Rethinking Circuit Evaluation: Do Circuits Explain Model Errors?

Mechanistic interpretability (MI) aims to explain a model's behaviour through analyzing its internal computations; circuit-based explanations aim to isolate these computations with compact subnetworks validated by ablating the rest of the model. We show that circuits validated this way may fail to recover the underlying mechanism of the model's behaviour by closely reproducing its successful…

arxiv.org ·2026-09-28

We haven't written up this one. arXiv cs.AI has the full story — the link below goes straight to it.

Read the original at arxiv.org →

More in AI

Unsurprisingly, Meta's new Muse AI agent blatantly ignores users permissions

In a completely unsurprising turn of events, Meta's Muse AI is stealing Apple Messages, past and present, and uploading the contents to its cloud, even if explicitly told not to.

2026-09-28 ·appleinsider.com

Meta hired MongoDB’s CEO to build its enterprise AI business — but Llama is missing

Meta announced on Monday that it is building a new enterprise business around its AI models and agents and has The post Meta hired MongoDB’s CEO to build its enterprise AI business — but Llama is…

2026-09-28 ·thenewstack.io

Microsoft contracts humans to review Copilot’s creepiest images

Copilot contractors determine which AI-generated She-Hulk upskirt best fulfills the user's prompt.

2026-09-28 ·avclub.com

AI agent startup Instinct hits $10 billion valuation following $1 billion round

2026-09-28 ·investing.com

Autoheal Raises $7.9M From Innovation Endeavors to Make Enterprise Software Heal Itself

Autoheal raises $7.9M led by Innovation Endeavors for self-healing AI agents that fix incidents, patch security flaws and cut AI coding costs at Nomura.

2026-09-28 ·hackernoon.com

Report: Progressive Disclosure of Agent Skills

Users of Workday's deployed LLM-based agents often request features which can be addressed by defining named procedures, also known as skills, in the LLM context, effectively augmenting agents' capabilities. However, as an agent's skills library grows in size, so does the agent's operational cost.

2026-09-28 ·arxiv.org