Open Problems in Mechanistic Interpretability

1 Introduction

• Understanding learned mechanisms can support AI control, deployment monitoring, and safety-critical trust, while enabling scientific insight into artificial minds and world representations learned by AI systems.Mechanistic interpretability aims to use knowledge of decision-making mechanisms to predict network behavior or achieve practical goals such as behavior control and improved design.Source: §1 · p. 4

1.1 The focus of this review: Open problems and the future of mechanistic interpretability

• Unlike prior reviews, this review takes a forward-looking stance, describing the current frontier and directions the field may benefit most from prioritizing.Source: §1.1 · p. 4

1.1.1 Why ‘mechanistic’ interpretability?

• The review defines mechanistic interpretability technically as investigating mechanisms underlying neural-network generalization, motivated by the hypothesis that generalization arises from shared computation.Relevant approaches examine component roles, circuits, information flow, representation subspaces, and representation vectors that generalize across task instances.Source: §1.1.1 · pp. 4–5

1.2 Types of open problems

• The review identifies goals spanning AI monitoring, behavior modification, prediction, model improvement, and latent-knowledge extraction, then examines open problems in methods, applications, and socio-technical challenges.It argues that substantial distance remains despite hopeful progress and discusses how technical progress might inform AI governance alongside social and philosophical challenges.Source: §1.2 · pp. 5–6

2 Open problems in mechanistic interpretability methods and foundations

• Reverse engineering identifies network-component roles, whereas concept-based interpretability searches for components corresponding to proposed concepts; open problems also span the interpretability pipeline and its automation.The section examines both approaches alongside cross-cutting questions about proceduralizing the pipeline and automating interpretability research.Source: §2 · p. 7

2.1.1 Reverse engineering is necessary because AI and humans use different representations and perform different tasks

• Because AI systems can outperform people on some tasks, fail at tasks children master, and use different heuristics, reverse engineering is needed to uncover their unknown concepts and mechanisms.Even on comparable tasks, models may rely on texture, dataset correlations, or strategies such as Fourier transforms that researchers understand only retrospectively.Source: §2.1.1 · p. 7

2.1.2 Reverse engineering step 1: Neural network decomposition

• Architectural units do not reliably expose network representations: neurons and attention heads can be polysemantic, individual layers may be too coarse, and representations can span layers.Representations distributed across heads or layers also pose challenges for applying sparse dictionary learning (SDL) and translating them into causal mechanism descriptions.Source: §2.1.2 · pp. 8, 12
• SDL is the leading unsupervised decomposition approach, but reconstruction errors can substantially reduce model performance, and applying SDL across model layers is costly.The section also notes that scaling costs remain unclear; error nodes capture discrepancies but may leave an uninterpretable “everything else.”Source: §2.1.2 · pp. 9–10
• Training models for interpretability as well as performance may be preferable to training solely for performance and interpreting them afterward.The section proposes developing models that are intrinsically decomposable while maintaining state-of-the-art or near-state-of-the-art performance.Source: §2.1.2 · pp. 11, 13
• SDL’s assumptions that concepts are linearly represented and sparsity indicates interpretability remain uncertain, while its feature-direction approach leaves potentially meaningful feature geometry unexplained.Feature splitting, absorption, and composition challenge sparsity as a proxy; whether global or only local geometry must be understood remains an open question.Source: §2.1.2 · pp. 11–12
• SDL identifies activation directions rather than mechanisms themselves, and its latents may omit useful concepts or depend on the training dataset and model architecture.For example, refusal-related latents were found in SAEs trained on chat-formatted data but generally not in those trained on pretraining data.Source: §2.1.2 · pp. 12–13
• The field still lacks clear formal foundations for features and decomposition; connecting theories of generalization or causality to interpretability could help identify networks’ functional units.The status of superposition as a fundamental account or a pragmatically useful hypothesis is unresolved, and alternative decompositions may incorporate feature geometry or development during training.Source: §2.1.2 · p. 13
• Training for sparse activations, sparse weights, or modularity could make networks easier to decompose, but interpretability incentives have often reduced performance or allowed superposition to persist.Proposed approaches include discrete codes, sparse activation functions, weight pruning, and geometrically local connections; finetuning with interpretability in mind may avoid training from scratch.Source: §2.1.2 · p. 14

