SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?

• URL: http://arxiv.org/abs/2609.09113v2
• Date: Fri, 11 Sep 2026 04:21:12 GMT
• Title: SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?
• Abstract summary: We introduce SAEScientist-Bench to evaluate whether AI agents can act as scientists utilizing SAE tools for autonomous mechanistic discovery.<n>Given a target concept, an agent designs contrastive probes and navigates a Gemma Scope dictionary of 131K+ features to discover the optimal feature.<n>Across 10 agent configurations and 20 tasks, frontier agents demonstrate genuine discovery capabilities but remain well behind the expert baseline.
• Score: 26.940600916534446
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: While research on recursive self-improvement (RSI) has predominantly automated model training pipelines, reliable autonomous development demands a missing pillar: post-hoc monitoring and auditing to understand what models learn and ensure safe alignment. Mechanistic interpretability tools are essential to bridge this gap, among which Sparse Autoencoders (SAEs) serve as a cornerstone by isolating interpretable features for model inspection and steering. In this paper, we introduce SAEScientist-Bench to evaluate whether AI agents can act as scientists utilizing SAE tools for autonomous mechanistic discovery. Given a target concept, an agent designs contrastive probes and navigates a Gemma Scope dictionary of 131K+ features in Gemma-2-9B-IT to discover the optimal feature, evaluated against curated expert reference features anchored on Neuronpedia across activation rank, concept selectivity on contrastive texts, and causal steering. Across 10 agent configurations and 20 tasks, frontier agents demonstrate genuine discovery capabilities and lead different evaluation dimensions, but remain well behind the expert baseline, approaching expert levels on separating target concepts from contrastive controls while lagging substantially in causal generation steering. Further analysis reveals that although agents can design contrasts to rule out spurious candidates, they frequently misinterpret experimental measurements. These results establish experimental model understanding as a measurable capability for closed-loop autonomous AI R&D. Our code is available at https://github.com/Trae1ounG/SAEScientist.

Related papers

• From Knowing to Acting: Benchmarking Self-Awareness Capability of LLM Agents [69.07911499189129]
  We introduce KAPRO, a framework that evaluates cognitive-behavioral alignment by decoupling an agent's metacognitive judgment (Knowing) from its spontaneous execution (Acting)<n>Experiments show that self-awareness capability is strongly correlated with task success but degrades sharply in internal-capability settings.<n> proprietary and reasoning-oriented models demonstrate more reliable cognitive gating.
  arXiv  Detail & Related papers  (2026-06-09T17:17:08Z)
• The Meta-Agent Challenge: Are Current Agents Capable of Autonomous Agent Development? [80.24951682268332]
  We introduce the Meta-Agent Challenge (MAC), an evaluation framework designed to test the capacity of frontier models for autonomous agent development.<n>To ensure evaluation integrity, this framework is secured by multi-layer defenses against reward hacking.<n>We demonstrate that meta-agents rarely match human-engineered baseline policies, and the few that do are dominated by proprietary frontier models.
  arXiv  Detail & Related papers  (2026-06-03T04:58:17Z)
• Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models [80.45129499188461]
  We introduce Qwen-Scope, an open-source suite of sparse autoencoders (SAEs) built on the Qwen model family.<n>We show that SAEs can go beyond post-hoc analysis to serve as practical interfaces for model development along four directions.
  arXiv  Detail & Related papers  (2026-05-12T10:01:06Z)
• Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment [59.536125286960186]
  Multi-agent frameworks assigning specialized roles are increasingly adopted to enable self-reflection and mutual auditing.<n>We find it simultaneously induces a human-like cognitive bias known as Actor-Observer Asymmetry.<n>ReTAS guides agents to synthesize conflicting viewpoints into an objective consensus.
  arXiv  Detail & Related papers  (2026-04-21T15:05:58Z)
• When the Coffee Feature Activates on Coffins: An Analysis of Feature Extraction and Steering for Mechanistic Interpretability [0.0]
  Recent work by Anthropic on Mechanistic interpretability claims to understand and control Large Language Models.<n>We conduct an initial stress-test of these claims by replicating their main results with open-source SAEs for Llama 3.1.<n>We find that feature steering exhibits substantial fragility, with sensitivity to layer selection, steering magnitude, and context.
  arXiv  Detail & Related papers  (2026-01-06T14:29:51Z)
• The Geometry of Benchmarks: A New Path Toward AGI [0.0]
  We introduce a geometric framework in which all psychometric batteries for AI agents are treated as points in a structured moduli space.<n>First, we define an Autonomous AI (AAI) Scale, a Kardashev-style hierarchy of autonomy grounded in measurable performance.<n>Second, we construct a moduli space of batteries, identifying equivalence classes of benchmarks that are indistinguishable at the level of agent orderings and capability inferences.<n>Third, we introduce a general Generator-Verifier-Updater (GVU) operator that subsumes reinforcement learning, self-play, debate and verifier-based fine-tuning
  arXiv  Detail & Related papers  (2025-12-03T21:34:09Z)
• Impatient Users Confuse AI Agents: High-fidelity Simulations of Human Traits for Testing Agents [58.00130492861884]
  TraitBasis is a lightweight, model-agnostic method for systematically stress testing AI agents.<n>TraitBasis learns directions in activation space corresponding to steerable user traits.<n>We observe on average a 2%-30% performance degradation on $tau$-Trait across frontier models.
  arXiv  Detail & Related papers  (2025-10-06T05:03:57Z)
• AgentThink: A Unified Framework for Tool-Augmented Chain-of-Thought Reasoning in Vision-Language Models for Autonomous Driving [31.127210974372456]
  Vision-Language Models (VLMs) show promise for autonomous driving, yet their struggle with hallucinations, inefficient reasoning, and limited real-world validation hinders accurate perception and robust step-by-step reasoning.<n>We introduce textbfAgentThink, a pioneering unified framework that integrates Chain-of-Thought (CoT) reasoning with dynamic, agent-style tool invocation for autonomous driving tasks.
  arXiv  Detail & Related papers  (2025-05-21T09:27:43Z)
• Interactive Autonomous Navigation with Internal State Inference and Interactivity Estimation [58.21683603243387]
  We propose three auxiliary tasks with relational-temporal reasoning and integrate them into the standard Deep Learning framework. These auxiliary tasks provide additional supervision signals to infer the behavior patterns other interactive agents. Our approach achieves robust and state-of-the-art performance in terms of standard evaluation metrics.
  arXiv  Detail & Related papers  (2023-11-27T18:57:42Z)
• RCAgent: Cloud Root Cause Analysis by Autonomous Agents with Tool-Augmented Large Language Models [46.476439550746136]
  Large language model (LLM) applications in cloud root cause analysis (RCA) have been actively explored recently. We present RCAgent, a tool-augmented LLM autonomous agent framework for practical and privacy-aware industrial RCA usage. Running on an internally deployed model rather than GPT families, RCAgent is capable of free-form data collection and comprehensive analysis with tools.
  arXiv  Detail & Related papers  (2023-10-25T03:53:31Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.