mlsanityData to 14 Oct 2026

Reading Task Failure Off the Activations: A Sparse-Feature Audit of GPT-2 Small on Indirect Object Identification

arXiv 2605.22719 ↗First posted 21 May 2026cs.LGcited by 1

On the map: Mechanistic interpretability → Sparse autoencoder training and evaluation

Abstract (arXiv)

We report a small, reproducible audit of which sparse-autoencoder (SAE) features of GPT-2 small fire differently on failed versus successful trials of the Indirect Object Identification (IOI) task. On 300 prompts, GPT-2 small reaches 79.7% accuracy; 146 of the 24,576 features in the layer-8 residual-stream SAE release of Bloom (2024) clear a Holm-corrected significance threshold and 105 reach a large effect size (|Cohen's d| > 0.8). The strongest single correlate of failure -- feature 17,491, d=+2.93, Neuronpedia label 'cryptographic keys' -- is essentially silent except when the prompt's transferred object is 'the keys,' on which GPT-2 small fails 93.3% of the time vs. 7.5% on the other seven objects (Fisher exact p = 8.79 x 10^-33). We put this correlate through three controls that a mechanistic claim should pass. (i) A causal ablation: zeroing feature 17,491 in the residual stream across all token positions of the 45 keys prompts does not restore accuracy (6.7% -> 4.4%); the feature is a correlate, not a sufficient cause at this layer. (ii) A representation baseline: a logistic regression on the raw 768-dimensional residual stream reaches 5-fold ROC AUC = 0.929, matching the top-100 SAE features (0.927); the SAE basis adds interpretability, not predictive power. (iii) A seed-robustness check: across five random seeds the keys-subset failure rate stays in 75.0--93.3% (the behavioural effect is real), but feature 17,491 is the top-|d| feature in only 1 of 5 runs. The methodological contribution is therefore the audit pipeline (cheap, model-agnostic, surfaces named correlates) rather than any single feature found through it. We release the code, the 300-prompt corpus, the 300x24,576 activation matrix, the ablation and baseline scripts, and the figures. The full pipeline runs on a laptop (Apple M3 Max, no discrete GPU).

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Sparse autoencoder training and evaluation, one of 13 directions in Mechanistic interpretability. Builds on 22 of its 39 references that are in the map. First posted on arXiv 21 May 2026.

Builds on

22 of 39 references.

Closest ancestors · 2

Same direction, newest first

  1. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 29 Aug
  2. 2022Toy Models of Superposition21 Sep

Other references · 20

  1. 2023SWE-bench: Can Language Models Resolve Real-World GitHub Issues?cited by 3,373
  2. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  3. 2023WebArena: A Realistic Web Environment for Building Autonomous Agentscited by 2,006
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  5. 2023AgentBench: Evaluating LLMs as Agentscited by 1,267
  6. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  7. 2021Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivitycited by 1,135
  8. 2023Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formattingcited by 817
  9. 2017Deep Reinforcement Learning that Matterscited by 754
  10. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  11. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  12. 2021Causal Abstractions of Neural Networkscited by 422
8 more
  1. 2020Improving Reproducibility in Machine Learning Research (A Report from the NeurIPS 2019 Reproducibility Program)cited by 322
  2. 2023State of What Art? A Call for Multi-Prompt LLM Evaluationcited by 310
  3. 2023How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language modelcited by 202
  4. 2024Transcoders Find Interpretable LLM Feature Circuitscited by 200
  5. 2023Localizing Model Behavior with Path Patchingcited by 195
  6. 2020Causal Mediation Analysis for Interpreting Neural NLP: The Case of Gender Biascited by 191
  7. 2016Linear Algebraic Structure of Word Senses, with Applications to Polysemycited by 165
  8. 2026What Twelve LLM Agent Benchmark Papers Disclose About Themselves: A Pilot Audit and an Open Scoring Schemacited by 6

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026SFT Leaves a Reproducible Signature in Activation Geometry: Calibrating Post-Training Interpretability Against Run-to-Run Noisecited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial Training
  2. 2026From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models
  3. 2024Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control
  4. 2026Query Lens: Interpreting Sparse Key-Value Features with Indirect Effects
  5. 2026From Geometric Recovery to Causal Validation: A Reproducible Audit of Sparse Autoencoder Features, from Superposition Geometry to Causal Inertness
  6. 2024Interpreting Attention Layer Outputs with Sparse Autoencoders
  7. 2025Steering CLIP's vision transformer with sparse autoencoders
  8. 2026From Input to Output: A Flexible Agent for Dual-End Interpretation of Sparse Autoencoder Features