2.1.3 Reverse engineering step 2: Describing the functional role of components

• Highly activating examples can mislead: human expectations and dataset bias shape interpretations, and plausible explanations can arise for arbitrary directions rather than identifying computational units.Bolukbasi et al. (2021) found that annotators assigned dramatically different meanings to BERT directions depending on the dataset used.Source: §2.1.3 · pp. 14–15
• Attribution methods aim to explain activation causes, but gradient approximations can be poor or model-independent, perturbations can leave the training distribution, and feature synthesis has uncertain interpretive value.Feature synthesis maximizes component activation subject to regularization, yet studies question whether synthetic inputs aid interpretation or reveal trojans.Source: §2.1.3 · pp. 15–16
• The logit lens and related projections reveal direct effects on output vocabulary but miss indirect effects through later hidden activations; tuned lenses improve decoding at the cost of faithfulness.Causal interventions can measure indirect effects, and future extensions might project onto intermediate downstream components.Source: §2.1.3 · p. 16
• Causal interventions, path patching, steering, and patchscopes probe component effects, but interventions can be expensive, gradient-based approximations offer faster alternatives, and chain-of-thought is not fully faithful.Path patching isolates a component’s effect on another component; incorporating model internals into chain-of-thought training or monitoring is proposed as a possible direction.Source: §2.1.3 · pp. 16–17

2.1.4 Reverse engineering step 3: Validation of descriptions

• Descriptions should be tested through predictions and counterfactuals, while model organisms and benchmarks can support validation but must not displace engineering goals or generalization to safety-relevant settings.Model organisms can provide infrastructure for rigorous hypothesis testing, but researchers lack consensus on suitable systems; benchmarks can compare interpretations with ground truth and assess pipeline steps.Source: §2.1.4 · pp. 17–19
• Interpretability tools need fair, non-cherry-picked comparisons on real-world tasks; current methods have not been consistently useful, and validation remains costly in time and cognitive labor.The authors argue that demonstrating competitive performance matters because evaluations limited to easy tasks may favor methods that fail under practical challenges.Source: §2.1.4 · p. 18

2.2.1 Concept-based probes

• A concept-based probe is a classifier trained on hidden representations to predict a labeled concept, so probing requires a labeling function that assigns concept values to inputs.Source: §2.2.1 · p. 19
• A probe trains a model to predict concept labels from hidden activations, and a linear probe localizes the concept as a latent-space vector.Probing is relatively simple to implement, but requires carefully chosen data and detects correlations rather than causal variables (Belinkov, 2022b).Source: §2.2.1 · p. 20

2.2.2 Probes need carefully chosen data for well-defined concepts

• Concept-based probes need carefully defined, labeled data, so they identify only sought concepts; even CCS requires clear positive and negative examples.Automatically developing probing datasets is proposed as a potential path forward (Shaham et al., 2025).Source: §2.2.2 · p. 20

2.2.3 Probes detect correlations, rather than causal variables, in hidden activations

• Probes identify correlations, not necessarily causal mediators, so their results should generate hypotheses that require further validation.High-dimensional activations risk spurious correlations, making out-of-distribution tests and better-understood regularization important; counterfactual and other causal methods can identify mediators but require more specialized data.Source: §2.2.3 · pp. 20–21

2.2.4 Concept-based intrinsic interpretability

• Training networks with concept bottlenecks or specified causal structures may make concept-assigned components more causally relevant.Because not all concepts or structures can be prespecified, combining this approach with concept-disentangling methods may be useful.Source: §2.2.4 · p. 21

2.3 Proceduralizing mechanistic interpretability into circuit discovery pipelines: A case study

• Circuit discovery’s decomposition, description, and validation choices form a pipeline, but current procedures have unresolved shortcomings.Common evaluation criteria include faithfulness, minimality, and completeness.Source: §2.3 · pp. 21–22
• Circuit discovery can miss individual-data behavior, rely on imperfect decompositions, and produce circuits whose faithfulness measures depend on interventions.Task-level circuits may approximate average performance poorly for individual examples; scalable gradient-based attribution is only a first-order approximation, and current methods can overlook backup or negative behavior and struggle on arbitrary tasks.Source: §2.3 · p. 22

2.4 Automating steps in mechanistic interpretability research

• Automation can generate and validate component descriptions or identify task-relevant subgraphs, but current pipelines do not automate all steps or yet yield satisfactory explanations.ACDC-like methods identify subgraphs for predefined tasks but do not describe component functions; description tests include predicting feature activations from examples.Source: §2.4 · p. 23

