Mechanistic analysis of a GPT-2–like model exploring the compositionality gap in transformers. Using Logit Lens and Causal Tracing, the study identifies and overcomes a deep-layer bottleneck via dataset enhancement addressing the stated Compositionality Gap (NeurIPS24).

To improve the adaptability of Large Language Models (LLMs) by examining and optimizing the storage paradigm within autoregressive transformer models. The emphasis is on pinpointing and editing the locations where factual associations are stored, ensuring that the models retain current and relevant information without requiring extensive retraining

Code for ACL 2026 “Multi-component Causal Tracing in Large Language Models”, introducing PGB-CT for identifying sparse sets of components that drive model behavior.

Code for the paper "Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models." Activation patching, knockout ablation, and mechanistic analysis of how VLMs resolve perception-knowledge conflicts, across Qwen 2.5 VL, LLaVA-NeXT, and PaliGemma 2.

Reactive harness layer for agent workflows. Describe automations in plain language, trace every decision, enforce policies, persist checkpoints. TypeScript. Zero dependencies.

Causal intervention framework for mechanistic interpretability research. Implements activation patching methodology for identifying causally important components in transformer language models.

Causal-tracing benchmark for neural-net interpretability: activation patching recovers the ground-truth circuit where correlational and gradient attribution each collapse -- proven by controls that toggle confounding and saturation on and off. Offline, numpy-only, no API keys.

Reactive harness layer for agent workflows. Describe automations in plain language, trace every decision, enforce policies, persist checkpoints. Python. Zero dependencies.

A small, extensible mechanistic-interpretability lab — logit lens & activation patching on GPT-2 and Qwen3 behind a unified backend adapter. Config-driven, tested, laptop-friendly.

GraphReFly — reactive harness layer for agent workflows. Describe automations in plain language, trace every decision, persist checkpoints. Cross-language spec. Implementations: graphrefly-ts (TypeScript), graphrefly-py (Python).