# ObserverBench: Testing Mechanistic Estimates for Intervention and Control

arXiv: https://arxiv.org/abs/2609.03026 · HTML: https://mlsanity.com/paper/2609.03026/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Circuits and mechanistic analysis, one of 13 directions in Mechanistic interpretability.
- Builds on 24 of its 37 references that are in the map.
- First posted on arXiv 2 Sep 2026.

## Placed in

- Mechanistic interpretability / [Circuits and mechanistic analysis](https://mlsanity.com/mechinterp/circuits-and-mechanistic-analysis/)

## Builds on (most cited first, 24 of 24)

- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Representation Engineering: A Top-Down Approach to AI Transparency](https://mlsanity.com/paper/2310.01405/) (2023)
- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2](https://mlsanity.com/paper/2408.05147/) (2024)
- [Smart "Predict, then Optimize"](https://mlsanity.com/paper/1710.08005/) (2017)
- [AI Control: Improving Safety Despite Intentional Subversion](https://mlsanity.com/paper/2312.06942/) (2023)
- [AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders](https://mlsanity.com/paper/2501.17148/) (2025)
- [Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representations](https://mlsanity.com/paper/2303.02536/) (2023)
- [Attribution Patching Outperforms Automated Circuit Discovery](https://mlsanity.com/paper/2310.10348/) (2023)
- [The Hydra Effect: Emergent Self-repair in Language Model Computations](https://mlsanity.com/paper/2307.15771/) (2023)
- [Are Sparse Autoencoders Useful? A Case Study in Sparse Probing](https://mlsanity.com/paper/2502.16681/) (2025)
- [AtP*: An efficient and scalable method for localizing LLM behaviour to components](https://mlsanity.com/paper/2403.00745/) (2024)
- [Copy Suppression: Comprehensively Understanding an Attention Head](https://mlsanity.com/paper/2310.04625/) (2023)
- [MIB: A Mechanistic Interpretability Benchmark](https://mlsanity.com/paper/2504.13151/) (2025)
- [pyvene: A Library for Understanding and Improving PyTorch Models via Interventions](https://mlsanity.com/paper/2403.07809/) (2024)
- [Explorations of Self-Repair in Language Models](https://mlsanity.com/paper/2402.15390/) (2024)
- [InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques](https://mlsanity.com/paper/2407.14494/) (2024)
- [Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models](https://mlsanity.com/paper/2605.11887/) (2026)
- [Activation Steering with a Feedback Controller](https://mlsanity.com/paper/2510.04309/) (2025)
- [The Bayesian Geometry of Transformer Attention](https://mlsanity.com/paper/2512.22471/) (2025)
- [The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching](https://mlsanity.com/paper/2606.27510/) (2026)
- [Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control](https://mlsanity.com/paper/2604.19018/) (2026)
- [Interpretability Can Be Actionable](https://mlsanity.com/paper/2605.11161/) (2026)
- [Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability](https://mlsanity.com/paper/2608.19338/) (2026)

## Built on by (most cited first, 1 of 1)

- [Replay-Gated Neural Execution: Decoupling Persistent Behavioral Specifications from Neural Realizations in Frozen Language Models](https://mlsanity.com/paper/2609.22243/) (2026)

## Abstract (arXiv)

Mechanistic interpretability is increasingly used to guide interventions such as activation steering, circuit removal, and safety monitoring. Yet an internal estimate that is accurate on average can still choose a poor action. We present ObserverBench, a benchmark framework for testing whether an internal estimator---an observer---is adequate for the intervention, control, or safety task it directs. Each task fixes the model, information boundary, allowed actions, decision rule, held-out cases, and loss. The benchmark reports estimation accuracy separately from the loss caused by the chosen action. Theory and experiments show why both are needed. In closed-loop control, observer errors matter at the starting point and along directions the allowed intervention can reach. On circuit-intervention tasks in GPT-2-small and Qwen2.5-7B, pairwise observers predict unseen effects more accurately without always choosing better actions; observers trained on action loss choose lower-loss actions. In safety triage, a score that perfectly separates violations can allocate a fixed intervention budget poorly when violations have different costs. Across Qwen2.5-7B, Gemma-2-9B-it, and prospectively frozen Qwen3.5-9B APPS tasks, AUROC can rank monitors differently from deployment loss, and the best information source changes across models. Sparse SAE readouts also trail their layer-matched dense controls on the reported Qwen panels, under disclosed activation-density or checkpoint mismatches. ObserverBench provides fixed task contracts, runnable baselines, and table-based submissions for evaluating interpretability methods through the actions they enable.