Reproducing J-Space Analysis to Examine Model Representations
MoonL88537 · x · 2026-07-18
The author shared a reproduction of Anthropic's J-space analysis, expanding it with two new experiments:
- Representation structure of Inkling 1T: It acts more like an "outlier" compared to other models, lacking clear separation between perception/workspace/action blocks and maintaining a near-identical geometric structure across all layers. The early-to-late layer CKA metric is around 0.8, whereas other models score around 0.5.
- Impact of Quantization: The J-space of @poolsideai's Laguna XS 2.1 was compared between bf16 and nvfp4, showing almost no difference. The quantized model shares the exact same JLens space as the unquantized version.
The author also noted that both J-space checkpoints have been uploaded to Hugging Face.
More from Research
- Deep Learning Course Adds Mechanistic Interpretability Lecture This Winter — CSProfKGD · 2026-10-26
- INCLUDE-2.0 benchmark evaluates regional knowledge across 91 languages and 86 countries — ABosselut · 2026-10-26
- CoRL paper: flow-matching VLAs can run 1-step Euler inference with no retraining — kaixhin · 2026-10-26
- CoVer at EMNLP 2026: Training Code LLMs to Generate and Self-Verify Simultaneously — algo_diver · 2026-10-26
- Robotic dexterity should be judged on speed, force control and tool use, not copying humans — DJiafei · 2026-10-26
- Turing-complete malware can run in PyTorch's backward pass via a zero-valued loss — retr0jirachi · 2026-10-26