3.1 Axes of mechanistic interpretability progress.

• Different goals demand different levels of component description and decomposition, task-distribution coverage, and understanding of mechanisms during rather than only after training.Some goals may need only specific components or narrow task distributions, whereas others may require understanding larger circuits or an entire model across its task distribution.Source: §3.1 · p. 24

3.2.1 Mechanistic interpretability-based evaluations could help us detect unsafe or unethical AI cognition

• White-box evaluations could reveal unsafe cognition that input-output tests miss, though validating coverage of concerning patterns remains difficult.Shallow, correlation-based descriptions may still flag concerning components, and internal monitoring or anomaly detection could help during deployment; interpretability may also strengthen red-teaming and safety cases.Source: §3.2.1 · pp. 25–27

3.2.2 Using mechanistic interpretability for better control of AI system behavior

• Interpretability-inspired control includes activation steering, unlearning, model editing, and finetuning analysis, but precise edits require better decomposition and descriptions.Activation steering has achieved moderate success and decomposition can narrow behaviors with fewer side effects; activation-based unlearning has not yet yielded competitive results (Farrell et al., 2024), and existing editing methods have struggled to identify appropriate components. Mechanistic work also suggests finetuning often makes shallow changes, which may make harmlessness training cheaply reversible (Gade et al., 2024; Lermen et al., 2024; Arditi et al., 2024).Source: §3.2.2 · pp. 27–28

3.3.1 Predicting behavior in novel situations

• Mechanistic understanding may predict behavior in novel situations, but strong claims and formal verification require advances beyond current interpretability capabilities.Mechanisms may expose jailbreaking, backdoors, adversarial examples, biases, hallucination signatures, or knowledge of facts; formal verification remains unresolved, with demonstrations limited to toy models and scaling to frontier systems uncertain.Source: §3.3.1 · pp. 28–29

3.3.2 Predicting capabilities that arise during training or finetuning

• Predicting emerging or elicitable capabilities requires tracing how mechanisms change during training and finetuning, including across multiple time steps.Research links in-context learning to induction heads (Olsson et al., 2022), while finetuning may enhance existing circuits or mask capabilities that can be revealed by simple changes (Prakash et al., 2024; Jain et al., 2024; Lee et al., 2025); current work largely studies single-step mechanisms.Source: §3.3.2 · pp. 29–30

3.4 Using mechanistic interpretability to improve our ability to perform inference, improve training, and make better use of learned mechanisms

• Mechanistic understanding could improve inference efficiency, training-data selection and monitoring, parameter-efficient training, reasoning interventions, and recombination of learned representations.Potential applications include skipping computations without changing outputs, identifying capability gains during training, removing reasoning bugs, and transferring representations across models.Source: §3.4 · p. 30

3.5 Using mechanistic interpretability for ‘microscope AI’

• “Microscope AI” uses neural networks to find potentially novel predictors in complex datasets, then applies interpretability to understand those patterns.This approach aims to leverage neural networks’ capacity to capture nonlinear relationships in massive datasets while making their discovered patterns scientifically interpretable.Source: §3.5 · p. 30
• Mechanistic interpretability can act as “microscope AI,” revealing chess concepts, facial effects on judgments, causal links in psychology, and immune-cell morphology (Tong et al., 2024b).As methods improve, deeper insights across more domains may become possible.Source: §3.5 · p. 31
• Most scientists cannot yet access microscope AI because it requires rare combined expertise in machine learning, interpretability, and domain knowledge.Wider scientific adoption and more automated, accessible interpretability could change this.Source: §3.5 · p. 31

3.6 Mechanistic interpretability on a broader range of models and model families

• Research has focused mainly on CNN image models and BERT- and GPT-based text models, leaving generalization to other architectures and multimodal models open.Alternatives include diffusion models (Sohl-Dickstein et al., 2015; Rombach et al., 2022), Vision Transformers (Dosovitskiy et al., 2021), RWKV (Peng et al., 2023), and state space models (Gu & Dao, 2024).Source: §3.6 · p. 31
• Whether interpretability methods transfer across architectures remains open, although some methods have transferred from CNNs to SSMs and from transformers to some SSMs (Paulo et al., 2024; Meng et al. (2022b); Sharma et al. (2024a); Wang et al. (2023); Ensign & Garriga-Alonso (2024)).Researchers may need to identify approaches that work across architectures.Source: §3.6 · p. 31
• Comparing model families could distinguish model-specific findings from general ones and test the universality hypothesis that networks learn similar features and circuits.The field’s concentration on transformers may leave such insights undiscovered.Source: §3.6 · p. 31

