Mechanistic Interpretability

Tools for inspecting and probing a model's internal computation directly – activation caching, linear probes, sparse autoencoders, circuit graphs – plus the GFlowNet-based non-goal-directed training objectives motivated by the Scientist AI research direction. More...

Files
file	activation_snapshot.hpp
	Self-contained, enumerable copy of one forward pass's cached activations.
file	circuit_graph.hpp
	Self-contained circuit-graph artifact – scored nodes and weighted edges.
file	detailed_balance_loss.hpp
	GFlowNet Detailed Balance loss (Bengio et al., "GFlowNet Foundations", arXiv:2111.09266).
file	gflownet_forward_policy.hpp
	Masked stochastic categorical policy over a logit-producing Module – P_F for GFlowNet training objectives.
file	gflownet_trajectory.hpp
	Rolls out one full HyperGrid episode under a GFlowNetForwardPolicy.
file	hypergrid_env.hpp
	HyperGrid – the standard minimal GFlowNet correctness-check environment.
file	learnable_scalar.hpp
	A single trainable float, updated by plain SGD – not a Tensor/Module parameter.
file	linear_probe.hpp
	Linear probe – trains one linear classifier to decode a binary concept from a layer's activations, answering "is this concept linearly represented here?".
file	sparse_autoencoder.hpp
	Sparse autoencoder – reconstructs an activation through an overcomplete, L1-penalized hidden layer, decomposing it into a larger, sparser basis.
file	subtb_loss.hpp
	GFlowNet SubTB(lambda) loss (Madan et al., "Learning GFlowNets from partial episodes for improved convergence and stability", arXiv:2209.12782).
file	trajectory_balance_loss.hpp
	GFlowNet Trajectory Balance loss (Malkin et al. 2022, arXiv:2201.13259).

Detailed Description

Tools for inspecting and probing a model's internal computation directly – activation caching, linear probes, sparse autoencoders, circuit graphs – plus the GFlowNet-based non-goal-directed training objectives motivated by the Scientist AI research direction.