Live intelligence

Latest AI updates

2 updates
Research

Causal Protocol Tests Whether Language-Model Workspaces Actually Compute

Researchers propose a causal protocol for determining whether modules in a global-workspace language model ignore, relay, or recombine broadcast information. The central test transplants a message from another input so that the three behaviors predict different answers. The protocol correctly identifies relay and recombination in constructed networks, and among trained workspace models it finds all three outcomes. In AFGW, a language model whose modules communicate only through a learned workspace, a transplanted message produces the recombination-predicted answer in 97% of cases, with its content redirecting what the receiver retrieves from memory.

Global Workspace Theorylanguage model interpretabilitycausal analysis
Research

Input-Dependent Activation Steering Outperforms Fixed Directions When Destination Varies Within Class

This paper investigates when activation steering edits should depend on the input, proposing that input-dependence is needed exactly when the destination varies within the requested class. The authors compare fixed-direction Contrastive Activation Addition (CAA) with a retrieval-based method that uses the mean of k nearest stored states, finding retrieval outperforms fixed steering on substitution corpora when within-class variation exists. The study also identifies four state-derived quantities that predict the appropriate regime.

activation steeringcontrastive activation additioninput-dependent editing