagnose failure patterns not anticipated during design is insufficiently validated. This raises concerns regarding its generalizability and applicability to complex, real-world environments.
  2. While closed-loop "steering" experiments demonstrate that manipulating SAE features influences robot behavior, showing that feature manipulation alters actions does not prove that a given feature is the actual cause of a failure. The proposed diagnostic output relies on a feature's contribution to a failure-detection score, which reflects correlation rather than direct causality. Although the authors report alignment with simulator ground-truth labels, the study lacks interventional experiments—such as feature correction—to confirm that the diagnosed features are indeed the true causal factors. This presents a critical limitation to the core claim that MIND achieves root-cause diagnosis rather than interpretable anomaly detection.
Questions for authors

Q1. In MIND, latent features extracted via SAEs are mapped to a predefined concept set describing robot states and decision-making. However, in real-world manipulation, failures often stem from unmodeled or unobserved factors. How does MIND behave when encountering a failure cause that falls outside this predefined concept vocabulary? Furthermore, while performance is evaluated on novel tasks and initial states, does the system maintain comparable detection accuracy and diagnostic quality for failure patterns that were entirely unobserved during concept definition and SAE training? Clarifying this is essential for assessing the framework's real-world generalizability.

Q2. The paper uses closed-loop "steering" experiments (feature clamping) to demonstrate that specific SAE features influence policy behavior. However, proving that feature manipulation alters behavior does not establish that feature anomalies are the root causes of runtime failures. Since MIND identifies diagnostic features based on their mathematical contribution to the anomaly score, how do the authors ensure these features are causally responsible for the failure rather than merely correlated? Specifically, do the authors have experimental evidence showing that corrective interventions (e.g., restoring or editing identified features during runtime) can prevent or recover from the failure?

Q3. The paper reports strong alignment between LLM-generated diagnostic notes and ground-truth failure causes in simulation. Could the authors clarify the precise metrics and protocol used to evaluate this diagnostic accuracy? Specifically:

  1. How were ground-truth failure labels established in the simulator, particularly for complex episodes involving multiple interacting causes?

  2. To what extent does the LLM generator introduce text-generation noise or bias into the evaluation?

  3. How does MIND’s diagnostic accuracy compare against an ablation baseline that directly diagnoses failure causes from dense internal activations without SAE feature decomposition?

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

The paper proposes MIND, a mechanistic-interpretability framework for detecting and explaining failures of visuomotor and VLA policies. MIND trains a sparse autoencoder on internal policy activations, assigns human-readable concepts to basis features, and uses closed-loop feature steering to validate whether decision-related features causally affect the policy's actions. The basis features are then used by a runtime monitor to detect failures and identify which internal concepts contributed to the alarm. The method is evaluated across OpenVLA, LingBot-VLA, and π0.5 in simulation, as well as π0.5 on a physical robot.

Overall, I find the problem important and the main idea interesting. In particular, using the same interpretable internal features for both failure detection and diagnosis is a useful direction, and the causal validation through feature steering makes the interpretation more convincing than relying only on correlations. The evaluation across several policies and a physical robot is also a strength. My main concerns are relatively focused and relate to scalability and the scope of the failure detector. The method relies on a predefined vocabulary of candidate concepts, so failures involving concepts outside this vocabulary may remain invisible or difficult to explain. The detection mechanism also relies on a number of fixed design choices and engineered temporal statistics, and the paper does not compare the detector against strong external failure-detection approaches. Addressing these points would make the generality and practical value of the method clearer.

Critical strengths
  1. The paper addresses an important and relatively underexplored problem: explaining why VLA policies fail, rather than only detecting that a failure occurred.

  2. The representation and naming of basis features is interesting. MIND uses sparse features and keeps a concept name only when it remains sufficiently precise on held-out episodes.

  3. The framework connects failure detection and explanation through the same interpretable features. This makes the diagnosis directly related to the signal that triggered the failure alarm.

  4. The use of causal validation through steering is a strong aspect of the paper. The authors do not rely only on correlations between features and concepts; they clamp decision-related features and test whether the robot's behavior changes accordingly.

  5. The experimental evaluation is broad for this type of work. Results are provided across three VLA policies in simulation and on a physical robot. The explanations are also evaluated for stability across bootstrap refits and consistency with the underlying concept record.

Critical weaknesses
  1. The method depends on a predefined concept vocabulary. Although the naming process itself is automatic, the set of possible concepts is specified in advance. This may limit scalability to new tasks and, more importantly, means that failures involving concepts outside the vocabulary cannot be directly diagnosed.

  2. The scope of the detector is limited by the named features and engineered temporal representation. The authors show that the dwell and switching statistics improve detection, but the current monitor can only reason over concepts that were successfully named and represented by these fixed statistics. The paper itself acknowledges that richer phase concepts are needed to distinguish additional failure types.

  3. The failure-detection evaluation lacks strong external baselines. MIND's named-feature monitor is primarily compared with the authors' own dense-activation monitor rather than existing failure-detection methods cited in the paper. This makes it difficult to assess how competitive the detection component is independently of the explanation capability.

Questions for authors
  1. Candidate concept vocabulary. MIND requires a predefined set (C) of state and decision concepts before feature naming. How much task-specific effort is required to construct this vocabulary for a new robotic task or domain? Could the authors discuss how the method would scale when the relevant concepts are not known in advance?

  2. Coverage of failures outside the concept vocabulary. Since only successfully named basis features enter the monitor, what happens when a failure is caused by an internal representation that does not correspond to any concept in (C), or when the relevant feature fails the naming criteria? How much of the possible failure space can therefore remain invisible to MIND?

  3. Long-range failures. Can the current fixed-window representation capture failures caused by dependencies over substantially longer horizons, where the relevant internal event may occur well before the failure becomes visible?

  4. External failure-detection baselines. The failure detector is mainly compared against a monitor operating on the full dense activations. Since the paper cites existing runtime failure-detection methods such as SAFE and related approaches, could the authors include comparisons against external failure detectors? Even simpler baselines such as action-based anomaly detection or a VLM-based failure judge would help establish the detection contribution independently of the interpretability contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.