3.7 Human computer interaction with model internals

• Interfaces to model internals could help engineers and researchers, enable auditors to find rare high-stakes errors, and help users calibrate trust and interact with systems.Proposed chatbot dashboards would show features influencing answers in real time and might flag errors or safety-relevant activations (Chen et al., 2024; Zou et al., 2023a; Viégas & Wattenberg, 2023); interpretability could also combine direct-manipulation and text interfaces.Source: §3.7 · pp. 31–32

4.1 Translating technical progress in mechanistic interpretability into levers for AI policy and governance

• Proposed governance levers span risk evaluation, deployment oversight, liability explanations, mitigation commitments and copyright protection, though translation into better governance remains uncertain.These consequential applications also carry social and philosophical ramifications.Source: §4 · p. 33, §4.1 · p. 33
• Model internals could improve frontier-risk evaluation and forecasting, from eliciting dangerous capabilities to informing threat models and giving policymakers more preparation time.Interpretability could guide red-team attempts, check for intentional underperformance on evaluations, and help assess competing accounts of catastrophic threats.Source: §4.1 · p. 33
• Interpretability may support incident reporting through inference monitoring, post-incident feature analysis and test-time monitors, and could help verify regulatory or treaty compliance.Investigating incident-related input features may guide further red-teaming and help reduce the likelihood of similar future incidents.Source: §4.1 · p. 33
• Mechanistic rationales could support rights to explanations, while model editing and interpretability could identify and remove memorized copyrighted works.The passage connects explanations to rights under the EU General Data Protection Regulation and removal to copyrighted works reproducible verbatim by generative models.Source: §4.1 · p. 34

4.2 Social and philosophical problems in mechanistic interpretability

• The field lacks paradigmatic clarity about its goals, success criteria, and definition of interpretability, including whether intrinsic understanding is a valid objective.Interpretability research has yielded few tools used to make state-of-the-art systems safer in the real world, while most systems are still trained, evaluated, monitored, and debugged without internal understanding.Source: §4.2 · p. 34
• Practical benchmarks can ground engineering goals, but safety applications must compare interpretability against alternatives; interpretability need not strictly outperform black-box methods to remain useful.The passage warns that omitting or misrepresenting comparisons can cause follow-up work to rest on false assumptions, while interpretability-based advantages may still matter when methods perform in a similar ballpark.Source: §4.2 · p. 34
• Interpretations can depend on deployment context, while selective transparency and corporate conflicts create risks of misleading communication and overstating interpretability’s safety progress.The passage notes that recent progress has been used to argue falsely that the black-box problem is resolved, potentially serving corporate interests at the expense of safety.Source: §4.2 · pp. 34–35

5 Conclusion

• Stronger theory, scalable decomposition methods, causal validation, and techniques for tracking training and interactions are needed to address major methodological challenges.Sparse dictionary learning remains promising but faces scaling limitations and conceptual challenges in its assumptions.Source: §5 · p. 36
• Interpretability could improve risk monitoring, model control, capability prediction, AI performance, and scientific insight, provided the field prioritizes real-world utility.The passage identifies more efficient architectures, better training procedures, and microscope AI approaches as possible applications.Source: §5 · p. 36
• Progress should pursue scientific and engineering goals together, establish better benchmarks, and compare interpretability-based approaches with non-interpretability baselines.Source: §5 · p. 36
• Interpretability tools could support governance and oversight, but their benefits require care about misuse and false assurance regarding AI safety.Potential uses include checking safety-standard compliance, detecting risks before deployment, and clarifying decision attribution.Source: §5 · p. 36
• As AI capabilities advance, understanding model decision-making becomes increasingly urgent, while the field’s open problems remain unresolved.Source: §5 · p. 36

