Embed‑TTT improves rule induction in ARC‑like tasks, authors say
The authors introduce Embed‑TTT, a two‑step test‑time training protocol for Vision ARC (VARC) that first fine‑tunes only a task‑specific embedding and then freezes it while fine‑tuning the backbone. They report that this approach, which updates less than 0.01% of model parameters, yields embeddings that align more closely with the intended underlying rules of ARC‑AGI‑1, ConceptARC and Mini‑ARC…
Key points
- Embed‑TTT fine‑tunes only task embeddings (<0.01% of model parameters) before backbone training.
- Method aligns embeddings with underlying rules and improves retrieval on ARC‑AGI‑1, ConceptARC, and Mini‑ARC datasets.
- Optimizing embeddings alone solves a non‑trivial fraction of tasks; full two‑step pipeline further raises accuracy.
Across these benchmarks, Embed‑TTT improves embedding‑based retrieval and enables accurate linear probing of known rules. Optimizing just the embeddings already solves a non‑trivial fraction of tasks, and the full two‑step pipeline further raises overall performance. The study also shows that the method recovers the geometric structure of parametric rules and learns compositional capabilities that support rule‑wise interpolation, though not extrapolation. The authors suggest that clearer separation of rule induction and execution could motivate better benchmarks that distinguish in‑distribution from out‑of‑distribution rules.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- OpenAI offers U.S. university labs free ChatGPT Business seats with Rosalind access · 2 src
- Anthropic reports small AI models may self-replicate on IoT devices within months · 29 src
- Researchers propose LLM-referenced C-test for evaluating supra-intelligence · 8 src
- Researchers study permutation diffusion LLMs for recommendation reranking · 1 src
- Researchers compare Non-Generative verifier to Astra LLM for RAG claim checks · 1 src
Comments
via GitHub Discussions