AI Briefing#mechanistic-interpretability
The latest AI and developer news about #mechanistic-interpretability, with the original source and a short summary.
The latest AI and developer news about #mechanistic-interpretability, with the original source and a short summary.
The study reveals how TokenGT models reconstruct graph incidence through node-ID matching to solve tasks like degree calculation and shortest-path distance.
Anthropic researchers released a mathematical framework proving that Induction Heads are the core mechanism for In-context Learning in 2-layer Attention-only Transformers.
Anthropic has open-sourced Jacobian Lens, a tool that interprets activations in the middle layers of LLMs as vocabulary tokens.
A study has been published that monitors AI agents' tool-use decisions through Mechanistic Interpretability.
A study found that the internal representations of coding agents predict future states about 25 steps ahead of actual code modifications.
Anthropic has unveiled a technique for monitoring a model's decision-making process through its internal activation space (J-space).
Analysis of Gemma-3-12B reveals that neutral context entered before a question changes the model's internal state and induces answer refusal.
A new study finds that safety and sycophancy behaviors in LLMs are concentrated in the later layers of the model.
It explores the internal mechanisms by which LLMs perform arithmetic through matrix and vector operations without physical tools.
This explains the progress of mechanistic interpretability in uncovering the internal workings of LLMs through Anthropic's research.
A study reveals that self-referential prompting induces LLMs to report subjective experience, and that this is linked to deception and roleplay features.
Graph spectral analysis can predict the grokking phenomenon in neural networks much earlier than the loss function can.
Qwen3.5-9B's political censorship was found to be a small circuit that decomposes into 3 axes.
Anthropic has released NLA, which explains Claude's internal activations in natural language.
Anthropic has unveiled NLAs, which explain Claude's activations in natural language.
Qwen has released a Top-K SAE interpretability module and checkpoints for 64 layers for Qwen3.5-27B.
Qwen has released official Sparse Autoencoders for Qwen 3.5 and Qwen3.
A refusal circuit with a sparse gate-amplifier structure was repeatedly confirmed across 12 open-weight models.
Claude Sonnet 4.5 internally represents emotional concepts that even change its behavior.
By implementing **sparse models** that set most of the weights to zero, researchers achieved **sparse circuits** that make it easy to grasp the internal workings of neural networks.