A.1.1 Reverse engineering: Identifying the roles of network components

• How networks should be decomposed into interpretable parts—including suitable isomorphisms, coarse-graining, and higher-level abstractions—remains open.Source: §A.1.1 · p. 75
• The field still needs to establish how often concepts are encoded linearly, how to characterize nonlinear representations, and what drives either outcome.Source: §A.1.1 · p. 75
• Researchers ask whether linear representations and superposition adequately frame computation, including superposition’s causes and forms in attention and across layers.Source: §A.1.1 · p. 75
• Sparse dictionary learning (SDL) raises open questions about reconstruction errors, sparsity as an interpretability proxy, scalability, and connecting features into circuits.The paper also asks whether SDL remains appropriate without linear representations and whether new methods can address its conceptual and practical limitations.Source: §A.1.1 · p. 75
• Open problems include how activation-space geometry explains network function and whether global or only local feature geometry is needed to understand computation.Source: §A.1.1 · pp. 75–76
• The section asks how interpretability can connect to generalization theories, distinguish mechanisms of generalization from memorization, and relate to robustness or deep-learning theory.Source: §A.1.1 · p. 76
• It remains open whether models can be made intrinsically interpretable, or converted after training, at little performance cost, and how training can promote human-understandable, controllable concepts.Source: §A.1.1 · p. 76
• Further priorities are less biased, mechanism-level interpretations; faithful attribution and causal-effect measurement; and validation that is tractable, predictive, and robust beyond best-case or in-distribution settings.The section proposes model organisms, known-ground-truth benchmarks, stress tests, and average- and worst-case evaluation as possible parts of a validation framework.Source: §A.1.1 · pp. 76–77

A.1.2 Concept-based interpretability: Identifying network components for given roles

• Probing must reliably distinguish causal from merely correlated features in neural networks.Source: §A.1.2 · p. 77
• Automated systems could generate high-quality probing datasets and reduce reliance on human effort.Source: §A.1.2 · p. 77
• Probes need regularization and validation techniques that prevent spurious correlations while finding generalizable features.Source: §A.1.2 · p. 77
• Probing methods need to improve for concepts without clear positive or negative examples.Source: §A.1.2 · p. 77

• Can lower-level methods be developed to provide deeper or more complete insights into neural networks?Source: §A.1.3 · p. 77
• Can methodological progress in decomposition improve circuit discovery’s faithfulness and reduce explanation description length?Source: §A.1.3 · p. 77
• Can circuit discovery move beyond concept-based task definitions, which may complicate mechanistic analysis?Source: §A.1.3 · p. 77
• Can practical problems involving negative and backup behavior be overcome in circuit discovery?Source: §A.1.3 · p. 77
• Will circuit discovery yield insights into arbitrary tasks, or only tasks that can be crisply defined?Source: §A.1.3 · p. 77

• Can automated feature description and validation improve by generating and testing arbitrary hypotheses, comparing features, or describing component interactions?Source: §A.1.4 · p. 78
• Can ACDC-like circuit discovery methods be improved?Source: §A.1.4 · p. 78
• Can other pipeline steps be automated, including conceptual interpretability research, decomposition method discovery, and ad hoc hypothesis validation?Source: §A.1.4 · p. 78
• Should researchers mitigate the risk of potentially misaligned AI systems sabotaging AI-automated interpretability?Source: §A.1.4 · p. 78

A.2.1 Using mechanistic interpretability for better monitoring and auditing of AI systems for potentially unsafe cognition

• Can “white box” evaluations detect concerning internal patterns without requiring understanding of the entire network?The subsection also asks how to validate that learned features capture all concerning reasoning patterns and distinguish features that recognize deception from mechanisms that generate it.Source: §A.2.1 · p. 78
• Can evaluations reliably distinguish features that recognize deceptive behavior from mechanisms that generate it?The subsection also asks how to validate that learned features capture all concerning patterns of reasoning.Source: §A.2.1 · p. 78
• Can evaluators identify which features are appropriately versus spuriously relevant?The passage raises this as an open question without specifying the context in the supplied text.Source: §A.2.1 · p. 78
• Can interpretability make red-teaming more efficient and help human red-teamers identify problematic inputs through feature attribution?The subsection leaves open how best to use interpretability insights for red-teaming and system testing.Source: §A.2.1 · p. 78
• Can test-time interpretability monitoring detect mechanistic anomalies and passively flag concerning internal patterns during deployment?The subsection asks whether effective monitoring can rely on feature-level understanding rather than deep mechanical insights.Source: §A.2.1 · p. 78
• Can interpretability support effective passive monitoring of concerning model internals during deployment?The subsection asks whether this can work as a test-time monitoring system.Source: §A.2.1 · p. 78

A.2.2 Using mechanistic interpretability for better control of AI system behavior

