5arXiv cs.CL (Computation and Language)·
CircuitKIT: unified toolkit for mechanistic interpretability circuit analysis
CircuitKIT is a new source-available Python library that integrates circuit discovery, evaluation, and intervention into a single typed, serializable workflow for mechanistic interpretability research. The library addresses fragmentation in the current ecosystem, where researchers must stitch together separate implementations for discovery algorithms, diagnostics, and downstream applications like pruning, editing, and steering. It includes declarative interfaces for mapping structured data into discovery tasks and is released by Lexsi Labs on GitHub.
Related guides (2)
Related events (8)
Understanding Neural Networks Through Sparse Circuits
OpenAI has published work on mechanistic interpretability using a sparse model approach aimed at understanding how neural networks reason internally. The research seeks to make AI systems more transparent by identifying sparse circuits within neural networks. This work is positioned as supporting safer and more reliable AI behavior through improved interpretability.
SchGen: LLM-Based PCB Schematic Generation via Semantic Code Representations
SchGen is presented as the first large language model system capable of generating editable PCB schematics from natural-language descriptions. The approach introduces a semantically grounded code representation that replaces verbose, geometry-heavy schematic formats with relative placement and pin-name-based wiring primitives, reframing the problem as a semantics-driven matching task. A large-scale dataset was constructed via a human-agent collaborative pipeline converting open-source hardware designs into the new representation. Experiments show SchGen outperforms alternative representations and larger general-purpose LLMs on wire connectivity accuracy and functional correctness.
Study disentangles task, language, and model roles in code model internal representations using circuit analysis
A new arXiv preprint investigates whether independently trained code language models converge on the same internal representations, using a 2x2 factorial design crossing Python/Rust with Qwen2.5-Coder-7B and DeepSeek-Coder-V1-6.7B. The study measures 58 Python and 57 Rust grammatical concepts via concept-circuit extraction across all four conditions. Key findings: which concepts earn dedicated circuits is consistent across models (task-determined), but where and how those circuits are organized differs substantially by model — Qwen processes concepts in late layers (~L17-19), DeepSeek in early layers (L6-7). The work reframes the 'circuit universality' question, concluding universality holds for representational content but not computational organization.
grapheme-kit: Open-source Python library for grapheme-level NLP metrics in multilingual settings
Researchers introduce grapheme-kit, an open-source Python library that extends lexical distance, similarity, and evaluation metrics to operate on Unicode grapheme clusters rather than raw code points. The library targets writing systems where a single grapheme spans multiple code points, with specific improvements for Tamil and Sinhala scripts. An OCR case study demonstrates that grapheme-level metrics yield more faithful evaluation of complex scripts than code-point-level metrics.
Phantom specialization in circuit discovery: structural differences don't imply distinct mechanisms
A new arXiv preprint challenges a core assumption in mechanistic interpretability: that structurally different circuits discovered for the same task imply distinct computational mechanisms. Using Literal Sequence Copying across token-frequency bands in five Pythia models (70M–1.4B), the authors extract 75 circuits and show that structurally distinct circuits implement the same computation, with band-specific edges transferring broadly and a shared core recovering ≥99% of circuit performance. The paper introduces the term 'phantom specialization' for this pattern and argues that standard source-level evaluation inflates apparent faithfulness, while edge-level evaluation and cross-condition transfer tests are needed to detect the many-to-one mapping from structure to function.
KnowledgeDebugger: GUI tool for knowledge localization and editing in Transformers
Researchers introduce KnowledgeDebugger, a no-code GUI tool for exploring how Transformers store and process knowledge and for applying knowledge editing methods. The tool wraps EasyEdit, a widely-used knowledge editing library, and is inspired by LM-Debugger, providing interactive access to state-of-the-art editing approaches without requiring code. It targets the exploratory phase of knowledge editing research, enabling case-study-level investigation before committing to large-scale experiments.
Graphify: open-source tool to convert codebases into queryable knowledge graphs for AI coding assistants
Graphify is a Python library that converts arbitrary code folders, SQL schemas, scripts, docs, and media into a queryable knowledge graph, designed to serve as a skill or context layer for AI coding assistants including Claude Code, Codex, Cursor, and Gemini CLI. The project has accumulated 72,438 GitHub stars with 504 added today, indicating strong community traction. It targets the problem of giving AI coding agents unified, structured access to heterogeneous project artifacts across code, schema, and infrastructure.
SkillKit: Portable Skills Layer for AI Coding Agents
SkillKit is an open-source TypeScript project that provides a portable skills abstraction for AI coding agents, enabling installation, translation, and sharing of skills across tools like Claude Code, Cursor, Codex, GitHub Copilot, and 40+ others. The project has accumulated 1,112 stars with 32 added today, indicating moderate community traction. It targets the interoperability gap between the growing ecosystem of AI coding assistants.