Aquin Labs is building Aleph, a universal predictor for applied ML, data science, and annotation.

Attention as Routing Graph: Circuit Extraction from Single Forward Pass

Treat attention as a routing map from one forward pass, keep a small subgraph aimed at the answer, and check it under ablation.

Finding Usable Weight Mechanisms with Tiled SVD

Extract mechanism mounts from weight sites as trigger, write, and strength triples. Pre-registered A/B/C eval on Gemma-2-2B: 182/182 GO.

axi

Auto RL self-learning agent for controlling developer environments like aq at full efficiency: coding and research with harness lessons, rejected paths, and refine loops.

aq

Developer environment for foundation models: frameworks, libraries, tools, and agents to design, test, and refine models with precision and intent.

World Models

Talking-avatar LoRA on MiniMax-H3, an action-conditioned JEPA for Mario with probe gates, and a 21MB pixel-art world model.

Interpretability

Applied research on AQIT: transformers and LLMs, attribution, sparse autoencoders, simulation, live training watch, embedding models, evals, security, and benchmarks.

Experimental Weight Editor

Agentic ROME on Pythia 2.8B: causal trace layer location, rank-one MLP updates, and a three-check validation loop that rolls back and retries on failure.

Simulating LLM Behaviour in Different Environments

Four environments for studying LLM behaviour: activation steering for bias reduction, Chess Agents Simulation, Echoes gossip RPG, and Among Us Agents Simulation.