• Can interpretability make activation steering more precise with fewer side effects and enable steering whole mechanisms rather than single features?The section frames both as open questions about improving steering methods.Source: §A.2.2 · p. 78
• Can mechanistic interpretability improve unlearning and editing by identifying network joints, evaluating unlearning, and determining feasible edits that preserve generalization?The section asks whether carving the network at its true joints can help, and which classes of edit are possible without undesirable damage to generalization.Source: §A.2.2 · pp. 78–79
• Can interpretability make finetuning more sample-efficient by targeting specific parameters and improve tools for analyzing feature- or mechanism-level differences?Both parameter targeting and analysis at feature or mechanism level are posed as open questions.Source: §A.2.2 · p. 79

Behavior prediction

• Can mechanistic understanding predict behavior in novel situations, including jailbreaks, safety bypasses, hallucinations, and behavior without evaluations, or reveal values, goals, or absent dangerous capabilities?The questions also ask whether internal signatures can predict specific failure modes.Source: §A.2.3 · p. 79

Formal verification

• Can formal verification scale from toy models to frontier systems, establish how much computation is verifiable symbolically, and provide guarantees in complex environments?The section also asks what level of mechanistic understanding meaningful verification requires.Source: §A.2.3 · p. 79

Model safety

• Can interpretability support rigorous safety claims by proving dangerous mechanisms absent, verifying claims about values and goals, and reliably enumerating relevant mechanisms?The section asks what kinds of safety claims current interpretability methods can support.Source: §A.2.3 · p. 79

Capability development

• Can interpretability predict capability development by identifying early signatures of emergence and explaining how mechanisms, circuits, and loss landscapes relate to capabilities?The section asks how mechanisms change during training and connect small-scale circuits to large-scale capabilities.Source: §A.2.3 · p. 79

Training data and capabilities

• Can researchers trace how training examples shape mechanisms and limitations, design data for desired capabilities, and understand data structure’s effects on in-context versus weights-based learning?Source: §A.2.3 · p. 79

Latent and maskable capabilities

• Can interpretability detect capabilities that prompting or finetuning might unlock or mask, analyze mechanisms spanning timesteps, and distinguish fundamental capabilities from superficially trained ones?Source: §A.2.3 · p. 79

A.2.4 Using mechanistic interpretability to improve our ability to perform inference, improve training and make use of learned representations

• Can mechanistic insights make inference more efficient, improve training, and help instill capabilities directly in networks?Questions span selecting training data, monitoring capability emergence, designing inductive biases and modular architectures, and transferring capabilities between models.Source: §A.2.4 · p. 80

A.2.5 Using mechanistic interpretability for ’microscope AI’

• Can interpretability reveal scientific patterns in models and make knowledge extraction accessible, automated, scientifically validated, and connected to domain concepts?The questions include detecting genuinely novel patterns, extending beyond simple correlational discoveries, and enabling use without deep machine-learning expertise.Source: §A.2.5 · p. 80

A.2.6 Mechanistic interpretability on a broader range of models and model families

• Can interpretability generalize across architectures and modalities, distinguish model-specific from universal mechanisms, and remain useful for novel and future models?The section asks how models trained on similar data compare, whether they learn similar features and circuits, and how architecture-specific methods should be balanced with general principles.Source: §A.2.6 · pp. 80–81

A.2.7 Human computer interaction with model internals

• Can interfaces make model internals understandable and useful to experts, auditors, policymakers, and end users?Questions cover intuitive visualizations and dashboards, failure and bias detection, trust calibration, and controls for steering model behavior.Source: §A.2.7 · p. 81

A.2.8 Governance

• Can mechanistic analysis identify and prevent failures, uncover dangerous or deceptive mechanisms, and help verify compliance?Questions include tracing causal mechanisms behind incidents and detecting mechanisms associated with dangerous capabilities.Source: §A.2.8 · p. 81

A.2.9a Translating technical progress in mechanistic interpretability into levers for AI policy and governance

• Can interpretability inform capability evaluations, forecasts of emerging capabilities, threat estimates, incident prevention, GPU workload verification, and copyright responses?Proposed incident levers include test-time monitors and evaluation tasks based on incident scenarios; copyright questions include detecting and removing memorized works.Source: A.2.9a Translating technical progress… · p. 82
• Can interpretability help verify GPU workloads and address copyright law, including detecting and removing memorized copyrighted works?Source: A.2.9a Translating technical progress… · p. 82

A.2.9 Open socio-technical problems in mechanistic interpretability

• How should the field define interpretability and success, balance scientific and engineering aims, and mitigate risks that research results could be misused?The questions also ask how research findings should be communicated to reduce misuse risks.Source: A.2.9a Translating technical progress… · p. 82